Computers & Operations Research 39 (2012) 17791789

Contents lists available at SciVerse ScienceDirect

Computers & Operations Research

journal homepage: www.elsevier.com/locate/caor

An efcient approach to incorporating interfraction motion in IMRT

treatment planning$
Chunhua Men a, H. Edwin Romeijn b,n, Anneyuko Saito c, James F. Dempsey d
a
Elekta Inc., 13723 Riverport Drive, Suite 100, Maryland Heights, MO 63043, United States
b
Department of Industrial and Operations Engineering, University of Michigan, Ann Arbor, MI 48109-2117, United States
c
Department of Radiology, Juntendo University Uryasu Hospital, Uryasu, Japan
d
ViewRay Inc., 2 Thermo Fisher Way, Village of Oakwood, OH 44146, United States

a r t i c l e i n f o abstract

Available online 31 October 2011 Intensity modulated radiation therapy (IMRT) is one of the most widely used delivery modalities for
Keywords: radiation therapy for cancer patients. A patient is typically treated in daily fractions over a period of 59
Interfraction motion weeks. In this paper, we consider the problem of accounting for changes in patient setup location and
IMRT treatment planning internal geometry between the treatment fractions, usually referred to as interfraction motion. The
Stochastic models conventional method is to add a margin around the clinical tumor volume (CTV) to obtain a planning
target volume (PTV). A uence map optimization (FMO) model is then solved to determine the optimal
intensity proles to deliver to the patient. However, a margin-based method may not adequately model
the changes in dose distributions due to the random nature of organ motion. Accounting for
interfraction motion in the FMO model essentially transforms the deterministic optimization problem
into a stochastic one. We propose a stochastic FMO model that employs convex penalty functions to
control the treatment plan quality and uses a large number of scenarios to characterize interfraction
motion uncertainties. Some effects of radiotherapy are impacted mainly by the dose distribution in a
given treatment fraction while others tend to manifest themselves over time and depend mostly on the
total dose received over the course of treatment. We will therefore formulate an optimization model
that explicitly incorporates treatment plan evaluation criteria that apply to the total dose received over
all treatment fractions and ones that apply to the dose per fraction. Particularly when many structures
fall into the former category, this can lead to signicant reductions in the dimension of the optimization
model and therefore the time required to solve it. We test an example of our model on ve clinical
prostate cancer cases, showing the efcacy of our approach. In particular, compared to a traditional
margin-based treatment plan, our plans exhibit signicantly improved target dose coverage and
clinically equivalent critical structure sparing at only a modest increase in computational effort.
& 2011 Elsevier Ltd. All rights reserved.

1. Introduction delivered as a series of daily treatment fractions over a period of 59

In this paper, we consider the problem of determining high-
quality intensity modulated radiation therapy (IMRT) treatment
plans for cancer patients. Since radiation therapy kills both cancer-
ous and normal cells, the treatment must be carefully planned so
that a clinically prescribed dose is delivered to cancerous cells while
sparing normal cells in nearby organs and tissues to the greatest
extent possible. Radiation therapy treatment plans are generally
$
This work was supported by the National Science Foundation under Grant no.
DMI-0457394/CMMI-0852727.
n
Corresponding author. Tel.: 1 734 7632238; fax: 1 734 7643451.
E-mail addresses: chunhua.men@elekta.com (C. Men),
romeijn@umich.edu (H.E. Romeijn), anyusaike@yahoo.co.jp (A. Saito),
jfdempsey@viewray.com (J.F. Dempsey).
weeks. For each treatment fraction, patients are treated using a
gantry-mounted radiation source that can deliver radiation from
several different directions. Moreover, the beam can be shaped
dynamically using a so-called multi-leaf collimator (MLC) system
that blocks part of the beam. Together, these tools allow for the
delivery of highly complex dose distributions that conform to the
clinical targets while adequately sparing the functionality of critical
structures. A uence map optimization (FMO) problem is formu-
lated and solved to identify a high-quality collection of intensity
proles, one for each beam direction. Such problems have been
extensively studied in the literature; we refer in particular to the
review papers by Shepard et al. [31] and Romeijn and Dempsey
[30]. More specically, Lee et al. [1618] studied mixed integer
programming approaches; Hamacher et al. [11] and Kufer et al. [14]
proposed a multi-criteria approach to the problem; and Romeijn
et al. [28,29] developed convex programming models.

0305-0548/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cor.2011.10.020
1780 C. Men et al. / Computers & Operations Research 39 (2012) 17791789
Traditionally, the design and evaluation of IMRT treatment
plans (i.e., the FMO model) is based on a static patient model
obtained using a single planning CT image set. During the course of
a fractionated radiation therapy the patient geometry may deviate
from the geometry observed in the image on which a treatment
plan is based. These deviations are caused by, for example, (i) an
inability to setup the patient in an identical location with respect to
the radiation source on each day of the treatment, (ii) movement
and shape changes of tumor as well as soft tissues and organs, and
(iii) patient weight loss or gain. Such changes in patient geometry
between fractions are generally referred to as interfraction motion.
(This distinguishes these sources of uncertainty to those caused by
intrafraction motion, which refers to internal organ motion occur-
ring during the actual radiation treatment due to breathing,
swallowing, etc.; see, e.g., [15].)
To account for organ motion, the conventional approach is to
add a margin around the clinical tumor volume (CTV) to obtain a
planning target volume (PTV) (see, e.g., [12,13]). The CTV contains
an area including and immediately surrounding the gross tumor
volume (GTV) that is at high risk for clinical extension of the
tumor. The PTV is an expansion of the CTV with a margin
accounting for both interfraction and intrafraction motion.
Several studies (see, e.g., [2,33,34,24]) have empirically derived
expressions for the required margin size. However, these
margin-based methods suffer from the fact that the margin may
not adequately model the changes in dose distribution due to the
uncertain deviations caused by interfraction and intrafraction
organ motion; in particular, it may not be necessary or appro-
priate to use a uniform margin for the entire CTV. Therefore, the
use of margins for treatment planning may lead to underdosing of
the CTV and/or overdosing of critical structures, causing reduced
cure rates or increased side-effects of treatment. Moreover, the
aggregate dose distribution received by those targets and critical
structures for which dose per fraction is important does not
accurately capture treatment plan quality. Although the cumula-
tive dose over the course of treatment is evaluated, the actual
dose received in each of the treatment fractions may signicantly
differ from the planned (or clinically prescribed) one.
Because of these problems associated with margin-based
methods, attention has shifted to the development of alternative
methods. It is expected that explicitly accounting for the uncer-
tainties of interfraction motion will lead to improved treatment
plans by implicitly identifying patient-dependent and non-
uniform margins to optimally hedge against geometry changes.
Hence, the conicting goals of tumor dose coverage and critical
structure sparing can be balanced more effectively. One approach
is to post process the static planned dose distribution by convol-
ving it with a probability density function which describes the
motion uncertainties (see, e.g., [19,20]). This dose-convolution
method assumes shift invariance of the dose distribution.
However, internal inhomogeneities and surface curvature may
lead to violations of this assumption (see [7,8]). In order to
overcome these drawbacks, Chu et al. [6] propose a robust
optimization model. However, their model not only views the
dose to the voxels as independent random variables, whereas in
practice there will be a strong dependence between these, but it
also focuses on the aggregate delivered dose over the entire
course of treatment. Alternatively, McShan et al. [21] proposed
the multiple instance geometry approximation (MIGA) technique.
In this approach, a relatively small number of scenarios (typically
about seven) for the patient geometry are selected and a full dose
calculation is performed for each of these scenarios. A treatment
plan is then determined based on the average or expected dose
distribution delivered to the patient in these scenarios. This
approach, although very promising, suffers from two potentially
serious drawbacks. Firstly, the number of scenarios that is used in
the MIGA approach may not be large enough to capture all
interfraction motion uncertainties, and signicantly increasing
the number of scenarios may make the optimization model
intractable. Secondly, while properties of a treatment plan that
depend on aggregate dose distribution are accurately assessed by
considering the expected or average dose delivered per fraction,
other properties that rely more heavily on the dose distribution
delivered in a single treatment fraction are ignored by the MIGA
approach. As McShan et al. [21] point out: Clearly, to believe that
the MIGA optimization is going to be clinically useful in this case,
the cost function for the optimization should be modied to
either (1) force dose/fraction constraints as well as the total dose
costs, or (2) specically include dose/fraction corrections into the
optimization through the use of linear/quadratic model estimates
of a bioeffect. In this paper, we will take the former approach.
In this paper, we will therefore introduce a new stochastic
model for FMO that accurately captures interfraction motion
uncertainties while remaining computationally tractable. Our
goals in this paper are to
 propose a new stochastic FMO model accounting for interfrac-
tion motion;
 empirically evaluate the ability of this model to nd high-
quality treatment plans when the model is used to account for
setup errors in particular;
 develop a new and appropriate way to evaluate treatment plan
evaluation criteria that depend on the dose distribution in a
single treatment fraction.
2. Incorporating interfraction motion
2.1. Deterministic FMO model
Our stochastic FMO model is based on a traditional determi-
nistic FMO model that we will rst introduce. We denote the set
of beamlets (which discretize a pre-determined and xed set of
beams from different directions) by N and the associated decision
variables representing the intensity of these beamlets by xi (iA N).
Furthermore, we denote the set of voxels by V, and associate a
decision variable zj with each voxel j A V that measures the dose
received by that voxel. We make the standard approximation that
expresses the vector of voxel doses as a linear function of the
intensities of the apertures through the so-called dose deposition
coefcients Dij : the dose received by voxel j A V in a given fraction
from beamlet iA N at unit intensity. Finally, we assume that the
clinician or physician has identied a collection of treatment plan
evaluation criteria given by the convex functions G : R9V9 -R
( A L) as well as associated criteria weights w ( A L). Then our
deterministic FMO model can be written as
X
minimize w G z
AL
subject to
X
zj Dij xi for j A V
iAN
xi Z 0 for i A N
(see, e.g., [28]). Of course an FMO model may also contain
constraints on certain treatment plan evaluation criteria. Our
approach can be extended to such models, but for ease of
exposition we will not consider this explicitly in this paper.
 C. Men et al. / Computers & Operations Research 39 (2012) 17791789 1781

2.2. Stochastic FMO model subject to

Integrating interfraction motion uncertainties into the FMO
model essentially transforms this deterministic optimization
problem into a stochastic one since the location of the patient is
no longer known with certainty. We use the convention that the
set of voxels discretizing the patient is constant relative to the
patient geometry, so that these move when the patient geometry
changes. This means that the values of Dij, the dose received by
each voxel at unit intensity, is now a random variable that we
denote by Dij . Since the current optimization models are already
challenging large-scale optimization problems, models that expli-
citly account for motion may quickly become intractable. Since
the number of fractions is generally relatively large, we could
consider simply replacing the coefcients Dij in the static model
by their expected values EDij . Implicitly, this assumes that the
relevant quantities are the total dose received by all voxels over
the course of the treatment. This assumption may not always be
appropriate; for example, for targets it is often considered
essential that the dose delivered to each voxel is close to constant
in each treatment fraction, typically 1.82.1 Gy. This means that
the amount of dose received in each individual fraction will be
important as well. We therefore decompose the collection of
treatment plan evaluation criteria into a set of criteria that
measure dose/fraction effects (indexed by Lfx ) and a set of criteria
that measure total dose effects (indexed by Ltotal L\Lfx ). The
former criteria are then evaluated at the actual dose in a fraction,
while the latter are evaluated at the expected dose in a fraction.
Denoting the dose received by voxel j A V in a given fraction by zj
(which is now a random variable), we then formulate our
stochastic FMO model as follows:
X X
minimize EG z G Ez
A Lfx A Ltotal
subject to
X scenarios. Since, for large numbers of scenarios, such clouds are
zj Dij xi for j A V
iAN
xi Z0 for i A N
In order to make the above stochastic model tractable, we
represent the random dose distribution coefcient matrices via
a nite number of scenarios, say S. We then let Dsij denote the dose
received by voxel j from beamlet i at unit intensity in scenario
s1,y,S, zsj denote the dose received by voxel j A V in scenario
s1,y,S, and Ezj  denote the expected dose received by voxel
j A V in any given fraction. This leads to the following so-called
deterministic equivalent model:
" #
X 1X S X
minimize G zs G Ez
fx
Ss1 total
AL AL
X s
zsj Dij xi for j A V
iAN
X
Ezj  EDij xi for j A V
iAN
xi Z 0 for i A N
It is easy to see that the number of scenarios S determines the
dimensionality of this problem. Moreover, since the deterministic
FMO problem is already a large-scale one, this model may become
intractable quickly. However, since dose/fraction effects generally
apply to only a relatively small set of structures, we only need to
incorporate the decision variables zsj for the voxels in such
structures, i.e., the ones that are subject to a dose/fraction-based
evaluation criterion. Therefore, this model will in fact generally be
much smaller than it appears. Note that the MIGA approach
proposed by [21] is a special case of our approach with (i) Lfx |,
P
(ii) S 7, and (iii) EDij  1=S Ss 1 Dsij .
2.3. Treatment plan evaluation
To evaluate treatment plans obtained by the stochastic opti-
mization model, one of the tools that we will use is the traditional
and clinically widely used dose-volume histograms (DVHs). These
histograms specify, for each target and critical structure, the
fraction of that structure that receives a certain dose level or
higher. In particular, for total dose effects, we consider the DVH
corresponding to the expected aggregate dose distribution deliv-
ered to the patient over all treatment fractions. (Our experiments
will validate that the number of fractions is typically large enough
to justify this.) However, to evaluate dose/fraction effects for
relevant structures we should evaluate the dose distribution
delivered in each fraction. One tool that is often used is a
so-called DVH-cloud, a collection of DVHs for individual
often hard to interpret due to the vast amount of information
displayed, we also propose a new evaluation technique. In parti-
cular, we propose to parameterize a corresponding criterion and
evaluate the probability that, in a given scenario, the parameter-
ized version of the criterion is satised. For example, in order to
evaluate target coverage, we may plot the probability that b% of
the target volume receives the prescription dose as a function of b.
3. Experiments
3.1. Clinical problem instances
To test our stochastic optimization model, three-dimensional
treatment planning data for ve prostate cancer patients was
exported from a commercial patient imaging and anatomy
segmentation system at the University of Florida Department of

Table 1
Model dimensions (full resolution; between parentheses: after downsampling).

Case # voxels # bixels

Total Target Rectum Bladder Femoral heads Normal tissue

1 205,911 (35,988) 1806 3132 1614 4639 194,720 (24,797) 2055

2 131,216 (22,330) 949 361 504 4545 124,857 (15,971) 2253
3 226,324 (38,731) 2888 3670 1364 4851 213,551 (25,958) 2653
4 162,628 (29,828) 1700 1429 4947 3011 151,551 (18,741) 2413
5 262,184 (40,636) 1502 1035 6205 6900 246,542 (24,994) 2736
1782 C. Men et al. / Computers & Operations Research 39 (2012) 17791789

Radiation Oncology. This data was then imported into the
University of Florida Optimized Radiation Therapy (UFORT) treat-
ment planning system and used to generate the data required by
the models described above (see, e.g., [28]).
For all ve cases, we design plans using nine equispaced
60
Co-beams (note that this does not affect the optimization
algorithm, which is, without modication, applicable to high-
energy x-ray beams as well; see, for example, [28] for results with
6 MV photon beams). The nine beams are evenly distributed
around the patient with angles 01, 401, 801, 1201, 1601, 2001,
2401, 2801, 3201, respectively. The nominal size of each beam is
40  40 cm2. The beams are discretized into bixels of size
1  1 cm2, yielding on the order of 1600 bixels. The coefcients
Dij are then obtained using a pencil beam model with a correction
for nonhomogeneous tissue densities. We then reduce the set of
beamlets that actually need to be considered in the optimization
by using the fact that the actual volume to be treated is usually
signicantly smaller. That is, for each beam, we identify a mask
consisting of only those bixels that can help treat the targets, i.e.,
Table 2
Criteria for prostate cancer.
Rectum Bladder
r 15% receives Z 75 Gy r 15% receives Z 80 Gy
we identify the bixels for which the dose deposition coefcient Dij
associated with at least one target voxel is nonzero in any of the
scenarios. For all cases we generated a voxel grid with voxels of
size 4  4  4 mm3 for the targets and critical structures. Dempsey
et al. [9] performed a Fourier analysis of required voxel resolution
for the case of 6 MV photon beams. Our resolution in this paper
was based on a similar (but unpublished) analysis of 60Co-beams.
Moreover, this resolution is a typical choice for prostate cancer
cases in clinical settings due to (i) the typical tumor size, and (ii)
the relatively regular shape of the prostate (for other sites a
different resolution may be required). For the optimization model,
we downsampled the voxels in normal tissue to a voxel grid of
8  8  8 mm3. However, the treatment quality of the obtained
plans was always evaluated on the full resolution voxel grid. There
is one target (prostate), with a prescription dose 73.8 Gy delivered
in 41 daily fractions. Moreover, there are three critical structures:
bladder, rectum, and femoral heads, in addition to normal tissue.
Table 1 shows the problem dimensions for the ve cases.
Both our deterministic and stochastic FMO models employ
treatment plan evaluation criteria that are quadratic one-sided
voxel-based penalty functions. Denoting the collection of targets
by T and the collection of critical structures by C, these penalty
functions are
Femoral heads 1 X 2
Gt z a
t  maxf0,T 
t zj g t AT 1
9V t 9 j A V
r 10% receives Z 52 Gy t

r 25% receives Z 70 Gy r 25% receives Z 75 Gy

r 35% receives Z 65 Gy r 35% receives Z 70 Gy 1 X
r 50% r 50%
Gt z at  maxf0,zj T t g2 t AT [ C 2
receives Z 60 Gy receives Z 65 Gy 9V t 9 j A V
t

100 100

80 80
% Scenarios

% Scenarios

60 60

40 40

20 20
Scenarios 1-25 Scenarios 1-50
Scenarios 101-200 Scenarios 101-200
0 0
70 75 80 85 90 95 100 70 75 80 85 90 95 100
% Volume % Volume

100 100

80 80
% Scenarios

% Scenarios

60 60

40 40

20 20
Scenarios 1-75 Scenarios 1-100
Scenarios 101-200 Scenarios 101-200
0 0
70 75 80 85 90 95 100 70 75 80 85 90 95 100
% Volume % Volume

Fig. 1. The probability that a specied percentage of the target received the prescription dose for case 1: (a) 25, (b) 50, (c) 75, and (d) 100 scenarios.
 C. Men et al. / Computers & Operations Research 39 (2012) 17791789 1783

where T 
t and T t are underdosing and overdosing thresholds and voxels is associated with a dose/fraction effects (i.e., lack of
at and at are criteria weights). For illustration purposes, our adequate tumor cell kill), while overdosing of voxels is associated
experiments deal with an example where underdosing of target with total dose effects (side-effects of treatment). That is, we
included overdosing-related penalties in the set Ltotal and under-
Table 3 dosing-related criteria in the set Lfx , i.e., Lfx T and Ltotal T [ C.
CPU running time (seconds). Since our model explicitly accounts for interfraction motion,
our target is the CTV with margins added for intrafraction motion;
Case Traditional model Stochastic model (100 scenarios)
we will denote the corresponding expanded CTV by CTV . In
1 45 271 order to compare the results of our stochastic model with a more
2 42 149 traditional margin-based approach, we also create a PTV by
3 71 375 expanding the CTV with a setup margin. Empirically, the setup
4 67 246
5 89 251
errors in each direction have been found to be approximately
normally distributed with a standard deviation of 3 mm (see, e.g.,

Case 1 Case 2
100 100

80 80
% Scenarios

% Scenarios
60 60

40 40

20 Traditional Model 20 Traditional Model

Stochastic Model Stochastic Model

0 0
70 75 80 85 90 95 100 70 75 80 85 90 95 100
% Volume % Volume

Case 3 Case 4
100 100

80 80
% Scenarios

% Scenarios

60 60

40 40

20 Traditional Model 20 Traditional Model

Stochastic Model Stochastic Model

0 0
70 75 80 85 90 95 100 70 75 80 85 90 95 100
% Volume % Volume

Case 5
100

80
% Scenarios

60

40

20 Traditional Model

Stochastic Model
0
70 75 80 85 90 95 100
% Volume

Fig. 2. The probability that a specied percentage of the target received the prescription dose: (a) case 1, (b) case 2, (c) case 3, (d) case 4, and (d) case 5.
1784 C. Men et al. / Computers & Operations Research 39 (2012) 17791789

[10,4,25,35]). Following Antolak and Rosen [2], we therefore set
the width of the setup error margin to about 1.65 times the
standard deviation and formed the PTV by adding a margin of
5 mm in each direction to the CTV . Clinical DVH criteria and
constraints were established to keep toxicity at acceptable levels,
and our optimized treatment plans were evaluated with respect
to these criteria. In particular, we used the RTOG [26] criteria for
rectum and bladder (see Table 2) which are commonly used by
physicians and researchers [27,5,23]. While femoral head criteria
Table 4
Target hotspot (% of volume exceeding 110% of the prescribed dose).
Case Traditional model
1 1.65
2 3.63
3 6.41
4 1.97
5 3.00
are not well-established, we used the most strict criteria that we
found in the literature [3]. For the traditional deterministic FMO
model, we required that (i) at least 95% of the PTV receives the
prescribed dose of 73.8 Gy, (ii) at least 99% of the PTV receives
93% of the prescribed dose, and (iii) no more than 10% of the PTV
receives 110% of the prescribed dose. Formalizing these criteria
yields a non-convex optimization model. Moreover, quantifying
the trade-off between these criteria is highly subjective. We
therefore chose to use the convex penalty-function based model
as described above without explicit incorporation of the DVH
criteria. We tuned the problem parameters of this model by
manual adjustment, taking the clinical DVH criteria into account
as well as the physicians clinical subjective trade-offs. This
Stochastic model
process was applied to a single patient case, and these parameters
2.07 were then used for all other cases as well. The performance of the
4.75 obtained treatment plans with respect to the clinical DVH criteria
4.70 is taken as evidence of the effectiveness of this approach.
1.46
2.17
We next generated 200 scenarios by perturbing the location of
the patient according to a normal distribution with a mean of

Case 1 (traditional)
100

80

60
% Volume

40

20

0
0 10 20 30 40 50 60 70 80
Dose (Gy)

Case 1 (stochastic)
100

90

80

70

60
% Volume

50

40

30

20

10

0
0 10 20 30 40 50 60 70 80
Dose (Gy)

Fig. 3. DVH clouds for the CTV for case 1: (a) traditional model and (b) stochastic model.
 C. Men et al. / Computers & Operations Research 39 (2012) 17791789 1785

0 and a standard deviation of 3 mm in each coordinate direction 3.2. Results

for each patient case. Of these scenarios, we used the rst 100 in
our optimization model (i.e., we set S 100), while we used the
remaining to validate our results. We performed an extensive
comparison of the solutions obtained with our stochastic FMO
model with the ones obtained using the traditional deterministic
FMO model.
We solved all optimization models using our in-house primal
dual interior point algorithm (see [1]). All experiments were
performed on a PC with 2.66 GHz Intel Quad CPU and 4 GB of
RAM, running under Windows Vista. Our algorithms were imple-
mented in Matlab 7.

Case 1 Case 2
(Traditional: dot; Stochastic: solid) (Traditional: dot; Stochastic: solid)
100 100
Femoral Head Femoral Head

Rectum Rectum
80 80
Normal tissue Normal tissue
% Volume

% Volume
Bladder Bladder
60 60

40 40

20 20

0 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
Dose (Gy) Dose (Gy)
Case 3 Case 4
(Traditional: dot; Stochastic: solid) (Traditional: dot; Stochastic: solid)
100 100
Femoral head Femoral head

Rectum Rectum
80 80
Normal tissue Normal tissue
% Volume

% Volume

Bladder
60 60 Bladder

40 40

20 20

0 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
Dose (Gy) Dose (Gy)
Case 5
(Traditional: dot; Stochastic: solid)
100
Femoral head

Rectum
80
Normal tissue
% Volume

Bladder
60

40

20

0
0 10 20 30 40 50 60 70 80
Dose (Gy)

Fig. 4. DVHs for critical structures (stochastic vs. traditional model: (a) case 1, (b) case 2, (c) case 3, (d) case 4, and (d) case 5).
1786 C. Men et al. / Computers & Operations Research 39 (2012) 17791789

3.2.1. Number of scenarios 3.2.3. Critical structure sparing

We start by testing the effects of varying the number of
scenarios that is included in the optimization models. A small
number of scenarios may be insufcient to capture the actual
uncertainties and may therefore result in unacceptable treatment
quality while a large number of scenarios may lead to an
unacceptably long CPU running time. By testing different number
of scenarios included in the models we can nd a balance
between computational effort and the treatment plan quality. In
particular, we have solved the stochastic optimization model with
25, 50, 75, and 100 scenarios, respectively. We then evaluated the
treatment plan quality based on the scenarios used in the
optimization model as well as on scenarios 101200, i.e., scenar-
ios outside the optimization model. Taking case 1 as an example,
Fig. 1 estimates the probability that b% of the target volume
received the prescription dose for the traditional and stochastic
models, respectively, as a function of b. We observe that the
actual treatment plan quality as measured by target coverage
based on scenarios 101200 is inferior to the quality that is
perceived by the models with 25, 50, or 75 scenarios. For
example, if we include 25 scenarios in the stochastic optimization
model, at least 97.5% of the target volume receives the prescrip-
tion dose with probability 97%. However, based on scenarios
101200 the actual probability is more accurately estimated by
79%. If we include 100 scenarios in the stochastic optimization
model the actual and perceived distribution of target coverage are
virtually indistinguishable from one another. A similar pattern
was observed for the other four cases, which led us to conclude
that 100 is an appropriate number of scenarios to use in the
stochastic optimization model.
Using 100 scenarios, the amount of time required, on average,
by our optimization algorithms to nd the solutions ranges from
about 2 to 6 min of CPU time for the stochastic optimization
model while it takes less than 2 min to solve the traditional
model. Table 3 records the CPU time for the two models.
Fig. 4 shows the DVHs for the critical structures for the
stochastic model as compared to the traditional model.
Tables 5 and 6 list the corresponding DVH values based on the
criteria, evaluated on scenarios 101200. We conclude from Fig. 4
and Tables 5 and 6 that our methods could obtain better target
dose coverage without a clinically signicant additional dose to
the critical structures.
3.2.4. Individual scenario DVHs
Our stochastic optimization model takes individual dose
distributions for the target into account and robust treatment
plans can be obtained while the traditional method does not have
this merit: in some extreme scenarios, the treatment plan quality
may be very bad using the traditional method. We would like to
take case 1 as an example, and check the target dose coverage and
critical structure sparing for Scenario 135. Note that this scenario
is not included in the optimization model. Fig. 5 shows the DVHs
for the stochastic model compared to the traditional model. We
can see that about 98% of the target receives the prescription dose
(73.8 Gy) by solving the stochastic model, while for the traditional
method that volume is only 92%. Critical structure sparing is
comparable for both sets of DVHs, and satisfy the guidelines
mentioned in Section 3.1.
3.2.5. Perceived vs. actual DVHs
Recall that the traditional method does not account for inter-
fraction motion in either the optimization model or its reporting
of the treatment plan quality. Hence, the perceived treatment
plan quality (i.e., the DVHs reported by the traditional model)
may differ from the actual one (i.e., the DVHs reported by the
Table 5
DVH criteria for rectum (%).

Case Model Rectum

3.2.2. Target coverage, hot spots, and cold spots @75 Gy @70 Gy @65 Gy @60 Gy
Fig. 2 shows the target coverage curves for all ve cases for both (r 15%) (r 25%) (r 35%) (r 50%)
the treatment plan obtained with the traditional model and the one
1 Traditional 8.37 11.91 14.78 17.40
obtained with the stochastic model using scenarios 1100, in both
Stochastic 9.29 12.35 15.16 16.92
cases evaluated on scenarios 101200. It is clear from these gures 2 Traditional 3.99 6.22 8.45 11.09
that, in general, the stochastic model obtains a much better target Stochastic 10.65 13.43 16.64 20.08
dose coverage than the traditional model does. Taking case 1 as an 3 Traditional 2.70 14.13 17.17 22.85
example, the probability of 95% target coverage is 99% for the Stochastic 2.77 9.69 16.06 20.49
4 Traditional 13.79 20.15 25.82 30.86
stochastic model as opposed to 91% with the traditional model. Stochastic 10.56 17.98 22.95 27.40
Besides estimating the probability that a specied percentage 5 Traditional 11.20 19.52 25.41 29.57
of the target receive the prescription dose, we would like to check Stochastic 8.21 15.46 20.29 25.60
another two important criteria: target coldspots (underdosing)
and hotspots (overdosing) which refer to the volume of target
which receives less than 93% and more than 110% of prescription
Table 6
dose, respectively. Table 4 shows the averaged hotspots for all ve DVH criteria for bladder and femoral heads (%).
cases. In most of the scenarios, 100% of the target receives at least
93% of the prescription dose for all ve cases, so there are no Case Model Bladder Femoral
unacceptable coldspots. heads
@80 Gy @75 Gy @70 Gy @65 Gy @52 Gy
A more traditional way of evaluating the dose distribution for ( r 15%) (r 25%) ( r 35%) (r 50%) ( r 10%)
structures for which dose/fraction effects are important is to use
so-called DVH clouds, i.e., a DVH for each potential fraction 1 Traditional 0.57 6.09 9.51 12.30 0.06
scenario. In our case, the optimization model only concerns itself Stochastic 0.68 10.47 15.06 17.97 1.21
2 Traditional 0.82 6.88 8.45 10.09 0
with individual fractions when evaluating the dose distribution in
Stochastic 7.7 13.12 16.57 20.16 5.44
the target, so we limit ourselves to DVH clouds for the CTV. Fig. 3 3 Traditional 0.79 11.36 16.82 24.95 0.20
conrms that the stochastic model achieves more consistent Stochastic 0 2.58 9.52 14.48 0.59
target coverage than the traditional model. It is interesting to 4 Traditional 0.42 4.65 6.87 8.71 0
note that this gain in target coverage does not seem to come at Stochastic 0.54 3.21 5.60 7.13 0
5 Traditional 0.29 4.45 6.93 8.99 0
the expense of additional hotspots: the upper tails of both DVH Stochastic 0.14 2.63 4.90 6.96 0.59
clouds appear very similar.
 C. Men et al. / Computers & Operations Research 39 (2012) 17791789 1787

Case 1, scenario 135 (traditional: dot; stochastic: solid) stochastic model, evaluated using the scenarios). To assess this
100 effect, we compared the perceived DVHs reported from the
traditional model with the actual DVHs that take into account
CTV+ the uncertainty using the 200 scenarios from the interfraction
80
Femoral head motion model. The results indicate that the perceived DVHs for
critical structures are very close to the actual ones (see examples
Rectum
% Volume

60 in Fig. 6). However, the perceived DVHs overestimate the dose

Normal tissue distribution to the targets (see results in Fig. 7). This further
underscores the importance of taking interfraction motion uncer-
40 Bladder
tainty into account explicitly.

20
3.2.6. Expected dose vs. aggregate dose over all fractions
Finally, note that an essential assumption that we make in our
0 model is that the aggregate dose delivered over 41 fractions
0 10 20 30 40 50 60 70 80
Dose (Gy)
(which is uncertain) can accurately be approximated by the
expected dose delivered over these 41 fractions. In order to
Fig. 5. DVHs for case 1 in Scenario 135 (stochastic vs. traditional model). validate this assumption, we simulated the optimized treatment

Case 1 (perceived: dot; actual: solid)

100

Femoral head

Rectum
80

Normal tissue

60 Bladder
% Volume

40

20

0
0 10 20 30 40 50 60 70 80
Dose (Gy)

Case 2 (perceived: dot; actual: solid)

100

Femoral head

Rectum
80

Normal tissue

60 Bladder
% Volume

40

20

0
0 10 20 30 40 50 60 70 80
Dose (Gy)

Fig. 6. Perceived and actual DVHs for critical structures for (a) case 1 and (b) case 2.
1788 C. Men et al. / Computers & Operations Research 39 (2012) 17791789

Case 1 Case 2
100 100
Actual Perceived
Perceived Actual
80 80

60 60
% Volume

% Volume
40 40

20 20

0 0
65 67 69 71 73 75 77 79 81 83 85 65 67 69 71 73 75 77 79 81 83 85
-20 -20
Dose (Gy) Dose (Gy)
Case 3 Case 4
100 100
Perceived Perceived
Actual Actual
80 80

% Volume
% Volume

60 60

40 40

20 20

0 0
65 67 69 71 73 75 77 79 81 83 85 65 67 69 71 73 75 77 79 81 83 85
-20 -20
Dose (Gy) Dose (Gy)
Case 5
100
Perceived
Actual
80

60
% Volume

40

20

0
65 67 69 71 73 75 77 79 81 83 85
-20
Dose (Gy)

Fig. 7. Perceived and actual DVHs for the target: (a) case 1, (b) case 2, (c) case 3, (d) case 4, and (d) case 5.

plans over a collection of ve sample random treatments of 41
fractions each, and plotted the corresponding realized DVHs.
Fig. 8 shows that, for case 1, the differences between the ve
simulated treatments were clinically indistinguishable from one
another. The results for the other four cases were similar.
4. Concluding remarks and future research directions
We propose a new stochastic optimization model for IMRT
treatment planning that accounts for interfraction motion and
distinguishes explicitly between total dose and dose/fraction
effects. For dose/fraction effects, we penalize the doses received
in each treatment fraction and evaluate the per-fraction DVHs,
while for total dose effects we consider the aggregate DVHs over
all treatment fractions. We illustrate the model by including dose/
fraction criteria for target coverage while including total dose
criteria for all targets and critical structures. Note, however, that
our proposed methodology applies equally well to other settings,
as long as the distribution of interfraction motion can accurately
be characterized. We show that robust treatment plans can be
obtained using only a modest number of scenarios. In particular,
the target dose coverage obtained with the stochastic model is far
superior to that obtained by a traditional model using xed
interfraction motion margins. Clearly, the quality of our results
depend on the correctness of the assumed organ motion model, so
 C. Men et al. / Computers & Operations Research 39 (2012) 17791789
Fig. 8. DVHs for ve simulated 41-fraction treatments for case 1 with the
treatment plan obtained with the stochastic model.
we should stress the importance of research on better motion
distribution models and estimation. The CPU time required for
solving the beamlet-based models is about 26 min, which is
feasible in a clinical setting.
Future research will focus on using a similar stochastic model
in a direct aperture optimization (DAO) setting (see [22], for
preliminary results). Moreover, we will investigate the benets of
a dynamic strategy in which we can reoptimize the treatment
plan for the remaining fractions at certain times during the
treatment (see also [32]).
References
[1] Aleman DM, Glaser D, Romeijn HE, Dempsey JF. A primaldual interior point
algorithm for uence map optimization in IMRT treatment planning. Physics
in Medicine and Biology 2010;55(18):546782.
[2] Antolak JA, Rosen II. Planning target volumes for radiotherapy: how much
margin is needed? International Journal of Radiation Oncology Biology
Physics 1999;44(5):116570.
[3] Bedford JL, Khoo VS, Oldham M, Dearnaley DP, Webb S. A comparison of
coplanar four-eld techniques for conformal radiotherapy of the prostate.
Radiotherapy & Oncology 1999;51(3):22535.
[4] Britton KR, Takai Y, Mitsuya M, Nemoto K, Ogawa Y, Yamada S. Evaluation of
inter- and intrafraction organ motion during intensity modulated radiation
therapy (IMRT) for localized prostate cancer measured by a newly developed
on-board image-guided system. Radiation Medicine 2005;23(1):1424.
[5] Chen S, Shah AP, Strauss JB, Turian JV, Kirk MC, Chu JCH, et al. Tomotherapy
treatment of the prostate and pelvic lymph nodes with a sequential cone-
down. International Journal of Radiation Oncology Biology Physics 2007;
69(3):S6923.
[6] Chu M, Zinchenko Y, Henderson SG, Sharpe MB. Robust optimization for
intensity modulated radiation therapy treatment planning under uncertainty.
Physics in Medicine and Biology 2005;50:546377.
[7] Craig T, Battista J, Van Dyk J. Limitations of a convolution method for
modeling geometric uncertainties in radiation therapy I: the effect of shift
invariance. Medical Physics 2003;30(8):200111.
[8] Craig T, Battista J, Van Dyk J. Limitations of a convolution method for
modeling geometric uncertainties in radiation therapy II: a nite number of
fractions. Medical Physics 2003;30(8):201220.
[9] Dempsey JF, Romeijn HE, Li JG, Low DA, Palta JR. A Fourier analysis of the dose
grid resolution required for accurate IMRT uence map optimization. Medical
Physics 2005;32(2):3808.
[10] El-Bassiouni M, Davis JB, El-Attar I, Studer GM, Lutolf UM, Ciernik IF. Target
motion variability and on-line positioning accuracy during external-beam
1789
radiation therapy of prostate cancer with an endorectal balloon device.
Strahlentherapie und Onkologie 2006;182(9):531636.
[11] Hamacher HW, Kufer K-H. Inverse radiation therapy planninga multiple
objective optimization approach. Discrete Applied Mathematics 2002;
118(12):14561.
[12] International Commission on Radiation Units and Measurements (ICRU).
Prescribing, recording and reporting photon beam therapy (Report 50); 1993.
[13] International Commission on Radiation Units and Measurements (ICRU).
Prescribing, recording and reporting photon beam therapy (Report 62); 1999.
[14] Kufer K-H, Scherrer A, Monz M, Alonso F, Trinkaus H, Bortfeld T, et al.
Intensity-modulated radiotherapya large scale multi-criteria programming
problem. OR Spectrum 2003;25(2):22349.
[15] Langen KM, Jones DTL. Organ motion and its management. International
Journal of Radiation Oncology Biology Physics 2001;50(1):26578.
[16] Lee EK, Fox T, Crocker I. Optimization of radiosurgery treatment planning via
mixed integer programming. Medical Physics 2000;27:9951004.
[17] Lee EK, Fox T, Crocker I. Integer programming applied to intensity-modulated
radiation therapy treatment planning. Annals of Operations Research 2003;
119(1):16581.
[18] Lee EK, Fox T, Crocker I. Simultaneous beam geometry and intensity map
optimization in intensity-modulated radiation therapy. International Journal
of Radiation Oncology, Biology, Physics 2006;64(1):30120.
[19] Lujan AE, Larsen EW, Balter JM, Ten Haken RK. A method for incorporating
organ motion due to breathing into 3D dose calculations. Medical Physics
1999;26(5):71520.
[20] McCarter SD, Beckham WA. Evaluation of the validity of a convolution
method for incorporating tumour movement and set-up variations into the
radiotherapy treatment planning system. Physics in Medicine and Biology
2000;45(4):92331.
[21] McShan DL, Kessler ML, Vineberg K, Fraass BA. Inverse plan optimization
accounting for random geometric uncertainties with a multiple instance
geometry approximation (MIGA). Medical Physics 2006;33(5):151021.
[22] Men C. Optimization models for radiation therapy: treatment planning and
patient scheduling. PhD thesis, Department of Industrial and Systems
Engineering, University of Florida, Gainesville, Florida; 2009.
[23] Moiseenko V, Liu M, Kristensen S, Gelowitz G, Berthelet E. Effect of bladder
lling on doses to prostate and organs at risk: a treatment planning study.
Journal of Applied Clinical Medical Physics 2007;8(1):5568.
[24] Parker BC, Shiu AS, Maor MH, Lang FF, Liu HH, White RA, et al. PTV margin
determination in conformal SRT of intracranial lesions. Journal of Applied
Clinical Medical Physics 2002;3(3):17689.
[25] Price RA. IMRT for prostate cancer. Technical Report: Fox Chase Cancer
Center, Philadelphia, Pennsylvania; 2006.
[26] Radiation Therapy Oncology Group (RTOG). RTOG0126: a phase III rando-
mized study of high dose 3D-CRT/IMRT versus standard dose 3D-CRT/IMRT
in patients treated for localized prostate cancer. /http://www.rtog.org/
members/protocols/0126/p0126.pdfS; 2004.
[27] Rodrigues G, Yartsev S, Chen J, Wong E, DSouza D, Lock M, et al. A
comparison of prostate IMRT and helical tomotherapy class solutions. Radio-
therapy & Oncology 2006;80(3):3747.
[28] Romeijn HE, Ahuja RK, Dempsey JF, Kumar A, Li JG. A novel linear program-
ming approach to uence map optimization for intensity modulated radia-
tion therapy treatment planning. Physics in Medicine and Biology 2003;
48:352142.
[29] Romeijn HE, Ahuja RK, Dempsey JF, Kumar A. A column generation approach
to radiation therapy treatment planning using aperture modulation. SIAM
Journal on Optimization 2005;15:83862.
[30] Romeijn HE, Dempsey JF. Intensity modulated radiation therapy treatment
plan optimization. TOP 2008;16(2):21543.
[31] Shepard DM, Ferris MC, Olivera GH, Mackie TR. Optimizing the delivery of
radiation therapy to cancer patients. SIAM Review 1999;41:72144.
[32] Sir MY, Epelman MA, Pollock SM. Stochastic programming for off-line
adaptive radiotherapy. Annals of Operations Research, doi:10.1007/s10479-
010-0779-x. In press.
[33] Stroom JC, de Boer HCJ, Huizenga H, Visser AG. Inclusion of geometrical
uncertainties in radiotherapy treatment planning by means of coverage
probability. International Journal of Radiation Oncology Biology Physics
1999;43(4):90519.
[34] van Herk M, Remeijer P, Rasch C, Lebesque JV. The probability of correct
target dosage: dose-population histograms for deriving treatment margins in
radiotherapy. International Journal of Radiation Oncology Biology Physics
2000;47(4):112135.
[35] Wang L, Jacob R, Chen L, Ma C, Movsas B, Feigenberg S, et al. Stereotactic
IMRT for prostate cancer: setup accuracy of a new stereotactic body
localization system. Journal of Applied Clinical Medical Physics 2004;5(2):
1828.