DISS. ETH NO. 23805
Towards clinical implementation of scanned
proton therapy of moving targets
A thesis submitted to attain the degree of
DOCTOR OF SCIENCES of ETH ZURICH
(Dr. sc. ETH Zurich)
presented by
Kinga Bernatowicz
MSc Nuclear Engineering,
EPF Lausanne - ETH Zürich
born on 05.05.1987
citizen of Poland
accepted on the recommendation of
Prof. Dr. Antony Lomax
Prof. Dr. Günther Dissertori
Prof. Dr. Per Munck af Rosenschöld
2016
This thesis has been supervised by
Prof. Dr. Antony Lomax
ETH Zürich, Department of Physics
Paul Scherrer Institute, Center for Proton Therapy
”Have no fear of perfection,
you’ll never reach it”
- Salvador Dalí
iv
Abstract
Respiratory motion compromises the accuracy and precision of proton pencil
beam scanned (PBS) treatments for tumors located in the thorax and abdomen.
Challenges occur at all stages of treatment, including difficulties in 4D imaging,
treatment planning, image guidance, efficient delivery and treatment verification. As such, PBS treatments at Paul Scherrer Institute (PSI) are currently
restricted to quasi-static tumors. However, with the help of effective 4D planning techniques, the accurate treatment of mobile cancer treatments will soon
become a reality.
In this context, the aim of this work has been to improve 4D treatment
planning and delivery for PBS proton therapy. A lot of effort has already been
done by our department and others to develop robust motion mitigation techniques for PBS, such as rescanning (the statistical averaging of motion effects by
scanning the target multiple times during delivery) and range-adapted PTV’s
(planning target volumes that compensate for proton range variations due to
motion). However, it is unclear to what extent PBS proton therapy still provides a clinical advantage over conventional therapies when such approaches are
used. As such, and as a first part of this work (chapter 2), PBS proton therapy using raITV’s and rescanning has been compared to photon-based VMAT
treatments using the standard ITV approach. Even though treatment margins
are substantially extended with raITV’s, proton treatments offer an advantage
over photons in terms of lower integral dose and second cancer risk.
The required magnitudes of rescanning, and the extent of ITV/raITV’s, necessary to mitigate motion however depend on accurate 4D imaging, and it is
methods to improve 4D imaging that are the subject of the second part of this
work. First (chapter 3), the impact of different respiratory-correlated image
acquisition techniques to reduce imaging artifacts in 4DCT have been studied, with prospectively-gated acquisition being shown to significantly improve
thoracic image quality, without an additional imaging dose burden. However,
v
vi
even with such improvements, 4DCT can only represent patient motion during
a single, averaged respiratory cycle and thus ITV/raITV’s may be under- or
over-estimated under conditions of free, and variable, breathing. To this aim,
we have also investigated 4DCT-MRI as an improved 4D imaging modality for
4D planning (chapter 4). This modality is shown to have the advantage of being
able to acquire breathing information from multiple breathing cycles without
additional imaging dose, whilst also providing the necessary density information
for accurate proton dose calculations.
Despite the presented improvements in 4D imaging however, the somewhat
simplistic combination of extended margins and rescanning, whilst being simple to implement, is perhaps not the most optimal 4D planning approach. As
such, this work proceeds with chapter 5 to investigate the potential of using
comprehensive motion data (as provided by 4DCT-MRI) directly at the plan
optimization stage in order to more optimally deal with motion effects (4D
optimization). Although this approach is shown to be able to achieve highly
conformal and homogenous dose distributions under conditions of motion (almost as good as for a static delivery), we also show that the resulting dose
distributions are very sensitive not only to the characteristics of motion, but
also to the exact delivery timeline of the individual proton pencil beams. Consequently, it is argued that methods for the dynamic compensation of motion
variability and delivery time-lines must be developed to make 4D-optimisation
a viable motion mitigation approach.
Finally, as much of the work presented in chapters 4 and 5 is based on a 4D
dose calculation (4DDC) algorithm developed at our institute, in chapter 6, a
first experimental validation of this is presented. As with the 4D-optimisation of
chapter 5, these measurements show the sensitivity of 4D deliveries to the exact
details of the motion and delivery timelines, as well as to the exact experimental conditions. Based on this, we conclude that advanced delivery techniques,
such as tracking or 4D optimization, whilst offering substantial dosimetric improvements remain currently challenging to realize clinically, due to their heavy
reliance on accurate monitoring of motion and exact knowledge of treatment
delivery dynamics. As such, the development of fast delivery adaptation (e.g.
using online intensity modulation) will be necessary in order to make them robust to the inevitable uncertainties and variations of fractionated PBS proton
therapy.
Zusammenfassung
Atembewegung beeinträchtigt die Genauigkeit und Richtigkeit der Behandlung
von Tumoren im Thorax und Abdomen mit gescannten Protonen-Nadelstrahlen
(PBS). Viele Glieder der Behandlungskette werden dabei von der Atembewegung beeinflusst: von Bildgebung – egal ob räumlich (3D) oder räumlich/zeitlich
aufgelöst (4D) – über Therapieplanung und deren Verifikation, bis hin zur (bildgeführten) Bestrahlung selbst. Deshalb beschränken sich PBS-Bestrahlungen
derzeit auf quasi-statische Tumore. Mithilfe orts- und zeitaufgelöster Therapieplanung wird die Behandlung bewegter Tumore jedoch bald verwirklicht werden.
Das Ziel dieser Arbeit ist es daher, die 4D Planung und Bestrahlung mit
gescannten Protonen-Nadelstrahlen zu verbessern. Auf diesem Gebiet wurden
bereits zahlreiche Studien durchgeführt mit dem Ziel, unerwünschte Effekte der
Atembewegung durch robuste Planung zu verringern. Dabei haben sich zwei
Haupttechniken herauskristallisiert: zum einen das Ausschmieren der Bewegungseffekte durch wiederholtes Bestrahlen des Zielvolumens mit reduzierter
Dosis und zum anderen das Anpassen der internen Zielvolumina (ITVs) an die
Reichweite der Protonen im Patienten – kurz raITV. Unter Einsatz dieser beiden
Techniken ist jedoch der klinische Vorteil von Protonentherapie gegenüber konventioneller Strahlentherapie fragwürdig. Deshalb werden im ersten Teil dieser
Arbeit modernste VMAT-Behandlungen (Tumorbestrahlungen, bei denen die
Photonenquelle um den Patienten rotiert) mit wiederholter PBS-Bestrahlung
des raITV verglichen (Kapitel 2). Obwohl sich der Sicherheitssaum des ITV
bei der Anpassung an die Protonenreichweite maßgeblich erhöht, weisen PBSBehandlungen immer noch eine geringere Gesamtdosis und ein damit verbundenes geringeres Risiko für Sekundärtumore auf.
Die optimale Anzahl an Bestrahlungswiederholungen sowie die Größe der
Zielvolumina (ITV und raITV) hängen von genauer 4D Bildgebung ab. Im
zweiten Teil dieser Arbeit werden deshalb Methoden zur Verbesserung der 4D
Bildgebung vorgestellt. Kapitel 3 beschäftigt sich zunächst mit der Reduzievii
viii
rung von Artefakten in der 4D Computertomographie (CT). Die Aufnahme mit
Atmentrigger liefert eine signifikant bessere Bildqualität im Brustbereich ohne
den Patienten zusätzlicher Strahlung auszusetzen. Doch 4DCTs geben lediglich Aufschluss über die mittlere Bewegung der Anatomie in einem Atemzyklus.
ITVs und raITVs können daher bei freier und irregulärer Atmung unter- oder
überschätzt werden. Deshalb haben wir aus statischen CTs und 4D Magnetresonanztomographien (MRTs) mehrperiodische 4DCT-MRTs rekonstruiert (Kapitel 4). Dieses Bildgebungsverfahren liefert Informationen zur Atmenbewegung
aus mehreren Zyklen, es enthält die für die Dosisberechnung relevante Dichteverteilung und seine Strahlenbelastung ist dabei identisch zu der eines einzelnen
CT-Scans.
Obwohl die Kombination aus erweitertem Sicherheitssaum und wiederholter
Bestrahlung einfach umzusetzen ist, stellt sie möglicherweise nicht den optimalen Ansatz für 4D Bestrahlungsplanung dar. Wir haben daher untersucht,
inwiefern ein detaillierter Datensatz der Atembewegung – gewonnen aus einer
Vielzahl an 4DCT-MRTs – die 4D Optimierung von Bestrahlungsplänen verbessern kann. Auf Grundlage dieses Datensatzes lassen sich zwar sehr gleichmäßige
und homogene Dosisverteilungen berechnen (beinahe vergleichbar zu statischen
Zielvolumina), diese reagieren jedoch äußerst empfindlich auf kleine Änderungen
in der Bewegung und der zeitlichen Abfolge, unter welcher die einzelnen Nadelstrahlen appliziert werden. Daraus leiten wir ab, dass die zeitliche Bewegungsund Bestrahlungscharakteristik dynamisch kompensiert werden muss, um mithilfe von 4D Optimierung Bewegungseffekte wirksam zu unterdrücken.
In Kapitel 6 präsentieren wir schließlich eine erste experimentelle Validierung der 4D Dosisberechnung (4DDC). Diese reagiert ebenso empfindlich auf
Änderungen in der Bewegungs- und Bestrahlungscharakteristik wie die 4D Optimierung aus Kapitel 5. Außerdem beeinflussen die experimentellen Rahmenbedingungen die Messresultate stark. Wir kommen zu der Schlussfolgerung,
dass fortschrittliche Behandlungsmethoden wie 4D Optimierung oder TumorNachverfolgung zwar dosimetrische Vorteile aufweisen, ihre klinische Umsetzung allerdings äußerst schwierig bleibt, da sie auf eine genaue Kenntnis der
Bestrahlungscharakteristik und eine präzise Überwachung der Atmenbewegung
angewiesen sind. Es müssen deshalb Methoden entwickelt werden, die es erlauben, die Bestrahlung unmittelbar an die Atembewegung anzupassen (z.B. durch
Atemtrigger oder Modulation der Intensität). Denn nur so sind diese Behandlungsmethoden robust gegenüber inhärenten Unsicherheiten und Schwankungen
der fraktionierten PBS-Bestrahlung.
Acknowledgements
This work would not have been possible without a joint effort of many motivated
people working together towards the ultimate goal of making scanned proton
therapy treatments available to a wider number of patients suffering cancer
disease. I express my gratitude to all employees of the Proton Therapy Center
at Paul Scherrer Institute (PSI), and all the national and the international
collaborators, who participated in proton projects in the last few years.
I would like express my biggest gratitude to Prof. Anthony Lomax for all
the support during my doctorate studies, for many interesting discussions about
the projects, giving me motivation in good and worse periods, and his patience
while correcting English in my publications (especially for adding many articles
that, as a Polish person, I tend to forget about). Moreover, I would like to
thank him for giving me the opportunity to develop my academic skills during
the doctorate work, especially through participating in the collaboration project
at the University of Sydney and completing the Medical Physics studies at ETH
Zurich.
A big thank you goes to Rosalind Perrin for making the late measurement
nights enjoyable (peanut-butter cupcakes!) and for many discussions on the
motion-related topics and phantom dosimetry. Next, I would like to thank
Marta Peroni for taking time in her busy schedule to teach me about the
image acquisition and processing techniques and for her positive attitude. Ye
Zhang showed me many details in handling big motion data and complex 4D
codes, for which I am very grateful. I would like to thank Jenny Dueck and
Carla Winterhalter for the nice atmosphere in our office and making sure that
there is always some chocolate around for those who need it. A great thanks to
Gilles Martin and David Oxley for their best support on computing resources
during my stay at PSI.
I would like to thank Prof. Paul Keall for being open to international
collaboration, many discussions on lung imaging and kind work atmosphere
ix
x
and I would like to acknowledge my colleagues from Sydney, John Kipritidis,
Emma Colvill, Sean Pollock, Chenyu Huang, Brendan Whelan, Andy
Shieh, Stefan Kolling, Pete King, Enid Eslick, Ricky O’Brien, who
made me feel very welcomed in their group and gave their best support by
organizing many career development meetings.
I would also like to thank Prof. Uwe Schneider for very stimulating discussions on second cancer risk modeling and uniting medical physicists in Switzerland by organizing the Jumbo meetings.
Thanks to many of my friends and colleagues, Francesca Belosi, Grischa
Klimpki, Francesca Albertini, Antje Knopf, Petra Trnkova, Dirk Boye,
Fabian Hennings, Abdel Hammi, Gabriel Meier, Serena Psoroulas,
Barbara Knäusl, Benedikt Krohn, Abhishek Saxena, Damar Wicaksono, Ferruccio Bolla, Matt Stark. You made my time in Switzerland a
great experience!
Many thanks to my parents and all family for believing in me from the start.
Last, but not least I would like to thank my dearest husband, Carles Gomà,
for being there for me and offering support whenever I needed.
List of publications
Peer reviewed publications
K. Bernatowicz, J. Lonsky, F. Albertini, A.J. Lomax D.C. Weber & U. Schneider (2016). ‘Comparison of radiation-induced complication risk in volumetric
modulated arc therapy and pencil beam scanned proton therapy for lung and
liver using motion-adapted margins’. Submitted.
K. Bernatowicz, M. Peroni, R. Perrin, D.C. Weber & A.J. Lomax (2016). ‘FourDimensional Dose Reconstruction for Scanned Proton Therapy Using Liver
4DCT-MRI’. International Journal of Radiation Oncology Biology and Physics
95(1):216-23.
S. Pollock, J. Kipritidis, D. Lee, K. Bernatowicz & P. Keall (2016). ‘The impact
of breathing guidance and prospective gating during thoracic 4DCT imaging:
an XCAT study utilizing lung cancer patient motion’. Physics in Medicine and
Biology 61(17)
K. Bernatowicz, P. Keall, A.J. Lomax, P. Mishra, A. Knopf & J. Kipritidis
(2015). ‘Quantifying the impact of respiratory-gated 4D CT acquisition on thoracic image quality: a digital phantom study’. Medical Physics 42(1):324-34.
R. Perrin, M. Peroni, K. Bernatowicz, D. Oxley, A. Mayor, D. Meer, T. Parkel,
S. Safai, D.C. Weber & A.J. Lomax (2016). ‘Rescanning measurements in a realistic lung phantom for evaluation of motion-mitigated, scanned-beam proton
therapy’. Manuscript in preparation.
S. Ehrbar, R. Perrin, M. Peroni, K. Bernatowicz, T. Parkel, I. Pytko, S. Kloeck,
xi
xii
M. Guckenberger, S. Lang, D.C. Weber & A.J. Lomax (2016). ‘Respiratory
motion-management in radiotherapy – a dosimetric comparison in an anthropomorphic lung phantom (LuCa). Submitted.
R. Perrin, M. Zakova, M. Peroni, K. Bernatowicz, C. Bikis, A.K. Knopf, S.
Safai, F. Fernandez-Camona, N. Tscharner, D.C. Weber & A.J. Lomax A.J
(2016). ‘An anthropomorphic breathing phantom of the thorax for testing new
motion mitigation techniques for pencil beam scanning proton therapy’. Submitted.
Conference Presentations
K. Bernatowicz, Y. Zhang, R. Perrin, D.C. Weber, A.J. Lomax, 4D optimization for proton pencil beam scanning – sensitivity to motion parameters, ICCR
2016, London, 27th-30th June 2016
K. Bernatowicz, M. Peroni, R. Perrin, D.C. Weber, A.J. Lomax, 4D proton
dose reconstruction in liver using 4DCT-MRI data sets, PTCOG 55, Prague,
26th-28th May 2016
K. Bernatowicz, Y. Zhang, R. Perrin, D.C. Weber, A.J. Lomax, 4D optimization
for proton pencil beam scanning – sensitivity to motion parameters, PTCOG
55, Prague, 26th-28th May 2016
K. Bernatowicz, M.Peroni, A.Bolsi, F. Albertini, R.S. Malyapa, D.C. Weber,
A.J. Lomax, What is the effect of fractionation in the treatment of mediastinum
tumours with pencil beam scanning? PTCOG 54, San Diego, USA, 18-23rd
May, 2015
K. Bernatowicz, P. Keall, P. Mishra, A. Knopf, A.J. Lomax, J. Kipritidis, The
impact of prospective respiratory-gated 4DCT acquisition on thoracic image
quality: a digital phantom study, 3rd ESTRO Forum, Barcelona, Spain, 24-28
April 2015
K. Bernatowicz, R. Perrin, M. Peroni, A.J. Lomax, Characterizing the effect of
density variation on proton water equivalent range in liver and lung as a result
of respiratory motion, 4D treatment planning workshop, London, UK, 28-29th
xiii
November, 2014
K. Bernatowicz, R. Perrin, A.J. Lomax, The effect of 4D-MRI motion mapping
to CT image for use in liver 4D dose calculations, ISMRM Workshop on Motion
Correction in MRI, Tromso, Norway, 11-14th July, 2014.
xiv
Contents
1 Introduction
1.1 Proton therapy for conformal cancer treatments . . . . . . . . . .
1.2 Clinical motivation for protons and PBS for mobile tumor indications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Respiratory motion and 4D imaging . . . . . . . . . . . . . . . .
1.3.1 Current limitations of 4D imaging . . . . . . . . . . . . .
1.4 Motion mitigation in proton therapy . . . . . . . . . . . . . . . .
1.5 Structure of the thesis . . . . . . . . . . . . . . . . . . . . . . . .
1
2
4
5
6
8
11
2 4D treatment planning using motion-adapted margins
13
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.1 Internal target volume definition . . . . . . . . . . . . . . 15
2.2.2 Treatment planning . . . . . . . . . . . . . . . . . . . . . 16
2.2.3 Volume and dose evaluation . . . . . . . . . . . . . . . . . 16
2.2.4 Evaluation of gITV and raITV for scanned proton therapy
of moving targets . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.5 Radiation-induced complication risk . . . . . . . . . . . . 18
2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Evaluation of gITV and raITV for scanned proton therapy
of moving targets . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.2 Comparison of proton raITV and photon gITV effects in
healthy tissue . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
xv
xvi
Contents
3 Respiratory-gated 4DCT acquisition to improve image quality
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1 Tumor motion data . . . . . . . . . . . . . . . . . . . . .
3.2.2 Modified XCAT phantom . . . . . . . . . . . . . . . . . .
3.2.3 4DCT simulation work-flow . . . . . . . . . . . . . . . . .
3.2.4 Generation of ground truth XCAT images . . . . . . . . .
3.2.5 Segmentation of lung and tumor structures . . . . . . . .
3.2.6 Image quality metrics . . . . . . . . . . . . . . . . . . . .
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1 Comparisons of thoracic image quality . . . . . . . . . . .
3.3.2 Comparisons of acquisition time and imaging dose . . . .
3.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
30
32
32
33
35
36
38
38
40
40
45
45
48
4 Image processing for 4D proton dose calculation on irregular
breathing patterns
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 4DCT-MRI concept . . . . . . . . . . . . . . . . . . . . .
4.2.2 Data sources . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.3 Motion extraction . . . . . . . . . . . . . . . . . . . . . .
4.2.4 MRI-CT liver correspondence . . . . . . . . . . . . . . . .
4.2.5 Comparing 4DCT-MRI and 4DCT . . . . . . . . . . . . .
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 4DCT-MRI subject-specific . . . . . . . . . . . . . . . . .
4.3.2 4DCT-MRI population-based . . . . . . . . . . . . . . . .
4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
50
51
51
51
52
53
53
56
56
58
58
62
5 Advanced treatment planning with 4D optimization
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . .
5.2.1 4D optimization approach and implementation . . . . .
5.2.2 Proof of principle and sensitivity to motion parameters
5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.1 Proof of principle . . . . . . . . . . . . . . . . . . . . . .
5.3.2 Fast 4D optimization . . . . . . . . . . . . . . . . . . . .
5.3.3 Sensitivity to motion parameters . . . . . . . . . . . . .
63
64
65
65
68
68
68
70
74
.
.
.
.
.
.
.
.
CONTENTS
5.4
5.5
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xvii
75
76
6 Dosimetric quantification of motion effects in scanned proton
therapy
79
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . 80
6.2.1 4D dose calculation . . . . . . . . . . . . . . . . . . . . . . 81
6.2.2 Measurement using a moving platform . . . . . . . . . . . 82
6.2.3 Measurement using anthropomorphic phantom . . . . . . 87
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.3.1 Comparison of the calculated and measured dose distributions using the moving platform . . . . . . . . . . . . . 88
6.3.2 Comparison of calculated and measured dose distributions
in the anthropomorphic phantom . . . . . . . . . . . . . . 97
6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7 Conclusion and outlook
103
8 Appendix
109
8.1 4DCT-MRI of lung . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.1.1 Generation of lung 4DCT-MRI . . . . . . . . . . . . . . . 109
8.1.2 Comparison of 4DCT and 4DCT-MRI . . . . . . . . . . . 113
xviii
Contents
Chapter 1
Introduction
Highly conformal radiation therapy is reliant on radiographic image guidance
for patient positioning, target localization and planning. However, tumor mobility due to respiratory lung motion can be a limiting factor and independent
of modality and method, 4D (time-resolved) radiographic x-ray image guidance
results in substantial dose to the patient (Murphy et al. 2007). Additionally,
generation of a representative breathing cycle from radiographic images assumes
breathing regularity, which result in imaging artifacts in 90 % of images (Yamamoto et al. 2008). Consequently, cycle-to-cycle and day-to-day variations in
breathing patterns are to be expected and should be considered for accurate
imaging and treatment planning. As such, in this thesis, state-of-art imaging
modalities that include the variability of motion to achieve high quality images
with reduced imaging dose compared to the conventional methods have been
investigated (chapters 3 and 4).
Another challenge in conformal radiotherapy for mobile targets is the treatment planning itself. In clinical practice, 4D images are used to expand the
treatment margins to account for motion. As a result, healthy tissue surrounding the tumor is unnecessarily irradiated. As such, the potential clinical impact
of this approach is investigated for Pencil Beam Scanned (PBS) proton therapy in comparison to state-of-the-art conventional radiation therapy modalities
(chapter 2), as have advanced mitigation methods based on 4D optimization, as
developed and evaluated in chapter 5. Finally, the validity of 4D dose calculations, required for accurate estimates of the delivered dose under conditions of
motion, have been experimentally investigated (chapter 6).
First however, this chapter presents a rationale for the project, by review1
2
Chapter 1. Introduction
Figure 1.1: Gantry 2 at Paul Scherrer Institute is a proton pencil beam scanned
delivery system equipped with an x-ray imaging device (Beam’s Eye View) mounted
on the treatment nozzle. The CT on rails is available for the in-room image guidance.
ing the current technical status and challenges in image-guidance, radiotherapy
planning and treatment delivery for the treatment of mobile tumors (e.g. liver
and lung) using PBS proton therapy.
1.1
Proton therapy for conformal cancer treatments
The main goal of radiation therapy (RT) is to achieve high dose conformity
i.e. deliver high dose to the tumor volume and as low as possible dose to the
surrounding tissue, and consists of several steps: disease diagnosis and imaging,
therapy planning, delivery and verification. Conventionally, treatments are delivered in several sessions (fractions), for example 30 fractions of 2 Gy, to exploit
differential radiobiological response effects i.e. improved repair capabilities of
healthy tissue compared to tumor cells. In addition, more targeted radiotherapy
improves outcomes: a 21.9 % improvement in survival at 3 years was observed
in patients receiving high-dose conformal RT (Nakayama et al. 2010) and a 1
Gy decrease in overall mean lung dose has been associated with a 2 % reduction
in pneumonitis (Marks et al. 2010). Hence, most of the efforts in RT physics
1.1. Proton therapy for conformal cancer treatments
3
research have been to improve dose conformity, while maintaining tumor control
in the presence of treatment uncertainties.
Proton therapy (PT) is an alternative modality to conventional (photon) RT.
It was first proposed by Robert Wilson (1946) and by the end of 2011, more than
80’000 patients had been treated with protons worldwide (http://ptcog.org).
A comprehensive summary on proton therapy history is presented in the book
of Paganetti (2011).
The basic physics of the proton beam penetrating matter can be described
with the Bethe-Bloch formula. It defines the energy deposition of charged particles as being proportional to the material density (and electron density) and
inversely proportional to particle velocity squared. As a result, the integral dose
deposited by a particle in a traversed material has a plateau region followed by
the well-known Bragg peak and a sharp distal fall-off. The width of the peak
depends on range straggling (stochastic characteristic of proton-electron interactions in medium) and the initial energy spectrum. The height of the Bragg
peak is determined by the particle weight (proton fluence), whereas the range is
determined by the initial energy of the particles and the stopping power of the
material. Deposition of the energy at a distinct depth is one of the advantages
of proton therapy.
The lateral spread of the proton beam results from multiple Coulomb scattering, where protons are deflected from their path by the heavy positively charged
nucleus, and increases with material depth. In a process of collision, protons
may also be lost from the beam, contributing to a loss of proton fluence at the
Bragg peak. Such collisions result from interactions with the atomic nucleus
from the traversed material. As a result, secondary neutrons or short lived radioisotopes like O-15 and C-11 are produced. These radioisotopes are positron
emitters and can be imaged during and after therapy using Positron Emission
Tomography (PET).
In practice, proton fields are adjusted to the depth and size of the tumor by
either passive scattering or pencil beam scanning (PBS). Passive scattering employs single or double scatterers to broaden the field, together with collimators
and compensators which conform the final dose to the target. Rotating wheels
(with varying thickness) or ridge filters are used to modulate the beam in depth
to assure proper dose coverage throughout the target (Koehler et al. 1977). In
contrast, in PBS the beam is steered laterally by magnets and different depths
are irradiated by energy modulation of the proton beam. As such, the target
is irradiated sequentially spot by spot (spot = pencil beam which can be visualized as a point corresponding to the Bragg Peak) with the deposited dose of
each beam being proportional to its weight (i.e. number of protons or fluence).
4
Chapter 1. Introduction
Pencil beam scanning has also been extended to deliver intensity modulated
proton therapy (IMPT) (Lomax 1999). In IMPT, each individual field delivers
an optimized and arbitrarily irregular dose pattern, which when combined with
multiple fields, allows for high flexibility in tailoring the dose distribution to
the target, whilst simultaneously avoiding selected critical structures. IMPT
however, is sensitive to treatment uncertainties (e.g. due to changing patient’s
anatomy) and may require additional measures to assure robust planning.
Recently at Paul Scherrer Institute, an advanced proton beam scanning system (Gantry 2) has been installed and the first patients were treated at the
end of 2013, see figure 1.1. This new gantry was built on the experience of
our previous system (Gantry 1) and offers extended technical features, such as
fast scanning, iso-centric layout and in-room image guidance. Amongst other
applications, this facility opens new horizons for the treatment of mobile tumors
(Pedroni et al. 2004).
1.2
Clinical motivation for protons and PBS for
mobile tumor indications
According to the World Health Organization, lung cancer is the most common
cause of cancer related death worldwide, and is estimated to be responsible for
nearly one in five of the total deaths (1.59 million deaths, 19.4 %). Because of its
high fatality (the overall ratio of mortality to incidence is 0.87) and the relative
lack of variability in survival in different world regions, the geographical patterns in mortality closely follow those of incidence (http://globocan.iarc.fr).
Liver cancer is the second most common cause of death from cancer worldwide,
estimated to be responsible for nearly 746,000 deaths in 2012 (9.1 % of the total). Despite improvements in imaging techniques for diagnosis, the prognosis
for liver cancer patients treated with conventional methods remains poor (Lepage et al. 2015), and similarly to lung cancer, geographical patterns in incidence
and mortality are similar (http://globocan.iarc.fr).
In Switzerland, more than 3’500 new cases of lung cancer are diagnosed each
year and a similar number are reported for liver cancer (http://nicer.org).
An increased incidence rate for lung cancer is largely attributed to smoking.
Non-smoker cases often being associated to a combination of genetic factors, as
well as occupational and environmental exposures (e.g. radon gas, asbestos, air
pollution, second hand smoke) (Boyle et al. 2008). The survival rate of lung cancer patients depends greatly on the time of diagnosis and the extent of disease.
1.3. Respiratory motion and 4D imaging
5
Because early stage lung cancer often has symptoms similar to those caused
by smoking, it is difficult to diagnose and most cancer patients are diagnosed
at a late stage. For instance, the survival of patients with locally advanced
non-small cell lung cancer (NSCLC) is poor, with a 5-year overall survival of
15 % for stage III NSCLC patients treated with concurrent chemo-radiotherapy
(Aupérin et al. 2010).
Recent clinical trial reports that lung treatments (x-ray RT) with increased
dose in the tumor, might be potentially harmful (Bradley et al. 2015). Scanned
proton therapy offers an alternative, allowing for more conformal treatments,
which could be beneficial for those patients. However, several challenges for
PBS delivery to moving targets need to be addressed.
1.3
Respiratory motion and 4D imaging
The most important muscle in respiration is the diaphragm, which, when it contracts, allows expansion of the chest cavity in the superior-inferior (SI) direction
during inhale. Likewise, the intercostal rib muscles contract during inhalation,
allowing for further chest expansion in the anterior-posterior (AP) direction.
Consequently, the lung expands and pressure drops, sucking the air through the
nasal cavity and mouth. During exhalation, respiratory muscles relax, decreasing the volume of the chest cavity and forcing air out of the lungs. Apart from
the autonomic breathing process, breathing can be controlled by the individual
– an important feature employed in RT motion mitigation such as breath-hold
or audio-visual guidance, as described in the next section.
The magnitude of lung and liver tumor motion varies from patient to patient
and is not correlated to tumor size or location (Stevens et al. 2001). Usually,
the motion in the SI direction is dominant, followed by the AP direction and
lastly the LR direction. Typical lung tumor motion ranges are: 0-50 mm in SI,
0-24 mm in AP direction and 0-16 mm in LR direction (Keall et al. 2006b) and
liver tumor motion ranges: SI 4-14 mm, AP 2-8 mm and LR 0-8 mm (Kitamura
et al. 2003).
Radiotherapy uses various imaging modalities for the delineation of the tumor and critical structures, treatment planning (plan definition and dose calculation), treatment guidance (beam setup, patient alignment) and patient followup. In addition, motion can be imaged using time-resolved (4D) imaging techniques such as computed tomography (4DCT), cone beam CT (4DCBCT) or
magnetic resonance imaging (4DMRI). In 4DCT, a set of 3D images of the patient at different motion states is acquired using a rotating x-ray tube (Keall
6
Chapter 1. Introduction
2004, Pan 2005). An external motion surrogate, such as a respiratory pressure
belt or optical surface tracking, collects signals during acquisition to provide the
temporal information for sorting images in the 4D reconstruction. As a result,
a set of 3D images representing patient anatomy at different breathing positions
in an averaged breathing cycle is obtained. The main advantage of this method
in the context of RT is the availability of the electron density information needed
for dose calculations, as x-ray imaging differentiates between anatomical tissues
according to their composition and photon attenuation. The main disadvantage is the imaging dose (250-400 mGy, Murphy et al. (2007)), which in adult
patients could cause an increase in lifetime risk of radiation-induced cancer of
between 2.7-12 % (Sodickson et al. 2009).
Imaging dose can be reduced however by as much as half using another
photon-based imaging technique – 4DCBCT (Sonke et al. 2005, Dietrich et al.
2006). An additional advantage of this approach is that CBCT equipment can
be installed on the delivery system, conveniently imaging the patient’s anatomy
immediately before treatment.
Alternatively, 4DMRI can be used to image respiratory motion. MRI provides a superior soft tissue contrast and does not expose the patient to the risks
of ionizing radiation. High resolution 4D data sets can be acquired over long periods of time upon use of respiratory correlated imaging strategies (Von Siebenthal et al. 2007). High speed acquisitions and good image quality are certainly
advantageous, however the main limitation of MRI usage in RT is the lack of
electron density information needed for accurate dose calculations.
1.3.1
Current limitations of 4D imaging
The problem of 4D imaging is quite complex and is associated with motion induced imaging artifacts. In 4DCT, these are mainly due to hardware limitations
and reconstruction issues when dealing with irregular motions, as the total acquisition time for a 4DCT study typically lasts over multiple breathing cycles.
Two different conventional CT acquisition techniques have to be mentioned:
cine and helical mode.
In cine-mode, a single gantry rotation corresponds to a finite number of acquired projections and spans only a part of the patient’s anatomy (’image slice’).
The full acquisition is then achieved by moving the patient couch and acquiring consecutive image slices. Additionally, to acquire a 4D image set, multiple
gantry rotations are performed in a single couch position, to capture the internal anatomy at different breathing states (assigned to different motion phases
or amplitudes). Consequently, images reconstructed from such projections con-
1.3. Respiratory motion and 4D imaging
7
tain only a part of the patient’s anatomy and have to be sorted according to
the recorded breathing state when they were acquired. In irregular motion
however, breathing states do not repeat often enough to be captured at all requested couch positions. Reconstructed image volumes can then contain image
slices from slightly different image states, causing the so-called duplication and
truncation artifacts as typically seen in clinical 4DCT images (Yamamoto et al.
2008).
In helical acquisition, image data is acquired with a continuous couch translation with a programmed pitch factor (the ratio between the distance the imaging
table translates in one gantry rotation and the width of the x-ray collimation
on the detector). In this mode, image slices can be reconstructed from multipledetectors and partial gantry rotations (e.g. 2/3 rotation in a half-scan (Thomas
et al. 2014)). In contrast to cine mode, images in this mode can come from more
than one breathing cycle and the finite CT image frequency causes blurring of
captured anatomy within the reconstructed breathing states (Pan 2005).
Regardless of the acquisition method, 4D reconstruction is done by sorting
image slices into breathing states (typically 8-10). Breathing states are extracted from an external surrogate signal (e.g. an abdominal pressure belt or
optical tracking system), translated into a breathing phase or motion amplitude. In irregular motion, sorting according to a breathing phase collects image
slices from breathing states with different motion amplitudes causing truncation and duplication artifacts in the reconstructed images. On the other hand,
clinical images reconstructed by sorting images by motion amplitude may not
be repeated enough to acquire all necessary projections, potentially resulting
in missing slices. Although these can be replaced with the closest possible motion amplitude, this also inevitably results in imaging artifacts. Additionally,
external surrogates extract only a 1D signal and do not guarantee the spatial
integrity of all (3D) anatomy between different volumes exhibiting the same
signal output.
Consequently, a number of motion mitigation methods to reduce such artifacts have been proposed in the literature. For instance, audiovisual breathing
guidance can improve patient breathing regularity (Kim et al. 2012) and improve
the quality of 4D image reconstruction. Alternatively, motion-compensation is
a popular approach that aims to interpolate missing or discontinuous image
data using deformable image registration (Ehrhardt et al. 2007). However none
of these techniques can prospectively avoid the acquisition of unnecessary or
unwanted data in the presence of irregular motion.
As already mentioned, 4DCT image sets consist of 8-10 image volumes representing the patient’s varying anatomy during breathing. For liver and lung
8
Chapter 1. Introduction
imaging, the largest motion component is in the superior-inferior direction and
can be approximated by the sin4 motion curve (Lujan et al. 2003). According to the Nyquist theorem however, in order to properly sample the frequency
component of a periodic function, the sampling rate has to be at least twice the
highest signal frequency, as shown in figure 1.2(a) (red points). Typical 4DCT
sampling (10 images) is shown in black, indicating that around the end-inhale
(Amplitude = 0 mm), amplitude errors of < 1 mm can additionally occur.
The finite resolution of CT data (e.g. 2 mm voxel size) is another important factor, as shown in figure 1.2(b). Here, amplitude errors occur both in
end-inhale and end-exhale phases, independently of the temporal sampling frequency. Additionally, 4DCT temporal sampling depends on the reconstruction
points. Figure 1.2(c) is similar to (b), but a different motion amplitude is
assigned to the starting phase. In 4DCT reconstruction therefore, a balance between the number of reconstructed images (=patient imaging dose) and motion
error should be achieved. As such, the triggering of image acquisition based on
external breathing surrogates and/or pre-defined amplitude gates could help.
Such approaches for improving 4DCT are investigated in chapter 3 of this work.
1.4
Motion mitigation in proton therapy
Due to its high flexibility and conformity, PBS is rapidly becoming the method
of choice for most new proton facilities (http://ptcog.org). However, socalled ‘interplay’ effects are a limiting factor in the treatment of mobile targets
with this approach. In short, interplay is the result of the inherently sequential
delivery of individual pencil beams interfering with the tumor motion. This
can ultimately lead to intra-fractional dose inhomogeneities within the target
volume (Phillips et al. (1992) and Bert et al. (2008)), an effect that cannot always
be completely compensated by fractionation (Seco et al. 2009). The interplay
effect is described in more detail in chapter 5. However, in addition to interplay,
there are other effects of motion as well, such as dose blurring and under-dosage
caused by geometrical (including range) target misses, all of which need to be
taken into account during the planning and treatment of mobile tumors.
There are different ways to mitigate motion effects for PBS proton therapy,
and sometimes a combination of them can be used. First, improving the 4D
imaging modalities (on-line and off-line) to monitor and record the respiratory
motion would help and examples of possible improvements will be addressed
in detail in this work in chapters 3 and 4. Second, a patient’s breathing can
be controlled during imaging and treatment delivery. For instance, audio-visual
Figure 1.2: Effects of temporal and spatial sampling: (a) The modeled motion curve (sin4 ) is shown in blue, the
ideal temporal sampling (based on the Nyquist theorem) in red and the black points correspond to the temporal
resolution for a typical 4DCT (10 images). (b) Ideal and 4DCT temporal sampling assuming 2 mm spatial resolution.
(c) Similar to (b), but with different starting phase.
1.4. Motion mitigation in proton therapy
9
10
Chapter 1. Introduction
feedback, in which a guiding waveform is calculated and displayed to the patient,
can improve the regularity and reproducibility of patient respiratory motion
and can be employed for all 4D imaging methods (Pollock et al. 2013, Kim
et al. 2012). For successful motion management however, volunteer studies
with such systems have stressed the importance of training prior to any clinical
interventions, which might be challenging, particularly for the elderly patients.
An alternative approach is to image and/or treat whilst the patient holds their
breath. This can be achieved using either an active breathing control (ABC)
device or by voluntary breath-holds. In the latter approach, the patient is asked
to breathe deeply and slowly, and then to hold their breath for a time that is
comfortable. Moreover, gating systems can be employed which acquire images
only if an external surrogate signal is within pre-calculated limits established
during a training period. Finally, for liver treatments, an abdominal pressure
plate can be used to substantially limit motion (Negoro et al. 2001).
Moreover, motion can be mitigated at the stage of treatment planning. Extending the planned target volume (PTV) margin to account not only for the
setup uncertainty, but also for intra-fractional target motion is an intuitive approach to avoid target misses. As such, ICRU 62 recommends the use of internal
target volumes (ITV) to artificially extend the irradiated volume to allow for
motion of the tumor. ITV’s are typically defined based on 4DCT images, and
can be effective (if not optimal) for a range of photon therapy applications in
the presence of breathing motion. However, the concept requires adjustments
for proton therapy in order to also account for motion induced range changes,
an effect studied in more detail in chapter 2. More recently a robust 4D planning strategy has also been proposed in the context of mitigating motion in
scanned particle therapy delivered using tracking (Eley et al. 2015). Indeed,
tracking can be a motion mitigation technique on its own, in which pencil beam
positions are adjusted during delivery to follow the target as it moves (Van de
Water et al. 2009, Zhang et al. 2014). However, exact monitoring of tumor
position is challenging and tissue deformation additionally affects pencil beam
dose contributions causing dose degradation. A simpler approach therefore is
rescanning, where the target volume is scanned several times each fraction to
achieve a statistically averaged dose that can significantly reduce the inhomogeneities caused by the interplay effect (Phillips et al. 1992, Bert et al. 2008,
Bernatowicz et al. 2013). In addition, rescanning can be particularly effective
when combined with gating (Bert et al. 2009, Mori et al. 2014), where the beam
is ‘triggered’ in synchronization with the breathing as monitored from a suitable
motion surrogate such as the motion of an external patient surface.
In this work however, we investigate the alternative approach of 4D optimiza-
1.5. Structure of the thesis
11
tion, in which the fluences of individual pencil beams of a plan are optimized
to provide the clinically required dose distribution under conditions of motion
(chapter 5).
1.5
Structure of the thesis
Although the standard internal target volume (ITV) approach is considered clinically acceptable for photon treatments, and adapted margins have been shown
to ensure proper target dose coverage and dose homogeneity in liver and lung
cancer patients with rescanned delivery (Bert & Rietzel 2007, Knopf et al. 2013),
it remains unclear to what extent the additional extension to range-adjusted ITV
(raITV) may mitigate the advantages of proton therapy in normal tissue dose.
In chapter 2 therefore, we first evaluate the impact of raITVs on surrounding
tissues and organs at risk (OAR). As such, VMAT and PBS treatment plans for
liver and lung cancer patients using motion adjusted ITV concepts have been
calculated and compared in terms of normal tissue complication probability
(NTCP) and second cancer risk.
The definition of adequate and effective raITVs however depends on accurate 4D imaging. As such, in chapter 3 methods to improve the image quality of
4DCT are described. In this, a digital anthropomorphic phantom has been used
to simulate three 4DCT acquisition modes: (i) “conventional” 4DCT that uses
a constant imaging and couch-shift frequency, (ii) “beam paused” 4D CT that
interrupts imaging to avoid oversampling at a given couch position and respiratory phase, and (iii) “respiratory-gated” 4DCT that triggers acquisition only
when the respiratory motion fulfills phase specific displacement gating windows
based on pre-scan breathing data. Image quality was evaluated by comparing
generated images with the ground truth images in terms of the overall image
quality, as well as lung and tumor volume errors. Additionally, the imaging dose
and acquisition time of the three 4DCT modes were compared.
Next, chapter 4 describes 4DCT-MRI - an image processing method to simulate multiple 4DCT data sets of a patient by combining a (static) patient CT
with motion extracted from acquired 4DMRI studies. 4DCT-MRI images have
been generated using different motion extraction approaches and compared with
4DCT data of liver patients in terms of image quality, extracted motion and proton dose calculation. This method extends the capabilities of motion modeling
for accurate 4D dose calculations by accounting for realistic and variable motion
patterns. 4DCT-MRI also enables not only generation of high quality raITVs,
but also plan robustness studies under different motion conditions and 4D dose
12
Chapter 1. Introduction
reconstruction of delivered plans for patient-specific quality assurance (QA).
Plans using raITVs and rescanning allow for adequate target dose coverage
and homogeneity, especially when generated on good quality 4D images, but
are typically not very conformal due to the inevitable dose-blurring effects at
the target edge. As such, in chapter 5 the development of a new technique
that could significantly improve dose conformation and homogeneity under conditions of motion - 4D optimization - is described. This approach has been
tested using a simulated moving phantom and in simulations on real cancer
patient anatomy. In this work, we show that, although 4D optimization allows for realistic calculations using irregular breathing patterns extracted from
the 4DCT-MRI technique (chapter 4), it is very sensitive to variations in motion. As such, a method to efficiently deliver 4D optimized plans employing 4D
intensity-modulated delivery is proposed.
Finally, the first experimental validations of a 4D dose calculation employed
throughout this work is presented in chapter 6. Measured doses are compared
with calculations and demonstrate the sensitivity of 4D deliveries to the exact
experimental conditions and motion characteristics, together with any variations
in the actually delivered treatments. The results of this work imply that the
clinical implementation of techniques such as tracking or 4D optimization will
be challenging, and in order to make such techniques robust, their delivery needs
to be carefully monitored and adapted retrospectively or in real time (online).
Chapter 2
4D treatment planning
using motion-adapted
margins
The following chapter is based on a recently submitted manuscript (Bernatowicz
2016). We acknowledge the ETH Master students, Julia Lonsky and Samuel
Stolz, who helped to run multiple simulations used in this work and enabled
performing part of the study at the Hirslanden Hospital in Zürich.
A possible method to account for respiratory motion in radiation therapy is
to enlarge the target volume according to the movement of the tumor. For proton therapy not only changes of the position of the target volume are relevant,
but also the related density changes. The range-adjusted internal target volume
(raITV) considers the geometrical deviations as well as the changes of the path
length of the particles. As a consequence, the raITVs can become strongly enlarged compared to the static target volume and result in an increase patient’s
integral dose. In this chapter, the impact of raITV has been studied in the
context of radiation-induced complication risk by calculating the integral dose,
organ equivalent dose for several organs and tissues and normal tissue complication probability (NTCP) for the lung pneumonitis for pencil beam scanned
(PBS) proton therapy and compared with the volumetric modulated arc therapy
(VMAT).
13
14
2.1
Chapter 2. 4D treatment planning using motion-adapted margins
Introduction
Motion is a major problem for radiotherapy and many mitigation techniques
have been proposed, including breath-hold, gating and tracking (Kubo & Hill
1996, Mah et al. 2000, Rietzel & Bert 2010). The most straight-forward approach is to extend the PTV margin to account not only for setup uncertainties,
but also for intra-fractional target motions. ICRU62 recommends the use of an
internal target volume definition (ITV) based on 4DCT images, which can be
effective for photon therapy applications in the presence of motion (Admiraal
et al. 2008, Nakagawa et al. 2012, Takahashi et al. 2013).
Motion is more complex in particle therapy due to the additional motion
induced changes in particle range. Several solutions have been proposed, e.g.
using compensator smearing (Engelsman et al. 2006) or extending the ITV to
the so-called range-adjusted ITV (raITV) (Bert & Rietzel 2007). Furthermore,
Knopf et al. (2013) showed that for pencil beam scanned (PBS) particle therapy,
it is necessary to combine the raITV with re-scanning to ensure adequate dose
homogeneity and target coverage in the presence of intra-fractional motion.
From a clinical perspective, the raITV plan can be created in a single CT
(from a 4DCT set) and standard (3D) treatment planning system. The target
contour is extended, but other volumes including OAR and critical structures
remain unchanged, which has an advantage for dose comparisons and evaluation. Plan robustness can then be evaluated with the help of deformable image
registration and 4D dose calculation algorithms to estimate the ’4D dose’ distributions accumulated on the planning CT (Boye et al. 2013a, Bernatowicz et al.
2016).
Although the standard ITV approach is considered clinically acceptable for
photon treatments, and adapted margins have been shown to ensure proper
target dose coverage and dose homogeneity in liver and lung cancer patients
(Bert & Rietzel 2007, Knopf et al. 2013), it remains unclear to what extent the
additional extension to raITVs may mitigate the advantages of proton therapy
in normal tissue dose. In this work, we aim to evaluate the impact of raITVs
on surrounding tissues and OARs. As such, VMAT and PBS treatment plans
for liver and lung cancer patients using motion adjusted ITV concepts were
computed and compared in terms of normal tissue complication probability
(NTCP) and secondary cancer risk.
2.2. Materials and Methods
15
Table 2.1: Maximum motion amplitude and size of planning volumes: clinical target
volume (CTV), geometric ITV (gITV) and range-adjusted ITV (raITV).
Patient
case
1
2
3
4
5
6
(liver)
(liver)
(liver)
(lung)
(lung)
(lung)
2.2
2.2.1
Max motion
[mm]
CTV
[cm3 ]
gITV [cm3 ]
(gITV/CTV ratio)
raITV [cm3 ]
(raITV/CTV ratio)
5
12
16
10
23
42
384.3
112.7
250.3
19.2
40.7
14.9
414.9 (1.08)
150.8 (1.34)
334.8 (1.34)
29.2 (1.52)
43.2 (1.04)
18.8 (1.26)
442 (1.15)
157.2 (1.39)
359.1 (1.43)
51.7 (2.7)
47.3 (1.18)
47.8 (3.21)
Materials and Methods
Internal target volume definition
Three liver patients and three lung patients, all with 4DCTs, were considered in
this study. The end-exhale phase was used to delineate all CTVs and OARs, and
as the reference phase for deformable image registration and dose accumulation.
Motion fields were obtained using an open source software (Plastimatch;
http://plastimatch.org) nd validated using marker position errors calculated
as the absolute difference between estimated motion (registration result) and the
ground truth motion (4DCT images). Marker position errors were of the order
of a single CT voxel (5/3 mm for the liver and lung patients respectively). Maximum motion amplitudes extracted at the tumor center relative to its position
in the reference phase are presented in table 2.1.
According to ICRU 62, the ITV considers deviations due to internal motion.
Photon plans were prepared by expanding the CTV margin using motion fields
extracted from each 4DCT phase. This ITV therefore contains all tumor positions during the respiratory cycle and will be referred to as the geometrical
ITV (gITV). Proton plans were prepared similarly, but additionally with rangeadjusted ITV’s (raITV) to account for changes in proton beam range due to
motion. raITV’s are specific to the treatment field and for each 4DCT phase,
a new volume is created by shifting the CTV contour by the difference of the
proton range relative to the reference CT phase. Finally, raITV is calculated as
a union of all volumes (Knopf et al. 2013).
16
2.2.2
Chapter 2. 4D treatment planning using motion-adapted margins
Treatment planning
To account for patient positioning uncertainties, different PTVs were defined as
isotropic enlargements of the gITV and raITV by 0 (theoretical), 7 and 10 mm
(figure 2.1). Treatment plans were normalized such that the target volume was
covered by the 95 % iso-dose (ICRU 62) and less than 2 % received ≥ 107 % of
the prescription dose.
VMAT photon plans were calculated using Eclipse v.10.0.28 (Varian Oncology Systems, Palo Alto, CA) using the AAA-algorithm. Planned dose was
assumed to be applied in a single gantry rotation and plans were created using
PTVs derived from patient-specific gITV’s.
PBS proton plans were calculated using the clinical treatment planning system PSIplan v.7.7.0. at PSI using the ray-casting model (Schaffner et al. 1999,
Lomax et al. 2004). Three co-planar fields per plan were used for all patients
(figure 2.1).
2.2.3
Volume and dose evaluation
To estimate the risk of second tumor induction, primary doses to the whole body
and all critical organs, as well as neutron scatter dose contributions, need to be
evaluated. However, as CT images were acquired only around the treatment site,
parts of the evaluated organs not included in the CT had to be compensated
by scaling their volume to a reference volume. For the whole-body, a reference
volume of 65’420 cm3 was used, assuming a mean density of the body of 1.07
g/cm3 and a mean weight of 70 kg (Valentin 2002). For the scaling of the liver for
the lung patients, a mean liver weight of 1’800 g, density of 57 HU (O’Riordan
et al. 2000) and a reference volume of 1’702 cm3 have been assumed. Lung
volumes were not scaled, due to the large uncertainty concerning the reference
volume (changes of volume and blood content during breathing) and because
only small portions of the lung were missing on the CT images of the liver
patients.
Dose calculated for the photon plans considered both primary and scattered
photon dose contributions (up to a cut-off distance of 1 7cm from the field edge).
However for the scaled volumes, the scatter dose received by the additional
volume had to be added to the dose volume histograms manually. For these
cases, a fixed value for the scatter dose of 240 mGy, which corresponds to the
absorbed dose at a distance of 20 cm from the border of the target volume, was
used as a conservative estimate (Hälg et al. 2012).
For the case of protons, the scattered dose was not included in the plan
2.2. Materials and Methods
17
Figure 2.1: Example of a geometrical and range-adjusted ITV for lung patient (a,
b). Planned dose calculated for photon (VMAT) and proton (PBS) therapy of liver
(c) and lung (d) cancer patients.
18
Chapter 2. 4D treatment planning using motion-adapted margins
calculation and had to be added in all evaluations. Scattered dose was attributed
to neutrons and hence the neutron dose was estimated with Dneutron =10−14
Sv/proton (Schneider et al. 2002). The number of protons per fraction and
beam was obtained as a result of the treatment planning and was of the order
of 1010 protons per beam depending on the size of the target volume. The
neutron dose contribution for the whole treatment volume can be estimated as
30x10−14 Sv/proton.
2.2.4
Evaluation of gITV and raITV for scanned proton
therapy of moving targets
In conventional radiation therapy (photon-based), effects of motion can be compensated by extending the target volume margins to ITV. However, expanding
margins does not guarantee the same effect for proton therapy. Combining ITV
approach with irradiating the volume multiple times with lowered dose per scan
(so-called rescanning) was demonstrated effective in mitigating motion effects
(Knopf et al. 2013).
The effectiveness of plans using extended margins (both gITV and raITV)
and rescanning was studied using PBS 4D dose calculation. Patient motion
was extracted from 4DCT using deformable registration, and the motion vectors were used for the ITV creation. Delivery dynamics of PSI Gantry 2 were
considered and 1, 3, 6 and 10 volumetric rescanning factors were simulated.
Resulting dose distributions were analyzed in terms of dose homogeneity, represented by D5 -D95 extracted from the dose-volume-histogram (DVH), and target
dose coverage was calculated as T C = VT,ref /VT . In this, VT is the target volume, VT,ref is the volume of the target receiving at least the reference dose and
Vref is the total volume receiving at least the reference dose. The reference dose
was defined as 95 % of the prescribed dose.
2.2.5
Radiation-induced complication risk
Despite uncertainties in estimating the radiation-induced complications, minimizing it is a major goal in modern radiotherapy. The dose-response relationship
for the risk of radiation-induced cancer in a tissue or organ is an essential tool
for an improved risk assessment. These relationships are based on mathematical models that incorporate biological mechanisms related to radiation exposure.
Normal Tissue Complication Probability (NTCP) describes the deterministic side effects of radiation and can be estimated using the Lyman-Kutcher-
2.2. Materials and Methods
19
Burman (LKB) model based on an sigmoid function of the following three parameters: the whole organ tolerance dose associated with 50 % complication
probability (TD50 ), the slope of the sigmoid curve (m) where the curve passes
through TD50 and an estimation of the volume effect (n). As described in Burman et al. (1991) and Emami et al. (1991), this model is derived for the conventional fractionation with 1.8-2 Gy per day, delivered 5 days a week. NTCP
parameters for liver failure are available from Burman et al. (1991) or Marks
et al. (2010). However, there is clinical evidence that these parameters overestimate the complication probability for IMRT and that liver can be irradiated
to much higher doses (Tao et al. 2016). Therefore, we decided to evaluate the
NTCP only in the lung with symptomatic pneumonitis chosen as the endpoint.
Model parameters were selected from the clinical data obtained from evaluations
of various institutions (Semenenko & Li 2007).
Second Cancer Risk describes the probability of radiation induced cancer
in terms of stochastic side effects. However, due to lack of relevant data, it is
not clearly defined for doses higher than 2 Gy. Cancer incidence is however
proportional to the Organ Equivalent Dose (OED) and therefore different plans
can be compared using this metric. As such, cancer induction risk has been
estimated using a combination of the linear-no-threshold model from the Abomb survivors in the low dose range (< 2 Gy) and cancer risk data of the longterm cancer survivors of the Hodgkin’s disease that were treated with radiation
therapy Schneider et al. (2011). In these models, carcinoma forming tissues
(liver and lung) follow the plateau dose-response curve (eq. 2.1) and sarcoma
forming tissues (bones and soft tissue) a sigmoidal dose-response relationship
(eq. 2.2). These different tissue responses can be associated with biological
effects such as cell killing, repair, repopulation/proliferation and fractionation
effects Schneider (2009).
0
RED(Di ) =
−α0 R
0
eα Di
(1 − 2R + R2 eα Di − (1 − R)2 eDi 1−R )
0
αR
(2.1)
0
RED(Di ) =
−α0 R
0
eα Di
(1 − 2R + R2 eα Di − (1 − R)2 eDi 1−R − α0 RDi )
0
αR
(2.2)
The parameter R describes in both equations the repopulation/repair capability of the tissue and Di refers to the dose bin i from the dose volume
histogram. α0 can be considered as cell killing parameter, and is given by
α0 = α + (β dT Di /DT ), where DT is the total prescribed dose and dT the dose
per fraction. Finally, α and β are the parameters from the linear-quadratic
model for the involved tissues. These were chosen similarly as described by
20
Chapter 2. 4D treatment planning using motion-adapted margins
Schneider et al. (2011), and β = α/(3 Gy) was selected for late responding liver
and lung.
In contrary to the liver and the lung, for which values of the parameter R
have been determined as 0.29 and 0.83 respectively, a good estimation of R for
sarcoma forming tissue is difficult due to lack of available data for these tissues
(Schneider et al. 2011). In order to calculate the OED for these tissues, three
different values of R have been used in the dose-response relationship model:
0.1, 0.5 and 1, for low, intermediate and full repopulation model respectively.
To compare the second cancer risk from different treatment techniques, we
will present the OED ratios of proton vs photon therapy plans. In this, we
include comparisons for different irradiated PTV volumes (derived from gITV
or raITV) and treatment sites (liver or lung), as well as for sarcoma inducement
in the bones and all soft-tissues. For the latter, all bones in each patient CT
have been identified by thresholding, and all soft-tissues by calculating the OED
for the whole patient outline.
2.3
2.3.1
Results
Evaluation of gITV and raITV for scanned proton
therapy of moving targets
The effect of rescanning on the target dose homogeneity is presented in figure 2.2.
Although interplay effects do not recover dose homogeneity in a linear manner, a
general trend can be observed, where dose homogeneity increases with a number
of rescans. 10 rescans are the most effective to recover dose quality of the static
plan. Moreover, the raITV are preferable strategy to assure proper target dose
homogeneity.
In figure 2.3, the target dose coverage is presented. Results show that rescanning alone does not guarantee a good target coverage (compare curves of
CTV to raITV) and margin expansion is necessary for good target coverage in
presence of motion. Moreover, it can be seen that for some cases the gITV
does not provide necessary target coverage and therefore raITV approach in a
combination with rescanning should be employed.
2.3. Results
21
Figure 2.2: Effect of rescanning on the dose homogeneity is quantified with the
D5 − D95 values calculated from the dose volume histogram of the clinical target
volume (CTV). The dashed lines represent to target scenario in motion and solid lines
correspond to plans calculated on a static target. The D5 − D95 values are determined
for the CTV, but the treatment plans are calculated for either the CTV (green), the
gITV (blue) or the raITV (red). The D5 − D95 values are normalized to the prescribed
dose.
22
Chapter 2. 4D treatment planning using motion-adapted margins
Figure 2.3: Effect of rescanning on the target coverage. The dashed lines correspond
to the target coverage for a moving target. The solid ones correspond to the target
coverage values, if the target is assumed to be static. The target coverage is calculated
for the CTV is target, but the treatment is planned on the CTV (green), the gITV
(blue) and the raITV (red).
2.3. Results
23
Figure 2.4: NTCP values evaluated in lung minus CTV of (a) liver and (b) lung
patients. Box plots show the median, whiskers define the NTCP range and box widths
correspond to the 75th percentile. Different margins were added to the range-adjusted
ITV for PBS plans and to the geometrical ITV for VMAT plans.
2.3.2
Comparison of proton raITV and photon gITV effects in healthy tissue
Normal tissue complication probability
Figure 2.4 compares the lung NTCP of proton PBS and VMAT plans using
different PTV margins, calculated for all patients. As would be expected, lung
NTCP’s for the liver patients are considerably smaller than for the lung patients,
and NTCP for all cases increases as a function of margin size. However, the
median NTCP is reduced by a factor of between 25-45 % by PBS proton therapy
when comparing plans with the same nominal margin size, despite the fact that
for all PBS proton plans a raITV has been used. For the liver patients, the lung
NTCP for the PBS plans with the largest margin is still smaller by a factor
of 10 % than that of the VMAT plan with no margin. However, it should be
noted that the lung NTCP is small for all liver cases, whatever the treatment
modality, at least if a margin of ≤ 7mm is used. For the lung patients, we
observe that using a 7 mm raITV for the PBS proton plans would provide a
similar median lung NTCP as with a 0 mm margin for VMAT, and a factor of
up to 45 % smaller than the VMAT plan calculated using the same margin size.
For all cases and margins, the variation in NTCP is substantially larger for the
VMAT plans than for the proton plans, indicating that the magnitude of the
advantage of proton therapy is case dependent.
24
Chapter 2. 4D treatment planning using motion-adapted margins
Figure 2.5: Organ equivalent dose ratios of proton and photon plans calculated in
(a) liver and (b) lung. Box plots show the median, whiskers define the range and box
widths correspond to the 75th percentile.
Second cancer induction
Liver and Lung Figure 2.5 shows the OED ratios of PBS proton and VMAT
plans for both liver and lung patients. As expected, for organs at risk at the
treated sites (i.e. liver in liver cancer patients and lung in lung patients), the
OED ratios increase with the PTV margin. However, proton PBS plans result
in consistently lower OED values (by at least 50 %) compared to photon VMAT
plans, implying that the second cancer risk could be substantially reduced using
PBS proton therapy, even allowing for the additional normal tissue involvement
due to use of raITV’s. With regards to the non-treated organs at risk (e.g.
liver in lung patients), the OED ratios vary substantially, indicating that the
magnitude of improvement when using PBS proton therapy will strongly depend
on the position and volume of the target volume being treated.
Bones and Soft Tissue As discussed above, OED’s for sarcoma induction
in bone (figure 2.6, a-c) and soft tissue (figure 2.6, d-f) have been calculated
assuming different repopulation parameters. For all studied cases, the OED
is lowest when low repopulation is assumed (R=0.1) and highest for the full
repopulation factor in the model. OED ratios in soft tissue have been found to
be similar for both lung and liver patient, whereas in bone, the ratios can be
quite different between the different lung patients. Overall, the OED’s of PBS
are consistently lower than that of VMAT by at least 45 % for bones and at least
55 % for soft tissues. Compared to the models which describe the OED of liver
2.4. Discussion
25
and lung, the low doses have a minor effect on the risk of sarcoma induction in
bones and the soft tissue, at least when assuming that the incidence of sarcoma
formation is equal among the A-bomb survivors and the normal population
(Schneider et al. 2002). Meanwhile, the increase of OED due to the additional
neutron dose from the PBS treatments in this study is only marginal.
2.4
Discussion
Proton therapy is known to result in lower integral dose when compared to
the photon therapy (Weber et al. 2004, Kristensen et al. 2015), but whether
the advantage of proton therapy remains when using clinically relevant margin
recipes is not clear. In this work, lung and liver treatment plans have been calculated for PBS proton and VMAT using modality specific margin recipes. The
resulting plans have been compared with a focus on healthy tissue complication
and second cancer risk to evaluate the potential clinical implications resulting
from the additional margin expansions required to compensate for proton range
changes due to motion.
Motion-adapted margins are a common way to mitigate free-breathing motion for lung and liver cancer patient treatments in photon treatments, but for
proton therapy, these have to be modified and expanded to also account for
variations in particle range due to motion. For the patient cases presented in
this study, a range of PBS proton 4D dose calculations assuming raITV’s have
been performed and compared with respect to target coverage and dose homogeneity to photon VMAT plans using conventional ITV margins. Despite the
larger margins required for PBS, we have found that a significant advantage for
proton therapy remains.
For lung NTCP (pneumonitis), and despite the use of raITV’s, significant
advantages are still predicted for PBS proton therapy in comparison to VMAT
planned on conventional ITV’s. On average, lung NTCP could still be reduced
by a factor of 53 % through the use of PBS proton therapy, if the same nominal
margin expansion is assumed for both modalities. For the liver patients, an
advantage for PBS proton therapy remains even if the expansion is reduced
for the VMAT plans, and also remains in the lung patients for PBS margin
expansions of ≤ 7 mm when compared to VMAT plans calculated with no
expansion.
In terms of liver and lung OED (a relative estimate of secondary cancer
risk), our work predicts that the secondary cancer risk could be reduced by
50 % with PBS proton therapy compared to VMAT, again when considering
26
Chapter 2. 4D treatment planning using motion-adapted margins
Figure 2.6: Organ equivalent dose ratios (active proton/photon) calculated in bones
(a-c) and soft tissues (d-f) assuming different repopulation/repair parameter R. Box
plots show the median, whiskers define the range and box widths correspond to the
75th percentile.
2.5. Conclusions
27
the larger margins necessary for proton therapy. Although the proton-photon
OED ratio increases as a function of margin size (and the advantage of proton
therapy reduces) this effect is marginal and is associated with the enlarged dose
region resulting from the larger target volumes, and not with any additional
neutron dose (in contrary to integral dose), as OED is based on non-linear doseresponse relationships (Schneider et al. 2011). In addition, OED for the bones
and soft tissues were evaluated using different repopulation parameters in the
dose-response model, which although influencing the absolute OED ratios, does
not affect the main observed trends, which indicate that the potential advantage
of protons remains substantial (from 40 % to 90 %) despite the enlarged motion
margin (raITV).
For the studied scenarios, a satisfactory result (dose differences <5 % compared to a static plan) was achieved only for raITV in a combination with
rescanning (Stolz et al. 2014). Evaluation of robustness against motion is difficult, as it strongly depends on selected motion scenarios and is specific to beam
delivery dynamics. The effectiveness of raITV for PBS proton plans were tested
on a single motion scenario extracted from 4DCT and assuming delivery parameters of PSI Gantry 2. As described further, in chapter 4), realistic motion
scenarios could be obtained with 4DCT-MRI. Such motion could be then employed to test the robustness of 4DCT-generated target volumes. Moreover, a
recent study in XCAT phantom showed that 4DCT-based motion margins can
lead to target under-dosage (Koybasi et al. 2014). It is believed that 4DCT-MRI
technique could help to improve delineation of both gITV and raITV to avoid
reported under-dosage.
2.5
Conclusions
Photon VMAT plans with gITV and proton PBS plans using raITV concepts
for treatments of liver and lung cancer patients were compared. The dosimetric
results were translated into lung NTCP and OED to different organs and tissues
to quantify the treatment-related side effects. The impact of raITV for proton
plans due to the enlarged high dose region and neutron dose is low and the NTCP
is smaller by at least 40 % when compared to the VMAT plans. Additionally,
it was found that the second cancer risk could be reduced by 50 % with PBS
compared to the VMAT treatment. Provided that gITV/raITV are robust
against treatment uncertainties and motion, using such margins is a clinically
valid approach.
28
Chapter 2. 4D treatment planning using motion-adapted margins
Chapter 3
Respiratory-gated 4DCT
acquisition to improve
image quality
In the previous chapter, the use of ITV and raITV have been studied for PBS
proton therapy planning of lung and liver cases. That work showed that the definition of such planning volumes can be a simple and effective, if not necessarily
optimal, method of mitigating motion induced range variations. However, the
accurate definition of both, ITV and raITV, depends on accurate and realistic
4D imaging of the patient, typically performed using 4DCT. In this and the following chapter, we will move on to investigating ways of improving 4D imaging
through either improved acquisitions approaches for 4DCT (this chapter) or the
use of 4DMRI (chapter 4).
The following chapter is based on Bernatowicz et al. (2015) and describes
different respiratory correlated 4DCT acquisition approaches and their impact
on the thoracic image quality. In this, a digital anthropomorphic phantom was
used to simulate three 4DCT acquisition modes: (i) “conventional” 4DCT that
uses a constant imaging and couch-shift frequency, (ii) “beam paused” 4D CT
that interrupts imaging to avoid oversampling at a given couch position and
respiratory phase, and (iii) “respiratory-gated” 4DCT that triggers acquisition
only when the respiratory motion fulfills phase-specific displacement gating windows based on pre-scan breathing data. Phantom was programmed to deform
according to free breathing motion data of 7 lung cancer patients. Image quality
29
30
Chapter 3. Respiratory-gated 4DCT acquisition to improve image quality
was evaluated by comparing generated images with the ground truth images in
terms of the overall image quality, as well as lung and tumor volume errors.
Additionally, the imaging dose and acquisition time of the three 4DCT modes
were compared.
3.1
Introduction
Respiration-correlated or ‘four dimensional’ computed tomography (4DCT) provides images resolved to different phases of the breathing cycle and is commonly
used to plan radiotherapy for tumor sites affected by respiratory motion e.g.
lung and liver (Keall 2004, Pan 2005). In the presence of irregular motion,
multi-slice cine mode 4DCT is prone to image artifacts – namely truncation
and duplication of critical structures owing to mismatches in respiratory phase
and displacement between adjacent couch positions. A study of 50 abdominal
and thoracic patients observed at least one artifact of magnitude > 4 mm for 90
% of 4DCT scans (Yamamoto et al. 2008). One mitigation method - prospective
respiratory-gated 4DCT – has been shown to reduce artifacts in simulated lung
tumor images by ≈50 % compared to conventional 4DCT (Langner & Keall
2008, 2010). However, no studies have quantified the impact of prospective
respiratory gated 4DCT on normal lung tissue images. To address this gap
in knowledge, we present the first quantitative analysis of normal lung image
quality in respiratory gated 4DCT using a deformable digital human phantom
driven by patient tumor motion patterns.
For lung tumors, it is known that 4DCT image artifacts contribute to target
delineation errors of up to 10 % between 4DCT phase images (Persson et al.
2010). A lesser-studied problem is the impact of 4DCT image artifacts on estimation of lung volume, which is essential to measuring treatment-related side
effects for normal lung tissue. Two systematic reviews (Marks et al. 2010, Rodrigues et al. 2004) have demonstrated strong correlations of radiation induced
lung toxicity to dose-volume parameters including the mean lung dose (MLD )
and lung volume irradiated to 6 20 Gy (V20 ). For example it is prudent to
limit MLD < 20 Gy and V20 < 30 % in order to keep the risk of pneumonitis
<20 % (Marks et al. 2010). However, within these data there are still variations within and between studies. Since MLD and V20 are calculated based on
lung volume, even a small (1 %) error in lung volume will add uncertainty and
noise masking the true dose-volume relationships. The use of treatment plans
based on single 4DCT phase images, for example using the mid-ventilation phase
(Wolthaus et al. 2006) or exhale phase (Keall et al. 2006a), only emphasizes
3.1. Introduction
31
the need for accurate representation of lung geometry in thoracic 4DCT. More
recently, accurate thoracic 4DCT images have become important for performing phase-matched attenuation correction of 4D-positron emission tomography
(PET) scans (Callahan et al. 2014), and applications of deformable image registration including adaptive radiation therapy (Guckenberger et al. 2011) and
CT-ventilation imaging for calculation of functionally-weighted dose-volume parameters of the lung (Castillo et al. 2012, Yamamoto et al. 2013).
Imaging artifacts in conventional 4DCT are mainly due to hardware limitations in dealing with irregular motion. Many mitigation methods have been
proposed, incorporating improved temporal image resolution (faster tube rotation speeds) and increased field of view (multiple detectors) (Thomas et al.
2014)) as well as audiovisual breathing guidance which can demonstrably improve patient breathing regularity (Kim et al. 2012). Motion-compensation is a
popular approach aiming to interpolate missing or discontinuous image data using deformable image registration (Ehrhardt et al. 2007). However none of these
techniques can prospectively avoid the acquisition of unnecessary or unwanted
data in the presence of irregular motion.
Prospective respiratory-gated 4DCT aims to reduce the number and magnitude of 4DCT image artifacts by limiting acquisition to regular breathing,
which can be defined variously in terms of real-time displacement, velocity
and/or phase criteria. Langner & Keall (2008) performed the first numerical
simulations to investigate the effectiveness of respiratory-gated 4DCT based on
real-time displacement and velocity for 103 respiratory motion patterns from
24 free-breathing lung cancer patients. Compared to conventional 4DCT, gating reduced the root mean square displacement error between consecutive axial
slices by up to 20 %, and reduced patient imaging dose by up to 50 % by
eliminating image oversampling. These same authors later simulated tumor images using a range of displacement gating schemes for 58 patient respiratory
tracesLangner & Keall (2010). Here, gating reduced the mean magnitude of
tumor image artifacts by ≈50 % compared to conventional 4DCT, but at the
cost of total acquisition time, which was increased by 20-100 %. Qualitative reduction in imaging artifacts were shown in proof-of-concept experiments using
a rigid-motion anthropomorphic lung slab (Keall et al. 2007).
In contrast to earlier studies, in this work we perform the first simulations
of respiratory-gated 4DCT with an emphasis on lung image quality. This is
achieved using the 4D eXtended Cardiac-Torso (XCAT) deformable digital human phantom (Segars et al. 2008, 2010), which has been recently modified to
synchronize with patient tumor motion patterns (Mishra et al. 2012). Our
simulations encompass three 4DCT acquisition modes featuring different levels
32
Chapter 3. Respiratory-gated 4DCT acquisition to improve image quality
of respiratory feedback (see figure 3.1(a)). At the lowest level is conventional
4DCT, which applies a constant imaging and couch-shift frequency based on
analysis of pre-scan breathing motion. At an intermediate level, we introduce
the concept of ‘beam paused’ 4DCT that interrupts acquisition where a kV projection has been previously acquired at a particular couch position and phase.
At the highest level is respiratory-gated 4DCT, which triggers acquisition only
where the motion fulfills pre-computed phase and displacement criteria based
on the pre-scan breathing motion. The study of beam paused 4DCT allows us
to separate the impact of displacement gating versus acquisition timing; to our
knowledge beam paused 4DCT has not been studied before.
We employ our 4DCT simulation framework to compare thoracic image
quality for conventional, beam-paused and respiratory-gated 4DCT acquisition
methods for seven lung cancer patients. In particular, we test the hypothesis
that respiratory-gated 4DCT can significantly reduce lung-imaging artifacts. A
specific novel aspect of our simulation framework is the generation of patientand phase-specific ground truth comparators representing the average XCAT
anatomy during the ‘beam-on’ time; image quality metrics include ground truth
comparisons in terms of mean square intensity difference, and clinically relevant measures of lung image quality including threshold-based lung volume
error, Dice similarity as well as false-positive and false-negative ratios. Our
simulations are benchmarked against earlier studies in terms of the impact of
respiratory-gating on tumor volume estimation, acquisition time and relative
imaging dose.
3.2
3.2.1
Materials and Methods
Tumor motion data
We observed tumor centroid positions measured for 30 lung cancer patients in
the Cyberknife Synchrony database (Suh et al. 2008). Tumor centroid positions
were originally estimated using a correlation between the external patient motion and internal fiducial locations with a sampling frequency of approximately
25 Hz. As in Suh et al. (2008), tumor motion patterns were visually assessed
to exclude data points showing no motion (indicating a pause in treatment or
reset of the external/internal correlation) or significantly large motion at the
beginning of the data (reflecting learning time of the tracking system). Data
was linearly interpolated to a constant sampling frequency of 33 Hz.
We selected 8 patients (14 fractions) with mean peak-to-trough tumor mo-
3.2. Materials and Methods
33
tion displacements >5 mm (SI) according to the AAPM TG76 guidelines for
explicit motion management (Keall et al. 2006b). The 5 mm cutoff was applied
to the entire motion trace. Table 3.1 summarizes the mean ± standard deviation
(STD) of peak-to-trough displacement in the SI direction, and breathing period
for the 14 fractions. These values are given separately for the entire motion and
the first ≈ 25 seconds corresponding to a pre-scan training period typical of conventional 4DCT. Note that over half of the tumor motion patterns fell below the
5 mm cutoff during training, but exceeded the 5 mm cutoff after training (underlined data in table 3.1). For each data point of the tumor motion patterns,
respiratory phase values (0-2π) were allocated retrospectively, spaced equally
in time between adjacent SI displacement maxima, corresponding to end-inhale
(bin 1). The end-exhale phase was the bin closest to the displacement minimum
(usually bin 6).
3.2.2
Modified XCAT phantom
The original XCAT phantom is based on realistic non-uniform rational basis
spline (NURBS) surfaces generated as a part of the Visible Human Project
(Segars et al. 2008). The phantom can be deformed based on parameterized
motion curves of two control points: one at the dome of the diaphragm (moving
in the superior-inferior direction, SI) and another on the chest wall (moving
in the anterior-posterior direction, AP). Additionally a spherical lesion (tumor)
can be introduced which by default is synchronized to the diaphragm and chest
motion curves (Segars et al. 2010). Mishra et al. (2012) recently expanded the
capabilities of the XCAT phantom to allow irregular tumor motion data as a
direct input to drive the diaphragm and chest motion. The modified XCAT
has been applied to simulate the dosimetric effects of conventional 4DCT based
margin selection in proton therapy (Koybasi et al. 2014) and to investigate the
generation of volumetric images from electronic portal image device (EPID)
images (Mishra et al. 2014).
The procedure to synchronize the anatomic and tumor motion is as follows.
First, the SI and AP components of tumor motion are copied into the XCAT
files containing the diaphragm and chest wall motion, respectively. Second,
the XCAT is programmed to generate a three-dimensional (3D) motion field
corresponding to the maximal respiratory displacement. Third, the motion
field is analyzed to obtain the SI (AP) displacement DSI (DAP ) at the tumor
location. Fourth, the diaphragm and chest wall motion curves are multiplied by
the scalar factors 1/DSI and 1/DAP yielding synchronized anatomic and tumor
motion. The synchronization procedure was successful for all fractions except
Chapter 3. Respiratory-gated 4DCT acquisition to improve image quality
34
Patient 1 (15)
Patient 2 (20)
Patient 3 (01)
Patient 4 (43)
Patient 4 (43)
Patient 4 (43)
Patient 4 (43)
Patient 4 (43)
Patient 5 (15)
Patient 6 (22)
Patient 6 (22)
Patient 6 (22)
Patient 7 (27)
Patient 8 (33)‡
Patient ID*
-
1
1
1
1
2
3
4
5
1
1
2
3
1
1
Fraction
number
±
±
±
±
±
±
±
±
±
±
±
±
±
±
0.21
0.30
0.17
0.16
0.30
0.25
0.22
0.33
0.19
0.47
0.46
0.45
0.19
0.23
0.81 ±0.28 (cm)
1.30
1.21
0.70
0.64
0.94
0.66
0.52
0.69
0.55
0.87
0.71
0.88
0.61
1.00
Mean ±
displacement
±
±
±
±
±
±
±
±
±
±
±
±
±
±
0.60
0.42
0.54
0.58
0.86
0.51
0.69
0.75
0.58
1.11
1.18
1.19
0.72
0.52
3.43 ± 0.73 (s)
3.59
2.59
2.59
3.40
3.98
3.52
3.15
3.62
3.37
4.12
3.24
3.53
3.62
3.75
Mean ±
breathing period
±
±
±
±
±
±
±
±
±
±
±
±
±
±
0.03
0.10
0.23
0.02
0.01
0.10
0.05
0.08
0.09
0.35
0.34
0.40
0.05
0.02
0.54 ± 0.25 (cm)
0.87
0.87
0.43
0.46
0.57
0.32
0.29
0.48
0.29
0.70
0.47
1.04
0.53
0.19
Mean ±
displacement†
±
±
±
±
±
±
±
±
±
±
±
±
±
±
0.60
0.08
0.25
0.15
0.80
0.33
0.17
0.58
0.26
0.94
0.66
0.74
0.63
0.57
3.19 ± 0.48 (s)
3.28
2.55
2.73
3.30
3.34
3.79
3.00
3.52
3.18
2.86
3.21
3.11
3.43
3.39
Mean ±
breathing period‡
Table 3.1: Patient tumor motion characteristics. The table indicates the mean ± standard deviation (STD) of
peak-to-trough tumor displacement in the SI direction and the mean ± STD breathing period. *Patient ID used in
this study (ID from Suh et al. (2008)). †Calculated for the training period only (25 s). Bold numbers indicate where
the mean displacement falls below the 5 mm cutoff. ‡Not used for XCAT simulation.
MEAN
3.2. Materials and Methods
35
Patient 8 where the range of AP motion (≈ 19 mm) exceeded the maximum value
allowed by XCAT. For all other fractions, visual inspection indicated successful
synchronization of tumor motion with the surrounding anatomy. Owing to the
large tumor displacements in our study (> 5 mm SI), the simulated lesion was
placed just above the dome of the right diaphragm for all patients. Following
the synchronization procedure, the user can generate 3D images at any time
point in the tumor motion curve, with time expressed as a fraction of the total
signal length.
3.2.3
4DCT simulation work-flow
All simulations were scripted in matlab, beginning with initialization of the
modified XCAT phantom. The first 25 s of tumor motion was treated as a prescan training period to calculate the average breathing period (Tav ), as well as
the mean and standard deviation (DAVG and DSTD ) of tumor SI displacements
for each of 10 respiratory phase bins. Following this, the simulations stepped
through each data point of the tumor motion pattern in sequence, as per the
schematic of figure 3.1(b). Here, white boxes represent processes common to all
three acquisition modes. Gray boxes represent processes for beam paused and
gated 4DCT, and black boxes represent a process for gated 4DCT only.
For conventional 4DCT, the cine period Tcine = (Tav + 1s) defines the (simulated) time spent at any one couch position. This approach represents current
practice (changed little since the introduction of 4DCT), with additional 1 s
added to account for the period variations that may occur during the scan (Rietzel et al. 2005). During Tcine , kV acquisition coincides with a simulated tube
rotation time of 0.33 s; that is a new acquisition is triggered every 0.33 s along
the tumor motion trace. Each kV acquisition requires the generation of a new
XCAT volume based on the tumor displacement; from this we extract 16 consecutive axial slices (thickness 2.5 mm) corresponding to the simulated couch
position. The cine time is followed by an ’idle time’ (one second) to simulate
couch movement where no kV acquisition occurs. This is followed by the onset
of a new cine period at the next couch position; this repeats until the end of
the couch is reached. We simulate 8 couch positions each separated by 40 mm
(16 image slices).
For the beam paused and respiratory gating simulations, we implement SI
displacement gating windows with boundaries located at (DAVG ± TL ×DSTD )
for each phase bin. Here TL is the gating tolerance; gated simulations used a
small tolerance (TL=1) whereas beam paused simulations used a large tolerance
(TL=300; essentially no displacement gating). Figure 3.2 shows the mean ±
36
Chapter 3. Respiratory-gated 4DCT acquisition to improve image quality
STD of upper and lower gating window limits averaged over all tumor traces for
the case of TL=1. Averaging over all the phase bins, the mean gating window
width was 1.9 ± 0.3 mm.
For the respiratory-gated simulations, kV acquisition only occurs if the respiratory signal falls within the corresponding phase-dependent gating window.
An additional constraint is that kV acquisition is paused where an image at
a given couch position and respiratory phase has already been acquired. Unlike conventional 4DCT, the beam paused and respiratory-gated methods are
allowed a maximum cine period Tcine = 20 s. Note that the acquisition times
for beam paused and respiratory-gated 4DCT are not necessarily longer than
for the conventional 4DCT, since an early couch move is triggered once images
for all 10 phase bins are acquired.
For each simulated 4DCT scan, we require 1200 kV acquisitions to obtain
a complete 4D sampling of the thorax geometry, that is 120 slices for each 4D
phase image. In the event that an acquisition misses slices at a particular couch
position and phase, the missing slices are replaced using slices at the same couch
position in the closest available phase bin. In the event where oversampling has
occurred (multiple images at the same couch position and phase), only the first
acquired image is selected; discarded images represent wasted imaging dose.
In general our simulations can be described as ‘quasi real-time’ in the sense
that: (i) the respiratory phase was pre-calculated based on a retrospective analysis of the entire motion trace, and (ii) the generation of each XCAT volume takes
about 60 seconds of processing time, much longer than the simulated tube rotation time. Additionally, the acquisition is considered to occur instantaneously
and any blurring due to finite tube rotation speed is not modeled.
3.2.4
Generation of ground truth XCAT images
One difficulty in assessing conventional 4DCT image quality is that the mean
tumor displacement may vary between the training and imaging periods (c.f.
table 3.1); thus any ground truth image based on the training period may not
provide a fair comparison. Rather, we obtain ground truth volumes by accumulating an average of all XCAT volumes generated for each phase bin. These
ground truth images exhibit some blurring due to intra-phase displacement variation, which can be significant and differs markedly between different acquisition
modes (see example in figure 3.3). Our image quality metrics compare each simulated 4D-CT phase image to the average XCAT anatomy for that particular
combination of acquisition mode and phase bin.
3.2. Materials and Methods
37
Figure 3.1:
4DCT simulation work-flow: (a) Conventional, beam paused and
respiratory-gated 4D CT use different levels of respiratory feedback. Filled dots indicate acquisition of a 2D image slice. (b) 4D CT simulation algorithm. Gray boxes
indicate processes used for both beam paused and gated simulations; the black box
indicates a process for gated 4D CT only.
38
Chapter 3. Respiratory-gated 4DCT acquisition to improve image quality
Figure 3.2: Average upper and lower limits for phase-specific displacement gating
windows (filled and hollow circles, respectively). Crosses show the mean displacement.
Results shown for gating tolerance TL=1.
3.2.5
Segmentation of lung and tumor structures
Lung volumes were calculated via simple threshold-based segmentation, incorporating all non-zero voxels with intensities less than half of that of fat tissue.
Tumor segmentation was performed by regenerating all images with the tumor
intensified by a factor 10; voxels with intensity greater than 20% of the intensified tumor intensity were counted towards the tumor volume. Figure 3.4 shows
sample segmentation results for the conventional ground truth image from figure 3.3.
3.2.6
Image quality metrics
Image quality metrics included comparisons of simulated and ground truth images in terms of mean square error (MSE) intensity difference, and in terms
of the Dice similarity coefficient (DSC), false-positive (FP) and false-negative
(FN) rates for segmented lung and tumor volumes. These metrics are compared
using two-tailed paired t-tests, after averaging over all fractions and phase bins.
The DSC values describe the volumetric overlap of a simulated structure
(Vsim ) with the ground truth (VGT ; Zou et al. (2004)). Meanwhile FP values
give the fraction of simulated structure not overlapping with the ground truth,
and FN values give the fraction of ground truth structure not overlapping with
3.2. Materials and Methods
39
Figure 3.3: Coronal XCAT images for Patient 6, fraction 2 (end-inhale phase). The
different rows show the simulated 4DCT image (upper), the corresponding ground
truth image (middle) and difference image (lower). Columns show different 4DCT
acquisition modes: conventional (left), beam paused (middle) and gated (right).
Figure 3.4: Threshold-based segmentation of lung and tumor volumes: ground
truth image used for lung segmentation (left); tumor-intensified ground truth image
(middle); lung and tumor segmentation superimposed on original ground truth image
(right).
40
Chapter 3. Respiratory-gated 4DCT acquisition to improve image quality
the simulated structure. In other words, DSC indicates the structure that is
correctly imaged or true-positives, whereas FP and FN relate to the incidence
of duplication and truncation artifacts, respectively. These are calculated as
follows:
DSC = 2|Vsim ∩ VGT |/(|Vsim | + |VGT |)
(3.1)
F P = (|Vsim | − |Vsim ∩ VGT |)/|Vsim |
(3.2)
F N = (|VGT | − |Vsim ∩ VGT |)/|VGT |
(3.3)
Acquisition time includes the sum of cine and idle times from all couch
positions. Meanwhile, the number of acquired images can be used as a measure
of relative imaging dose, and is normalized to the number of images acquired
by the conventional method.
3.3
3.3.1
Results
Comparisons of thoracic image quality
Intensity mean square error
The boxplot of figure 3.5(a) shows thoracic MSE averaged over all fractions,
as a function of respiratory phase bin. The conventional and beam paused
methods perform similarly with a mean ± STD of (2.5 ± 0.3)×10−6 and (2.4
± 0.3)×10−6 intensity units, respectively. Gating reduced MSE by 46%, with
an average (1.3 ± 0.1)×10−6 . Compared to conventional 4DCT, gating had a
significant impact on MSE (p ∼ 10−19 ), whereas beam paused acquisition did
not (p = 0.37). Using gating, MSE was smallest around phases 4-5, similar to a
previous study where breathing phases close to end-exhale were less subject to
motion artifacts (Vedam et al. 2001). The shaded bands of figure 3.5(b) show
the mean ± STD of cumulative intensity differences averaged over all fractions
and phase bins. Qualitatively, the distributions for conventional and beam
paused acquisitions tend to overlap, indicating no difference in image quality.
Meanwhile the distribution for gated 4DCT is distinct to the others, yielding a
larger fraction of voxels below any given error value.
Errors in lung volume and structure
Figure 3.6 compares simulated and ground truth segmented lung structures in
terms of (a) absolute volume error, (b) lung DSC, (c) false positive rate and (d)
3.3. Results
41
Figure 3.5: (a) Intensity MSE between simulated and ground truth 4D CT images.
(b) Cumulative distribution of intensity errors averaged over all fractions and phase
bins.
false negative rate. Compared to conventional 4DCT, gating reduced the (mean
± STD) of absolute volume error from (1.8 ± 0.7)% to (1.4 ± 0.6)%, reduced
false positives from (4.0 ± 0.7)% to (2.6 ± 0.4)% and reduced false negatives
from (2.7 ± 0.6)% to (1.3 ± 0.2)%. Gating also improved the mean lung DSC
slightly from (0.97 ± 0.01) to (0.98 ± 0.01). For each of these comparisons,
respiratory gating had a significant impact (p < 10−10 ). Corresponding results
for beam paused 4DCT were not significantly different to conventional 4DCT
(p > 0.70 in all cases).
Errors in tumor volume and structure
Figure 3.7 compares simulated and ground truth segmented tumor structures
in terms of (a) absolute volume error, (b) DSC, (c) false positive rate and
(d) false negative rate. Compared to conventional 4DCT, gating reduced the
mean ± STD of absolute volume error from (30.2 ± 3.3)% to (18.0 ± 3.0)%,
false positives from (5.2 ± 7.4)% to (1.1 ± 2.7)% and false negatives from
(33.0 ± 3.1)% to (18.9 ± 3.0)%. Gating also improved the average tumor
DSC from (0.77 ± 0.05) to (0.88 ± 0.03). For each of these comparisons, the
effect of respiratory gating was statistically significant (p < 10−18 in all cases).
Beam paused acquisition was significantly different to conventional 4DCT only
in terms of the average false positive rate of (2.2 ± 3.4)% with p ∼ 10−9 .
In general, the volume-based metrics in figures 3.6 and 3.7 performed worse
42
Chapter 3. Respiratory-gated 4DCT acquisition to improve image quality
Figure 3.6: Comparison of simulated and ground truth segmented lung volumes for
all studied cases. (a) Mean percentage volume difference, (b) mean Dice similarity
coefficient, (c) mean false positive rates (% volume), and (d) mean false negatives (%
volume).
3.3. Results
43
Figure 3.7: Comparison of simulated and ground truth segmented tumor structures
in terms of: (a) mean percentage volume difference, (b) mean Dice similarity coefficient, (c) mean false positives (% volume), and (d) mean false negatives (% volume).
for tumor structures than for lung structures. This is due to the large amount
of intra-phase displacement variation (c.f. Fig. 2) leading to blurred/expanded
tumor structure in the ground truth images (c.f. figure 3.4). The blurring effect
appears relatively larger for tumor structure than for lung volumes. Compared
to the specified spherical lesion volume of 5.5 cm3 , the (mean ± STD) of ground
truth tumor volumes were (8.6 ± 0.7) cm3 for conventional 4DCT, (8.8 ± 0.5)
cm3 for beam paused 4DCT and (7.0 ± 0.5) cm3 using respiratory-gating. Thus
gated 4DCT not only performed better in terms of the ground truth comparisons
of figures 3.6 and 3.7, but the ground truth images themselves suffered from less
blurring as a result of the displacement gating.
44
Chapter 3. Respiratory-gated 4DCT acquisition to improve image quality
Figure 3.8: (a) Comparing segmented lung volumes for different fractions of patient
4 (end-inhale). Segmented lungs are divided into regions corresponding to true positive
(green), false positive (yellow) and false negative (red). White indicates the intensified
tumor. (b) MSE conventional to gated ratio versus the gating window (STD of SI
displacement).
Inter-fraction variability of image quality
Figure 3.8(a) compares sagittal views of segmented lungs between different fractions and acquisition modes for Patient 4 (end-inhale phase). Here the segmented lungs are divided into true positive (green), false positive (yellow) and
false negative (red). White indicates the intensified tumor. We see that gating
has had a larger effect for fraction 2 (displacement gating window 1.5 mm) than
for fraction 3 (1.2 mm). Interestingly, the conventional 4DCT image for fraction
2 shows significant shearing due to AP motion, which is reduced using SI-based
displacement gating.
Figure 3.8(b), shows the ratio of conventional-to-gated MSE as a function
of the SI gating window DSTD across all patients and fractions. We found that
91% of cases exhibited a ratio > 1 indicating reduced MSE as a result of gating.
Calculating linear correlations above and below the median DSTD of 0.68 mm
we obtained r (p) values of 0.10 (0.41) and -0.26 (0.04) respectively. In other
words, the effectiveness of gating decreased with increasing DSTD up to around
1 mm but there was not a clear trend outside of this range.
3.4. Discussion
3.3.2
45
Comparisons of acquisition time and imaging dose
Finally, figure 3.9 shows the acquisition times and relative imaging dose (i.e.
number of acquired image slices) for the three 4DCT imaging scenarios for the
first fraction of each patient. Note this data includes patient 8, for which XCAT
volumes were not generated. Compared to conventional 4DCT, the beam paused
method resulted in slightly longer acquisition times, with a (mean ± STD)
increase of (8.2 ± 10.0) % and an average reduction of the image dose by (7.7 ±
7.1) %. The beam paused method was able to achieve ideal 4D sampling, that
is it acquired exactly 1200 images across the entire range of couch positions
and respiratory phases. In contrast, the conventional method has imperfect
4D sampling, with at least 63 % of scans involving acquisition of at least one
duplicate slice.
Respiratory-gating increased the average acquisition time by (84.3 ± 39.1)%
compared to conventional 4DCT, with a maximum increase of 135 %. Gating
led to a relative imaging dose reduction of (11.1 ± 9.2) %. This is larger than the
dose reduction for beam paused 4DCT, the difference being that gated 4DCT
suffers from some missing slices. In two cases (patients 2 and 6) all three methods achieved the same relative dose. This is a result of the average breathing
period being such that the conventional method performs exactly 10 acquisitions
in each cine duration period. This is the same number of acquisitions performed
for the beam paused and respiratory-gated methods at each couch position. In
general, dose reduction is higher for patients with longer breathing period.
3.4
Discussion
In this work, we developed a simulation framework to compare thoracic image
quality between conventional, beam-paused and prospective respiratory-gated
4DCT. Our simulations have been benchmarked against earlier numerical studies of gating effectiveness by Langner & Keall (2010); similar to those studies,
we found that respiratory-gating reduced tumor volume differences by around
60 % and relative image dose by up to 17 %, but at the cost of scanning time,
which increased by up to 150 %.
Our simulations support the hypothesis that respiratory-gated 4DCT reduces normal lung image artifacts compared to conventional 4DCT. Averaged
over all simulations and phase bins, the most significant impact of respiratorygating was on the thoracic MSE, which was reduced by 46 % (p∼ 10−19 ). The
impact of gating on the accuracy of lung volume and structure was small, but
46
Chapter 3. Respiratory-gated 4DCT acquisition to improve image quality
Figure 3.9: (a) Acquisition time and (b) dose for conventional, beam paused and
gated 4D CT acquisition.
still significant: gating reduced lung volume errors from 1.8 % to 1.4 %, reduced false positives from 4.0 % to 2.6 % and false negatives from 2.7 % to
1.3 %. These differences exhibited p-values <0.02. For a typical 6L lung, these
percentage reductions correspond to respiratory-gating reducing image artifacts
by up to 90 cm3 of lung tissue compared to conventional 4DCT. This could
have important dosimetric implications where the image quality improvement
is observed near the tumor.
Compared to conventional 4DCT, beam paused 4DCT did not significantly
impact image quality or acquisition time, but still yielded similar dose reductions as for respiratory-gating. This is best observed by comparing the level
of diaphragm-blur in the ground truth images of figure 3.3. From this figure
it is apparent that beam-paused 4DCT produces different ground truth images to conventional 4DCT, which is a result of the altered acquisition timing.
However, only the respiratory-gated method reduces the diaphragm-blur (and
subsequently, reduces image artifacts in the simulated 4DCT acquisitions). This
is because respiratory-gating reduces the range of intra-phase tumor displacements during which the anatomy is actually imaged.
We observed that respiratory-gated 4DCT produced the best image quality (smallest MSE) around the end-exhale phase. In light of this, a potential
additional advantage of respiratory-gated 4DCT, though not studied here, is
the ability to limit image acquisition (and thus imaging dose) to selected respiratory phases known to exhibit better image quality. As an example, if we
3.4. Discussion
47
can prospectively gate the 4DCT acquisition at maximal exhale during normal free-breathing, then we can obtain the equivalent of a 4DCT exhale phase
image without having to acquire the other phase images. This would be useful for reducing the patient imaging dose required to perform gated treatment
planning.
A few issues still need to be addressed to achieve prospective respiratorygated (or beam paused) 4DCT in practice. One challenge is that the benefit
of respiratory gating appears to vary with the tumor motion characteristics;
we found that the gating-effectiveness decreased with increasing width of the
displacement gating window in the range 0-1 mm. Thus respiratory-gated 4DCT
could potentially benefit from phase-specific tolerance factors, providing gating
windows <1 mm across all phase bins.
Another challenge is that, in the case where respiratory motion becomes
irregular/unstable outside of the training period or undergoes a baseline shift,
then pre-calculated displacement gating windows may spontaneously become
too narrow to allow acquisition within an acceptable time-frame. A range of
mitigation methods may be possible, for example adaptive recalculation of gating windows over the course of acquisition, or adaptive-switching between the
beam paused and respiratory-gated methods, which differ only in the choice of
gating threshold and cine duration times. In cases where AP motion is larger
than SI motion, adaptive selection of the gating direction based on the direction
of largest motion could also be considered.
Beyond these considerations, we point out a few limitations of this study.
One limitation is that we not investigated the impact of different tumor locations
on the synchronized XCAT motion. In cases where motion data from upper lobe
tumors was mapped to the dome of the diaphragm, the resulting diaphragm and
chest motion may have been under-estimated. As a result our simulations could
represent a lower bound for the effectiveness of respiratory-gating.
Another limitation is that our calculations of breathing phase were based
on retrospective analysis of the respiratory motion motion. In other words, our
simulation did not model the possible effects of system latency that may impact
real-time phase calculation in a practical implementation. The use of a respiratory signal based on internal fiducials could also be considered a limitation in
that the majority of clinical respiratory monitoring systems are based on external tracking (for example via spirometry, mechanical pressure/strain belts, or
optical monitoring of surfaces/reflective blocks) that may suffer reduced positional accuracy or reduced correlation with respect to the internal tumor motion
(Ernst et al. 2012, Ionascu et al. 2007, Wu et al. 2008). It will be important
to understand the impacts of system latency, positioning accuracy and external
48
Chapter 3. Respiratory-gated 4DCT acquisition to improve image quality
/ internal motion correlation on the effectiveness of respiratory gated 4DCT.
This will require additional modification of the XCAT to encompass deformation based on various external control points and may be the subject of a future
study.
Ultimately, the question of ‘which 4DCT acquisition method is best’ is subjective to the extent that it requires some judgment of the relative importance
of image quality, imaging dose and acquisition time in a clinical work-flow. For
lung cancer radiation therapy, improved thoracic 4DCT image quality can improve not only the delivery of conformal and homogeneous dose to the tumor,
but also the accuracy of lung dose-volume parameters used to minimize lung
toxicity. We expect that these combined benefits will be worth the small potential increase in scan-time. In cases where gating avoids the need for re-scanning
due to poor image quality, the total imaging time may even be reduced.
3.5
Conclusion
We have developed a framework for simulating thoracic image quality in respiratory gated 4DCT. Compared to conventional 4DCT, respiratory gating can
reduce imaging artifacts by 1-2 % of the total lung volume, whilst also reducing patient imaging dose. However the degree of image quality improvement is
closely related to the definition of the gating window and to the tumor motion
characteristics. The potential benefits of improved normal lung imaging must
also be balanced against the potential drawback of increased acquisition time.
As an alternative, beam paused 4DCT is a simple strategy to reduce imaging
dose without sacrificing acquisition time.
Chapter 4
Image processing for 4D
proton dose calculation on
irregular breathing patterns
In the previous chapter, respiratory gated 4DCT was studied as a potential
approach to improving 4DCT image quality and reducing dose to the patient.
However, even with such improvements, motion artifacts could not be completely eliminated and the problem of breathing variability during image acquisition and therapy remains. In this chapter therefore, we investigate the
potential of a different 4D imaging modality (4DMRI) for motion imaging for
proton therapy.
This chapter is based on Bernatowicz et al. (2016) and describes 4DCT-MRI
- an image processing method to simulate many 4DCT data sets by combining
patient CT with motion extracted from 4DMRI studies. 4DCT-MRI images
have been generated using different motion extraction approaches and compared with 4DCT data of liver patients in terms of image quality, extracted
motion and proton dose calculation. This method extends the capabilities of
motion modeling for accurate 4D dose calculations by accounting for realistic
and variable motion patterns. 4DCT-MRI enables plan robustness studies to
different motion conditions and 4D dose reconstruction of delivered plan for
patient-specific QA.
49
50
4.1
Chapter 4. Image processing for 4D proton dose calculation on irregular
breathing patterns
Introduction
4DCT imaging is widely used in radiotherapy, but does not represent inter-cycle
respiratory motion variability, may contain artifacts when acquired under conditions of irregular breathing (Yamamoto et al. 2008, Pan et al. 2007, Bernatowicz
et al. 2015) and can lead to target delineation errors and systematic treatment
uncertainties (Persson et al. 2010). In particular, artifacts can have a substantial effect on the accuracy of particle therapy treatments, where the calculated
range of the particles may be adversely affected. Various motion mitigation
techniques have been developed, to either improve breathing regularity or limit
motion (Langner & Keall 2008, Damkjær et al. 2013, Mittauer et al. 2015),
at the cost of patient comfort and relying on patient compliance. Radiation
therapy under free breathing nevertheless remains a relevant goal and can even
be an advantage for rescanned particle therapy, which is dependent on sufficient statistical averaging (Knopf & Lomax 2014), or if tumor tracking can be
performed (Eley et al. 2014).
Studies show that 4D-MRI imaging is capable of resolving the irregular characteristics of organ motion (Dinkel et al. 2009). Two 4D-MRI methods have been
developed so far, both of which do not rely on an external surrogate for image
sorting, but on a navigator slice (Von Siebenthal et al. 2007), or on image content only (Paganelli et al. 2015). While being a promising technique for motion
evaluation, MR imaging alone does not provide density information for accurate
dose calculations, with significant dose differences being reported compared to
CT based calculations for photons when replacing segmented MR anatomy with
bone and water-equivalent tissue (Eilertsen et al. 2008).
The conversion of MRI to density data can be done using either voxel-based
or anatomy-based segmentation methods. In the first case, voxels are directly
characterized by different tissue classes, but discriminating between air and bone
is still a major difficulty (Catana et al. 2010). On the other hand, anatomy-based
methods match reference density datasets (either the patient’s CT or a MRI-CT
atlas) using deformable registration to a new patient (Dowling et al. 2012). Such
methods were recently extended to 4D using motion models derived from 4DCT
(Ehrhardt et al. 2011) or 4D-MRI (Marx et al. 2014). The latter approach is
particularly interesting, as it extends the capabilities of 4DCT imaging by also
being able to represent multiple breathing cycles without extra imaging dose.
In this work, we expand on the study already published by Boye et al.
(2013a), where motions extracted from 4D-MRI of volunteers were used to deform a single phase CT data in order to generate what we call 4DCT-MRI
images. 4DCT-MRI is compatible with dose calculation algorithms, but can
4.2. Materials and Methods
51
also represent multiple, variable breathing cycles. Clinical applications of this
approach include pre-treatment quality assurance procedures to test the robustness of a particular mitigation technique to realistic breathing motions and
reconstruction of delivered ‘4D dose of the day’ based on the subject-specific
motion model. As such, two different motion extraction approaches have been
tested, subject-specific and population-based, on six liver patient cases. The
results have been compared to 4DCT using three different metrics: (1) spatial
accuracy of the images and extracted motions, (2) comparison of Hounsfield
Units (HU) and proton range, and (3) scanned proton 4D dose calculations
(4DDC).
4.2
4.2.1
Materials and Methods
4DCT-MRI concept
The concept of 4DCT-MRI has been previously proposed by Boye et al. (2013a),
and involves an originating single phase CT that is deformed using separate motion vector fields to form one or more simulated 4DCT data sets. The advantage
of this is that the two information streams (density and motion) can be obtained
from modalities that are best suited for these purposes. For instance, CT data
is the modality of choice for imaging density, whereas motion can be more conveniently imaged using 4D-MRI - a no-dose imaging technique, which allows for
the acquisition of motion data over multiple breathing cycles of patients and/or
volunteers (Von Siebenthal et al. 2007).
4.2.2
Data sources
CT data
For one patient, both 4DCT (SI motion 10 mm) and 4DMRI images have been
analyzed to demonstrate the potential of subject-specific 4DCT-MRI. The reference CT was defined as the end-inhale phase of 4DCT. As this patient was
treated at our institute for a non-liver tumor, a simulated target volume was
drawn in the right liver lobe (CTV= 38 cm3 ) for 4DDC.
In addition, 4DCT-MRI data has been derived for 5 liver patients with different tumor sizes [range: 70-403 cm3 ; mean: 191 cm3 ] and motion characteristics
[range: 9-17 mm; mean: 13 mm]. 4DCT studies were acquired for all cases and
all had fiducial markers implanted in the liver. The end-exhale phase has been
used as the reference CT data for 4DCT-MRI generation.
52
Chapter 4. Image processing for 4D proton dose calculation on irregular
breathing patterns
Motion data
Motions have been extracted from two sources: (1) original 4DCTs from the
patients, used as a reference to test our methodology and the accuracy of the
deformation algorithms, and (2) from a library of liver motions, acquired either
from 4D-MRI of 13 healthy volunteers (Von Siebenthal et al. 2007) (for the
population-based approach) or from the same patient (for the subject-specific
approach). The motion library is a collection of XYZ displacements of corresponding liver points (see section 4.2.4) extracted from multiple breathing cycles
of each imaged volunteer/patient (mean: 504; range: [221,851]). In total more
than 6000 motion cycles are available in this library.
In order to make a fair comparison to the original (reference) 4DCT data sets,
a ’most similar’ motion to that of the 4DCT must be chosen from the library.
For the subject-specific approach, the motion extracted from the patient’s 4DCT
was compared to the ’most-similar’ motion found in the 464 breathing cycles
available in the 4DMRI data of the same patient, whereas for the populationbased approach, 4DCTs of five patients were compared with the ‘most-similar’
motions found from the volunteer library.
4.2.3
Motion extraction
4D-MRI was generated by retrospective sorting of the MRI sequences, based
on the liver state determined from simultaneously acquired navigator slices
(Von Siebenthal et al. 2007). Motion was then extracted using a combination
of a multi-resolution affine registration and a B-spline non-rigid registration.
The correlation coefficient was used as a similarity measure without additional
smoothing of the deformation fields, leading to registration accuracies of the
order of one voxel or smaller (Von Siebenthal et al. 2007). Finally, this motion data was converted to match the number of phases in the corresponding
reference 4DCT.
The reference motion fields were obtained by registering patient 4DCT images using the open source software (http://plastimatch.org). Five stages
of multi-resolution optimization with gradual refinement of the grid spacing between control points in the B-spline grid have been used with a mean square
error cost function and regularization lambda of 0.05 (Shackleford et al. 2010,
2012). Registration was validated using marker position errors (MPE), calculated as the absolute difference between estimated motion (i.e. registration
result) and the ’ground truth’ motion (i.e. displacement of implanted fiducial
markers obtained directly from the 4DCT images).
4.2. Materials and Methods
4.2.4
53
MRI-CT liver correspondence
Mapping motion information from 4D-MRI to CT images was done based on
a pre-defined correspondence between livers (Boye et al. 2013a). This assumes
that certain corresponding points will move similarly across different livers, e.g.
points on the anterior surface will show an SI-AP motion, as they slide along the
abdominal wall. As such, a fine mesh (>500 points) are automatically created
by gradual refinement of a coarse liver prototype to the segmented liver surface
of each patient, see Von Siebenthal et al. (2007).
In order to determine the ’most-similar’ motion, a direct comparison of motion vectors at corresponding points was performed. Once found, a continuous vector field was generated by a B-spline based interpolation using Insight
Segmentation and Registration Toolkit software (ITK; US National Library of
Medicine; http://itk.org).
4.2.5
Comparing 4DCT-MRI and 4DCT
Spatial evaluation
4DCT and 4DCT-MRI images were compared visually (see figure 4.1) and then
using Summed-Square Differences (SSD), calculated for displacements over all
breathing phases for points inside and on the liver surface. SSD was used as the
metric for assessing ‘most similar’ motion from the motion library. Additionally,
for the population-based approach, motion data was evaluated at the fiducial
markers positions. Marker position errors (MPE) have been calculated as an
absolute difference between the motions extracted from 4D-MRI and the ’ground
truth’ motion extracted from the originating 4DCT data (figure 4.2(b)).
Density information
Deforming a static CT with externally derived motion vectors may result in
somewhat different density information than for 4DCT acquisition (figure 4.1).
Therefore, density changes for different incident field directions were quantified
using water-equivalent range (WER) calculations through both the reference
and 4DCT-MRI data sets. WERs were calculated by converting HU values to
proton stopping power using a clinically validated calibration curve and were
evaluated in the CTV region only.
54
Chapter 4. Image processing for 4D proton dose calculation on irregular
breathing patterns
Figure 4.1: Comparison of 4DCT images with generated 4DCT-MRI for (a) patientspecific and (b) population-based approach. Note that biggest differences occur outside
the liver volume (marked with red contour).
4.2. Materials and Methods
55
Figure 4.2: (a) Reference CT (end-exhale [green]) overlaid with the moving phases
deformed according to motion extracted from 4D-MRI (magenta). Matching regions
are shown in gray scale. (b) Box plot of marker position errors calculated for motion
extracted from the generated 4DCT-MRI images of 5 liver patients. Boxes correspond
to the 75th percentiles of the MPE at different breathing phases.
56
Chapter 4. Image processing for 4D proton dose calculation on irregular
breathing patterns
Dosimetric comparison
Finally, 4D dose calculations (4DDC) for scanned proton beam have been performed for both 4DCT and 4DCT-MRI, using the deforming dose-grid approach
(Boye et al. 2013a, Zhang et al. 2013) and ray-casting model (Schaffner et al.
1999), as used by the clinical treatment planning system at Paul Scherrer Institute (Lomax et al. 2004). In this algorithm, the positions of calculation points
are adjusted by consideration of time-dependent displacements in directions orthogonal to the central beam axis, with lateral displacements obtained from
extracted motion fields, and water-equivalent depths from pre-calculated density maps at each motion phase. For all cases, a breathing period of 5 s was
assumed and a planning target volume (PTV) was defined using an isotropic expansion of 10 mm around the CTV. 4DDC’s have been calculated for four plans,
three single field (P1 anterior-posterior, P2 right lateral and P3 anterior-superior
oblique) and one three field plan (P4, combining P1-P3). For all 4DDC’s, an
energy switching time of 80 ms has been assumed (Pedroni et al. 2004, Bernatowicz et al. 2013).
4.3
4.3.1
Results
4DCT-MRI subject-specific
4DCT and the ‘most similar’ 4DCT-MRI for the subject-specific approach are
compared in figure 4.1(a): motion excursion of 4DCT-MRI resembles that of
the original 4DCT. When comparing corresponding points in the liver between
the two, SSD differences as low as 0.8 mm were found.
On comparison of HU differences, a good agreement was observed within
the liver volume (figure 4.1(b)). However, larger differences appeared near ribs,
in particular at the anterior site subjected to respiratory motion. Calculated
WER’s between 4DCT and 4DCT-MRI were found to be <2 mm.
Figure 4.3 compares dose distributions calculated for 4DCT and subjectspecific 4DCT-MRI. Mean dose difference are <1 %, and 93(±8) % of CTV
points pass the gamma evaluation (3%/3mm).
In summary, for the subject-specific approach, 4DCT-MRI has been found
to be an excellent surrogate for modeling motion.
4.3. Results
57
Figure 4.3: 4D dose calculation results for 4DCT and patient-specific 4DCT-MRI
for a single-field plan (P3) and three-field plan (P4); insert shows 3%/3mm gamma
analysis – agreement is marked in blue and solid white line defines CTV.
58
4.3.2
Chapter 4. Image processing for 4D proton dose calculation on irregular
breathing patterns
4DCT-MRI population-based
Examples of 4DCT and 4DCT-MRI for the population-based approach are presented in figure 4.1(b). Accuracy of matching of the ’most similar’ motion from
the 4D-MRI motion library (compared to 4DCT extracted motion) varied both
on breathing phase (figure 4.2(a)) and from patient to patient, with SSD’s within
the liver ranging from 1 to 12 mm (mean 5.5 mm).
For the inserted fiducials, mean MPE’s were less than 4 mm in 4 of 5 cases,
figure 4.2(b), which is slightly larger than the voxel size of the CT (voxel diagonal
dimension ≈3 mm). For one case however (patient 4) MPE of 6 mm and a
maximum of 9 mm were found.
Example of HU differences are shown in figure 4.1(b), and WER analysis
for five patients and all phases of their 4D data in figure 4.4. Mean absolute
WER deviations were 2 mm with a maximum difference of 25 mm (34 % of endexhale WER). For maximum ranges, mean differences were 2 mm ranging to a
maximum of 24 mm, whereas for minimum range, mean differences were 3 mm,
but differences of up to 63 mm were also observed. All worst case differences
were found for the same patient (patient 4).
Figure 4.5 shows a comparison of 4DDC results for 4DCT and populationbased 4DCT-MRI. Mean dose differences were about 1 %, except for case 4,
where they were up to 4 %, see figure 4.6(a). On average 79(±14) % of points
within the CTV met the gamma criteria of 3%/3mm, figure 4.6(b). However, for
3/5 cases with ‘most similar’ SSD’s <2 mm, mean dose differences were lower
than 1 % and more than 80 % of points met the gamma criteria of 3%/3mm for
single field plans.
4.4
Discussion
We envisage three main clinical applications for 4DCT-MRI: (1) to evaluate
plan robustness under different breathing, for instance using an appropriate
model to select different percentiles of motion vectors from the library, (2) to
improve volume delineation, employing the superior MRI contrast in liver and
realistic motion excursions, particularly needed for ITV definition, and (3) to
reconstruct the ’4D dose of the day’, based on the time-resolved delivery logfiles (Tyagi et al. 2012) and motion reconstructed from a surrogate combined
with a motion model (Zhang et al. 2013). Detected dose errors could then be
compensated by plan adaptions, and applied in subsequent treatment fractions.
Additionally, online plan QA is an interesting idea, where tools similar to those
4.4. Discussion
59
Figure 4.4: (a) Mean, (b) maximum, and (c) minimum water-equivalent range
(WER) differences between 4DCT and 4DCT-MRI images, calculated in the CTV of
5 liver patients, at 10 respiratory phases, and for 3 different field directions (from
anterior, from the right and a non-coplanar field in between). (d) An example of
original and warped image liver data (patient 2).
60
Chapter 4. Image processing for 4D proton dose calculation on irregular
breathing patterns
Figure 4.5: Example of the 4D dose calculation results for 4DCT and 4DCT-MRI
for case 2; insert shows 3%/3mm gamma analysis – agreement is marked in blue.
Figure 4.6: Comparison of 4DCT and 4DCT-MRI: (a) mean dose differences and (b)
CTV volume meeting the 3%/3mm gamma criteria. C1-C5 refer to the five different
cases (patients) studied and P1-P4 to treatment plans.
4.4. Discussion
61
mentioned above (but very fast) would allow treatment intervention in case of
erroneous dose delivery. Although, there are many challenges to be addressed
before implementing such sophisticated methods, similar concepts are currently
being developed for the photon therapy applications (Ravkilde et al. 2013).
We have shown that the subject-specific 4DCT-MRI provides an excellent
surrogate to 4DCT with the big advantage being 4DCT-MRI data can be generated for multiple and different breathing cycles. Although results for the
population-based approach are not as good, such an approach could nevertheless still be of use for a-priori robustness tests of the sensitivity of plans to a
range of different, realistic breathing scenarios.
However, all results are dependent on the quality of the reference CT image, as seen for example in case 4 of the population-based approach, where
the quality of the 4DCT images was poor to start with, with high-intensity
streaking artifacts, invalidating deformable registration results. Moreover, such
data should not be used for realistic planning and dose calculations. In general,
the end-exhale phase is shown to have the least motion artifacts (Pan et al.
2007) and is preferable for the 4DCT-MRI generation. However, if good image
quality is sustained, other breathing phases can be used (as demonstrated on
end-inhale with the subject-specific approach). In clinical practice, 4DCT-MRI
images could be also generated based on a reference CT acquired with a fast CT
scan or a gated-CT protocol, which has a clear advantage in terms of imaging
dose compared to 4DCT acquisition.
The current model and analysis is however specific to the liver. The method
could potentially be extended to other sites/organs if tissue properties are preserved. This could be a challenge in lung tissue, where WER changes may be
larger: Mori et al. (2009) reported WER changes of 1.5 cm in lung between different CT scans. First results on generation of the lung 4DCT-MRI is presented
in appendix 8.
4DCT-MRI images were matched and evaluated based on the current clinical ’gold standard’ - 4DCT images. 4DCTs represent phases of a single, averaged breathing cycle, which may underestimate the target excursion and exhibit
imaging artifacts that present challenges to deformable registration algorithms
(Persson et al. 2010). However, it is difficult to know whether motion extracted
from 4DCT or 4D-MRI is the best representation of the ground truth motion.
This is one of the reasons why validation of 4D medical image registration is
very challenging. The use of sophisticated, anthropomorphic moving phantoms
(Perrin et al. 2014) could be a first-step towards development and evaluation of
advanced motion imaging techniques.
62
4.5
Chapter 4. Image processing for 4D proton dose calculation on irregular
breathing patterns
Conclusion
The 4DCT-MRI method extends the capabilities of 4DCT imaging by generating
multiple-breathing cycle image sets of the patient. It represents the variability
of deformable liver motion on high-quality images with few discontinuities and
relevant density information. The potential clinical applications of 4DCT-MRI
include, but are not restricted to, plan robustness evaluations under variable
motion conditions and patient-specific QA for moving targets. 4DCT-MRI can
also be directly employed in clinical research studies on advanced 4D treatment
planning and delivery techniques taking into account patient movement.
Chapter 5
Advanced treatment
planning with 4D
optimization
In the previous chapters 3 and 4, methods for improving 4D imaging were investigated, whilst in chapter 2, the use of motion specific margin expansions
for PBS proton therapy were studied. However, the use of ITV’s/raITV’s, although simple and effective, is not necessarily the optimal approach to motion
mitigation. For instance, if accurate and realistic motion data is available from
advanced 4D imaging, this opens the door to more advanced mitigation approaches. As such, the purpose of this chapter is to determine whether a new
four-dimensional (4D) optimization approach for scanned proton beams could
reduce dose to critical structures in proximity to moving targets, while maintaining effective target dose homogeneity and coverage. The proposed approach
has been tested using both a simulated phantom and a clinical liver cancer
case, and allows for realistic 4D calculations and optimization using irregular
breathing patterns extracted from e.g. 4DCT-MRI (see chapter 4). Moreover,
techniques to improve the robustness of delivery of such an approach by true
intensity-modulated delivery are discussed.
63
64
5.1
Chapter 5. Advanced treatment planning with 4D optimization
Introduction
In proton therapy, two different delivery techniques are currently employed: (1)
passive scattering (PS), where the whole tumor volume is irradiated simultaneously and dose conformity is achieved using collimators and compensators, and
(2) pencil beam scanning (PBS), where a proton beam is scanned sequentially
across the tumor volume. Up to recently, most particle treatment institutions
employed PS, but PBS had become the preferred technique in most planned and
currently developed centers (http://ptcog.org) due to its inherent flexibility
and ability to deliver intensity-modulated proton therapy (Lomax et al. 2004).
However, motion is a major problem for many forms of external beam radiotherapy and is particularly challenging for pencil beam scanned (PBS) proton
therapy. Currently, most proton institutions treat mobile tumors (e.g. in liver
or lung locations) are using PS, where range uncertainties are accounted for by
extending the proximal and distal edge of the target volume (Fukumitsu et al.
2009, Nichols Jr et al. 2012). Similarly for PBS, the concept of the field-specific
PTV has been proposed to compensate for range uncertainties (Park et al. 2012,
Knopf et al. 2013, Schuemann et al. 2014). However, the major concern of PBS
when treating mobile tumors is the ’interplay effect’, an effect resulting from the
interference between the sequential delivery of individual pencil beams with tumor motion, which can severely deteriorate the dose distribution (Phillips et al.
1992, Dowdell et al. 2013). This effect cannot be corrected by margin expansion and consequently, several mitigation techniques (e.g. rescanning, gating,
tracking or combinations therewith), are currently under investigation.
Both simulations and experiments have shown that rescanning has the capability to mitigate the interplay effect, with the degree of mitigation being
related to the patient’s geometry, motion and the speed of the delivery system
(Knopf et al. 2011, Bernatowicz et al. 2013, Grassberger et al. 2015). Despite
its effectiveness against the interplay effect however, statistical averaging of the
dose has the disadvantage of still compromising the delivered plan quality in
terms of dose conformity and sparing of healthy tissues. Alternatively, gated
delivery has been shown to be attractive in simulation studies, as using small
gating windows allows to retrieve dose homogeneity close to static plans, if only
at the cost of substantially increased treatment times (Bert & Durante 2011,
Zhang et al. 2015). However, gating does not always provide a good result for
PBS deliveries, as the simplistic inclusion of a 1D motion signal (as provided by
most motion monitoring systems) does not always allow for proper handling of
motion induced tissue deformations, a problem also found for tumor tracking,
Van de Water et al. (2009)). This can cause discrepancies between the actual
5.2. Materials and Methods
65
anatomy and the one predicted from the motion signal, which directly translate
into dosimetric errors (Zhang et al. 2015).
As such, an intriguing idea is to ensure plan integrity even in the presence
of motion, by including motion into the plan optimization stage in order to recover dose conformity and homogeneity for moving targets without the need for
tumor tracking. This idea has been pursued in simulations and experiments for
different delivery modalities of photon therapy by a number of authors (Keall
2004, Schlaefer et al. 2005, Ma et al. 2009, Falk et al. 2010, Li et al. 2011, Trofimov et al. 2005), but only a few publications have focused on 4D optimization
methods for scanned particle therapy. Graeff et al. (2013) developed a 4D optimization method, in which the beam weights are optimized in multiple sectors
of the target, with each sector being assigned to an individual motion phase.
Alternatively, it has been proposed to assign each pencil beam (PB) to a specific
motion phase (Graeff 2014). However, both approaches require synchronization
of delivery with the patient’s motion phase and as such, a fast gating or ’multigating’ approach has been proposed (Graeff 2014). Eley et al. (2014) modified
the 4D optimization approach by adjusting PB weights for all pencil beams at
each motion phase, allowing for good dose coverage and tissue sparing. As such,
the necessary synchronization with delivery was demonstrated by combining 4D
optimization with tracking.
In this work, we propose an alternative 4D optimization method for scanned
particle therapy. The presented concept is not necessarily bound to the motion
phase definition and incorporates both irregular motion patterns and the delivery dynamics of the treatment machine into the plan optimizer. For this, fast
4D proton dose calculations have been used, based on a deforming calculation
grid, together with varying motion patterns extracted from 4DCT-MRI (Boye
et al. 2013a, Bernatowicz et al. 2016). This approach has been used to study
the sensitivity of 4D optimized treatment plans to variations in the motion
characteristics that could occur during delivery.
5.2
5.2.1
Materials and Methods
4D optimization approach and implementation
A gradient based optimization algorithm used in our in-house planning system
for PBS proton therapy (Lomax 1999, Albertini et al. 2009) has been modified
to work with the 4D dose calculation described in Chapter 6. This is based on
a cost function F (ω) as follows.
66
F (ω) =
Chapter 5. Advanced treatment planning with 4D optimization
LT
X
i
2
gT,i
(PT,i − Di )2 +
LX
OAR
2
gOAR,i
(POAR,i − Di )2 H(POAR,i − Di ) (5.1)
i
The first part of equation 5.1 defines an objective function that minimizes
the difference between the prescribed (PT,i ) and calculated dose (Di ) at a grid
point i for all grid points in the target volume (LT ). The second part relates
to organ at risk (OAR) constraints and a step function ‘weighting’ H, with
H = 1 when the calculated dose in the OAR is greater than the prescription
constraint (POAR,i ). Additionally, gi is an associated importance factor for the
OAR constraint in relation to the importance assigned to target coverage. For
the target volume, g = 1 and different values (<> 1) can be assigned to different
OARs in order to adjust their relative ‘importance’ during the optimization.
Finally, (w) are the pencil beam fluences which are updated in every iteration
and contribute to the newly calculated 4D dose as follows:
Di = D4D =
M
X
d4D,i,j ωj
(5.2)
j
In this di,j is the dose contribution from a pencil beam j of all pencil beams
M to grid point i.
The optimization proceeds iteratively using a gradient based approach, similar to that proposed by Bortfeld et al. (1990), and modified with a damping
factor introduced by Lomax et al. (1996) to guarantee fast convergence. The
following equation summarizes the final update solution:
PL 2 2 Pi
i gi di,j Di
ωj (k + 1) = ωj (k) PL
2 2
i gi di,j
(5.3)
In this work, our 4D dose calculation has been implemented in the optimization routine by incorporating the time dependent offsets and WERs required to
calculate the dose contributions on each iteration of the optimization process
(see details in chapter 6, equation 6.2). An important feature of this, is that
the time dependent information represents motions away from the reference geometry, with the final solution then representing a single, 4D treatment plan,
which is easy to evaluate using a single set of clinically delineated volumes. In
addition, by the use of a deforming dose grid (rather than multiple instances
of the patients geometry in the form of time resolved CT data sets), our 4D
5.2. Materials and Methods
67
Figure 5.1: (a) Lateral profile through an iso-energy layer in the target volume.
The 4D optimization adjusts the beam weights (represented by Gaussians) in every
iteration to minimize differences between the prescribed dose (black) and 4D calculated dose (red) accumulated on the reference geometry. (b) Evolution of the cost
function F (ω) with iterations of the 4D optimization. Every few iterations, the delivery time-line is adjusted according to the updated beam weights (peaks). During the
optimization, differences between the updated cost function get smaller and begin to
stabilize.
dose calculation is fast, with the dose calculation being the same as for a 3D
calculation, which is an important feature for such an iterative solution.
In the description so far, the optimization routine assumes that the time-line
of delivery does not change as part of the optimization procedure. However, as
the fluences of each pencil beam are adjusted in each iteration, then also the
time to deliver each pencil beam will also change (assuming the beam intensity
remains the same, which is currently the case for delivery on Gantry 2 at our
facility). Therefore, another level must be added to the optimization routine,
with the time of delivery of each pencil beam being updated between iterations
to reflect the changing beam-on time resulting from the optimization process.
A graphical representation of the 4D optimization and corresponding evolution of the cost function is shown in figure 5.1.
68
5.2.2
Chapter 5. Advanced treatment planning with 4D optimization
Proof of principle and sensitivity to motion parameters
As a proof of principle of this 4D optimization approach, a 4D optimized plan
has been calculated for a simple digital phantom, consisting of a spherical target
moving rigidly with two different motion amplitudes (5 mm and 10 mm) and a
motion period of 4.5 s (see results in figure 5.2). In addition, it has also been
applied to a liver cancer case with deformable motion having been extracted
either from 4DCT (repeated single breathing cycle with an amplitude of 14
mm) or from 4DCT-MRI with irregular and multiple breathing cycles. Maximum displacement of the latter was 16 mm (see results in figure 5.4). A 4.5 s
motion period was used in 4DCT simulation, extracted based on the patient’s
respiratory signal recorded during the imaging session.
The resulting, optimized plans were then further analyzed under conditions
of potential breathing and delivery variations. As such, 4D dose calculations
were performed with breathing patterns that were systematically changed from
those used for the optimization. Lastly, the effect of different numbers of 4D
iterations, together with updating the time-line of delivery in the optimization
routine, has also been studied.
5.3
5.3.1
Results
Proof of principle
For the digital phantom, a 3D optimized plan is severely compromised when
simulated under conditions of motion using the 4DDC. However, when optimized in 4D, the resulting 4DDC distribution provides a highly conformal and
homogenous dose distribution (figure 5.2), indicating the potential of 4D optimization for motion mitigation, at least if the motion characteristics during
delivery are the same as that assumed during the optimization.
The potential advantages of 4D optimization also remain for the more clinically realistic liver case, where deformable and both regular (4DCT) and irregular motions (4DCT-MRI) have been used during the optimization process, see
figure 5.4. However, a dose ‘hot-spot’ at the field entrance has been observed
in these plans, similar to those reported for tumor tracking with PBS particle
therapy (Bert et al. 2010, Zhang et al. 2015) - and referred to as the ‘inverse
interplay’ effect.
From these preliminary examples, it is clear that 4D dose distributions resulting from our 4D optimization can achieve almost the same quality as static
5.3. Results
69
Figure 5.2: Comparison of dose distributions from a simulation in a moving software
phantom. The 4D optimized solution provides an excellent result in presence of motion.
70
Chapter 5. Advanced treatment planning with 4D optimization
(a) phantom
(b) patient 4DCT
(c) patient 4DCT-MRI
Figure 5.3: Dose-volume histograms of target volumes obtained from 4D dose calculations based on: (a) simulated phantom 4DCT, (b) liver cancer patient 4DCT and
(c) liver patient 4DCT-MRI.
plans, independently on the studied geometry/anatomy or selected motion (regular and irregular), as shown in figures 5.3 and 5.5. For instance, when comparing the 4D optimization results in presence of motion, with the static (3D
optimized) plans, an excellent dosimetric agreement can be observed in the target volume. In addition, the 4D optimized plans results in lower dose to the
heart, but an ‘inverse interplay’ effect could be observed for some field directions. However, it is expected that these effects will be significantly reduced if
a realistic multiple field plan is considered (typically three fields per plan are
used clinically).
5.3.2
Fast 4D optimization
The inclusion of detailed delivery time-lines and dynamics is a novel feature
of the 4D optimization procedure described here. As such, the effect on the
speed of optimization (e.g. number of iterations required to converge) has been
studied as a function of how these time-lines are updated during the optimization
process. For this, the liver patient with motion extracted from 4DCT has been
used. The DVHs obtained from the 4D dose calculation, based on exactly the
same motion as used during the optimization, are presented in Figure 5.6.
As described above 5.2.1, updating the delivery time as a function of the
new beam weights resulting from the optimization results in different pencil
beam offsets at the time of delivery of the particular pencil beam even though
motion of the anatomy remains the same. However, with an increasing number
5.3. Results
71
Figure 5.4: Comparison of dose distributions for the liver cancer patient. 4DCT
information was used as a representation of an averaged, regular breathing cycle and
4DCT-MRI for optimization/calculation under irregular breathing conditions.
72
Chapter 5. Advanced treatment planning with 4D optimization
Figure 5.5: Dose differences between 4D optimized treatment fields calculated using
optimized motion and the same field calculated assuming no motion (3D optimisation
and delivery). Different fields were evaluated: F0: anterior-posterior (0 deg), F2:
right-left (90 deg) and F1: in between F0 and F2 (45 deg)
of iterations, these differences reduce and the solution converges (figure 5.6(a)),
and after 600 iterations in this example, a substantially improved DVH can be
obtained.
Interestingly however, when performing the 4D dose calculation using the
same 4D optimized beam weights, but keeping the time-line of delivery fixed
to that of the original, unoptimized plan, the solution converges faster (figure 5.6(b)). In this case, after only 60 iterations (its), a very conformal plan for
the motion scenario can be obtained, with the 4D optimization time for a single
field being <3 minutes on a 32-bit Linux computer.
In addition, time-line adapted optimization (600 its) resulted in much higher
beam weights than in the initial plan, as shown in Table 5.1. Moreover, larger
differences can also be observed between pencil beams (PB), indicating steeper
inter-PB dose gradients which could decrease overall plan robustness. In contrast, the time-line fixed approach (60 its) resulted in a solution closer to the
initial beam weight distribution. Although the consequences of this observation
at the moment are unclear, it nevertheless shows that the inclusion of exact
time-line dynamics into the optimization process adds an additional degree of
freedom to the optimization problem which could be potentially exploited in
the future to further enhance the effectiveness of 4D optimizations.
5.3. Results
73
(a) Updated time
(b) No time update
Figure 5.6: The effect of number of optimization iterations and time-line updates
on the 4D optimization represented by dose-volume histograms from 4D calculation
results performed on the same motion as that used during optimization.
Table 5.1: Beam weights of 4D vs 3D optimized plan of liver cancer patient.
3D (60 its)
4D (60 its)
4D (600 its)
Maximum BW
Sum of BWs
0.1081
0.1371
0.4146
47.2447
49.7794
51.2927
74
Chapter 5. Advanced treatment planning with 4D optimization
(a) phantom
(b) phantom
(c) patient 4DCT
Figure 5.7: Sensitivity of 4D optimized fields to changes in motion amplitude and
period. D5 − D95 represents target dose inhomogeneity.
5.3.3
Sensitivity to motion parameters
So far, we have looked into the performance of 4D optimization under the conditions that the same motion pattern as used in the optimization routines was also
used for computing the final 4D dose calculation. In this section, we now investigate the robustness of the 4D optimization to potential differences in motion
characteristics during delivery to those assumed during optimization.
Two 4D optimized plans in the phantom (optimized on 5 mm and 10 mm
motions) were tested using the 4D dose calculation with varying motion amplitudes and periods, each being varied one at a time (figure 5.7(a-b)). Although
it is intuitive that optimizing on small motion does not guarantee a good plan
quality for larger motions, our results also suggest that optimizing on larger
motion amplitudes does not necessarily help if the actual motion is smaller.
Moreover, respiratory period differences larger than 20 ms of those assumed in
the optimization already result in loss of target dose homogeneity by more than
5 %. Figure 5.7(c) shows results of the 4D calculation for the liver case using
motion from 4DCT, but again with varying motion period with respect to the
optimized motion scenario. Similar sensitivities to those for the phantom study
are observed.
In summary, based on this sensitivity analysis, an acceptable dose homogeneity (<5 %) could only be found for differences in amplitude of up to ≤1 mm
and for changes in the breathing period of <20 ms in both cases in comparison
to the motions used during optimization.
5.4. Discussion
5.4
75
Discussion
This chapter has reported on the development of a fast and flexible 4D optimization approach for proton pencil beam scanned therapy. In contrary to
previously published work, our approach allows for optimization on variable
breathing patterns and inherently includes the detailed temporal dynamics of
PBS proton therapy delivery. As such, it can be considered a first 4D ‘direct
optimization’ approach. In addition, by effectively de-coupling the delivery dynamics from assumptions of synchronization with the breathing cycle at the
optimization stage, we have added flexibility to the 4D optimization process.
For instance, with this approach, we can easily incorporate other motion mitigation processes, such as re-scanning, gating and tracking, directly into the
optimization process, opening up a whole new area of potential improvements
for motion mitigation at the planning stage.
In this work, a first and preliminary implementation of the 4D optimizer has
been tested on a variety of breathing curves, phantoms and patient geometries
as a proof of principle study of the method. The resulting 4D optimized plans
have been found to be as conformal and homogenous as statically optimized
and applied plans, but have also been found to be very sensitive to variations
in motion between the planned/optimized scenario and possible variations in
the motion characteristics on the day of delivery. As such, 4D optimization
only provided good results if the daily motion would not vary more than a
millimeter in tumor excursion and tens of milliseconds in respiratory period or
phase shifts. Although these results show that these initial 4D optimized plans
are very sensitive to motion variations, perhaps when used in combination with
breathing-control approaches such as audio-visual patient guidance and training,
together with low-latency motion monitoring systems (to reduce phase shifts),
sufficient agreement between optimized and actual motion could be achieved
to make this clinically interesting. This line of thought needs however to be
pursued in more detail.
Nevertheless, and as mentioned above, the flexibility of our 4D optimization
approach will also allow for the future incorporation of other motion mitigation
techniques, such as re-scanning, gating and tracking, directly into the optimization process, potentially allowing for more robust 4D optimized plans. However,
one should also keep in mind that all 4D calculations implemented in the optimization routine and 4DDC’s used here are also subject to errors due to 4D
imaging, deformable image registration and uncertainties in the true dynamics
and time-lines of the delivery. We will return to these uncertainties in the next
chapter of this thesis.
76
Chapter 5. Advanced treatment planning with 4D optimization
The optimization algorithm used here is based on the algorithm used clinically at PSI and hence exhibits similar limitations. For instance, it is based
on the gradient descent algorithm, where the solution might not always reach
a global minimum. Given the extra ‘dimension’ in the optimization of a changing delivery time-line between iterations, it is possible that the solution space
becomes non-simple and has multiple local-minima, and more sophisticated optimization approaches (such as simulated annealing) may be required. In addition, the optimization uses a damping factor to facilitate fast convergence,
which effectively restricts large increases in the weight of low weighed pencil
beams, and with increasing number of iterations, high-weighted PB’s are further enhanced, with low weighted ones eventually being discarded as they drop
below the minimal fluences that can be effectively applied.
On the other hand, we have observed that the 4D optimization solution converges much faster when delivery time-line is assumed to be constant. This is
also a feature of this approach, which could be further exploited. After all, the
delivery time-line is something that can be varied by the treatment machine,
for instance by varying the beam intensity during delivery of each pencil beam
(a feature that has just been incorporated into the treatment control system
of Gantry 2 at PSI, and is under investigation at other centers (Lüchtenborg
et al. 2011)) and/or adjusting scanning speed between spots (an approach also
being investigated at PSI for continuos and intensity modulated line scanning,
see Klimpki et al. (2016)). Indeed, such ‘temporal-modulation’ of delivery has
further potential applications in relation to 4D optimization and motion mitigation. For instance, it could potentially also be used to compensate for variations
in the periodicity of motion, by keeping the synchronization of PB delivery to
the phase of the breathing cycle, by speeding-up or slowing-down delivery based
on motion surrogate measurements.
5.5
Conclusions
4D optimization can reduce the maximum dose to critical structures and healthy
tissue near a moving target while maintaining target dose coverage and our
preliminary results suggest that similar dose conformity as 3D plans could be
obtained for moving targets. Currently however, the presented method is very
sensitive to variations in motion between the optimized and tested scenario. A
combination of 4D optimization, with flexible delivery that is synchronized to
the motion could improve this. In addition, the ‘direct optimization’ approach
of our method, which inherently incorporates delivery dynamics into the op-
5.5. Conclusions
77
timization process, opens the door to implementing other motion mitigation
techniques directly into the optimization process.
78
Chapter 5. Advanced treatment planning with 4D optimization
Chapter 6
Dosimetric quantification of
motion effects in scanned
proton therapy
Much of the analytical work performed in this thesis (and other work published
by our group) is dependent on an efficient and novel 4D dose calculation (4DDC)
(Boye et al. 2013a, Zhang et al. 2013). But the quality of this algorithm, and the
validity of the resulting 4D dose distributions, can only be determined through
experimental validation. As such, the following chapter describes a first experimental validation of this algorithm, which quantifies dose distributions delivered
to moving targets with pencil beam scanned (PBS) proton therapy using different 4D measurement techniques. 4DDC simulations are then compared with
this measured phantom data to assess the accuracy of the 4D dose calculation
to predict dosimetric effects of motion in-vivo.
6.1
Introduction
In clinical practice, respiratory motion is accounted for by using 4D computed
tomography (4DCT) images, representing the motion of a patient’s anatomy
during an averaged respiratory cycle. Treatment plans are then typically computed on a representative, but static (3D), data set extracted from this 4D data.
Different strategies for creating the planning image are available, for example:
79
80
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
(i) an average CT, representing the mean intensities calculated from all phases
of the 4DCT images, (ii) maximum intensity projection image from the 4DCT
or (iii) a representative mid-ventilation or mid-position CT (Jeong et al. 2013,
Wang et al. 2013). In proton therapy, the simple delineation of the tumor on
the planning image is not enough however, and an additional overwriting of the
tumor density is often required to account for possible range variations that can
additionally occur as a result of anatomical motions.
Independent of how the treatment plan is created, the most popular solution to evaluate plan robustness to motion is to recalculate the plan on each
of the individual 3DCT data sets representing each phase of the 4DCT data
and to then accumulate the dose on a reference CT using deformable image
registration. Similar approaches are now commercially available in treatment
planning systems. However, for PBS proton therapy, it is both anatomical motion and the inherently sequential delivery of individual proton pencil beams
which make the method susceptible to dosimetric errors. Therefore an accurate
4D calculation for PBS proton therapy should consider as much as possible the
following: motion information, beam delivery parameters, delivery sequence and
exact timings. At the Centre for Proton Therapy at the PSI, a 4D dose calculation (4DDC) method has been developed that takes all these into account (Boye
et al. 2013a) and which has been subsequently used in many simulation-based
studies (Knopf et al. 2011, Bernatowicz et al. 2013, Zhang et al. 2013). The 4D
dose calculation however is based on several assumptions, and in this chapter,
its accuracy has been experimentally investigated.
Performing accurate dose measurements in moving geometries is a complex
problem. Difficulties start with a proper phantom design, which is a tradeoff
between being able to perform an accurate (and reproducible) measurement
and creating a complex geometry and motion that represents the patient as
realistically as possible. As such, different phantom designs have been reported
in the literature (Followill et al. 2007, Steidl et al. 2012, Haas et al. 2014)
and at our center, an anthropomorphic thorax phantom has been developed
(Perrin et al. 2016), capable of performing dose measurements in realistic patient
geometries and with complex motions.
6.2
Materials and Methods
The PSI 4DDC has been compared with measurements using two different phantom setups: (1) a plexi glass phantom mounted on a movable support, and (2)
the anthropomorphic thorax phantom described in (Perrin et al. 2016). The
6.2. Materials and Methods
81
former allows for dosimetric quantification in a simpler and rigidly moving geometry, whereas the latter exhibits complex geometry, large density differences
and deformable motions.
6.2.1
4D dose calculation
Dose algorithm
Proton dose calculations at PSI are performed using the ray-casting model
(Schaffner et al. 1999, Pedroni et al. 2005), as used by the clinical treatment
planning system at the Paul Scherrer Institute (Lomax et al. 2004). This is
based on CT data which is converted to relative proton stopping powers using
the stoichiometric calibration as described by Schneider et al. (1996), and from
which water-equivalent ranges (WER), along the field direction, of all pencil
beams and dose calculation grids are calculated (Siddon 1985). The dose contributions to all relevant dose grid points are then calculated for each pencil
beam, with the contributions from pencil beams being added into the final 3D
dose distribution, as described in equation 6.1.
d3D (x, y, z) =
−y 2
−x2
DW ER(x,y,z)
· e 2σx (z)2 · e 2σy (z)2
2πσx (z)σy (z)
(6.1)
The 4D dose calculation (4DDC) used in this work is an extension of this
model, and uses a deforming dose-grid approach (Boye et al. 2013a, Zhang et al.
2013). In this algorithm, the positions of dose calculation points are adjusted by
considering time-dependent displacements of the dose grids orthogonal to the
central beam axis (dx(t) and dy(t)), with corresponding adjustments to their
depth being calculated from 4DCT/4DCT-MRI data. For example, lateral displacements are obtained from the extracted motion fields, and water-equivalent
depths from pre-calculated density maps extracted from each motion phase.
The dose delivered by a single proton pencil beam to a dose grid point (x, y, z)
at a reference geometry can then be calculated from equation 6.2.
d4D (x, y, z) =
−yt (t)2
DW ER(xt (t),yt (t),z) −xt (t)22
· e 2σx (z) · e 2σy (z)2
2πσx (z)σy (z)
(6.2)
The term before the exponents is the central axis dose at the WER of the
calculation point at time t, whilst the two exponents describe the lateral fall-off
at distances x and y away from the central axis. Meanwhile, xt (t) = x + dx(t)
and yt (t) = y + dy(t) are the time dependent deformations of the point away
from its nominal position at the time of PB delivery.
82
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
From treatment plan to delivery
A crucial part of the 4D dose calculations is to be able to relate the patient’s
motion with the delivery time-line – a time resolved description of the delivery
of every pencil beam of the treatment. As such, a time stamp is assigned to the
(simulated) delivery of each pencil beam by utilizing the field information from
the 3D treatment plan (i.e. pencil beam sequence, corresponding weight and
field dose etc), together with the characteristics of the delivery, such as scan
path, beam position adjustment time and dose rate. For this work, 4D dose
calculations were based on approximated delivery parameters extracted from
the original specifications of Gantry 2 at PSI (Pedroni et al. 2004). In clinical
practice however, beam position and the number of protons (monitor units) for
each pencil beam need to be translated into the exact machine delivery parameters. For instance, beam positions are effectively translated to proton energy
to reach the desired depth, and lateral positions are achieved by setting proper
current settings of the steering magnets mounted at the nozzle. In this process,
so-called ‘steering files’ (machine control files) are generated and the additional
delivery information resulting from this process can affect the final machine settings and pencil beam delivery. As such, steering files could potentially be used
as a more accurate estimation of a-priori delivery parameters than have been
used in this work. We will come back to this point later on.
6.2.2
Measurement using a moving platform
Experimental setup and treatment planning
As a first step to validate this 4D dose calculation, a simple measurement setup
has been designed (see setup 1 and 2 in figure 6.1(a)). Both phantoms were
considered in the 4D dose calculation, and proton PBS plans with a single field
and dose of 1 Gy were created using the PSI clinical treatment planning system.
An example of the CT image and corresponding treatment plan is shown in
figure 6.1(b).
Two Gafchromic EBT3 films (Sorriaux et al. 2013) were placed in the phantom to measure dose in proximal and distal layers of the field. Both phantoms were then also placed on the scintillation – charged coupled device (CCD)
camera dosimetry system (Boon et al. 1998, 2000), such that the readout corresponds to the bottom film layer, figure 6.1(a). The whole setup was then
mounted on the Quasar respiratory motion platform (Moduc Medical Devices,
London, Ontario, Canada). A similar experimental setup, but without the films,
Figure 6.1: (a) Scheme of measurement setups using plexi glass phantom and (b) planned proton field calculated
based on CT image of setup 2.
6.2. Materials and Methods
83
84
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
has been used by Schätti et al. (2014) to evaluate the effectiveness of gating and
rescanning.
Different experiments were performed to measure the planned field with and
without motion. The moving platform was first set to a periodic sin4 motion
(4 s) with a peak-to-peak amplitude of 10 mm. Numerous studies have shown
the sensitivity of 4D dose distributions to the exact phase of motion when the
delivery starts (Knopf et al. 2011, Zhang et al. 2013, Bernatowicz et al. 2013). As
such, the starting phase of delivery was controlled using an in-house developed
optical tracking system (OTS), similar to that described by Fattori et al. (2014).
First, the OTS was used to monitor phantom motion prior to delivery. Then, a
displacement gate was selected (see gate ON in figure 6.2) and this signal was
used only at the beginning of delivery to control the starting phase of motion
on repeated measurement with sub-mm accuracy.
In figure 6.2(b), statistics on the monitored phantom motion are shown. The
phantom setup had a very stable reproducibility of the motion amplitude (submm) and motion period reproducibility (50 ms for the 25th-75th percentiles). It
is worth to mention however, that initial motion reproducibility of the phantom
was much worse than this and a rail with wheels was added to the QUASAR
platform, allowing for smoother motion under the increased weight of the phantom.
Setup 2 was deliberately designed to have sharp edges and abrupt changes
in material thickness in order to model a large density heterogeneity as may
be encountered around a rib or at the edge of a tumor within the lung, and
these features were exploited to help to identify dose inhomogeneities caused by
a moving phantom.
Dose measurement and calibration procedure
To characterize the delivered 4D doses, relative dosimetry measurements using Gafchromic films and a scintillator-CCD dosimetry system were performed.
Both provide excellent spatial resolution, but also exhibit a quenching effect
(i.e. under-response) in the Bragg peak region due to the increased linear energy transfer (LET) of protons at low energies. For instance, quenching effects
of up to 20 % for EBT film and scintillating dosimetry have been reported when
measuring mono-energetic proton beams (Zhao & Das 2010, Beddar 2015). However, for a 10 cm SOBP measured in solid water (irradiated with proton beam
energies 50-160 MeV), EBT film under-response has been observed to be considerably lower (maximum 10 %), but increased from the middle of the SOBP
(5cm from the distal end) to the end of the range (Zhao & Das 2010). In ad-
6.2. Materials and Methods
Figure 6.2: Effects of temporal and spacial sampling of a sin4 motion curve.
85
86
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
dition, to truly compare calculations with measurements, absolute doses should
be evaluated and dosimetric films are well documented as being sensitive to the
exact evaluation (scanning) procedure used (Butson et al. 2006, Reinhardt et al.
2012) for estimating dose. For all these reasons, a careful handling protocol has
been developed and applied to all films in the same manner.
As such, CCD camera and Gafchromic film response was measured at different dose levels by applying the standard ‘box-field’ (0.25 – 3.5 Gy) used for the
daily Quality Assurance (QA) program on Gantry 2 and for energies similar to
those used in the treatment plan calculated for the phantom (150 MeV + 4cm
pre-absorber). Response in the films was measured at the middle of the QA box
and compared to the absolute dose from the QA field as measured daily using
a calibrated ionization chamber. From this, a linear dose-response was found
for the scintillator-CCD system and a 3-degree polynomial calibration curve for
the red channel intensity from the scanned EBT3 films.
Simulated delivery using 4D dose calculations
The PSI 4DDC was used to assess the dosimetric effect of the plan delivered to
the moving phantom, as described in subsection 6.2.1. For planning, 2 CTs were
acquired, one each for setup 1 and 2. Since the whole setup was much higher
above the table than a typical patient in the direction perpendicular to the
couch, CT scans were acquired of the PMMA phantom only, without the CCD
and QUASAR, in order to ensure that the phantom fitted into the useful fieldof-view of the scanner. 4DCT images were then simulated by rigidly shifting this
CT in relation to its original position using displacements calculated from a sin4
curve with 10 mm amplitude and 4 s period. 10 such CT’s were generated with
equal time spacing and as a result, 10 individual CT images were generated (see
figure 1.2(c)) and formed the input to the 4DDC in a similar way as a normal
4DCT.
Beam data as used clinically has been considered, together with the estimated delivery dynamics of Gantry 2. However, as mentioned above, the
planned PB distributions and weights are translated into machine steering files
(e.g. PB sequence, beam energies and steering currents) and in this process, the
actual delivery dynamics requested from the machine might be different from
that estimated from the idea delivery conditions. Hence, we have also performed
4D dose calculations using more accurate beam delivery dynamics as extracted
from field specific steering files.
6.2. Materials and Methods
6.2.3
87
Measurement using anthropomorphic phantom
Experimental setup and treatment planning
Measurements have been compared with 4D dose calculations for deliveries of
single fractions in a sophisticated 4D anthropomorphic phantom (Lung Cancer; Luca), developed for 4D dosimetric measurements at PSI. The first prototype has recently been enhanced (Perrin et al. 2016) and used to evaluate
the effectiveness of rescanning (Perrin et al. 2015) and provides realistic patient
geometries, motions and non-rigid deformations.
The LuCa phantom consists of an inflatable foam compartment representing
the lung of a patient, with a wooden ball (’tumor’) which can be inserted in
the lung and follows the movement of the foam (figure 6.3). Dosimetric films
can also be inserted in this ‘tumor’ to provide high resolution dosimetric measurements. The lung is then surrounded by an anthropomorphically accurate
skeleton and intra-costal muscle tissue substitutes, together with two possible
skin substitutes, representing either a male or female patient. More details on
the phantom were reported in publications by Perrin et al. (2014, 2016).
Two proton plans, both planned to deliver 2 Gy to the target, were calculated
for the LuCa phantom with SI tumor motions of 4 and 8 mm. Target contours
were delineated on a slow CT (representing the time-averages tumor position
during the imaged period) and this target contour was expanded with additional
setup margins of 7 and 10 mm to form PTV’s to which the dose was planned.
Each plan consisted of three fields, planned according to the single field uniform
dose (SFUD) approach, with gantry angles of 0, 45 and 110 degrees. Such an
arrangement was selected to avoid field directions exactly perpendicular to the
film orientation in the tumor (placed as close as possible to the coronal plane
of the patient) to limit the film under-response effect associated with the high
LET region in the distal end of the fields.
Dose measurement and simulated delivery
Relative film measurements were performed using Gafchromic EBT3 films calibrated in the red channel for doses from 0.5 – 10 Gy, and the green channel
for higher doses (up 35.5 Gy). This choice matches reported film response over
similar dose ranges (Reinhardt et al. 2012). Absolute dose calibration was performed as described in 6.2.2. 4DDC were then performed using 4DCT images
of Luca phantom acquired on our in-house Siemens CT and reconstructed using
the motion signal from a pressure belt. For the calculations, treatment delivery
parameters were taken from the Gantry 2 specifications, as described by Pe-
88
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
Figure 6.3: Anthropomorphic phantom with the wooden tumor ball. Note the
Gafchromic films inserted in two (coronal) planes.
droni et al. (2004). Phantom deformation was extracted using the open source
software (http://plastimatch.com).
Dose measurement and simulated delivery in anthropomorphic phantom
In order to study the accuracy of 4D dose calculation in more complex geometry,
a simple delivery with three proton fields was planned to the tumor contoured on
the averaged CT, extended with an additional margin of 4 and 8 mm depending
on the motion scenario. 4DCT images of the anthropomorphic phantom with
the maximum tumor excursion of 4 and 8 mm were then acquired and used for
the 4DDCs. This plan (2 Gy) was then delivered to the static phantom and
then to the moving phantom.
6.3
6.3.1
Results
Comparison of the calculated and measured dose
distributions using the moving platform
Figure 6.4 presents calculation and measurement results of a static delivery to
experimental setup 2, for the bottom (a-f) and the top measured planes (g-i).
6.3. Results
89
Dosimetric agreement was assessed comparing dose differences analyzed in the
region defined by the 90 % iso-dose of calculated distribution.
Although 1 Gy was planned to the whole target, high dose ‘strips’ can be seen
in the measured dose distributions. These regions correspond to the sharp edges
of the phantom (see setup 2 in figure 6.1) and considering the CT resolution of
2x2x5 mm, partial volume effects can occur, which limit the accuracy of calculated water equivalent ranges and dose. Moreover, the calculated plan consists
of spots with very different energies to compensate for the change in penetration
depth around the sharp density heterogeneity. Such spots have different lateral
penumbras and the calculated dose distribution is not as homogenous as in currently treated clinical cases, where such abrupt changes and sharp edges are
extremely uncommon. Qualitatively, dose distributions measured on the film
and CCD are in good agreement (differences < 0.05 Gy) but larger differences
are observed when comparing with the calculated distribution (up to about 0.2
Gy). Similar observations were noted for the upper measurement film plane.
The relative measurement dose error (point-to-point differences) from the
CCD and Gafchromic film were further compared in both static and moving experiments (figure 6.5(a)). Although, the film measurements with motion result
in slightly higher doses than the CCD in the 4D experiments, in general, the
25th and the 75th percentiles of dose differences are distributed around zero, so
for the investigated scenarios, both measurement techniques have been deemed
to be equally suitable for 4D dose measurements. It has to be mentioned however, that each of the measured values has a measurement uncertainty – dose
uncertainty with Gafchromic films have been reported as being as high as 2 %
due to variations in the daily scanner alone.
Dose errors between the calculated and measured doses were also quantified
using point-to-point differences for the static delivery (figure 6.5(b)). Measured
doses were consistently lower compared to the calculated prescribed dose, with
median differences of 3 % and 5 % of the prescribed dose in the CCD and
film, respectively. This is in fact consistent with clinical experience at PSI with
Gantry 2, where field specific dosimetry typically results in 3-4 % too low dose,
which is then corrected by ’boosting’ the MU of the field. Such ’boosting’ step
was not performed with the fields applied here, thus perhaps being the reason
for this systematic under-dosage on the measurement.
Figure 6.6 now compares measurements for delivery to setup 2 under conditions of motion. Sensitivity to measurement setup and delivery conditions were
reduced by repeating the measurements twice (experiments #1 and #2). Once
again, the measured dose distributions are similar across film and CCD from a
single experiment, where the error is quite uniformly distributed through the
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
90
Figure 6.4: Calculated and measured dose in setup 2 under static condition for the bottom measured plane (a-f)
and the top measured plane (g-i).
6.3. Results
91
(a)
(b)
Figure 6.5: (a) Relative measurement error (point-to-point) in static and moving
target experiments. Numbers #1 and #2 correspond to the repeated experiment. Only
the bottom measurement plane was considered. (b) Point-to-point dose differences
(calculation-measurement) in static delivery.
measurement plane and the output of the film is higher than for the CCD. However, substantial dose differences (in both directions) appear between different
experiments, in both film and CCD (see for instance dose difference film#1 film#2). Meanwhile, differences between measured and calculated 4D dose distributions are even higher and appear in different locations of the target plane.
Quantitative comparisons of point-to-point dose differences between calculated and measured distributions are presented in figure 6.7. Similarly to the
static results, calculated doses are systematically higher with the 25th and 75th
percentiles showing differences in range of -8 % to 20 % of the prescribed dose.
Between different repeated measurements, median dose differences were similar,
but slightly lower for the second experiment, indicating that dose for the second experiment was slightly higher (by about 2 % of prescribed dose). This is
confirmed both from the CCD and the film readout.
Part of the dose uncertainties can be attributed to the limited spatial resolution of the CT, dose calculation and the sharp edge of the phantom (setup 2). In
order to confirm this, the same plan was delivered to setup 1, with equal PMMA
depth across the phantom. Figure 6.8 compares the dosimetric results for delivery to setup 1 under conditions of motion. As expected, both film and CCD
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
92
Figure 6.6: Point-to-point dose differences between calculated and measured distributions delivered to moving
setup 2. Numbers #1 and #2 correspond to the repeated experiment.
6.3. Results
93
Figure 6.7: Box plot of point-to-point dose differences between calculated and
measured distributions delivered to moving setup 2. Numbers #1 and #2 correspond
to the repeated experiment.
94
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
measurement are similar. 1 Although qualitatively, the agreement between
measurement and 4D calculation is now better, the observed point-to-point differences between calculated and measured dose have a similar magnitude as for
setup 2. This suggests that whilst the limited spatial resolution may be a factor,
this alone cannot explain the differences.
Delivery time-line uncertainties
Dosimetric differences between the calculated and measured values are higher
in the moving target experiment than in the static case, with the 25th-75th
percentile region being 2-3 times bigger (figure 6.7). Dose differences are now
both positive and negative, but the median is still centered on the positive side
(i.e. calculated dose is higher).
There are multiple sources of errors which contribute to differences between
calculated and measured dose. We can assume that in the motion scenario,
similar dose errors due to the measurement technique and calculation accuracy
will appear as for the static case 2 . In addition, motion specific errors can also
come from incorrect modeling of the spatial and temporal sequence of both
target/geometry motion and delivery sequence.
The 4D dose calculations used here were based on the approximated delivery
parameters extracted from the machine specifications of Gantry 2 and do not
represent the post-processing occurring during the steering files generation. For
example, the calculation time necessary to move from one spot to the next is
fixed in the current calculation implementation, independent of whether the dose
has been deposited at that spot or not. In practice, pencil beams with fluences
below the deliverable threshold are simply skipped, which can rearrange the PB
sequences and timing.
As such, we have recently modified our 4D calculation code such that the
delivery time-line is directly derived from the steering files, the results of which
are presented in figure 6.9. Certainly, a quite different 4D dose distribution is
obtained, indicating how sensitive 4D dose calculations are to the exact delivery
time-line of the spots. In the future, a more detailed investigation of the steering
file evolution and its delivery time-line will be performed at PSI.
Finally, when looking at the repeat experiments, the measured dose distributions have been found to have median differences of 5̃ % (0.05 Gy) of the planned
1 Note that some differences appear in the un-irradiated region. Films were calibrated in
the range 0.2-2 Gy, for which extracted dose values are most accurate.
2 which is not entirely true for example, in moving geometry the effective quenching effects
are more complex
Figure 6.8: Calculated vs. measured dose delivered to moving setup 1.
6.3. Results
95
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
96
Figure 6.9: Point-to-point dose differences between 4D calculated dose using steering file data and the measured
dose distributions delivered to moving setup 2. Numbers #1 and #2 correspond to the repeated experiment.
6.4. Discussion
97
dose (figure 6.7). Part of observed differences between calculated and measured
dose in repeated measurements may be due to daily machine performance and
the accuracy of beam position, but such large discrepancies would not generally
be expected. In a recent work, the quality of Gantry 2 clinical treatments at
PSI was analyzed using treatment log files, which record the actual measured
parameters for each delivered pencil beam and dose accuracies of ± 1 % of the
nominal were reported (Scandurra et al. 2016).
It becomes clear then, that even the best guess of the delivery timing (an
accurate steering file information) might still not be the best representation of
what is the exact delivered dose. Prospectively therefore, it perhaps makes more
sense to look into a range of 4D dose calculations (each dose value + uncertainty
due to delivery parameters) to assess the sensitivity of a plan to motion and to
only retrospectively try to reconstruct the actual 4D dose based on an accurate
record of the actual delivery (from delivery log files) together with accurately
monitored motions.
6.3.2
Comparison of calculated and measured dose distributions in the anthropomorphic phantom
Figure 6.10(a) compares film doses measured in the anthropomorphic phantom
with calculated doses from the 4DDC. Two different planes of the dose distribution are compared, corresponding to the film placement in the wooden tumor
insert. The 25th and 75th percentiles show differences in range of -5 % to 8
% of the prescribed dose for both films, but in most of the measured area, the
4D dose calculation overestimates the dose with respect to the measurement. In
addition, the calculated dose is more inhomogeneous than the measured dose, as
seen in figure 6.10(b). However, the dose differences are not equally distributed
over the measured plane, suggesting that the 4DDC result cannot truly identify
the position of the cold or hot spots.
6.4
Discussion
In this chapter, calculation and measurements of proton treatment plans delivered with PBS to two different 4D phantoms have been compared. It was found
that in its current form, the 4D dose calculation overestimated local doses by
up to 20 % when compared with measurements, independent on the phantom
used. For reference, in the static calculations, point-to-point dose differences
were substantially lower, i.e. maximum differences of 10 %. Similar findings
98
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
(a)
(b)
Figure 6.10: (a) Calculated vs measured dose from a plan delivered to anthropomorphic breathing phantom and the corresponding point-to-point dose differences. (b)
Dose-area histograms of the corresponding plane dose.
6.4. Discussion
99
have been reported when comparing measurement with another 4D dose calculation (Schätti et al. 2013, 2014). Although, such results are conservative in the
radiotherapy context, it is important to be able to model 4D dose distributions
accurately, especially in the context of advanced delivery techniques such as
tracking or 4D optimization (chapter 5). Several uncertainties in analyzed dose
values have to be mentioned and these are associated both with the calculation
procedure, as well as with the measurement technique.
In the 4DDC, dose errors most likely originate from: (1) motion extraction
(temporal and spatial resolution due to 4DCT), (2) the exact timing and delivery
of pencil beams, (3) daily variations in machine output, estimated < 2 % (Scandurra et al. 2016), and (4) the 4D dose calculation algorithm itself. The first
point can be addressed using simple geometries and motion with good quality
4DCT images. Points two and three can be improved when using actual delivery
information from machine log-files (Meier et al. 2015). Uncertainties due to the
dose calculation algorithm can result for instance from the fact that the geometry is assumed to be fixed during the delivery of each pencil beam. Hence, the
largest differences to the ground truth could be observed for the high-weighted
spots, which are the main contributors to the final dose distribution. In addition, the very sharp density heterogeneity used in the non-anthropomorphic
phantom will also be extremely challenging to the ray-casting algorithm used
in our 4D calculation, and this could also contribute to the differences observed
between experiment and calculation. Although this is certainly a problem of
the dose calculation algorithm used for the 4DDC (ray-casting), it should also
be noted that such sharp and long density interfaces are rarely a problem in
real clinical cases. This is confirmed by previous validation work we have performed of the 3D version of the same algorithm, where good agreement has been
shown, at least relatively, with film dosimetry (Dietlicher et al. 2014) and also
Monte Carlo calculations (Tourovsky et al. 2005, Winterhalter et al. 2016) in
complex phantoms and real patient geometries respectively. Nevertheless, the
experimental results presented here may indicate potential problems with the
algorithm for lung patients, where the large density gradients between the water
like tumor and surrounding, low density lung, could cause problems. Indeed,
it was due to this potential problem in the lung that setup 2 in this work was
used.
In addition to limitations in the dose calculation however, our 4D experimental results (in both measurement and calculation) also strongly depend on
the motion parameters exhibited by the phantom geometry (motion amplitude,
period, starting phase) as well as the 4D imaging and motion extraction procedures used. Problems related to proper motion sampling were introduced in
100
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
section 6.2.1 and limitations with 4DCT image quality were discussed in chapter 3. Additionally, motion extraction using deformable image registration is
also a challenging area. For instance, even the best set of 4D images do not help
to extract the ground truth motion in regions with a uniform intensity (e.g.
the foam lung compartment in the LuCa phantom). As such, the 4DCT-MRI
technique described in chapter 4, where motion is extracted from MRI imaging
with its superior soft tissue contrast, could be an important development for
more accurate motion modeling in the future.
The observed differences could also of course be associated with underresponse of the dosimetric film at the end of the depth-dose curve (similar
effect is observed for the scintillator screen of the CCD setup), as mentioned in
section 6.2.2. 10 % dose differences were reported in proton SOBP field measured with film and the correction factor for different depths could be calculated
(Zhao & Das 2010). Meanwhile, it is difficult to quantify the quenching correction factors in inhomogeneous fields and complex (moving) geometries. For
instance, in the anthropomorphic phantom used in our study, dose deviations of
15 % of the planned distribution were recently reported for the ’best-case’ static
delivery with no positioning error (Perrin et al. 2014). It would seem therefore
that the quenching effect in film and CCD dosimetry should be corrected for
in order to improve the measurement accuracy, although this could be quite a
challenge in complex PBS plans, where dose at any point can be delivered from
many different pencil beams with varying energies and LET contributions.
All of the above points indicate that experimental validation of 4D treatments is very challenging, and the results of this chapter show that much remains
to be done to improve both the measurement devices used and the experimental
set-ups to be able to more accurately reproduce the exact conditions of motion
and treatment delivery. Indeed, all these issues, together with the sensitivity of
the 4D optimized plans presented in the previous chapter, put into question the
validity and meaning of any, single 4D dose calculation. Being so sensitive on
so many factors, and given that it is very unlikely that all these factors at the
time of delivery will be exactly those of the simulation, then it is highly unlikely
that any a-priori calculated 4D dose distribution will actually be delivered.
Thus, 4D dose calculations, as mentioned elsewhere in this work, may either
be best employed as a posteriori calculations, based on measured motions during
delivery and the exact delivery time-lines used (as extracted from delivery logfiles), or, if used a priori, many such calculations should be performed in order to
explore the solution-space of 4D distributions relevant to the patients treatment.
As such, a fast 4D algorithm, as used in this work, is a must in order to be able
to adequately search this.
6.5. Conclusions
101
However, on a more optimistic note, although the agreement between measurement and calculation for single deliveries is rather poor, at a more general
level, it has been previously shown both in 4DDC (Knopf et al. 2011, Bernatowicz et al. 2013) and experimentally (Schätti et al. 2013, Perrin et al. 2015),
that re-scanning (relying on the multiple application of plans under conditions
of motion) can mitigate the interplay effect at a global level. This indicates,
that 4DDC’s, although not currently able to predict local doses or spatial distributions of dose inhomogeneities under conditions of motion, can provide some
confidence in the effectiveness of statistical solutions such as rescanning or fractionation. Their validity however for more sophisticated applications such as
4D optimization or tumor tracking remains unclear.
6.5
Conclusions
In conclusion, the 4D dose calculation used in this work is an efficient tool
for studying global dosimetric effects in moving geometries/anatomies. Several
dosimetric measurements have been performed to validate existing software and
hardware tools and we have shown that simulations and measurements agree in
estimating the severity of motion effects, and in assessing the effectiveness of rescanning to mitigate motion induced dose inhomogeneities. However, large local
differences between simulated and measured doses have been found, depending
on the studied geometry, scenario or measuring device. The exact prediction
of under-/over- dosage regions due to motion remains important for advanced
4D treatment applications, such as prospective and retrospective plan adaptation or 4D optimization (introduced in chapter 5). It appears that, for these
techniques to be effective, more accurate knowledge on motions and delivery
time-lines may have to be known.
102
Chapter 6. Dosimetric quantification of motion effects in scanned proton
therapy
Chapter 7
Conclusion and outlook
This dissertation focused on the development of 4D imaging and treatment
planning techniques for PBS proton therapy.
To begin with (chapter 2), we have demonstrated that PBS proton therapy
using raITV’s and rescanning can be used as an effective strategy to achieve
good (similar or better than static) target coverage and dose homogeneity in the
presence of respiratory motion, confirming earlier work already performed in our
group (Knopf et al. 2013). In addition however, we found that despite the larger
margins resulting from the use of raITV’s and necessary for effective proton
delivery, proton plans still resulted in lower integral dose and secondary cancer
risk when compared to VMAT photon plans planned using the conventional
ITV approach.
ITV’s and raITV’s are typically generated from 4DCT images, but in the
presence of imaging artifacts, these are difficult to accurately delineate, causing
errors in treatment planning and dosimetric quantification. In an attempt to
improve image quality, different 4DCT acquisition methods have been compared
in chapter 3. We found that prospectively-gated 4DCT acquisitions improved
image quality (small MSE compared with the ground truth) and that the best
image quality was achieved for the end-exhale phase. In light of this, a potential
additional advantage of respiratory-gated 4DCT can be explored by acquiring
images only for phases which provide the best image quality without having
to acquire the other phase images. The exact effect of reducing the number
of reconstructed 4DCT images should be further studied, with the benefit of
lowering the imaging dose being balanced against the ability of maintaining
good quality 4D plans. Additionally, lowering imaging dose could be exploited
103
104
Chapter 7. Conclusion and outlook
in the context of adaptive therapy. Nowadays, as a standard practice, only
a single 4DCT is acquired per patient treatment due to the substantial dose.
Such low dose 4DCT could however be repeated over the course of therapy to
update information on changes occurring in the patient’s anatomy and internal
motions.
A few issues still need to be addressed to achieve prospective respiratorygated (or beam paused) 4DCT acquisition in practice however. One challenge
is that, in the case where respiratory motion becomes irregular or unstable, or
undergoes a baseline shift, the pre-calculated displacement gating window may
become too narrow to allow acquisition within an acceptable time-frame. A
range of mitigation methods against this may be possible however, for example
by the adaptive recalculation of gating windows over the course of acquisition,
or adaptive-switching between the beam paused and respiratory-gated methods which differ only in the choice of gating threshold and cine duration times.
In addition, in cases where AP motion is larger than SI, adaptive selection
of the gating direction based on the direction of largest motion could also be
considered. Alternatively, the patient’s breathing could be also controlled. For
instance, a recent study demonstrated that audiovisual biofeedback can substantially improve breathing regularity of lung cancer patients and help to further
reduce imaging artifacts in 4DCT acquisition (Pollock et al. 2016).
Due to the inherent limitations of 4DCT, proton treatment planning on
ITV’s derived from a single 4DCT has been shown to result in under-dosage of
the tumor by up to 25.7 % over 3 min of treatment (Koybasi et al. 2014). Hence,
the definition of adequate ITV’s/raITV’s depends on accurate 4D imaging, especially in the presence of irregular respiratory motion. Although prospectivelygated 4DCT acquisitions offer potential improvements by increasing image quality and lowering the imaging dose, they still represent in the end only a single
(averaged) respiratory breathing cycle. As such, a validation of 4DCT-MRI as
a potential imaging modality for 4D treatment planning of PBS proton therapy has been presented in chapter 4. In comparison to conventional 4DCT, it
has the advantage of being able to capture breathing information from multiple
breathing cycles without additional imaging dose, while providing the necessary
electron density information for accurate proton dose calculations. In addition,
this modality could be used to generate more effective raITV’s by including
breathing variability into the raITV design.
Two methods of calculating 4DCT-MRI data sets have been investigated;
patient-specific and population-based. The first is a step towards highly personalized 4D treatments, whilst the latter approach was shown to be a less
accurate approximation of 4DCT data from the point of view of 4D proton dose
105
calculations. Nevertheless, the population based approach could certainly be
improved by extending the amount of patient data included into the motion library and the population-based motion library is anyway interesting for testing
treatment plan robustness under variable motion conditions.
One of the important limitations of the 4DCT-MRI approach described in
this work however, is that it is specific to the liver. As such, in the appendix 8,
we also present preliminary results on 4DCT-MRI in the thorax, where problems
with the population-based modeling of patient’s lungs were observed due to large
deformations, indicating that such cases will likely require subject-specific modeling. Additionally, the motion extracted from the 4DCT of these patients was
very different from that from 4DMRI and such a situation needs to be treated
with care. As such, similar external signals should be acquired during the acquisition of 4DCT and 4DMRI to evaluate if motion differences present in the
reconstructed 4D images are due to large irregularities in patient’s breathing or
problems with the acquisition/reconstruction protocols used for the generation
of 4DCT-MRI data. If motion can be accurately imaged however, this opens
up the possibilities of more sophisticated 4D treatment planning and delivery
techniques.
One such – 4D optimization - has been investigated in chapter 5. 4D optimization has been shown to be a powerful motion mitigation technique when
exact information on the motion is available, but has also been shown to be very
sensitive to variations in these motion conditions. Thus, in order to best exploit
the potential of 4D optimization, it will be crucial to apply techniques such as
audio-visual guidance (to improve breathing regularity, i.e. period and amplitude) and/or low-latency motion monitoring systems (to reduce phase shifts) to
reduce motion variations to within the ranges estimated in this work. However,
the flexibility of our 4D optimization approach opens the door for the incorporation of other motion mitigation techniques, such as rescanning, gating and
tracking, directly into the optimization process, potentially allowing for more
robust 4D optimized plans.
Alternatively, the delivery of 4D optimized plans could be directly synchronized to the patient’s motion. As such, an interesting approach could be to
perform an online ‘temporally-modulated’ delivery, where the speed of delivery
(e.g. scanning speed and/or beam intensity) is adjusted according to the motion
signal. For instance, when the tumor moves faster than assumed at the planning/optimization stage, the speed of delivery could be increased to deliver the
requested spot at the correct phase or amplitude of the breathing cycle. The
potential of this approach needs to be studied in detail however through comprehensive 4DDC simulations for real (and variable) breathing patterns using
106
Chapter 7. Conclusion and outlook
(e.g.) 4DCT-MRI data sets.
Finally, in chapter 6, the first experimental validations of the PSI 4D dose
calculation have been performed. Similarly to the results of the sensitivity analysis of 4D optimization, the results of this indicate the sensitivity of 4D deliveries
to the exact experimental conditions, motion characteristics and delivery timelines actually used. Thus, it becomes clear that, although advanced delivery
techniques such as tracking or 4D optimization offer substantial dosimetric improvements, they remain currently challenging to realize clinically, relying as
they do on the accurate monitoring of motion and delivery timelines. Although
in controlled experiments, geometry and phantom motion can be controlled with
sub-millimeter and millisecond accuracy, it is not always so easy to obtain apriori and accurate information on the exact delivery time-lines, especially when
unexpected events such as delivery interlocks occur. As such, the use of delivery log-files, in combination with monitored motion information, is potentially
the most accurate way to reconstruct the actually delivered 4D dose. Based on
previously published work on log-file based dose calculations (Meier et al. 2015,
Scandurra et al. 2016) such studies will be the natural extension of this work.
In addition to uncertainties in the exact delivery time-lines however, there
are also inevitable limitations in the accuracy of the dose calculation itself. The
4DDC used in this work uses the same dose calculation algorithm as that used
clinically at PSI (the ray casting model). However, the phantom used in the
experimental validation of the algorithm will certainly challenge the capabilities
of this algorithm. For instance, a recent study has shown, that in comparison
to Monte Carlo calculations, analytical calculations (albeit of a different nature
to that used in this work) systematically overestimate the target dose and underestimate the dose to normal lung (Grassberger et al. 2014). In addition, all
4D calculations are also subject to errors in 4D imaging and deformable image registration. Certainly, the validation of 4D medical image registration is
very challenging, but perhaps the use of sophisticated, anthropomorphic moving
phantoms (Perrin et al. 2014) could be a first-step towards the development and
evaluation of such advanced motion imaging techniques.
In summary, in this work we have developed and investigated methods for
4D planning, imaging and optimization of PBS proton therapy. We have also
shown that the delivery of plans based on these techniques remains a challenging
task for the future. As such, the next steps should be clinical comparisons of
robust (e.g. raITV/rescanning) vs optimal (4D optimization) solutions, and
further experimental validation of all treatment approaches through end-to-end
tests on 4D anthropomorphic phantoms. In short, although we are improving
our understanding of motion, and developing ever more effective methods of
107
imaging and mitigating it during treatment, much interesting and exciting work
remains to be done.
108
Chapter 7. Conclusion and outlook
Chapter 8
Appendix
8.1
4DCT-MRI of lung
4DCT-MRI technique was presented earlier in detail (chapter 4). However, as
mentioned, the method is organ specific and earlier work has focused on the
liver modeling. In the following section, preliminary results of the 4DCT-MRI
method in the lung region are shown.
8.1.1
Generation of lung 4DCT-MRI
In the last 4 years, 4DMRI images of over 20 volunteers and 1 lung cancer patient
were acquired at the Proton Therapy Center at PSI. For the imaged patient,
both 4DCT and 4DMRI images were available. Collected data was analyzed to
characterize motion mechanics under free breathing and breath-hold conditions
(Kaeser 2016). Firstly, similar approach to liver methodology was pursued to
generate the lung meshes, with lung model and points as reported by Boye et al.
(2013b). Such approach was efficient and worked for most of the volunteer cases,
however meshing problems appeared while generating a population based model.
One of the challenges was that the huge tumor lesion present in the posterior
part of patient’s anatomy, overlapped substantial part of the lung and made it
difficult to match the patient-specific mesh to the population-based model due
the distorted appearance of patient mesh.
For this patient, a more straight forward subject-specific approach was applied instead. Deformation fields were extracted from 4DCT and 4DMRI images
of the same patient upon establishing common coordinate system. In order to
109
110
Chapter 8. Appendix
(a)
(b)
(c)
Figure 8.1: (a) A 4DCT image (end-exhale) of lung cancer patient. (b) 4DMRI
image (end-exhale) of the lung cancer patient. (c) Motion at 6 different points in lung,
extracted from 4DCT using deformable image registration (lu - left lung points, tu points in the tumor volume).
Figure 8.2: Motion at 6 different points in lung, extracted from 4DMRI using deformable image registration
(lu - left lung points, tu - points in the tumor volume), extracted from 4DMRI images using deformable image
registration.
8.1. 4DCT-MRI of lung
111
112
Chapter 8. Appendix
(a)
(b)
(c)
Figure 8.3: (a) Reference 4DCT image, and (b-c) simulated 4DCT-MRI images at
different respiratory phases (time step 3 and 5).
capture both lungs, 4DMRI images were acquired with the maximum possible
field of view (FOV). However, parts of the anatomy e.g. left and right side of
the body (around rib area), were not captured. 1 In clinical practice only a
part of anatomy surrounding the tumor volume is relevant and perhaps a more
optimal FOV could be found to include all possible beam entrance paths. Nevertheless, extracted deformation fields were processed using a lung mask and
motion outside of the mask was set to zero (e.g. no motion in ribcage).
An example of 4DCT images and extracted motion of selected points in the
lung and inside the tumor volume are show in figure 8.1 and 4DMRI images
and motion is shown in figure 8.2. It can be seen that motion mechanics of
1 Note that the smaller FOV, the faster the acquisition. Acquisition time is crucial for patient’s comfort, 4D acquisition sequence is long, because patient’s anatomy is excited multiple
times causing heating of the tissue and some pauses are necessary for patient cooling.
8.1. 4DCT-MRI of lung
113
Figure 8.4: Comparison of the 4D dose calculation result from the 4DCT vs 4DCTMRI in lung cancer patient.
4DCT and 4DMRI is different, although it comes from the same patient and was
acquired within the span of a week apart. In general, displacements extracted
from 4DMRI are smaller than the 4DCT data. Motion extracted from the first
cycle from the 4DMRI study was then used to warp the reference (end-exhale)
image from 4DCT and generate the 4DCT-MRI, see figure 8.3.
8.1.2
Comparison of 4DCT and 4DCT-MRI
4D dose calculations were performed on the original 4DCT images and on 4DCTMRI images, see figure 8.4. Dose differences higher than ± 20 % in the tumor
volume were observed, which were expected due to large differences between
original and simulated motion. Generation of 4DCT-MRI images in lung remains challenging and careful evaluation of all image processing steps (including 4DMRI image reconstruction) should be performed to pick up the origins
of these substantial differences. Additionally, it would be interesting to validate
the 4DCT-MRI method experimentally, for example using the anthropomorphic
Luca phantom (Perrin et al. 2016).
114
Chapter 8. Appendix
Bibliography
M. A. Admiraal, et al. (2008). ‘Dose calculations accounting for breathing
motion in stereotactic lung radiotherapy based on 4D-CT and the internal target
volume’. Radiotherapy and Oncology 86(1):55–60.
F. Albertini, et al. (2009). ‘Planning and optimizing treatment plans for actively
scanned proton therapy’. Biomedical Mathematics: Promising Directions in
Imaging, Therapy Planning, and Inverse Problems .
A. Aupérin, et al. (2010). ‘Meta-analysis of concomitant versus sequential radiochemotherapy in locally advanced non–small-cell lung cancer’. Journal of
Clinical Oncology 28(13):2181–2190.
S. Beddar (2015). ‘Real-time volumetric scintillation dosimetry’. In Journal of
Physics: Conference Series, vol. 573, p. 012005. IOP Publishing.
K. Bernatowicz (2016). ‘Comparison of radiation-induced complication risk in
volumetric modulated arc therapy and pencil beam scanned proton therapy for
lung and liver using motion-adapted margins.’. XXX .
K. Bernatowicz, et al. (2015). ‘Quantifying the impact of respiratory-gated 4D
CT acquisition on thoracic image quality: a digital phantom study’. Medical
physics 42(1):324–334.
K. Bernatowicz, et al. (2013). ‘Comparative study of layered and volumetric
rescanning for different scanning speeds of proton beam in liver patients’. Physics
in medicine and biology 58(22):7905.
K. Bernatowicz, et al. (2016). ‘Four-Dimensional Dose Reconstruction for
Scanned Proton Therapy Using Liver 4DCT-MRI’. International Journal of
Radiation Oncology* Biology* Physics 95(1):216–223.
115
116
Bibliography
C. Bert & M. Durante (2011). ‘Motion in radiotherapy: particle therapy’.
Physics in medicine and biology 56(16):R113.
C. Bert, et al. (2009). ‘Motion management in scanned particle therapy: beam
gating &amp; tracking’. In World Congress on Medical Physics and Biomedical
Engineering, September 7-12, 2009, Munich, Germany, pp. 345–348. Springer.
C. Bert, et al. (2010). ‘Dosimetric precision of an ion beam tracking system’.
Radiation Oncology 5(1):1.
C. Bert, et al. (2008). ‘Quantification of interplay effects of scanned particle
beams and moving targets’. Physics in medicine and biology 53(9):2253.
C. Bert & E. Rietzel (2007). ‘4D treatment planning for scanned ion beams’.
Radiation Oncology 2(1):1.
S. Boon, et al. (2000). ‘Performance of a fluorescent screen and CCD camera as
a two-dimensional dosimetry system for dynamic treatment techniques’. Medical
physics 27(10):2198–2208.
S. Boon, et al. (1998). ‘Fast 2D phantom dosimetry for scanning proton beams’.
Medical physics 25(4):464–475.
T. Bortfeld, et al. (1990). ‘Methods of image reconstruction from projections applied to conformation radiotherapy’. Physics in Medicine and Biology
35(10):1423.
D. Boye, et al. (2013a). ‘Mapping motion from 4D-MRI to 3D-CT for use in 4D
dose calculations: a technical feasibility study’. Medical physics 40(6):061702.
D. Boye, et al. (2013b). ‘Population based modeling of respiratory lung motion
and prediction from partial information’. In SPIE medical imaging, pp. 86690U–
86690U. International Society for Optics and Photonics.
P. Boyle, et al. (2008). World cancer report 2008. IARC Press.
J. D. Bradley, et al. (2015). ‘Standard-dose versus high-dose conformal radiotherapy with concurrent and consolidation carboplatin plus paclitaxel with
or without cetuximab for patients with stage IIIA or IIIB non-small-cell lung
cancer (RTOG 0617): a randomised, two-by-two factorial phase 3 study’. The
Lancet Oncology 16(2):187–199.
BIBLIOGRAPHY
117
C. Burman, et al. (1991). ‘Fitting of normal tissue tolerance data to an analytic function’. International Journal of Radiation Oncology* Biology* Physics
21(1):123–135.
M. Butson, et al. (2006). ‘Scanning orientation effects on Gafchromic EBT
film dosimetry’. Australasian Physics &amp; Engineering Sciences in Medicine
29(3):281–284.
J. Callahan, et al. (2014). ‘High-resolution imaging of pulmonary ventilation and
perfusion with 68Ga-VQ respiratory gated (4-D) PET/CT’. European journal
of nuclear medicine and molecular imaging 41(2):343–349.
R. Castillo, et al. (2012). ‘Spatial correspondence of 4D CT ventilation and
SPECT pulmonary perfusion defects in patients with malignant airway stenosis’.
Physics in medicine and biology 57(7):1855.
C. Catana, et al. (2010). ‘Toward implementing an MRI-based PET attenuationcorrection method for neurologic studies on the MR-PET brain prototype’. Journal of Nuclear Medicine 51(9):1431–1438.
S. M. Damkjær, et al. (2013). ‘Reduced lung dose and improved inspiration
level reproducibility in visually guided DIBH compared to audio coached EIG
radiotherapy for breast cancer patients’. Acta Oncologica 52(7):1458–1463.
I. Dietlicher, et al. (2014). ‘The effect of surgical titanium rods on proton
therapy delivered for cervical bone tumors: experimental validation using an
anthropomorphic phantom’. Physics in medicine and biology 59(23):7181.
L. Dietrich, et al. (2006). ‘Linac-integrated 4D cone beam CT: first experimental
results’. Physics in medicine and biology 51(11):2939.
J. Dinkel, et al. (2009). ‘4D-MRI analysis of lung tumor motion in patients with
hemidiaphragmatic paralysis’. Radiotherapy and Oncology 91(3):449–454.
S. Dowdell, et al. (2013). ‘Interplay effects in proton scanning for lung: a 4D
Monte Carlo study assessing the impact of tumor and beam delivery parameters’.
Physics in medicine and biology 58(12):4137.
J. A. Dowling, et al. (2012). ‘An atlas-based electron density mapping method
for magnetic resonance imaging (MRI)-alone treatment planning and adaptive
MRI-based prostate radiation therapy’. International Journal of Radiation Oncology* Biology* Physics 83(1):e5–e11.
118
Bibliography
J. Ehrhardt, et al. (2011). ‘Statistical modeling of 4D respiratory lung motion
using diffeomorphic image registration’. Medical Imaging, IEEE Transactions
on 30(2):251–265.
J. Ehrhardt, et al. (2007). ‘An optical flow based method for improved reconstruction of 4D CT data sets acquired during free breathing’. Medical Physics
34(2):711–721.
K. Eilertsen, et al. (2008). ‘A simulation of MRI based dose calculations on the
basis of radiotherapy planning CT images’. Acta Oncologica 47(7):1294–1302.
J. G. Eley, et al. (2014). ‘4D optimization of scanned ion beam tracking therapy
for moving tumors’. Physics in medicine and biology 59(13):3431.
J. G. Eley, et al. (2015). ‘Robustness of target dose coverage to motion uncertainties for scanned carbon ion beam tracking therapy of moving tumors’.
Physics in medicine and biology 60(4):1717.
B. Emami, et al. (1991). ‘Tolerance of normal tissue to therapeutic irradiation’.
International Journal of Radiation Oncology* Biology* Physics 21(1):109–122.
M. Engelsman, et al. (2006). ‘Four-dimensional proton treatment planning for
lung tumors’. International Journal of Radiation Oncology* Biology* Physics
64(5):1589–1595.
F. Ernst, et al. (2012). ‘Correlation between external and internal respiratory
motion: a validation study’. International journal of computer assisted radiology
and surgery 7(3):483–492.
M. Falk, et al. (2010). ‘Real-time dynamic MLC tracking for inversely optimized
arc radiotherapy’. Radiotherapy and oncology 94(2):218–223.
G. Fattori, et al. (2014). ‘Commissioning of an integrated platform for timeresolved treatment delivery in scanned ion beam therapy by means of optical
motion monitoring’. Technology in cancer research &amp; treatment 13(6):517–
528.
D. S. Followill, et al. (2007). ‘Design, development, and implementation of the
radiological physics center’s pelvis and thorax anthropomorphic quality assurance phantoms’. Medical physics 34(6):2070–2076.
BIBLIOGRAPHY
119
N. Fukumitsu, et al. (2009). ‘A prospective study of hypofractionated proton
beam therapy for patients with hepatocellular carcinoma’. International Journal
of Radiation Oncology* Biology* Physics 74(3):831–836.
C. Graeff (2014). ‘Motion mitigation in scanned ion beam therapy through
4D-optimization’. Physica Medica 30(5):570–577.
C. Graeff, et al. (2013). ‘A 4D-optimization concept for scanned ion beam
therapy’. Radiotherapy and Oncology 109(3):419–424.
C. Grassberger, et al. (2014). ‘Quantification of proton dose calculation accuracy
in the lung’. International Journal of Radiation Oncology* Biology* Physics
89(2):424–430.
C. Grassberger, et al. (2015). ‘Motion mitigation for lung cancer patients treated
with active scanning proton therapy’. Medical physics 42(5):2462–2469.
M. Guckenberger, et al. (2011). ‘Potential of adaptive radiotherapy to escalate
the radiation dose in combined radiochemotherapy for locally advanced non–
small cell lung cancer’. International Journal of Radiation Oncology* Biology*
Physics 79(3):901–908.
O. C. Haas, et al. (2014). ‘IGRT/ART phantom with programmable independent rib cage and tumor motion’. Medical physics 41(2):022106.
R. A. Hälg, et al. (2012). ‘Systematic measurements of whole-body imaging dose
distributions in image-guided radiation therapy’. Medical physics 39(12):7650–
7661.
D. Ionascu, et al. (2007). ‘Internal-external correlation investigations of respiratory induced motion of lung tumors’. Medical physics 34(10):3893–3903.
H. Jeong, et al. (2013). ‘Compensation method for respiratory motion in proton
treatment planning for mobile liver cancer’. Journal of Applied Clinical Medical
Physics 14(2).
J. Kaeser (2016). ‘Lung motion modelling from 4DMRI datasets’. Master Thesis
ETH Zurich.
P. Keall (2004). ‘4-dimensional computed tomography imaging and treatment
planning’. In Seminars in radiation oncology, vol. 14, pp. 81–90. Elsevier.
120
Bibliography
P. Keall, et al. (2006a). ‘The clinical implementation of respiratory-gated
intensity-modulated radiotherapy’. Medical Dosimetry 31(2):152–162.
P. Keall, et al. (2007). ‘Respiratory regularity gated 4D CT acquisition: concepts
and proof of principle’. Australasian Physics &amp; Engineering Sciences in
Medicine 30(3):211–220.
P. J. Keall, et al. (2006b). ‘The management of respiratory motion in radiation
oncology report of AAPM Task Group 76a)’. Medical physics 33(10):3874–3900.
T. Kim, et al. (2012). ‘Audiovisual biofeedback improves diaphragm motion
reproducibility in MRI’. Medical physics 39(11):6921–6928.
K. Kitamura, et al. (2003). ‘Tumor location, cirrhosis, and surgical history
contribute to tumor movement in the liver, as measured during stereotactic
irradiation using a real-time tumor-tracking radiotherapy system’. International
Journal of Radiation Oncology* Biology* Physics 56(1):221–228.
G. Klimpki, et al. (2016). ‘Fast dose modulation in proton therapy with continuous line scanning’ .
A.-C. Knopf, et al. (2013). ‘Adequate margin definition for scanned particle
therapy in the incidence of intrafractional motion’. Physics in medicine and
biology 58(17):6079.
A.-C. Knopf, et al. (2011). ‘Scanned proton radiotherapy for mobile targets—
the effectiveness of re-scanning in the context of different treatment planning
approaches and for different motion characteristics’. Physics in medicine and
biology 56(22):7257.
A.-C. Knopf & A. J. Lomax (2014). ‘In the context of radiosurgery–Pros and
cons of rescanning as a solution for treating moving targets with scanned particle
beams’. Physica Medica 30(5):551–554.
A. Koehler, et al. (1977). ‘Flattening of proton dose distributions for large-field
radiotherapy’. Medical physics 4(4):297–301.
O. Koybasi, et al. (2014). ‘Simulation of dosimetric consequences of 4D-CTbased motion margin estimation for proton radiotherapy using patient tumor
motion data’. Physics in medicine and biology 59(4):853.
BIBLIOGRAPHY
121
I. Kristensen, et al. (2015). ‘Comparative Proton and Photon Treatment Planning in Pediatric Patients with Various Diagnoses’. International Journal of
Particle Therapy 2(2):367–375.
H. D. Kubo & B. C. Hill (1996). ‘Respiration gated radiotherapy treatment: a
technical study’. Physics in medicine and biology 41(1):83.
U. Langner & P. Keall (2008). ‘Prospective displacement and velocity-based
cine 4D CT’. Medical physics 35(10):4501–4512.
U. W. Langner & P. J. Keall (2010). ‘Quantification of artifact reduction with
real-time cine four-dimensional computed tomography acquisition methods’. International Journal of Radiation Oncology* Biology* Physics 76(4):1242–1250.
C. Lepage, et al. (2015). ‘Survival in patients with primary liver cancer, gallbladder and extrahepatic biliary tract cancer and pancreatic cancer in Europe 1999–
2007: Results of EUROCARE-5’. European Journal of Cancer 51(15):2169–
2178.
X. Li, et al. (2011). ‘A 4D IMRT planning method using deformable image
registration to improve normal tissue sparing with contemporary delivery techniques’. Radiation Oncology 6(1):1.
A. Lomax (1999). ‘Intensity modulation methods for proton radiotherapy’.
Physics in medicine and biology 44(1):185.
A. Lomax, et al. (1996). ‘3D treatment planning for conformal proton therapy
by spot scanning’. In Proceedings of the 19th LH Gray Conference. London:
BIR Publishing, pp. 67–71.
A. J. Lomax, et al. (2004). ‘Treatment planning and verification of proton
therapy using spot scanning: initial experiences’. Medical physics 31(11):3150–
3157.
R. Lüchtenborg, et al. (2011). ‘Experimental verification of a real-time compensation functionality for dose changes due to target motion in scanned particle
therapy’. Medical physics 38(10):5448–5458.
A. E. Lujan, et al. (2003). ‘A method for incorporating organ motion due to
breathing into 3D dose calculations in the liver: sensitivity to variations in
motion’. Medical physics 30(10):2643–2649.
122
Bibliography
Y. Ma, et al. (2009). ‘Four-dimensional inverse treatment planning with inclusion of implanted fiducials in IMRT segmented fields’. Medical physics
36(6):2215–2221.
D. Mah, et al. (2000). ‘Technical aspects of the deep inspiration breath-hold
technique in the treatment of thoracic cancer’. International Journal of Radiation Oncology* Biology* Physics 48(4):1175–1185.
L. B. Marks, et al. (2010). ‘Use of normal tissue complication probability models
in the clinic’. International Journal of Radiation Oncology* Biology* Physics
76(3):S10–S19.
M. Marx, et al. (2014). ‘Simulation of spatiotemporal CT data sets using a
4D MRI-based lung motion model’. International journal of computer assisted
radiology and surgery 9(3):401–409.
G. Meier, et al. (2015). ‘Independent dose calculations for commissioning,
quality assurance and dose reconstruction of PBS proton therapy’. Physics in
medicine and biology 60(7):2819.
P. Mishra, et al. (2012). ‘Adaptation and applications of a realistic digital
phantom based on patient lung tumor trajectories’. Physics in medicine and
biology 57(11):3597.
P. Mishra, et al. (2014). ‘An initial study on the estimation of time-varying
volumetric treatment images and 3D tumor localization from single MV cine
EPID images’. Medical physics 41(8):081713.
K. E. Mittauer, et al. (2015). ‘Monitoring ABC-assisted deep inspiration breath
hold for left-sided breast radiotherapy with an optical tracking system’. Medical
physics 42(1):134–143.
S. Mori, et al. (2014). ‘Amplitude-based gated phase-controlled rescanning in
carbon-ion scanning beam treatment planning under irregular breathing conditions using lung and liver 4DCTs’. Journal of radiation research 55(5):948–958.
S. Mori, et al. (2009). ‘Effects of interfractional anatomical changes on waterequivalent pathlength in charged-particle radiotherapy of lung cancer’. Journal
of radiation research 50(6):513–519.
M. J. Murphy, et al. (2007). ‘The management of imaging dose during imageguided radiotherapy: report of the AAPM Task Group 75’. Medical physics
34(10):4041–4063.
BIBLIOGRAPHY
123
K. Nakagawa, et al. (2012). ‘4D registration and 4D verification of lung tumor
position for stereotactic volumetric modulated arc therapy using respiratorycorrelated cone-beam CT’. Journal of radiation research p. rrs058.
H. Nakayama, et al. (2010). ‘High-Dose Conformal Radiotherapy for Patients
With Stage III Non–Small-Cell Lung Carcinoma’. International Journal of Radiation Oncology* Biology* Physics 78(3):645–650.
Y. Negoro, et al. (2001). ‘The effectiveness of an immobilization device in conformal radiotherapy for lung tumor: reduction of respiratory tumor movement
and evaluation of the daily setup accuracy’. International Journal of Radiation
Oncology* Biology* Physics 50(4):889–898.
R. C. Nichols Jr, et al. (2012). ‘Proton therapy for lung cancer’. Thoracic Cancer
3(2):109–116.
E. O’Riordan, et al. (2000). ‘Dual phase hepatic CT: influence of scanning
direction on liver attenuation’. American Journal of Roentgenology 174(5):1417–
1421.
C. Paganelli, et al. (2015). ‘Liver 4DMRI: A retrospective image-based sorting
method’. Medical physics 42(8):4814–4821.
H. Paganetti (2011). ‘Proton Therapy Physics. Series in Medical Physics and
Biomedical Engineering’.
T. Pan (2005). ‘Comparison of helical and cine acquisitions for 4D-CT imaging
with multislice CT’. Medical physics 32(2):627–634.
T. Pan, et al. (2007). ‘Improvement of the cine-CT based 4D-CT imaging’.
Medical physics 34(11):4499–4503.
P. C. Park, et al. (2012). ‘A beam-specific planning target volume (PTV) design
for proton therapy to account for setup and range uncertainties’. International
Journal of Radiation Oncology* Biology* Physics 82(2):e329–e336.
E. Pedroni, et al. (2004). ‘The PSI Gantry 2: a second generation proton
scanning gantry’. Zeitschrift für medizinische Physik 14(1):25–34.
E. Pedroni, et al. (2005). ‘Experimental characterization and physical modelling
of the dose distribution of scanned proton pencil beams’. Physics in medicine
and biology 50(3):541.
124
Bibliography
R. Perrin, et al. (2015). ‘OC-0484: Rescanning measurements in a 4D anthropomorphic phantom for evaluation of motion-mitigated, PBS proton therapy’.
Radiotherapy and Oncology 115:S238–S239.
R. Perrin, et al. (2014). ‘SU-D-BRE-01: A Realistic Breathing Phantom of the
Thorax for Testing New Motion Mitigation Techniques with Scanning Proton
Therapy’. Medical Physics 41(6):111–111.
R. Perrin, et al. (2016). ‘An anthropomorphic breathing phantom of the thorax
for testing new motion mitigation techniques for pencil beam scanning proton
therapy’. Physics in medicine and biology .
G. F. Persson, et al. (2010). ‘Deviations in delineated GTV caused by artefacts
in 4DCT’. Radiotherapy and Oncology 96(1):61–66.
M. H. Phillips, et al. (1992). ‘Effects of respiratory motion on dose uniformity
with a charged particle scanning method’. Physics in medicine and biology
37(1):223.
S. Pollock, et al. (2016). ‘The impact of breathing guidance and prospective
gating during thoracic 4DCT imaging: an XCAT study utilizing lung cancer
patient motion’. Physics in Medicine and Biology 61(17):6485.
S. Pollock, et al. (2013). ‘Audiovisual biofeedback improves motion prediction
accuracy’. Medical physics 40(4):041705.
T. Ravkilde, et al. (2013). ‘Time-resolved dose reconstruction by motion encoding of volumetric modulated arc therapy fields delivered with and without
dynamic multi-leaf collimator tracking’. Acta Oncologica 52(7):1497–1503.
S. Reinhardt, et al. (2012). ‘Comparison of Gafchromic EBT2 and EBT3 films
for clinical photon and proton beams’. Medical physics 39(8):5257–5262.
E. Rietzel & C. Bert (2010). ‘Respiratory motion management in particle therapy’. Medical physics 37(2):449–460.
E. Rietzel, et al. (2005). ‘Four-dimensional computed tomography: image formation and clinical protocol’. Medical physics 32(4):874–889.
G. Rodrigues, et al. (2004). ‘Prediction of radiation pneumonitis by dose-volume
histogram parameters in lung cancer—a systematic review’. Radiotherapy and
oncology 71(2):127–138.
BIBLIOGRAPHY
125
D. Scandurra, et al. (2016). ‘Assessing the quality of proton PBS treatment
delivery using machine log files: comprehensive analysis of clinical treatments
delivered at PSI Gantry 2’. Physics in medicine and biology 61(3):1171.
B. Schaffner, et al. (1999). ‘Dose calculation models for proton treatment planning using a dynamic beam delivery system: an attempt to include density
heterogeneity effects in the analytical dose calculation’. Physics in medicine
and biology 44(1):27.
A. Schätti, et al. (2013). ‘Experimental verification of motion mitigation of
discrete proton spot scanning by re-scanning’. Physics in medicine and biology
58(23):8555.
A. Schätti, et al. (2014). ‘The effectiveness of combined gating and re-scanning
for treating mobile targets with proton spot scanning. An experimental and
simulation-based investigation’. Physics in medicine and biology 59(14):3813.
A. Schlaefer, et al. (2005). ‘Feasibility of four-dimensional conformal planning
for robotic radiosurgery’. Medical physics 32(12):3786–3792.
U. Schneider (2009). ‘Mechanistic model of radiation-induced cancer after
fractionated radiotherapy using the linear-quadratic formula’. Medical physics
36(4):1138–1143.
U. Schneider, et al. (2002). ‘Secondary neutron dose during proton therapy
using spot scanning’. International Journal of Radiation Oncology* Biology*
Physics 53(1):244–251.
U. Schneider, et al. (1996). ‘The calibration of CT Hounsfield units for radiotherapy treatment planning’. Physics in medicine and biology 41(1):111.
U. Schneider, et al. (2011). ‘Site-specific dose-response relationships for cancer
induction from the combined Japanese A-bomb and Hodgkin cohorts for doses
relevant to radiotherapy’. Theor Biol Med Model 8:27.
J. Schuemann, et al. (2014). ‘Site-specific range uncertainties caused by dose
calculation algorithms for proton therapy’. Physics in medicine and biology
59(15):4007.
J. Seco, et al. (2009). ‘Breathing interplay effects during proton beam scanning: simulation and statistical analysis’. Physics in medicine and biology
54(14):N283.
126
Bibliography
W. Segars, et al. (2008). ‘Realistic CT simulation using the 4D XCAT phantom’.
Medical physics 35(8):3800–3808.
W. Segars, et al. (2010). ‘4D XCAT phantom for multimodality imaging research’. Medical physics 37(9):4902–4915.
V. Semenenko & X. Li (2007). ‘Lyman NTCP Model Parameters for Radiation Pneumonitis and Xerostomia Based on Combined Analysis of Published
Clinical Data’. International Journal of Radiation Oncology* Biology* Physics
69(3):S605–S606.
J. A. Shackleford, et al. (2010). ‘On developing B-spline registration algorithms
for multi-core processors’. Physics in medicine and biology 55(21):6329.
J. A. Shackleford, et al. (2012). ‘Analytic regularization of uniform cubic bspline deformation fields’. In Medical Image Computing and Computer-Assisted
Intervention–MICCAI 2012, pp. 122–129. Springer.
R. L. Siddon (1985). ‘Prism representation: a 3D ray-tracing algorithm for
radiotherapy applications’. Physics in medicine and biology 30(8):817.
A. Sodickson, et al. (2009). ‘Recurrent CT, cumulative radiation exposure,
and associated radiation-induced cancer risks from CT of adults 1’. Radiology
251(1):175–184.
J.-J. Sonke, et al. (2005). ‘Respiratory correlated cone beam CT’. Medical
physics 32(4):1176–1186.
R EBT3 films characterisJ. Sorriaux, et al. (2013). ‘Evaluation of Gafchromic
tics in therapy photon, electron and proton beams’. Physica Medica 29(6):599–
606.
P. Steidl, et al. (2012). ‘A breathing thorax phantom with independently programmable 6D tumour motion for dosimetric measurements in radiation therapy’. Physics in medicine and biology 57(8):2235.
C. W. Stevens, et al. (2001). ‘Respiratory-driven lung tumor motion is independent of tumor size, tumor location, and pulmonary function’. International
Journal of Radiation Oncology* Biology* Physics 51(1):62–68.
Y. Suh, et al. (2008). ‘An analysis of thoracic and abdominal tumour motion
for stereotactic body radiotherapy patients’. Physics in medicine and biology
53(13):3623.
BIBLIOGRAPHY
127
W. Takahashi, et al. (2013). ‘Verification of planning target volume settings in
volumetric modulated arc therapy for stereotactic body radiation therapy by
using in-treatment 4-dimensional cone beam computed tomography’. International Journal of Radiation Oncology* Biology* Physics 86(3):426–431.
R. Tao, et al. (2016). ‘Ablative radiotherapy doses lead to a substantial prolongation of survival in patients with inoperable intrahepatic cholangiocarcinoma:
A retrospective dose response analysis’. Journal of Clinical Oncology 34(3):219–
226.
D. Thomas, et al. (2014). ‘A novel fast helical 4D-CT acquisition technique to generate low-noise sorting artifact–free images at user-selected breathing phases’. International Journal of Radiation Oncology* Biology* Physics
89(1):191–198.
A. Tourovsky, et al. (2005). ‘Monte Carlo dose calculations for spot scanned
proton therapy’. Physics in medicine and biology 50(5):971.
A. Trofimov, et al. (2005). ‘Temporo-spatial IMRT optimization: concepts,
implementation and initial results’. Physics in medicine and biology 50(12):2779.
N. Tyagi, et al. (2012). ‘A real time dose monitoring and dose reconstruction tool
for patient specific VMAT QA and delivery’. Medical physics 39(12):7194–7204.
J. Valentin (2002). ‘Basic anatomical and physiological data for use in radiological protection: reference values: ICRP Publication 89’. Annals of the ICRP
32(3):1–277.
S. Van de Water, et al. (2009). ‘Tumour tracking with scanned proton beams:
assessing the accuracy and practicalities’. Physics in medicine and biology
54(21):6549.
S. Vedam, et al. (2001). ‘Determining parameters for respiration-gated radiotherapy’. Medical physics 28(10):2139–2146.
M. Von Siebenthal, et al. (2007). ‘4D MR imaging of respiratory organ motion
and its variability’. Physics in medicine and biology 52(6):1547.
N. Wang, et al. (2013). ‘Evaluation and comparison of New 4DCT based strategies for proton treatment planning for lung tumors’. Radiation Oncology 8(1):1.
128
Bibliography
D. C. Weber, et al. (2004). ‘A treatment planning comparison of intensity
modulated photon and proton therapy for paraspinal sarcomas’. International
Journal of Radiation Oncology* Biology* Physics 58(5):1596–1606.
C. Winterhalter, et al. (2016). ‘Monte Carlo model of a gantry for proton pencil
beam scanning’. In SASRO/SSRMP annual meeting.
J. W. Wolthaus, et al. (2006). ‘Mid-ventilation CT scan construction from fourdimensional respiration-correlated CT scans for radiotherapy planning of lung
cancer patients’. International Journal of Radiation Oncology* Biology* Physics
65(5):1560–1571.
H. Wu, et al. (2008). ‘Gating based on internal/external signals with dynamic
correlation updates’. Physics in medicine and biology 53(24):7137.
T. Yamamoto, et al. (2013). ‘4D CT lung ventilation images are affected by the
4D CT sorting method’. Medical physics 40(10):101907.
T. Yamamoto, et al. (2008). ‘Retrospective analysis of artifacts in fourdimensional CT images of 50 abdominal and thoracic radiotherapy patients’. International Journal of Radiation Oncology* Biology* Physics 72(4):1250–1258.
Y. Zhang, et al. (2013). ‘Deformable motion reconstruction for scanned proton
beam therapy using on-line x-ray imaging’. Physics in medicine and biology
58(24):8621.
Y. Zhang, et al. (2014). ‘Online image guided tumour tracking with scanned
proton beams: a comprehensive simulation study’. Physics in medicine and
biology 59(24):7793.
Y. Zhang, et al. (2015). ‘Improving 4D plan quality for PBS-based liver tumour
treatments by combining online image guided beam gating with rescanning’.
Physics in medicine and biology 60(20):8141.
L. Zhao & I. J. Das (2010). ‘Gafchromic EBT film dosimetry in proton beams’.
Physics in medicine and biology 55(10):N291.
K. H. Zou, et al. (2004). ‘Statistical validation of image segmentation quality based on a spatial overlap index 1: Scientific reports’. Academic radiology
11(2):178–189.

Ensure to take tramadol buy pelosta-5 properly and

A Monte Carlo based dose calculation and evaluation method for

CHEMOTHERAPY

IG Sub-Q Injection Referral Form

WedgeCheck

ClearCT™ Lung

4 EIIERGENCY EXPOSURE SITUA ONS - OurPlanet-TV

In Vivo Dosimetry: A Model for Correcting Dose Measurements of

39241 - SENSORY BLOCK LEVELS DURING CSE IN LABOR

Antibiogram 2014 - ILH Interim LSU Hospital