Development and Testing of an Automatic Lung
IMRT Planning Algorithm
by
Wei Zhu
Medical Physics Program
Duke Kunshan University and Duke University
Date:
Approved:
Q. Jackie Wu, Supervisor
David Huang
Anuj Kapadia
Thesis submitted in partial fulfillment of the requirements for the degree of
Master of Science in the Medical Physics Program
in Duke Kunshan University and Duke University
2016
Abstract
Development and Testing of an Automatic Lung IMRT
Planning Algorithm
by
Wei Zhu
Medical Physics Program
Duke Kunshan University and Duke University
Date:
Approved:
Q. Jackie Wu, Supervisor
David Huang
Anuj Kapadia
An abstract of a dissertation submitted in partial fulfillment of the requirements for
the degree of Master of Science in the Medical Physics Program
in Duke Kunshan University and Duke University
2016
c 2016 by Wei Zhu
Copyright All rights reserved except the rights granted by the
Creative Commons Attribution-Noncommercial Licence
Abstract
Knowledge-based radiation treatment is an emerging concept in radiotherapy. It
mainly refers to the technique that can guide or automate treatment planning in
clinic by learning from prior knowledge. Different models were developed to realize
its application, such as the model proposed by Yuan et al. at Duke University Medical Center for lung intensity modulated radiation therapy (IMRT) planning. This
model can automatically determine both beam configuration and optimization objectives with non-coplanar beams based on patient-specific anatomical information.
Although plans automatically generated by this model demonstrate equivalent or
better dosimetric quality compared to clinically approved plans, its validity and generality are limited due to the empirical assignment to a coefficient called angle spread
score defined in the beam efficiency index used for beam ranking. To eliminate these
limitations, a systematic study on this coefficient is needed to acquire evidence for
its optimal value.
To achieve this purpose, eleven lung cancer patients with complex tumor shapes
with non-coplanar beams adopted in clinically approved plans were retrospectively
studied in the frame of the automatic lung IMRT treatment algorithm. The primary
and boost plans used in three patients were treated as different cases due to the
different target size and shape. A total of 14 lung cases, thus, were re-planned using
the knowledge-based automatic lung IMRT planning algorithm by varying angle
spread score from 0 to 1 with increment of 0.2. A modified beam angle efficiency
iv
index was used to navigate the beam selection. Great effort was made to assure the
quality of plans associated with every angle spread score was as good as possible.
Important dosimetric parameters for planning target volume (PTV) and organs-atrisk (OARs), quantitatively reflecting the plan quality, were extracted from the DVHs
and analyzed as a function of angle spread constraint for each case. Comparisons
of these parameters between clinical plans and model-based plans were evaluated
by two-sampled Students t-tests, and analysis was performed on a composite index
built on the percentage errors between dosimetric parameters in the model-based
plans and those in the clinical plans as a function of angle spread score.
Results show that model-based plans generally have equivalent or better quality
than clinically approved plans, both qualitatively and quantitatively. All dosimetric
parameters except those for lungs in the automatically generated plans are statistically better or comparable to those in the clinical plans. On average, more than 15%
improvement on the conformity index and homogeneity index for PTV and V40 , V60
for heart while an 8% and 3% deterioration on V5 , V20 for lungs, respectively, are
observed. The intra-plan comparison among model-based plans demonstrates that
plan quality does not change much with angle spread score larger than 0.4. Further
examination on the variation of the composite index as a function of angle spread
score shows that 0.6 is the optimal value that can result in statistically the best
achievable plans.
v
Contents
Abstract
iv
List of Tables
viii
List of Figures
ix
List of Abbreviations and Symbols
xi
Acknowledgements
xiii
1 Introduction
1
1.1
Intensity modulated radiation therapy . . . . . . . . . . . . . . . . .
1
1.2
IMRT treatment planning procedure . . . . . . . . . . . . . . . . . .
2
1.2.1
Beam angle selection . . . . . . . . . . . . . . . . . . . . . . .
4
1.2.2
Structure constraint determination . . . . . . . . . . . . . . .
6
1.3
Automatic IMRT treatment planning . . . . . . . . . . . . . . . . . .
7
1.4
Study purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2 Materials and Methods
9
2.1
Polynomial factors in the BAS model . . . . . . . . . . . . . . . . . .
9
2.2
Study design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
2.2.1
Plan template generation . . . . . . . . . . . . . . . . . . . . .
11
2.2.2
Objective template generation . . . . . . . . . . . . . . . . . .
13
2.2.3
Automatic plan generation . . . . . . . . . . . . . . . . . . . .
14
Plan evaluation and data analysis . . . . . . . . . . . . . . . . . . . .
18
2.3
vi
2.3.1
Plan evaluation . . . . . . . . . . . . . . . . . . . . . . . . . .
18
2.3.2
Data analysis . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
3 Results
21
3.1
Qualitative results . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
3.2
Quantitative results . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
4 Discussion
29
4.1
BAS model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
4.2
OAR DVH prediction model . . . . . . . . . . . . . . . . . . . . . . .
31
5 Conclusion
34
A Beam configurations and DVHs
35
Bibliography
39
vii
List of Tables
2.1
Duke clinical protocol for lung cancer . . . . . . . . . . . . . . . . . .
13
2.2
The planning objectives used in lung IMRT planning . . . . . . . . .
15
3.1
The mean of the dosimetric parameters over all plans . . . . . . . . .
25
viii
List of Figures
2.1
The user interface of BAS model. By the courtesy of Yang Sheng . .
12
2.2
The user interface of OAR DVH prediction model.
. . . . . . . . . .
14
2.3
The flowchart of automatic plan generation. . . . . . . . . . . . . . .
15
2.4
Ring-body used to control entrance dose. . . . . . . . . . . . . . . . .
17
2.5
NTO setup in Eclipse TPS . . . . . . . . . . . . . . . . . . . . . . . .
18
3.1
Comparison diagram of beam configuration for a lung case . . . . . .
22
3.2
Comparison diagram of dose distribution between a clinical approved
plan and a model-based plan for a lung case . . . . . . . . . . . . . .
23
Comparison diagram of DVHs for PTV and OARs between a clinical
approved plan and model-based plans for a lung case . . . . . . . . .
24
3.3
3.4
Dose metrics boxplots of clinical approved plans and model-based plans 26
3.5
Bar plot of the percentage error of dose metrics between the averages
of clinical and model-based plans . . . . . . . . . . . . . . . . . . . .
26
3.6
Dose metrics as a function of angle spread constraint k for one case .
27
3.7
Average of dose metrics as a function of angle spread constraint k over
all plans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
3.8
Plot of composite index as a function of angle spread constraint . . .
28
4.1
Plot of beam spread term as a function of angle separation with δ
taking various values . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
A.1 Beam configurations for all lung cases . . . . . . . . . . . . . . . . . .
36
A.2 Comparison diagram of DVHs for PTV and OARs between a clinical
approved plan and model-based plans for all cases(1) . . . . . . . . .
37
ix
A.3 Comparison diagram of DVHs for PTV and OARs between a clinical
approved plan and model-based plans for all cases(2) . . . . . . . . .
x
38
List of Abbreviations and Symbols
Symbols
The following symbols are used to define dosimetric parameters.
Vx
Dx :
the percentage of the structure volume exceeding x Gy or x% of
the prescription dose.
dose received by x% of the organ volume.
Abbreviations
3D-CRT
3 Dimensional Conformal Radiation Therapy
BEV
Beam Eye View
BAS
Beam Angle Selection
CI
Conformity Index
CT
Computed Tomography
DVH
Dose Volume Histogram
FMO
Fluence Map Optimization
HI
IMRT
IRB
MLC
NSCLC
NTO
Homogeneity Index
Intensity Modulated Radiation Therapy
Institutional Review Board
MultiLeaf Collimator
Non-Small Cell Lung Cancer
Normal Tissue Objectives
xi
OAR
OF
Organ at Risk
Objective Function
OVH
Overlap Volume Histogram
PCA
Principal Component Analysis
PDD
Percent Depth Dose
QA
QUANTEC
RTOG
TPS
Quality Assurance
Quantitative Analyses of Normal Tissue Effects in the Clinic
Radiation Therapy Oncology Group
Treatment Planning System
xii
Acknowledgements
I would like to first express my sincere appreciation to Dr. Jackie Wu, my supervisor,
for her constructive opinions and patient instructions on my research, hands-on guidance on my clinical treatment planning practice as well as her concern and support
for my career during my precious study at Duke. Sincere gratitude is also given to
Dr. Lulin Yuan for his resourcefulness, patience and generosity in finding time from
his busy residence schedule to help me solve my research problems. Great thanks
should also go to Dr. Fang-fang Yin, who has provided the best resources and offered
great help for my research. I would also like to thank Sheng Yang, Chunhao Wang,
Lei Zhang, Yawei Zhang for their valuable suggestions and advice about my research.
I am also grateful for Dr. James Bowsher’s advice and critique of my thesis work.
I would also like to thank Dr. David Huang and Dr. Anuj Kapadia for being my
defense committee members.
xiii
1
Introduction
As one of the most common cancers occurring worldwide, lung cancer has the highest
mortality rate at 1 in 5 of all cancer deaths (1). Various types of treatments have been
developed to eliminate lung cancers, one of which is the treatment using radiation
beams. For example, for the locally advanced non-small cell lung cancer (NSCLC),
three-dimensional conformal radiation therapy (3D-CRT) and intensity modulated
radiation therapy (IMRT) are the most common treatment types. For decades 3DCRT has dominated the radiation treatment for NSCLC, but IMRT, after more
than a decade of development, has gained regular adoption in clinics due to its lower
dosemetric toxicity to normal tissues compared to 3D-CRT (2; 3).
1.1 Intensity modulated radiation therapy
IMRT inherits many tools from 3D-CRT, such as three dimensional computed tomography (CT) images for contouring, external collimators to shape the radiation
beam to fit the target, and the three dimensional dose calculation algorithm (4).
In addition, IMRT has several other radical features. First, instead of the uniform
yield of radiation from each field in 3D-CRT, intensity within a field in IMRT is
1
further modulated, resulting in a non-uniform energy distribution, while the total
dose deposited in the target from all the beams is kept homogeneous. Specifically,
a radiation beam is divided into small beamlets, the intensity of which is adjusted
based on an optimization objective function. This feature enables IMRT to generate
complex (e.g., concave) but conformal dose distributions, better achieving the goal
of delivering sufficient dose to kill the tumor while sparing the normal tissues as
much as possible. To realize radiation intensity modulation, multi-leaf collimator
(MLC) plays a significant role. Two MLC motion schemes are commonly used: sliding window(5) and static step-and-shoot(6). In the former method, MLC moves as
the irradiation is on while the MLC in the later mode only moves as the beam is off.
The movement of MLC is controlled by the leaf motion algorithm. Another feature of
IMRT is the inverse planning process (5), which utilizes the expected dosimetric constraints for structures as inputs to iteratively optimize the intensity or fluence map
of each field via minimizing an objective function. The target and critical structure
constraints are either dose-based or dose-volume-based and each of them is assigned
a priority score which indicates the relative importance of these parameters.
Although IMRT for lung cancer shows a superiority over other techniques, it
presents new challenges and creates additional work in current clinical practice as
well, e.g., the complex and time-consuming treatment planning process, the accuracy
of radiation delivery using MLC and the lengthy IMRT QA procedures(4). Any
improvement of these processes can greatly benefit the current IMRT practice as
well as patient care and this study focuses on the improvement of IMRT treatment
planning for lung cases.
1.2 IMRT treatment planning procedure
General IMRT treatment planning starts with contouring of target and critical organs
based on 3D CT or MRI images. Based on target shape, size and relative location to
2
other critical organs, multiple radiation beams are carefully chosen and placed around
the target to minimize unnecessary penetration through normal tissues. With the
beam configuration defined, fluence optimization for each beam can be iteratively
performed after dose prescription is specified and certain constraints for target and
critical organs are given. The setup of these structure constraints reflects the tradeoffs between dose coverage to the target and dose sparing to organs-at-risk (OARs)
and is highly dependent on the planner’s experience. After the optimal fluence map
is obtained, leaf motion sequence and dose calculation are performed and a plan is
generated. The plan quality is then evaluated in terms of the dose distribution and
the DVH curves. If the plan is clinically acceptable, it can be approved to treat
patients; if not, the planner needs to return to one of the previous steps to tune
either the structure objectives or beam angles and repeat the whole optimization
process until the final plan meets the requirements.
The IMRT treatment planning, therefore, is a complex and time-consuming procedure and highly depends on the planner’s experience and knowledge. Further,
these challenges become even more significant for lung cases. First, lung tumors can
occur in different positions in the thorax and its geometrical relationship with other
critical organs (spinal cord, lungs, heart, esophagus as well as some less affected
organs such as the stomach and liver) varies from case to case. This causes more
difficulties in beam setup compared to other tumor types such as prostate where the
beam configuration is relatively stable. Moreover, several studies show that noncoplanar beams can improve the dosimetric quality in many lung cases (7; 8), but
the expanded beam angle searching space (from 2D to 3D) significantly increases the
challenge in beam angle selection. Second, the involvement of multiple organs-at-risk
(OARs) raises the complexity of embracing critical structure constraints. Although
Radiation Therapy Oncology Group (RTOG) protocols (0617 and 0813) and Quantitative Analysis of Normal Tissue Effects in the Clinic (QUANTEC) (9) provide
3
guidelines, they are population-based and offer only the minimum requirements.
The optimal or near optimal objective setup can only be determined through a timeconsuming trial-and-error process. Furthermore, dose distribution of all the organs
are closely correlated and many optimums probably exist corresponding to different
clinical concerns. Thus, the planner’s experience is important in this step in terms
of both efficiency and quality. Finally, due to the large thorax size and the tissue
inhomogeneity, hot spots emerge easily so that extra effort is required to eliminate
them.
As IMRT becomes a routine clinical radiotherapy practice for lung cancers, there
is a clinical need to reduce the planning time as well as enhance the plan quality and
consistency, especially in two areas: beam angle selection (BAS) and patient-specific
structure constraints determination.
1.2.1
Beam angle selection
Beam angle selection is an essential step in the IMRT treatment planning process.
Clinically, either a standard radiation beam configuration is used (e.g., seven to nine
equally spaced coplanar beams for head and neck and five to seven fixed angles for
prostate), or a manual exploration is conducted for intracranial, thoracic, abdominal
tumors, or cases where non-coplanar beams are employed.
Mathematically, BAS is typically considered a combinatorial optimization problem, finding an optimal combination of beam angles from a large discrete angle
searching space (10). Many approaches have been developed to automatically select
the beams, which can generally be divided into two categories. In the first category
a global optimization strategy is adopted in which the BAS and fluence map optimization (FMO) are solved simultaneously. For example, stochastic search methods
such as simulated annealing (11; 12) and genetic algorithm (13; 14; 15) are applied to
BAS to obtain an optimized beam ensemble in consideration of the beam orientation
4
combinations and the associated beam profiles (beamlet weighting); some research
groups, however, reduce the complexity by iteratively adding (16; 17; 18) beams to
or eliminating (19; 20) beams from a radiotherapy plan. A recent study (21) further
accelerates the iterative BAS based on the score of an objective function after a limited number of FMO iterations or a projected gradient attained in the first iteration.
Whereas the BAS runtime reduces dramatically with the help of some acceleration
methods (17; 21), approaches incorporating FMO with BAS are still computationally expensive because of the requirement of dose calculation and beamlet weight
optimization for every beam angle set. Things get even tougher for complex cases
where the number of candidate beams becomes larger and/or non-coplanar beams
are proposed.
The other strategy to deal with BAS is to treat beam selection and intensity
optimization separately. Pugachev et al.(22) and Meyer et al.(23) create different
scalar score functions to rank radiation beams in terms of the dose to both target and
OARs along the transmission path by introducing a pseudo beam’s-eye-view (pBEV)
technique and a BEV ray-tracing method, respectively. Bangert and Oelfke (24),
employing a similar idea, construct a radiological quality index for each potential
angle and target voxel and classify the voxels into several clusters of which the
centers express the optimal angles. Potrebko et al.(25) extract pure geometrical
information from the analysis of polygonal surface mesh data of contoured patients
anatomy, indicating the optimal beams are those that tangentially bisect the target
and OARs while parallel to the polygon flat surfaces of the target. Also, a research
group from Duke Medical Center (26) suggests a set of standardized beam bouquets
for lung IMRT planning using cluster analysis by learning the beam configuration
features from previous high quality plans. Apparently, the calculation time of the
above methods in this class is competitive, but there is evidence indicating that the
plans generated by these methods may be no better or may even be worse than the
5
clinical plans or those generated by other methods (27).
Another notable point upon reviewing the literature is that few of these publications apply their methods to lung cases. This is likely due to the complicated
thoracic geometry and the involvement of so many critical organs, which increase the
uncertainties of BAS. Thus, to face this challenge, another BAS model which can
automatically determine beam configurations involving non-coplanar beams in lung
IMRT planning is being developed at Duke Medical Center, the details of which are
introduced in Chapter 2.
1.2.2
Structure constraint determination
With selected radiation beams, manually or algorithm-aided, IMRT optimization
needs to be performed to acquire beamlet intensity for each direction. An objective
function (OF) is iteratively calculated to guide this IMRT optimization after dosebased constraints and/or dose-volume (DV)-based constraints are specified. Dosebased constraints are straightforward which specify prescription dose to target while
zero to normal tissues and OARs, but they may lead to undesired dose distributions
for multi-objective optimization problems such as those in lung IMRT because of
the hard constraints; DV-based constraints, in contrast, reserve more tolerance room
for different organs so that an optimal overall plan can be obtained after the optimization (5; 28). Therefore, DV-based constraints are most commonly used in clinic.
Nevertheless, the determination of a set of structure constraints leading to an optimal dose distribution for a particular beam setup for a specific patient is generally
non-trivial. Sometimes, a time-consuming trial and error process cannot be avoided
even for experienced planners.
To address this issue, knowledge-based treatment planning is proposed with the
assumption that previous knowledge that lay behind the high quality plans can be
extracted and applied to new cases to make the planning more efficient but at the
6
same time, with no deviation to plan quality. Several methods to automatically determine the structure constraints are reported. By building up a large database of
prostate plans, Chanyavanich et al. (29) match the new patient’s anatomy with the
old one from the database based on mutual information between the BEV projections
of the reference 3D images from different angles. Similarly, Wu et al. (30) compare
the overlap volume histogram (OVH) indicating the spatial interrelations between
target and OARs of a new patient with those in the database to generate dose volume histogram (DVH) objectives as the initial optimization goals for the new case.
However, these two methods highly rely on the size and quality of the plans in the library. To eliminate the database reliability, Appenzoller et al. (31) develop an OAR
DVH prediction model based on the correlation of expected dose to the minimum
distance from a voxel to the PTV surface while Yuan et al. (32) exploit a stepwise
multiple regression method to find the link between dosimetric and anatomical features extracted from DVHs and distance-to-target histograms (DTHs), respectively,
using principle component analysis (PCA).
1.3 Automatic IMRT treatment planning
Equipped with automatic BAS tools and patient-specific structure constraint determination tools, a fully automatic IMRT treatment planning method can be achieved.
Zhang et al. (33; 34) first report a methodology for automatic lung IMRT planning
and build an external plug-in system named mdaccAutoPlan to help generate IMRT
plans for lung cases. Breedveld et al. (17; 35) also introduce a multicriterial plan
optimization algorithm called iCycle to help design coplanar or non-coplanar IMRT
plans. However, these methods are either database dependent in selecting initial
beam set (33) or time-consuming by using an iterative method (17). In order to
improve planning efficiency, consistency and quality, a novel knowledge-based automatic IMRT treatment planning algorithm is suggested by Yuan et al. at Duke
7
Medical Center. This algorithm combines a new automatic BAS method with the
automatic determination of dosimetric objectives. The automatic determination of
structure objectives for fluence map optimization takes advantage of the OAR DVH
prediction model (32) while beam configuration in the new BAS model is determined by analyzing patient-specific anatomical features and learning from previous
planning experiences. More specifically, beam angles are ranked and chosen based
on a beam efficiency index which takes into account both the dose contributions
from individual beams and the combined effect of multiple beams. However, several
polynomial factors in the efficiency index reflecting clinical tradeoffs among different
OARs are assigned empirically.
1.4 Study purpose
The empirical values of those polynomial factors appearing in the BAS model make
the whole algorithm less robust. To further validate the effectiveness and generality
of this automatic lung IMRT treatment planning algorithm, the values of these polynomial factors need to be systematically studied. Our hypothesis is that these factors
are independent on the patient specific anatomical information, and the optimal values for these factors exist. That is, once this optimal set of factors is determined, the
BAS model is supposed to generate an optimal or near optimal beam configuration
for any specific patient which can further lead to a high quality plan. To verify our
hypothesis, the relationship between plan quality and those polynomial factors was
investigated by applying this automatic lung IMRT treatment planning algorithm to
clinical patient cases in this study.
This thesis is structured as follows: in Chapter 2, the polynomial factors in the
BAS model are introduced in detail and the method used to study and obtain the
optimal values of these factors are discussed. The results are shown and analyzed in
Chapter 3. The discussions on the results are found in Chapter 4 while conclusions
8
are summarized in the final Chapter 5.
9
2
Materials and Methods
2.1 Polynomial factors in the BAS model
In the BAS model developed by Yuan et al. at Duke, a beam efficiency index is
constructed for each candidate beam direction α specified by the couch and gantry
angles with a full consideration of the potential device collisions. Its mathematical
expression is as follows:
qα
N1
¸
P
qαν
(2.1)
α ν Sα
PTV
where ν expresses the original voxel on the target surface SPα T V through which the
radiation goes into the target from direction α, Nα is the number of such voxels and
qαν is defined as the ratio of the weighted sum of the dose deposited in the OARs and
normal tissues to that deposited in the target along the beam path through voxel ν
in direction α and is expressed as
qαν
°
P
n Structure
dνtarget
where dνn is the dose deposited in structure n
wn dνn
P
(2.2)
Structure, which can be target,
lung, esophagus, spinal cord, heart, liver, kidney, stomach or normal tissues such as
10
muscles and bones, wn is the corresponding weighing factor assigned to each structure. The dose deposition along the beam path is calculated based on a tabulated
percent depth dose (PDD) curve for photon beams with corrections for lung tissue
inhomogeneity. The weighting factors, representing the trade-offs among different
structures, are determined empirically based on the prior knowledge. Therefore,
Eq. 2.1 for a candidate beam angle is the average of dosimetric score 2.2 for each
beamlet in that direction. Another way to construct the index is to take the mediant
of all the qαν without simplification:
q1
α
°°
°P
ν
n Structure
ν
ν dtarget
wn dνn
(2.3)
The denominator is the total dose deposited to target in beam direction α while
the numerator is the total dose received by all the critical structures. Beams are
selected as those that can deliver more dose to the whole target and less dose to the
nearby normal tissue at the same time. Compared with qα , qα1 emphasizes the total
dosimetric effect of radiation beams and it is used in this study.
In addition, as it is demonstrated that beam separation has an influence on the
dosimetric quality, especially on the PTV dose conformity (23; 36), a beam separation
term which is inversely proportional to the angle difference between each angle pair
is added to the beam efficiency index. For a beam configuration containing m beams,
the beam efficiency index (EI) is
EI
¸
m
i 1
qαi
¸
2k
1 cosαij
ij
δ
(2.4)
where αij is the angle separation between angles αi and αj , δ is a small number (! 1)
to keep the denominator from 0 and the positive coefficient k denotes the relative
importance of beam separation in beam selection. Based on the EI values, the beam
angles for a particular case are chosen using a greedy algorithm.
11
By examining Eq. 2.4, there are two sets of polynomial factors that deserve further
explanation: one set is a group of weighting factors in the first term, representing the
relative weight of the dose contributions from different structures; the other set is the
beam angle spread score k in the second term, characterizing the combined effect of
multiple beams. From a clinical aspect, these factors are actually embodiments of a
series of trade-offs between dose sparing goals of different OARs and beam separation
extent. Hence, the assignment of these factors is of significance. As we suppose these
factors are independent of patient’s anatomy, the optimal values exist. To find these
values, the relationship between plan quality and these factors should be explored.
2.2 Study design
This study involved eleven lung cancer patients with clinically approved IMRT plans
using non-coplanar beams and was approved by the Duke University Medical Center
Institutional Review Board (IRB). Among these patients, three of them used primary
and boost plans. As the primary PTV and the boost PTV for each of these three
patients were different, the primary and boost plans were treated as different cases.
Therefore, a total of fourteen lung IMRT plans with prescription doses ranging from
60 Gy to 70 Gy and tumor sizes from 151.48 to 521.20 cm3 were retrospectively
analyzed. In these clinical plans, six to eleven beams including non-coplanar ones
are used with an average of eight beams.
2.2.1
Plan template generation
Based on patient images and structure contours, beam angles were reselected using
the BAS model coded in MATLAB. Fig. 2.1 shows the user interface of this model.
As the picture shows, there are five user inputs: weighting factors, angle spread
constraint, max number of non-coplanar angles, number of angles and case ID. Case
ID indicates the plan number for a certain patient. Number of angles is set to eight
12
Figure 2.1: The user interface of BAS model. By the courtesy of Yang Sheng
which is the average beam number used for lung cancer treatment in clinic. Max
number of non-coplanar angles is four, which is consistent with the clinical practice.
The weighting factors and angle spread constraint are the polynomial factors in the
beam efficiency index with their optimal values need to be determined. Due to the
large number of factors, the control variable method was adopted, i.e. only one
factor was tuned to generate a series of plans for each patient while others were kept
unchanged within reasonable empirical values. In this study, the set of weighting
factors as shown in the figure was found empirically based on the dosimetric tolerance
of different organs, i.e. spinal cord lung esophagus and heart other organs and
normal tissues. In addition, the weighting factor for PTV was negative, indicating the
inverse relationship between PTV and EI. With a set of fixed weighting factors, the
only variable was the non-negative factor angle spread constraint, or more precisely,
angle spread score k. Further testing showed no significant changes happened when
this factor took values larger than one for different cases. Hence, the range of angle
spread score was r0, 1s with an increment of 0.2. For each patient case, six beam
13
Table 2.1: Dosimetric constraints for lung IMRT used in Duke University Cancer
Center
PTV Constraints
Prescription Dose
Volume
100%
¡ 95%
95%
Minimum
120%
Maximum
OAR Constraints
OARs
Spinal cord
Spinal cord+3mm
Lung
Lung
Esophagus
Esophagus
Heart
Heart
Heart
Dose
45 Gy
50 Gy
5 Gy
20 Gy
20 Gy
60 Gy
20 Gy
24 Gy
40 Gy
Volume
Max
Max
50%
30%/Mean
50%
25%
100%
Mean
50%
angle sets then were generated and a total of 84 sets were produced. The information
of beam angles was stored in .xml files which could be plugged into the Eclipse TPS
as plan templates.
2.2.2
Objective template generation
After the beam configuration was determined, fluence map optimization for each
angle was performed in the Eclipse planning system, identical to the clinical planning
platform. Typically the dose constraints were prescribed by physicians following
either the RTOG guidelines or an institute template such as Duke clinical protocol for
lung cancer as show in Table 2.1. To achieve these goals, dose- or dose-volume-based
objective points were manually added and adjusted in a trial and error process by
a physicist during fluence optimization. In this study, however, instead of manually
setting up structure constraints for each patient, a well-tested knowledge-based DVH
prediction model (32) was employed which correlates patient anatomical features
with dosimetric features by learning from prior plans. Before using this model for
prediction, it should first be trained with high quality lung IMRT plans. Dozens
of approved lung IMRT cases randomly chosen from the Duke database were used
14
Figure 2.2: The user interface of OAR DVH prediction model.
for the training. After training and validation, the model can work on new cases
independently from the database. Fig. 2.2 shows the user interface of this model.
To predict dose parameters for a new patient, DICOM files with structure contours were first imported and read into this software. DVH prediction then could
start once PTV and OARs were matched and dose prescription was specified. From
the predicted OAR DVH curves, specific dose-volume points were extracted and
stored in .xml files as objective templates. As to the dose coverage objectives for PTV
and priorities for each objective, they were determined based on previous planning
knowledge and they were not modified during the planning. The complete planning
objectives used for lung IMRT planning in this study are shown in Table 2.2.
2.2.3
Automatic plan generation
By incorporating the DVH prediction model into the BAS model, IMRT planning was
automatically performed for all patients with minimum or no human intervention.
15
Table 2.2: The planning objectives and the corresponding weighting factors used in
lung IMRT planning
OARs/PTV Lungs Esophagus Heart
V5
V20
V60
Indices
V10
V60
V20
Weighting
70
70
70
Cord
D50
D2
D0
Cord+3mm
D50
D2
D0
30, 80, 150 30, 80, 150
PTV
D0
D95
D98
D100
120
Figure 2.3: The flowchart of automatic plan generation.
The overall flowchart of automatic plan generation is shown in Fig. 2.3. As can
be seen in the flowchart, a set of patient images with contouring serves as the sole
input for the BAS model as well as for the DVH prediction model. Beam selection
and objective determination were parallel computed to produce plan templates and
objective templates respectively, which could be plugged into Eclipse TPS to guide
the plan generation.
Ideally, to obtain the relationship between plan quality and beam separation
score, the latter should be the only independent variable of the former, i.e. there
should be no other variables that could impact plan quality. Some additional effort
was put in place to achieve the clinically best plans:
• Optimization objective adjustment
16
Although favorable evidence (37; 38) has proved the validity of the OAR
DVH prediction model, dose-volume objectives automatically derived from this
model are not necessarily the absolute best because of the limitations of the
”best” available knowledge from the past (32). Therefore, adjustments to these
objectives were implemented for some cases in this study to achieve a tighter
dose distribution. The whole process was a trial-and-error process.
• Rings
Rings are supplementary structures commonly used and manually delineated
by physicists in clinic to acquire a sharp dose falloff around the targets and
reduce hot spots in healthy tissues. Current Eclipse TPS (Varian Medical
Systems, Palo Alto, CA) can also achieve these goals by using the Normal
Tissue Objectives (NTO) during photon beam optimization. The shape of the
NTO is calculated as a function of the distance away from the target border.
To compare the effectiveness of rings and NTO in dose control, plans using
only rings with different widths and locations and plans using only NTO with
proper dosimetric setup were generated and compared for several patients. It
turned out that for plans using rings, two rings were sufficient to result in high
quality plans: one with width of 1 cm and 1 cm away from PTV (ring-PTV);
the other with width of 2 cm and 3 cm away from PTV while 1 cm inward from
the body border (ring-body). Plans using NTO, however, showed better dose
conformity, but less control to the entrance dose in regions far from PTV, i.e.
hot spots were more likely to appear in these plans. In view of the advantages of
both methods, a combination of NTO with ring-body was used for dose control
in this study, as shown in Fig. 2.4 and 2.5. The dose constraint for ring-body
was that no volume would be above 85% of the prescription dose. Parameters
for the NTO were set as 0.5, 105, 50, 0.4 for distance from the target border
17
Figure 2.4: Ring-body used to control entrance dose.
(cm), start dose (%), end dose (%) and fall-off, respectively. The priorities for
both ring-body and NTO are equal to that for PTV. These parameters were
determined by experience.
• Dose structures
After the previous steps, near optimal plans were achieved for most of the
cases. For some cases with complex tumor shape, nonetheless, dose structures
converted from isodose levels were used to further eliminate hot spots and limit
the dose spread out. This control region was created by subtracting PTV with
2 cm expansion from the 95% of prescription dose region. The dose constraint
was 95% of the prescription dose with the same priority of PTV. This post
planning step is performed in clinical plans routinely and hence implemented
in this study as well.
In addition, clinical planning steps, such as fixed jaws, structure resolution adjustment, fluence smoothing after optimization were also used for some cases.
With these efforts, automatically generated plans had the best dosimetric quality
and the variation of the quality only depended on the beam angle spread score in
18
Figure 2.5: NTO setup in Eclipse TPS
the BAS model. Plan evaluation and statistic analysis were performed.
2.3 Plan evaluation and data analysis
2.3.1
Plan evaluation
Two categories of metrics were used to assess the quality of the automatically generated plans: qualitative analysis and quantitative analysis. Qualitative analysis
involves the check of spatial dose distributions on each image slice and DVH curves
for each structure. In qualitative analysis, indexes such as dose coverage, dose uniformity, dose gradient, hot spots, etc. for every plan are evaluated and recorded.
As to the quantitative analysis, important dosimetric parameters are selected based
on RTOG protocols (0617 and 0813) and literature (39) for plan evaluation, among
which conformity index (CI), homogeneity index (HI), dose spillage (R50 ) and D2cm
19
are used to estimate the PTV dose coverage while other dose-volume data, such as
Dmax for spinal cord, V5 , V20 , Dmean for lungs, V60 , Dmean for esophagus, V40 , V60 ,
Dmean for heart, etc. are used for assessment of OAR dose sparing. The meaning of
these parameters are listed as follows:
Vx : the percentage of the structure volume receiving at least x Gy or x% of the
prescription dose.
Dx : dose delivered to at least x% of the organ volume.
Dmean{max : mean or maximum dose deposited to a certain organ.
CI: conformity index (CI) is defined as the ratio of volume enclosed by the prescription isodose line to the PTV volume, as shown in Eq. 2.5. Since in this
study all plans were normalized to 100% prescription dose covering 95% of the
PTV, 0.95 is the best value for CI, i.e. the closer to 0.95, the more conformal
of the dose distribution to the PTV.
CI
VV100%
(2.5)
PTV
HI: homogeneity index (HI) is the ratio of the difference between maximum dose
and minimum dose received by the PTV to the prescription dose. However, as
the true maximum and minimum dose is not reliable because of the frequent
occurrence of the high dose gradient in IMRT, minimum dose to 2% of the
PTV (D2% ) and maximum dose to 98% of the PTV (D98% ) are used(40), as
shown in Eq. 2.6. Since HI describes the dose distribution in the target volume,
a lower HI indicates a more homogeneous distribution.
HI
D2%D D98%
Rx
20
(2.6)
R50 : R50 or dose spillage is the ratio of the target volume receiving 50% of the
prescription dose to the target volume, as shown in Eq. 2.7. The lower the R50 ,
the smaller the region of the dose spread out.
R50
VV50%
(2.7)
PTV
D2cm : D2cm is the maximum dose in the region 2cm away from the PTV in all directions.
2.3.2
Data analysis
Dosimetric parameters as defined above were extracted from the DVH curves for all
plans. They were analyzed in two steps. First, automatically generated plans, as a
whole group, were compared with the clinical plans. The p values with two-sided
Students t-test were calculated for each dosimetric index based on the assumption
of normal distribution of the data. Then dose metrics as a function of angle spread
score were investigated. To help determine the statistically optimal value of the angle
spread score, a composite index was further constructed based on the percentage error
between these metrics in the model-based plans and clinical plans as follows
I
¸
wi Ei
(2.8)
i
where Ei is the percentage error of the ith index and wi is the related weighting.
Based on the significance of the improvement to plan quality for each index, 1, 1, 1,
1, 0.8, 0.5, 0.5, 0.8, 0.5, 0.5, 0.5, 0.8, 0.5, 0.5 were assigned to CI, R50 , HI, D2cm ,
Dmax for cord and cord +3mm, V5 , V20 , Dmean for lungs, V60 , Dmean for esophagus,
V40 , V60 , Dmean for heart, respectively.
21
3
Results
The average size of the targets for all 14 lung cases was 336.0 109.8 cc (range from
151.48 to 521.20 cc), representing a relative large and complex target. Clinically,
the average planning time for these complicated cases were hours. By means of the
knowledge-based automatic IMRT planning algorithm, the planning efficiency was
greatly improved; typically a plan was completed within one hour. To examine the
difference on beam angle selections between manual work and the BAS model, clinical
and model-generated beam angle sets were compared, as shown in the example in
Fig. 3.1. Arrows with longest length represent radiation beams (nine beams) in
the clinical plan and others with various lengths identified by different colors are
beam angle sets (eight beams) derived from the model by changing the angle spread
constraint k. Although deviation exists, beams in all plans distribute over the same
regions, i.e. anterior directions from
400 to 500 and posterior directions from 1200
to 2400 . This is consistent with the intuition that candidate beams should minimize
the penetration through the lungs and other critical organs as shown in Fig. 3.2.
In addition, a beam set with angle spread score k
0, unsurprisingly, has the
most compact distribution while the angles in other sets are more disperse. Angle
22
Figure 3.1: Comparison diagram of beam configuration for a lung case
differences among sets with k
¡ 0 are relatively small. Beam configurations for other
cases also share the same features and they can be found in Fig. A.1.
To further evaluate which beam set is preferable, the corresponding plans were
generated and effort was spent to make the plans as good as possible. Qualitative
analysis and quantitative analysis then were performed to assess the plan quality.
3.1 Qualitative results
Fig. 3.2 is an example that shows dose distributions for a clinically approved plan
and a model-based plan with angle spread score k
0.6, respectively.
In this case,
100% isodose line (red) in the model-based plan (left) is more conformal to the target
(green area) than the clinical plan in the right. The dose falloff outside of the PTV
is also faster in the model-based plan. Further, hot spots are severer in the clinical
23
Figure 3.2: Comparison diagram of dose distribution between a clinical approved
plan and a model-based plan for a lung case
plan. However, the low dose region, such as the area enclosed by 50% isodose line, is
larger in the model-based plan, which indicates a dose increase to the lungs. Since
lung volume receiving 20% of the prescription dose (V20 ) is a significant index for the
occurrence of radiation pneumonia (3), determination of a preferable plan between
these two cannot be made without further inspection of the lung DVH curves.
Fig. 3.3 shows DVHs (dash lines) of PTV, lungs, heart, spinal cord, as well as
esophagus for each automatically generated plan and the comparison with those
(solid lines) in the clinical plan. All of the model-based plans show the same features
compared to the clinical plan: a sharper dose falloff, a much lower maximum dose
received by the spinal cord, equivalent dose distributions for the heart and esophagus
but a slight dose increase in the low dose region for the lungs. Therefore, taking both
the spatial isodose distribution and the 2D DVH curves into account, the model-based
plans are more acceptable in clinic. DVHs for other cases can be found in the Fig. A.2
and Fig. A.3.
24
Figure 3.3: Comparison diagram of DVHs for PTV and OARs between a clinical
approved plan and model-based plans for a lung case
3.2 Quantitative results
The values of dosimetric parameters for each structure were extracted from the clinical and model-generated plans. Based on the data, we explore (1) the effectiveness of
this automatic lung IMRT treatment planning algorithm, and then (2) the optimal
value of angle spread score in the beam efficiency index. To achieve the first goal, all
plans are divided to two categories: clinical plans and model-based plans. The values of selected dose parameters are averaged in these two categories respectively and
compared with each other by paired two-sample Student’s t-tests. Table 3.1 shows
the mean and standard deviation of the dosimetric parameters over all plans in each
category. P-values are calculated for comparison between clinical plans and model
plans. According to the data, model-based plans have much lower values on dose
metrics for the PTV, heart and cord while larger values for other organs especially
for lungs. On average, we observed more than 15% reduction on the conformity
index and homogeneity index for PTV and V40 , V60 for heart while an 8% and 3%
25
Table 3.1: The mean of the dosimetric parameters over all plans
Structure
PTV
Dose Metrics
CI
R50
HI p%q
D2cm
Cord
Dmax (Gy)
Cord+3mm Dmax
Lungs
V5 (%)
V20 (%)
Dmean (Gy)
Esophagus V60 (%)
Dmean (Gy)
Heart
V40 (%)
V60 (%)
Dmean (Gy)
Clinic
1.57 0.36
7.15 2.16
14.35 4.25
109.32 6.42
41.20 7.44
46.47 6.34
54.03 11.35
26.79 6.24
15.39 3.29
26.2615.40
32.569.65
17.23 12.35
4.86 5.74
17.7511.08
Model
1.26 0.22
6.44 1.42
12.00 2.22
104.959.18
40.83 6.16
47.59 6.01
57.95 9.94
27.63 6.11
15.25 3.06
26.00 15.55
32.769.76
13.37 13.00
3.81 4.45
16.3810.12
p
0.001
0.011
0.001
0.001
0.718
0.234
0.016
0.372
0.767
0.910
0.895
0.047
0.176
0.398
increase on V5 , V20 for lungs, respectively. P-values for PTV metrics are all less than
0.05, indicating statistically the significance of the improvement to PTV, whereas
those for critical organs are larger than 0.05 except V5 for lungs and V40 for heart.
The corresponding box plots for each dosimetric pair are shown in Fig. 3.4. A more
intuitive expression of these differences is shown in the bar plot 3.5.
To achieve the second goal, i.e. to find the best value of angle spread score
k, we explore the variation of dose metrics as a function of k, as shown in the
example in Fig. 3.6. From the graphs we can see that some metrics such as CI,
HI, D2cm for PTV and dosimetric parameters for lung, esophagus and heart, have
clear trends, while the others behave in a more complex manner. Even for those
metrics with explicit trends, it is still hard to draw any solid conclusion once they
were considered synthetically. As the model is supposed to be useful for any case, the
integrated information over all model-based plans is utilized. Profiles for the average
of each dosimetric index over all plans are shown in Fig. 3.7. From the picture each
26
Figure 3.4: Dose metrics boxplots of clinical approved plans and model-based plans
Figure 3.5: Bar plot of the percentage error of dose metrics between the averages
of clinical and model-based plans
27
Figure 3.6: Dose metrics as a function of angle spread constraint k for one case
index shows a clearer variance as k changes. Indexes for PTV and heart seem to
monotonically decrease, those for lungs and esophagus monotonically increase and
for cord, however, a decrease-increase-decrease trend is observed. Yet, the dilemma
of determination of k still exists. To solve this issue, the composite index defined in
Eq. 2.8 is calculated and plotted in Fig. 3.8. The minimum point lies at k
which is considered as the best value for angle spread score.
28
0.6,
Figure 3.7: Average of dose metrics as a function of angle spread constraint k over
all plans
Figure 3.8: Plot of composite index as a function of angle spread constraint
29
4
Discussion
Within the framework of a knowledge-based automatic lung IMRT treatment planning algorithm, a polynomial coefficient reflecting the combined effect of multiple
beams was systematically studied and its statistically optimal value was obtained.
Although only 11 patients and 14 planning cases were involved (primary and boost
plans were regarded as different cases), they were so complex that much energy and
time was invested to obtain clinically acceptable plans. Plan regeneration using this
new algorithm is encouraging, with planning time dramatically reduced while quality
is ensured. This is attributed to the validity of beam configurations determined by
the BAS model and the efficacy of optimization objectives predicted by the OAR
DVH prediction model.
4.1 BAS model
Instead of Eq. 2.1, a modified beam efficiency index 2.3 was employed in this study.
The former emphasizes the relative energy absorption along a beamlet path to target
and normal tissues and is an average index value of Eq. 2.2. The Eq. 2.3, in contrast,
focuses on the total dose distribution in one direction. It can be considered as the
30
Figure 4.1: Plot of beam spread term as a function of angle separation with δ
taking various values
radiation beam in a particular direction contains only one beamlet with the size of
the tumor or it can also be thought of as the sum of the beamlet weighted by the
total target dose deposition in a particular direction. Hence, it is a more direct way
to select beams, i.e. those that result in the largest dose accumulation to the target
are selected.
Moreover, the introduction of the beam separation term in the EI is effective.
According to the qualitative plan evaluation and the curves of the dose metrics as
a function of angle spread score k, beam configurations derived based only on the
dosimetric information, i.e. k
0, lead to inferior plans compared to either clinical
plans or other model-based plans. However, due to the nonlinearity of this term as
shown in Fig. 4.1, it can dominate the EI in the case of the proximity of two beams
31
if δ is not properly chosen. The beam configurations for each case also show that the
minimum angle difference is about 5 degree as shown in Fig. A.1. Thus there is a
risk that the best beam set might be missed. Also, the model only adjusts the beam
spread in the transverse plane, but this can cause over-adjustment especially when
non-coplanar beams are used. In addition, the model currently does not consider
the optimization of the collimator rotation, which is supposed to reduce the leakage
dose to the patient via MLC.
Several uncertainties exist in this model as well. First, model-based plans only
use 8 beams with at most 4 non-coplanar beams. These two numbers are utilized
based on clinical experience, but they are not necessarily the best for a particular
patient. The literature (41; 42) has some discussions on this issue: Ehrgott et al.
(42) claim that IMRT plans are definitely improved by adding more beams; Bortgeld
et al. (41) conversely state that the required beam number can be derived based
on the measurement of dose variability. Another uncertainty in this model is the
values of a set of weighting factors defined in the EI. They were set empirically in
this study based on prior knowledge of the dosimetric tolerance of the critical organs.
Theoretically, the best value of angle spread score can only be achieved with the best
set of weighting factors while the weighting factors can be set optimally when the
angle spread score takes the best value. For this dilemma, empirical determination
of some parameters is a good starting point. Finally, these two sets of factors can
potentially be reduced to one set by absorbing the angle spread score in the weighting
factors.
4.2 OAR DVH prediction model
As mentioned before, the OAR DVH prediction model used to acquire dose- and
dose-volume-based objectives for beamlet optimization does not necessarily result in
absolute best dose sparing constraints. Manual adjustment, albeit modest, to these
32
constraints are needed, which is a time-consuming process. Even so, the set of dose
objectives from this model is a reliable guide and it effectively reduces the planning
time. To generate best plans for each angle set, rings in regions far from the target
and dose structures are used. They effectively reduce hot spots and ensure PTV
dose conformity, but slightly increase the dose to the low dose regions for OARs. In
view of these trade-offs, generally used are rings with 2cm width and 3cm far from
the target and dose structure converted from 95% isodose line.
Owing to the BAS model as well as the OAR DVH prediction model, plans can
be produced automatically with little or no human intervention. First, beam configurations coincide with the prior knowledge for beam selection. As shown in Fig. A.1,
beam configurations from all model-based plans and clinical plans show regional
consistency in angle distribution. Second, automatically generated optimization objectives reflect the ”best” prior available knowledge on dosimetric constraint setup.
These two models functioning together result in plans with comparable or improved
quality compared to clinical plans, especially on dose metrics for the PTV.
However, although the mean values of the dosimetric parameters shown in the
Table 3.1 show improvement on dosimetric quality for PTV, the corresponding standard deviations degrade this conclusion to some degree, showing less strong evidences
of the improvement of the plan quality. Nevertheless, the large standard deviation in
either the clinical plans or model-based plans should be mostly ascribed to the variation of the patient’s anatomy, including the PTV size, shape, location, and the lung
size as well. Moreover, the case by case comparisons of those parameters, isodose lines
(Fig. 3.2) and DVH curves (Fig. A.2 and Fig. A.3) between clinical plans and modelbased plans show the quality improvement in the model-based plans. In addition,
most of the p values for critical organs in Table 3.1 are not less than the significance
level of 0.05, indicating a limitation of the models to the dosimetric improvement to
critical organs. The DVHs shown in Fig. A.2 and Fig. A.3 also demonstrate the vari33
ation of dosimetric quality for critical organs among different cases. The geometrical
differences of OARs to PTV probably cause the variation. For example, DVHs for
the esophagus in the clinical plan for case lung-032 are much lower than the clinical
threshold which indicates a large separation between the esophagus and the target.
Therefore, the model tends to increase the dose to the organ within the threshold so
that other organs can gain improvements. However, the model behaves differently in
lung-059 where the esophagus is in proximity to the target. From this point of view,
the model can balance the dose deposition to different OARs based on the target
location. Due to the limited number of the patient cases, expanding the study with
more cases will likely improve the statistical power.
The optimal beam angle spread score is determined by minimizing the composite
index Eq. 3.8 which is a sum of the weighted percentage error of the dosimetric
indexes. The weights are set empirically based on the significance of these indexes
described in RTOG 0617, i.e. Dmax for cord, V20 for lungs and V40 for heart have a
relative higher weight except the dose metrics for PTV. However, the derived beam
angle spread score is only optimal in statistics. Depending on patient anatomy, the
optimal score may be slightly different, but Fig. A.2 and Fig. A.3 show that when
beam spread score varies from 0.4 to 0.8, plan quality does not change much.
Another limitation is the patient database used in this study. Only a limited
number of patients with similar target locations (all tumors are located in mediastinum with some extended to lungs) treated by non-coplanar beams were involved.
Nonetheless, the tumors in these cases were large and complex, causing planning
difficulties in clinic. We assume that if the algorithm can work well on these complicated cases, it will show positive efficacy on easier ones. Moreover, more tests on
clinical plans are needed to verify the effectiveness of this algorithm.
34
5
Conclusion
Knowledge-based lung IMRT planning algorithm further developed and tested in this
study demonstrates the effectiveness in producing high quality treatment plans even
for complex tumor cases with a need for non-coplanar beams. The BAS model using
a modified beam efficiency index can provide statistically the best beam configuration
based on patient-specific anatomy with angle spread score taking the optimal value
derived from this study. The application of OAR DVH prediction model for fluence
map optimization then leads to an optimal or near-optimal dose distribution for the
best beam angle set. This algorithm can potentially improve the planning efficiency
while maintain the plan quality for IMRT plans with non-coplanar angles in which
the planners usually have less experience. However, the generality and robustness of
this algorithm should be further verified by more tests.
35
Appendix A
Beam configurations and DVHs
36
37
Figure A.1: Beam configurations for all lung cases
Figure A.2: Comparison diagram of DVHs for PTV and OARs between a clinical
approved plan and model-based plans for all cases(1)
38
Figure A.3: Comparison diagram of DVHs for PTV and OARs between a clinical
approved plan and model-based plans for all cases(2)
39
Bibliography
[1] Worldwide cancer statistics,
URL:
http://www.cancerresearchuk.org/health-professional/
worldwide-cancer-statistics
[2] N. Jegadeesh, Y. Liu, T. Gillespie, F. Fernandez, S. Ramalingam, J. Mikell,
J. Lipscomb, W. J. Curran and K. A. Higgins, Evaluating Intensity Modulated
Radiation Therapy in Locally Advanced Non-Small Cell Lung Cancer: Results
from the National Cancer Data Base, Clinical Lung Cancer (2016), pp. –,
URL:
http://www.sciencedirect.com/science/article/pii/
S1525730416300043
[3] S. Chun, C. Hu, H. Choy, R. Komaki, R. Timmerman, S. Schild, J. Bogart,
M. Dobelbower, W. Bosch, J. Galvin, V. Kavadi, S. Narayan, P. Iyengar,
C. Robinson, R. Wynn, A. Raben, M. Augspurger, R. MacRae, R. Paulus
and J. Bradley, Comparison of 3-D Conformal and Intensity Modulated
Radiation Therapy Outcomes for Locally Advanced Non-Small Cell Lung
Cancer in {NRG} Oncology/RTOG 0617, International Journal of Radiation
Oncology*Biology*Physics 93 (2015) (3, Supplement), pp. S1 – S2,
URL:
http://www.sciencedirect.com/science/article/pii/
S0360301615007415, proceedings of the American Society for Radiation
Oncology 57th Annual MeetingASTRO’s 57th Annual Meeting
[4] G. A. Ezzell, J. M. Galvin, D. Low, J. R. Palta, I. Rosen, M. B. Sharpe,
P. Xia, Y. Xiao, L. Xing and C. X. Yu, Guidance document on delivery,
treatment planning, and clinical implementation of IMRT: Report of the IMRT
subcommittee of the AAPM radiation therapy committee, Medical Physics 30
(2003) (8), pp. 2089–2115,
URL: http://scitation.aip.org/content/aapm/journal/medphys/30/8/
10.1118/1.1591194
[5] Q. Wu and R. Mohan, Algorithms and functionality of an intensity modulated
radiotherapy optimization system, Medical Physics 27 (2000) (4), pp. 701–711,
URL: http://scitation.aip.org/content/aapm/journal/medphys/27/4/
10.1118/1.598932
40
[6] Y. Wu, D. Yan, M. B. Sharpe, B. Miller and J. W. Wong, Implementing multiple static field delivery for intensity modulated beams, Medical Physics 28
(2001) (11), pp. 2188–2197,
URL: http://scitation.aip.org/content/aapm/journal/medphys/28/11/
10.1118/1.1406515
[7] X. Zhang, X. Li, E. M. Quan, X. Pan and Y. Li, A methodology for automatic
intensity-modulated radiation treatment planning for lung cancer, Phys Med Biol
56 (2011) (13), pp. 3873–93,
URL: http://www.ncbi.nlm.nih.gov/pubmed/21654043
[8] H. H. Zhang, S. Gao, W. Chen, L. Shi, W. D. D’Souza and R. R. Meyer, A
surrogate-based metaheuristic global search method for beam angle selection in
radiation treatment planning, Phys Med Biol 58 (2013) (6), pp. 1933–46,
URL: http://www.ncbi.nlm.nih.gov/pubmed/23459411
[9] L. B. Marks, R. K. T. Haken and M. K. Martel, Guest Editor’s Introduction
to QUANTEC: A Users Guide, International Journal of Radiation Oncology*Biology*Physics 76 (2010) (3, Supplement), pp. S1 – S2,
URL:
http://www.sciencedirect.com/science/article/pii/
S0360301609033021, quantitative Analyses of Normal Tissue Effects in
the Clinic
[10] M. Bangert, P. Ziegenhein and U. Oelfke, Characterizing the combinatorial beam
angle selection problem, Physics in Medicine and Biology 57 (2012) (20), p. 6707,
URL: http://stacks.iop.org/0031-9155/57/i=20/a=6707
[11] J. Stein, R. Mohan, X.-H. Wang, T. Bortfeld, Q. Wu, K. Preiser, C. C. Ling
and W. Schlegel, Number and orientations of beams in intensity-modulated
radiation treatments, Medical Physics 24 (1997) (2), pp. 149–160,
URL: http://scitation.aip.org/content/aapm/journal/medphys/24/2/
10.1118/1.597923
[12] A. Pugachev, J. G. Li, A. L. Boyer, S. L. Hancock, Q.-T. Le, S. S. Donaldson
and L. Xing, Role of beam orientation optimization in intensity-modulated radiation therapy, International Journal of Radiation Oncology*Biology*Physics
50 (2001) (2), pp. 551 – 560,
URL:
http://www.sciencedirect.com/science/article/pii/
S0360301601015024
[13] Q. Hou, J. Wang, Y. Chen and J. M. Galvin, Beam orientation optimization for
IMRT by a hybrid method of the genetic algorithm and the simulated dynamics,
Medical Physics 30 (2003) (9), pp. 2360–2367,
URL: http://scitation.aip.org/content/aapm/journal/medphys/30/9/
10.1118/1.1601911
41
[14] Y. Li, J. Yao and D. Yao, Automatic beam angle selection in IMRT planning
using genetic algorithm, Physics in Medicine and Biology 49 (2004) (10), p.
1915,
URL: http://stacks.iop.org/0031-9155/49/i=10/a=007
[15] D. Nazareth, S. Brunner, M. Jones, H. Malhotra and M. Bakhtiari, Optimization of beam angles for intensity modulated radiation therapy treatment
planning using genetic algorithm on a distributed computing platform, Journal
of Medical Physics 34 (2009) (3), pp. 129–132,
URL: http://www.jmp.org.in/article.asp?issn=0971-6203;year=2009;
volume=34;issue=3;spage=129;epage=132;aulast=Nazareth;t=6
[16] S. Das, T. Cullip, G. Tracton, S. Chang, L. Marks, M. Anscher and J. Rosenman, Beam orientation selection for intensity-modulated radiation therapy
based on target equivalent uniform dose maximization, International Journal of
Radiation Oncology*Biology*Physics 55 (2003) (1), pp. 215 – 224,
URL:
http://www.sciencedirect.com/science/article/pii/
S0360301602038178
[17] S. Breedveld, P. R. M. Storchi, P. W. J. Voet and B. J. M. Heijmen, iCycle:
Integrated, multicriterial beam angle, and profile optimization for generation
of coplanar and noncoplanar IMRT plans, Medical Physics 39 (2012) (2), pp.
951–963,
URL: http://scitation.aip.org/content/aapm/journal/medphys/39/2/
10.1118/1.3676689
[18] R. A. Popple, I. A. Brezovich and J. B. Fiveash, Beam geometry selection using
sequential beam addition, Medical Physics 41 (2014) (5), 051713,
URL: http://scitation.aip.org/content/aapm/journal/medphys/41/5/
10.1118/1.4870977
[19] K. Engel and E. Tabbert, Fast Simultaneous Angle, Wedge, and Beam Intensity
Optimization in Inverse Radiotherapy Planning, Optimization and Engineering
6 (2005) (4), pp. 393–419,
URL: http://dx.doi.org/10.1007/s11081-005-2065-3
[20] X. Zhang, X. Li, E. M. Quan, X. Pan and Y. Li, A methodology for automatic intensity-modulated radiation treatment planning for lung cancer, Physics
in Medicine and Biology 56 (2011) (13), p. 3873,
URL: http://stacks.iop.org/0031-9155/56/i=13/a=009
[21] M. Bangert and J. Unkelbach, Accelerated iterative beam angle selection in
IMRT, Medical Physics 43 (2016) (3), pp. 1073–1082,
URL: http://scitation.aip.org/content/aapm/journal/medphys/43/3/
10.1118/1.4940350
42
[22] A. Pugachev and L. Xing, Pseudo beams-eyeview as applied to beam orientation
selection in intensity-modulated radiation therapy, International Journal of
Radiation Oncology*Biology*Physics 51 (2001) (5), pp. 1361 – 1370,
URL:
http://www.sciencedirect.com/science/article/pii/
S0360301601017369
[23] J. Meyer, S. M. Hummel, P. S. Cho, M. M. Austin-Seymour and M. H. Phillips,
Automatic selection of non-coplanar beam directions for three-dimensional conformal radiotherapy, The British Journal of Radiology 78 (2005) (928), pp.
316–327,
URL: http://dx.doi.org/10.1259/bjr/13015047, pMID: 15774592
[24] M. Bangert and U. Oelfke, Spherical cluster analysis for beam angle optimization
in intensity-modulated radiation therapy treatment planning, Physics in Medicine
and Biology 55 (2010) (19), p. 6023,
URL: http://stacks.iop.org/0031-9155/55/i=19/a=025
[25] P. S. Potrebko, B. M. C. McCurdy, J. B. Butler and A. S. El-Gubtan, Improving
intensity-modulated radiation therapy using the anatomic beam orientation
optimization algorithm, Medical Physics 35 (2008) (5), pp. 2170–2179,
URL: http://scitation.aip.org/content/aapm/journal/medphys/35/5/
10.1118/1.2905026
[26] L. Yuan, Q. J. Wu, F. Yin, Y. Li, Y. Sheng, C. R. Kelsey and Y. Ge, Standardized
beam bouquets for lung IMRT planning, Physics in Medicine and Biology 60
(2015) (5), p. 1831,
URL: http://stacks.iop.org/0031-9155/60/i=5/a=1831
[27] M. Bangert, P. Ziegenhein and U. Oelfke, Comparison of beam angle selection
strategies for intracranial IMRT, Medical Physics 40 (2013) (1), 011716,
URL: http://scitation.aip.org/content/aapm/journal/medphys/40/1/
10.1118/1.4771932
[28] D. Djajaputra, Q. Wu, Y. Wu and R. Mohan, Algorithm and performance of a
clinical IMRT beam-angle optimization system, Physics in Medicine and Biology
48 (2003) (19), p. 3191,
URL: http://stacks.iop.org/0031-9155/48/i=19/a=007
[29] V. Chanyavanich, S. K. Das, W. R. Lee and J. Y. Lo, Knowledge-based IMRT
treatment planning for prostate cancer, Medical Physics 38 (2011) (5), pp.
2515–2522,
URL: http://scitation.aip.org/content/aapm/journal/medphys/38/5/
10.1118/1.3574874
43
[30] B. Wu, F. Ricchetti, G. Sanguineti, M. Kazhdan, P. Simari, R. Jacques,
R. Taylor and T. McNutt, Data-Driven Approach to Generating Achievable DoseVolume Histogram Objectives in Intensity-Modulated Radiotherapy Planning,
International Journal of Radiation Oncology*Biology*Physics 79 (2011) (4),
pp. 1241 – 1247,
URL:
http://www.sciencedirect.com/science/article/pii/
S0360301610007492
[31] L. M. Appenzoller, J. M. Michalski, W. L. Thorstad, S. Mutic and K. L. Moore,
Predicting dose-volume histograms for organs-at-risk in IMRT planning, Medical Physics 39 (2012) (12), pp. 7446–7461,
URL: http://scitation.aip.org/content/aapm/journal/medphys/39/12/
10.1118/1.4761864
[32] L. Yuan, Y. Ge, W. R. Lee, F. F. Yin, J. P. Kirkpatrick and Q. J. Wu, Quantitative analysis of the factors which affect the interpatient organ-at-risk dose
sparing variation in IMRT plans, Medical Physics 39 (2012) (11), pp. 6868–
6878,
URL: http://scitation.aip.org/content/aapm/journal/medphys/39/11/
10.1118/1.4757927
[33] X. Zhang, X. Li, E. M. Quan, X. Pan and Y. Li, A methodology for automatic intensity-modulated radiation treatment planning for lung cancer, Physics
in Medicine and Biology 56 (2011) (13), p. 3873,
URL: http://stacks.iop.org/0031-9155/56/i=13/a=009
[34] E. M. Quan, J. Y. Chang, Z. Liao, T. Xia, Z. Yuan, H. Liu, X. Li, C. A. Wages,
R. Mohan and X. Zhang, Automated Volumetric Modulated Arc Therapy
Treatment Planning for Stage {III} Lung Cancer: How Does It Compare
With Intensity-Modulated Radio Therapy?, International Journal of Radiation
Oncology*Biology*Physics 84 (2012) (1), pp. e69 – e76,
URL:
http://www.sciencedirect.com/science/article/pii/
S0360301612002295
[35] P. W. Voet, M. L. Dirkx, S. Breedveld, D. Fransen, P. C. Levendag and
B. J. Heijmen, Toward Fully Automated Multicriterial Plan Generation:
A Prospective Clinical Study, International Journal of Radiation Oncology*Biology*Physics 85 (2013) (3), pp. 866 – 872,
URL:
http://www.sciencedirect.com/science/article/pii/
S0360301612005627
[36] G. Amit, T. G. Purdie, A. Levinshtein, A. J. Hope, P. Lindsay, A. Marshall,
D. A. Jaffray and V. Pekar, Automatic learning-based beam angle selection for
thoracic IMRT, Medical Physics 42 (2015) (4), pp. 1992–2005,
44
URL: http://scitation.aip.org/content/aapm/journal/medphys/42/4/
10.1118/1.4908000
[37] A. Fogliata, P.-M. Wang, F. Belosi, A. Clivio, G. Nicolini, E. Vanetti and
L. Cozzi, Assessment of a model based optimization engine for volumetric modulated arc therapy for patients with advanced hepatocellular cancer, Radiation
Oncology 9 (2014) (1), pp. 1–13,
URL: http://dx.doi.org/10.1186/s13014-014-0236-0
[38] A. Fogliata, F. Belosi, A. Clivio, P. Navarria, G. Nicolini, M. Scorsetti,
E. Vanetti and L. Cozzi, On the pre-clinical validation of a commercial
model-based optimisation engine: Application to volumetric modulated arc
therapy for patients with lung or prostate cancer, Radiotherapy and Oncology
113 (2014) (3), pp. 385 – 391,
URL:
http://www.sciencedirect.com/science/article/pii/
S0167814014004873
[39] J. Li, J. Galvin, A. Harrison, R. Timmerman, Y. Yu and Y. Xiao, Dosimetric
Verification Using Monte Carlo Calculations for Tissue Heterogeneity-Corrected
Conformal Treatment Plans Following {RTOG} 0813 Dosimetric Criteria for
Lung Cancer Stereotactic Body Radiotherapy, International Journal of Radiation Oncology*Biology*Physics 84 (2012) (2), pp. 508 – 513,
URL:
http://www.sciencedirect.com/science/article/pii/
S0360301611036601
[40] Q. Wu, R. Mohan, M. Morris, A. Lauve and R. Schmidt-Ullrich, Simultaneous
integrated boost intensity-modulated radiotherapy for locally advanced head-andneck squamous cell carcinomas. I: dosimetric results, International Journal of
Radiation Oncology*Biology*Physics 56 (2003) (2), pp. 573 – 585,
URL:
http://www.sciencedirect.com/science/article/pii/
S0360301602046175
[41] T. Bortfeld, The number of beams in IMRTtheoretical investigations and implications for single-arc IMRT, Physics in Medicine and Biology 55 (2010) (1),
p. 83,
URL: http://stacks.iop.org/0031-9155/55/i=1/a=006
[42] M. Ehrgott, A. Holder and J. Reese, Beam selection in radiotherapy design,
Linear Algebra and its Applications 428 (2008) (56), pp. 1272 – 1312,
URL:
http://www.sciencedirect.com/science/article/pii/
S0024379507002480, special Issue on Linear and Nonlinear Models and
Algorithms in Intensity-Modulated Radiation Therapy
45