Optimization of Treatment Plans, Inverse Planning 207

17 Optimization of Treatment Plans, Inverse Planning

Thomas Bortfeld and Christian Thieke

CONTENTS to shape dose distributions in 3D. Optimization of

beam directions in fully 3D was not really possible
17.1 Introduction 207
with computers of that time.
17.2 Optimization Principles 208
17.2.1 Optimization Criteria 208 The situation changed drastically with the inven-
17.2.1.1 Physical Dose Criteria 209 tion of intensity-modulated radiation therapy (see
17.2.1.2 Physical and Technical IMRT Chap. 23) by Brahme and colleagues in 1982, along
Delivery Criteria 210 with exponential increase of computer power. The
17.2.1.3 The Objective Function 210
IMRT has a much greater potential to shape complex
17.2.2 Variables to Be Optimized 211
17.2.2.1 Intensity Maps 211 spatial dose distributions. Interestingly, the first ap-
17.2.2.2 Number of Beams 211 proaches to calculate intensity maps for IMRT were
17.2.2.3 Beam Angles 212 not based on optimization techniques. They instead
17.2.2.4 Number of Intensity Levels 212 inverted the underlying integral equation analyti-
17.2.2.5 Beam Energy 212
cally, which means that the spatial dose distribution
17.2.3 Optimization Algorithms 212
17.3 Inverse Planning in Practice 213 was prescribed, and the beam intensity distribution
17.3.1 General Approach 213 that would precisely yield this dose distribution was
17.3.2 Clinical Example 215 calculated. To find an exact solution to this inverse
17.4 Multicriteria Optimization in problem, several assumptions and approximations
Inverse Planning 216
had to be made, and solutions could only be found for
17.5 Conclusion 218
References 218 very simple symmetrical cases. Nevertheless, based
on this inversion idea, IMRT planning is often still
called inverse planning. Other early attempts at solv-
ing the inverse problem of radiation therapy plan-
17.1 ning used deconvolution techniques that were bor-
Introduction rowed from image processing (Brahme et al. 1982;
Holmes et al. 1991). Again, the solutions were always
Computerized optimization of treatment plans has approximations.
been proposed in radiation therapy since comput- A true solution of the inverse problem cannot
ers became available to hospitals in the 1960s (Hope be found in any but the simplest hypothetical cases,
et al. 1967; Bahr et al. 1968; Redpath et al. 1975; because the beam intensities can never become
McDonald and Rubin 1977). The first prototypes negative and therefore we can never subtract dose.
were presented at that time but did not find their Mathematically, the problem to generate a desired
way into clinical practice. The main reason was that dose distribution with rays of radiation of variable
those early approaches were basically restricted to intensity is equivalent to the problem of creating an
the optimization of beam weights, sometimes in com- arbitrary drawing using uniform straight lines drawn
bination with wedges, which have a limited potential with a pencil and ruler on paper. This problem was
mathematically analyzed by Birkhoff (1940). He
found that the problem can only be solved if an
T. Bortfeld, PhD eraser is also available. Unfortunately, we do not have
Department of Radiation Oncology, Massachusetts General a physical dose eraser at hand in radiation therapy.
Hospital, 30 Fruit Street, Boston, MA 02114, USA
Therefore, other mathematical methods were soon
C. Thieke, MD, PhD
Department of Radiation Oncology, German Cancer Research applied to this problem. One of them is called feasi-
Center (DKFZ), Im Neuenheimer Feld 280, 69120 Heidelberg, bility search (Censor et al. 1988; Starkschall et al.
Germany 2000). Here the idea is that, even though there is no
208 T. Bortfeld and C. Thieke

exact solution to the inverse problem that will yield variables (e.g., beam intensities and angles) that yield
the ideal dose distribution to the tumor with no dose the best scoring treatment for an individual patient.
to normal structures, we may still be able to find a In IMRT the number of variables (the intensities
plan that fulfills all clinical and physical constraints. for all beam elements for all beams) can be in the
This approach can be very useful if the constraints thousands, so we are typically dealing with a highly
are carefully chosen: we do not want to make it too dimensional optimization problem; however, it is im-
easy to fulfill the constraints because that may result portant to note that the score is a single number. The
in suboptimal plans. We also do not want to over- issues that are related with this are discussed later, but
constrain the problem, because then there will be no it is immediately clear that it is a formidable task to
solution at all. Feasibility search could be particularly characterize the goodness of a radiation treatment
successful for clinical cases that do not vary too much plan with a single number.
between individuals, such as prostate cancer. Here it The basics of optimization in radiotherapy have
is possible to find class solutions of constraints that been discussed in many recent reviews and book
will yield good treatment plans for many patients. chapters (Bortfeld 1999; IMRT 2001; Bortfeld
Presently, the most commonly used approach 2003; Censor 2003), so we are brief on this subject.
in IMRT planning is the optimization approach, in
which one tries to find the best physically and tech-
nically possible treatment plan with respect to given 17.2.1
physical and clinical criteria. Among the first pub- Optimization Criteria
lished optimization techniques were those by Webb
(1989) and Bortfeld et al. (1990). It was shown that The definition of the objective function and con-
with the optimization approach complex dose dis- straints (for which we use the combined term “op-
tributions could be generated with a limited num- timization criteria” in the sequel) is arguably the
ber (less than ten) of beam directions (Bortfeld most challenging part of IMRT optimization, and
et al. 1990). In most other previous approaches the requires a lot of thought. Optimization criteria that
use of an unrealistically large number of beams was are not clinically relevant or not complete can lead to
assumed. This paved the way for implementation of treatment plans that are mathematically optimal but
IMRT at conventional linear accelerators equipped clinically useless. Mathematical optimality is not a
with a multileaf collimator. Even though we focus on guarantee for a good treatment plan at all. In the defi-
IMRT optimization, many of the concepts that we dis- nition of the objective function one has to translate
cuss in this chapter can be applied to general radio- the clinical experience gained over the past 100 years
therapy optimization with other modalities. or so into mathematical terms, which in turn needs to
We first discuss the principles of current optimiza- be translated into computer language. The key ques-
tion approaches in IMRT. In section 3 we discuss the tion is: What characterizes the “best” treatment plan?
clinical aspects of IMRT optimization, and in section Clinicians often find it difficult to formulate a com-
4 we give an outlook on future developments in this plete, unique, quantitative set of optimization (scor-
field. ing) criteria for radiotherapy planning, even though
they feel capable of ranking individually prepared
plans (Langer et al. 2003). This is somewhat differ-
ent from other applications of optimization, e.g., in
17.2 economics, in which the objective is often quite clear
Optimization Principles (e.g., minimize the cost).
It may seem obvious to define the optimization
Optimization is mathematically defined as the maxi- criteria in terms of treatment outcome, say, maxi-
mization or minimization of a score (a number) un- mize the probability of curing the patient without
der certain constraints. The scoring function (also exceeding tolerance limits of side effects in normal
called “objective function”), which yields the score, is tissues. The problem is that there exists no commonly
a function of all variables that need to be optimized. accepted and validated model that translates the
In general mathematical optimization the problem is physical dose distribution into normal tissue compli-
to find the set of variables that yield the maximum cation probabilities (NTCP) or tumor control prob-
(or minimum) score while fulfilling all constraints. abilities (TCP), let alone cure rates. In 1968 Alexander
Specifically, in radiotherapy optimization the prob- Solzhenitsyn wrote in his novel Cancer Ward: “There
lem is to find the treatment plan with all its associated was no formula for calculating the right intensity of
Optimization of Treatment Plans, Inverse Planning 209

v
dose, for knowing how much would be most lethal tribution di(b ). One often sets Dmin=Dmax= prescribed
for an individual tumor yet least harmful to the rest dose in the target volume. In critical structures the
of the body.” Even though many clinical studies have first term is omitted by setting u=0 (or equivalently,
been performed since then, and a lot of data has been setting the minimum dose constraint to 0). [x]+stands
collected, this basic problem still persists. for x if x>0 and 0 otherwise. This quadratic deviation
The most common approach in practical radio- approach has been criticized because it does not suf-
therapy optimization at present is therefore to use ficiently penalize dose cold spots in the target volume
optimization criteria that are defined in terms of or hot spots in critical structures. For example, if the
the physical dose distribution. In many cases these dose in the target volume becomes zero in only one
criteria are not absolutely fixed, but are adjusted by voxel, F may still be in an acceptable range, as long
the treatment planner in a “human iteration loop,” as as most voxels receive the prescription dose. Higher-
the optimization progresses. The clinical aspects of order measures of the deviation (such as differences
this are explained in detail in the subsequent section. taken to the power of 4, 6, …) have been suggested
Suffice it to say that optimization using these physical (Fraass 2002) but are not common in clinical IMRT
criteria is not expected to yield the best clinical plan optimization.
in one step. What is more important is that these cri- One way to enforce dose homogeneity is to put
teria provide good “steerability” of the plan: If, after hard constraints on the actual minimum and maxi-
the first optimization run, the plan is not satisfactory mum dose in the target volume (as opposed to the
from a clinical point of view, an adjustment of the “soft” deviation approach above). For example, an op-
criteria should allow to steer the plan in the desired timization constraint might require to keep the target
direction in a straightforward manner using only a dose within –5% and +7% of the prescription dose,
few of the “human iteration loops.” as recommended by the International Commission
We now discuss optimization criteria that are used on Radiation Units and Measurements (ICRU 1993).
in various commercial IMRT optimization systems. A possible optimization objective is to maximize the
The physical dose criteria discussed below can be minimum dose value in the target volume (Langer et
used as either constraints or as objectives. In the con- al. 1996). Maximum dose limitations (but, of course,
straint formulation one merely requests that the cri- not minimum dose limitations) make sense in critical
teria stay within certain limits. If they are defined as structures as well. In structures with a serial orga-
(part of) the “objective function,” the goal is to maxi- nization, such as the spinal cord, the complication is
mize or minimize these quantities. correlated with the maximum dose. For example, in
the spinal cord the maximum dose should be limited
17.2.1.1 to, for example, 45 Gy.
Physical Dose Criteria The mean dose is a useful descriptor of the dose
effect in the target volume as long as deviations from
A widely used criterion is the quadratic deviation of the mean are not too big (Brahme 1984; Levegrun
the actual dose from a prescribed dose level Dmin (in et al. 2001). The mean dose is also a useful clinical
the target volume) or tolerance dose level Dmax (in parameter in critical structures that are organized in
both the target volume and critical structures), which parallel, such as the lung (Kwa et al. 1998). It should
is to be minimized. Different weight factors (also be noted, however, that different regions in the lung
called penalty factors or importance factors) can be have different radiation sensitivities. The mean dose
used for underdose and overdose. The objective func- overall can therefore only be a very coarse predictor
tion (“costlet”) for a particular organ is then: of side effects. It should also be emphasized that dif-

( )
r N r 2 r ferent kinds of complication can occur in the same
F (b ) = ∑ u ⎡⎣ Dmin − di (b ) ⎤⎦ + w ⎡⎣ di (b ) − Dmax ⎤⎦
2
organ, e.g., in the lung, such that different criteria may
+ +
i =1 have to be used for one organ in the optimization.
where u and w are the weight
v factors for under- and Another common method to represent dose effects
overdose, respectively. di(b ) is the dose at voxel i as a in parallel critical structures is to use dose-volume
function of the beam element (bixel) intensities and histogram (DVH) constraints (Langer and Leong
other variables (e.g.,
v beam angles) combined into the 1989; Langer et al. 1990; Niemierko 1992; Bortfeld
argument vector b . Dmin and Dmax are the prescribed et al. 1997; Spirou and Chui 1998; Gustafsson and
minimum and maximum doses that should ideally Langer 2000), which take into consideration the
not be exceeded. Note that Dmin and Dmax are not the volume dependence to some degree. The DVH con-
minimum and maximum doses of the actual dose dis- straints can be formulated as “no more than Vmax% of
210 T. Bortfeld and C. Thieke

the volume should receive more than a dose of Dmax”. ity. All organs can be characterized by just one param-
They can be visualized as a barrier with a corner at eter. The EUD has recently been implemented into
the point Dmax, Vmax on the DVH plot. The use of mul- experimental IMRT optimization systems (Thieke
tiple DVH constraints may also be indicated in some et al. 2002; WU et al. 2002). A convenient feature of
cases (Carol et al. 1997), and DVH constraints can the EUD is that it is a convex function of the dose
be used for the target volume as well. distribution, and therefore of the beam intensities,
One general advantage of physical criteria, such as for a * 1 and for a ) 0, i.e., for almost all organs. This
dose-volume criteria, is that, because they are simply facilitates optimization and makes the use of projec-
and clearly defined, they can be used easily in clinical tion onto convex sets (POCS) methods possible (Cho
protocols. Such IMRT protocols using a combination et al. 1998; Thieke et al. 2002).
of various physical dose criteria have been developed
and distributed by the Radiation Therapy Oncology 17.2.1.2
Group (RTOG, www.rtog.org). A well-known exam- Physical and Technical IMRT Delivery Criteria
ple is their protocol 022 for IMRT of oropharyngeal
cancer. The most obvious physical limitation of this type is
A potential mathematical problem with dose-vol- that the beam intensities must not be allowed to take
ume constraints is that they are non-convex (Deasy negative values. This does not require any further
1997). As a consequence of this, optimization based discussion.
on DVH constraints may get trapped in local minima; Another important aspect of optimized treat-
however, it has been shown that this is mainly a theo- ment plans is that they have to be deliverable in an
retical problem, which seems to be of little relevance acceptable amount of time. Also, small errors in the
in practical optimization (Wu and Mohan 2002; delivery should not affect the resulting dose distribu-
Llacer et al. 2003). tion too much, i.e., the plan should be robust. Both
The final physical dose criterion that we discuss of these requirements can be fulfilled with intensity
is the equivalent uniform dose (EUD). It is defined as maps that are smooth. If there are multiple solutions
the uniform dose that would create the same biologi- to the optimization problem, which is frequently the
cal effect in a specific organ as the actual non-uni- case, one should obviously choose the solution with
form dose distribution (Brahme 1984; Niemierko the smoothest (easiest to deliver) intensity maps. If
1997). Because it involves the biological effect, it is one has the choice between a highly conformal plan
often interpreted as a biological rather than physical with extremely complex intensity maps, and a plan
dose criterion; however, according to a more recent that is almost as good but much easier and safer to
definition by Niemierko (1999) the EUD is simply deliver, one would probably choose the latter plan.
the generalized mean (a-norm) of the physical dose Basically there are two ways to ensure smooth inten-
distribution: sity maps. First, one can smooth the maps using a fil-
1/ a ter. Median filters can be used for this purpose (Webb
⎛ ⎞
EUD = ⎜ ∑ vi ⋅ dia ⎟ et al. 1998; Kessen et al. 2000). Secondly, one can add
⎝ i ⎠ a smoothing term to the objective function (Alber
where vi is the volume of voxel i divided by the total and Nüsslin 2000).
volume of the organ. With this definition and with
a = 1 the EUD is in fact the mean dose and with a = ' 17.2.1.3
it equals the maximum dose. Of course, the value The Objective Function
of a is organ specific. Values between 1 and infinity
represent organs with a mixture between parallel and The criteria (or “costlets”) mentioned in the previ-
serial structures. It is interesting to note that the value ous subsections must somehow be combined to form
of a can be derived from the well-known power-law the total objective function. An alternative is to use
relationship of the tolerance dose TD as a function only one criterion in the objective function and set
of the relative treated volume v: all other criteria as constraints. The most common
approach is to add some criteria using weight fac-
TD(1)
TD(v) = . tors that reflect the importance of the different cri-
vn teria. Multiplication of costlets has also been sug-
One finds that a = 1/n. gested (Fraass 2002). A typical approach is to add
For target volumes, the value of a is negative. The the deviations from the dose prescription (tumor)
beauty of this approach is its simplicity and general- and maximum doses (critical structures) as defined
Optimization of Treatment Plans, Inverse Planning 211

in subsection 2.1.1 using organ-specific weight fac- Besides the common two-step approach, it has also
tors uk and wk. According to this definition the opti- been suggested to avoid the intermediate step of us-
mum treatment is the one with the smallest overall ing intensity maps altogether and directly optimize
deviation from the prescription. In the plan steering MLC shapes (apertures) and their weights. This ap-
process (the “human iteration loop”) it is not a priori proach has been suggested by DeNeve et al. (1996),
clear by how much to change the weights to achieve among others. The MLC shapes can be determined
the desired effect on the dose distribution. Suitable manually based on the geometry (anatomy) of the
weight factors have to be determined by trial and problem, or they can be directly optimized together
error, which can be quite time-consuming and is dis- with the weights of the segments. The latter method
cussed further below. Some researchers argued that has recently been published (Shepard et al. 2002).
optimized “inverse” IMRT planning is just replacing The direct optimization of MLC shapes and weights
the conventional manual trial-and-error search to find is mathematically a difficult, non-convex problem.
the best beam parameters by a manual trial-and-er-
ror search for suitable weight factors and constraints. 17.2.2.2
While there may be some truth to this statement, it is Number of Beams
also true that IMRT allows one to deliver highly con-
formal concave dose distributions that are impossible The question of how many intensity-modulated
to achieve with conventional 3D techniques. beams should be used is highly relevant for the practi-
cal delivery of IMRT. As a consequence of the analogy
between image reconstruction in computed tomogra-
17.2.2 phy (CT) and inverse radiotherapy planning, the early
Variables to Be Optimized theoretical approaches to inverse planning assumed
a very high number of coplanar beams (Brahme et
After determining the objective function and the set al. 1982; Cormack and Cormack 1987). Later it was
of constraints, the next question is about the variables recognized that quite acceptable results can also be
to be optimized. Even with the most powerful mod- achieved with a moderate number of beams. In fact,
ern computers it is impossible to optimize all treat- some publications claim that one generally does not
ment parameters (variables). In practice, a subset of need more than three intensity-modulated beams to
the treatment parameters, including, for example, the obtain results that can hardly be improved any more
beam angles, need to be manually pre-selected. (Söderström and Brahme 1995). These issues have
been discussed frequently in the literature (Brahme
17.2.2.1 1993; Brahme 1994; Mackie et al. 1994; Mohan and
Intensity Maps Ling 1995; Mohan and Wang 1996; Söderström
and Brahme 1996).
In IMRT the main variables to be optimized are obvi- In principle, it is clear that the higher the number
ously the intensity maps for each beam. Each beam is of beams, the higher is the dose conformation poten-
typically subdivided into beam elements (bixels) of tial; however, the incremental improvement of the
5×5 to 10×10 mm2. The intensity (fluence) for each conformity of the total dose distribution diminishes
of the bixels is optimized. The total number of bixels as more beams are added. The real question about the
for all beams is typically of the order of 1000–10,000. “optimum” number of beams in IMRT is therefore:
Because there is no way to deliver intensity-modu- What is the number of beams beyond which one does
lated photon beams directly with a linac, the intensity not see any practically relevant improvement of the
maps are then converted to a series of multileaf col- treatment plan? Several authors have found indepen-
limator (MLC) shapes (segments) in a second almost dently that it is hardly ever necessary to use much
independent step, which is called leaf sequencing. Of more than ten intensity-modulated beams to achieve
course, there has to be some link between optimi- results that are close to optimum (Bortfeld et al.
zation and sequencing. For example, the optimizer 1990; Webb 1992; Söderström and Brahme 1995;
must know the leaf width of the MLC and should use Stein et al. 1997).
that as the bixel size in one dimension. More thor- On the other extreme, with a number of beams as
ough approaches for the consideration of delivery small as three or four, the conformity of the result-
constraints in intensity map optimization have also ing dose distribution is considerably reduced when
been suggested (Cho and Marks 2000; Alber and compared with a plan incorporating ten beams;
Nüsslin 2001). however, thanks to intensity modulation, it is still
212 T. Bortfeld and C. Thieke

possible to imprint the desired shape on at least one with step-like beam profiles as well (Bortfeld et al.
isodose curve, e.g., the 80% isodose. This finding is 1994; Gustafsson et al. 1994; DeNeve et al. 1996).
related to image reconstruction of homogeneous ob- In fact, using a moderate number of stair steps with
jects using only three or four projections (Natterer five to seven “intensity levels” in each beam profile,
1986); hence, in cases where the tolerance of the criti- the results are almost as good as with continuous
cal structures is not too low as compared with the modulation (Keller-Reichenbecher et al. 1998).
required target dose, one may get away with very Consequently, it is not necessary to go to a fully dy-
few intensity-modulated beams (Söderström and namic treatment mode to perform IMRT with an
Brahme 1995). MLC. The IMRT can instead be realized in a “step-
and-shoot” mode, i.e., by the successive delivery
17.2.2.3 of a number of static MLC-shaped beam segments
Beam Angles from each direction of incidence. The total number
of beam segments or “subfields” to be delivered in
The question about the optimum beam angles is re- this way is of the order of 100. Modern treatment
lated to the number of beams. Clearly, if very many machines can deliver such a sequence of subfields
beams (more than ten) are used, they can be placed automatically and quickly. A comparison of the fea-
at evenly spaced angular intervals and there will be tures of dynamic vs step and shoot IMRT has been
no need to optimize beam orientations. Even with a published by Chui et al. (2001).
moderate number of beams of the order of seven or
nine, one may often use evenly spaced beams without 17.2.2.5
compromising the dose distribution (Bortfeld and Beam Energy
Schlegel 1993); however, it has been shown that
this is case dependent, and complex cases, such as The choice of the beam energy is less critical in IMRT
head and neck, sometimes benefit from beam ori- than in conventional radiotherapy. In fact, it was sug-
entation optimization even for nine or more beams gested that very low energies around 1 MeV or less
(Pugachev et al. 2001). With very few beams, such suffice in IMRT. The reason is once again that in IMRT
as four or less, very careful and time-consuming op- one tends to spread the beams more evenly around
timization of beam orientations is always essential; the patient, and the depth-dose fall-off is therefore
otherwise, one will not be able to achieve acceptable not very relevant.
results.
It should be noted that optimal orientations of
intensity-modulated beams are generally different 17.2.3
from those of uniform beams: in IMRT it is not gen- Optimization Algorithms
erally necessary and often not even advantageous to
avoid beam directions through organs at risk, because The objective function as a function of the variables
these can be spared by reducing the intensity for the to be optimized, in combination with the optimiza-
corresponding rays (Stein et al. 1997). Also, for simi- tion constraints, defines the optimization problem.
lar reasons, non-coplanar beams are rarely used in As for its solution, many mathematical algorithms
IMRT. Another point worth mentioning is that paral- have been developed to solve optimization problems
lel opposed beams should be avoided in IMRT, such of various kinds. The algorithm of choice depends on
that, for evenly spaced beams, the number of beams the type of the problem (linear/non-linear, convex/
should be odd. The reason for this is that a parallel non-convex). We cannot describe these algorithms
opposed beam adds much less beam-shaping poten- in any detail in this chapter. The interested reader is
tial than a slightly angled beam. Clearly, if the attenu- referred to other reviews (Bortfeld 1999; Shepard
ation were zero, parallel-opposed beams would be et al. 1999).
completely useless. The most common approach in commercial IMRT
optimization systems is the non-linear (often qua-
17.2.2.4 dratic) problem formulation. The algorithm used for
Number of Intensity Levels finding the solution is usually a variant of the “gradi-
ent” technique. It converges rapidly and can be ap-
Most IMRT planning methods assume a continuous plied to a wide range of optimization problems; how-
modulation of the intensity. Several investigations ever, in the non-linear case one does not usually let
have shown that promising results can be achieved the algorithm converge to the numerical optimum,
Optimization of Treatment Plans, Inverse Planning 213

but stops after an acceptable number of iteration There are also clinical cases where even complex
steps. Statements about the proximity to the true op- shaped targets can safely be treated with conventional
timal solution are more difficult to make than in the techniques by skillfully combining different radiation
linear case, or impossible. modalities (photons and electrons), wedges, individ-
There is no doubt that in general optimization the ual blocks, and non-coplanar beam angles (Yajnik
biggest experience exists in the field of linear opti- et al. 2003); however, the planning and delivery of
mization algorithms (“linear programming”). Even those treatments might be even more time-consum-
though many of the IMRT objectives and constraints ing than an IMRT treatment with a comparable dose
discussed above are non-linear, a linear model or its distribution. So even from an economic point of view
variants can approximate the problem sufficiently there can be an indication for IMRT. But of course, in
well to yield useful solutions (Langer et al. 1996; general IMRT is more time-consuming and more ex-
Romeijn et al. 2003). An advantage of the linear prob- pensive than conventional radiotherapy and should
lem formulation is that the optimality of the solution only be applied if a clinical benefit for the patient can
can be proven. be expected.
Step 3. From here on we assume that open fields
did not lead to a satisfying dose distribution or the
geometry of the planning problem obviously is too
17.3 complex for open field treatment. The IMRT module
Inverse Planning in Practice has to be started from within the TPS. It imports the
CT data, the beam configuration, and the organ con-
17.3.1 tours from the TPS.
General Approach Step 4. Now the parameters needed for the optimi-
zation have to be defined, namely the dose prescrip-
Most planning systems used in clinical practice pres- tion and weight factors for each structure. Figure 17.1
ently were not designed from ground up for inten- shows the input window of the inverse planning pro-
sity-modulated radiotherapy. Instead, the IMRT opti- gram KonRad (distributed by Siemens OCS). Please
mization engine is usually implemented as a separate note that a priority is assigned to structures which
program that communicates with the conventional have some overlap with other structures. In the re-
three-dimensional forward treatment planning sys- gion of overlap the optimization engine is using the
tem (TPS). Since many steps of the complete planning settings of the structure with the higher priority.
process are the same for both forward and inverse Step 5. Then the optimization is carried out.
planned treatments, this is not necessarily a disad- Depending on the specific case and on the program
vantage. used, this process takes from less than a minute up to
A general approach to radiotherapy planning, in- hours. In the end the inverse planning program pres-
cluding IMRT, can be formulated as follows: ents the resulting treatment plan. It is evaluated with
Step 1. The contours of target structures and organs regard to the dose distribution in the target struc-
at risk are outlined in the TPS (using the appropriate tures and the organs at risk using DVHs and full 3D
imaging modalities such as CT, MRI, and PET). dose information in form of isodoses or color-wash
Step 2. In some cases, at this stage it might be un- displays. If the actual treatment plan is not accept-
clear whether conventional unmodulated fields are able in one or more parts, the optimization has to be
sufficient or whether IMRT is necessary to achieve restarted with modified start parameters. The steer-
a satisfying dose distribution. In these cases one ing parameters are both the dose constraints and the
should first try to find a good treatment plan with weight factors. In the case of step-and-shoot deliv-
open fields by varying number, direction, and weight ery, the planner also has to check the number of seg-
of the beams and their individual shape defined by ments of the plan. Too many segments can lead to
the MLC leafs. E.g., for intracranial meningiomas a unacceptable delivery times, making it necessary to
3D-planned, conventional treatment with uniform reduce the number of intensity levels or to activate
fields usually is sufficient. But some meningiomas profile smoothing.
are located directly next to several critical structures, Hunt et al. (2002) have systematically investigated
such as the optical nerves and the brain stem, so it the interplay of the optimization parameters for a
may turn out that curative target doses with open simple test case consisting of a concave-shaped tar-
fields would lead to unacceptable doses in those or- get structure (PTV) around an organ at risk (normal
gans at risk, and IMRT should be used instead. tissue, NT) shown in Fig. 17.2a. The evaluation crite-
214 T. Bortfeld and C. Thieke

Fig. 17.1. Example of maximum and minimum dose prescriptions and the corresponding weight factors (penalties) for a prostate
case, as entered into a commercial inverse planning program (KonRad by Siemens OCS)

a b c

Fig. 17.2. a Geometry of the test case. b Variation of the NT dose constraints for PTV-NT distances of 5–13 mm. c Variation of
the NT weight factor for different NT dose constraints at a fixed PTV-NT distance of 5 mm. (From HUNT et al. 2002)

rion for the PTV was the Uniformity Index defined direction of increasing weight factors: higher weight
as maximum dose divided by 95% dose level, and for factors led to worse PTV dose uniformity in all cases,
the NT the maximum dose was used. Figure 17.2b but only for the 10% dose constraint did they also
shows the impact of the NT dose constraint onto improve the dose to the normal tissue. Of course, the
the PTV: especially for small distances between PTV results depend on the specific inverse planning sys-
and NT, a demanding NT dose constraint can greatly tem, but it is a general observation in inverse plan-
deteriorate the target dose. Note that, in this figure, ning that the sensitivity of a solution, i.e., the impact
a large PTV uniformity index corresponds with a of changes to the optimization parameters onto the
highly non-uniform dose. In Fig. 17.2c, the influence optimization result, greatly varies and can only be
of the weight factors is shown: for different NT dose found out by trial and error.
constraints from 10 to 70% of the prescribed target Based on the analysis of the above test case and
dose, the associated weight factor was varied from their experience in clinical practice, Hunt et al. (2002)
10 to 1000. The arrowhead in Fig. 17.2c indicates the formulated a strategy for inverse planning (Fig. 17.3).
Optimization of Treatment Plans, Inverse Planning 215

Fig. 17.3. Strategy of inverse plan-

ning. (From HUNT et al. 2002)

Depending on PTV dose uniformity or normal tissue dissemination; and the macroscopic tumor volume
dose, they propose different changes to the constraint (“GTV”). Both parotid glands are very close to the
and penalty settings. In clinical practice there will be target contour. Also the eyes with optic nerves and
modifications to this scheme, depending on the plan- chiasm, the brain stem, the spinal cord, the esopha-
ning system actually used, but it gives a good general gus, and both lungs are in immediate vicinity to the
overview over the parameters used for steering the target volume and have therefore been considered in
optimization result. the IMRT optimization.
Step 6. Once an acceptable IMRT plan has been The right side of Fig. 17.4 shows an exemplary CT
found, the treatment plan can be exported to the de- slice of the final IMRT plan obtained with the inverse
livery machine for verification and application pur- planning program KonRad. One can see an additional
poses. Depending on the accuracy of the dose calcu- contour called “Target + 10 mm.” This contour was
lation used for the IMRT optimization, it might be defined solely for better control of the dose gradient
necessary to recalculate the dose in the TPS. between the high-dose area and the normal tissue
during the optimization (such pseudostructures can
be used whenever an IMRT plan shall be altered in
17.3.2 circumscribed areas of the irradiated volume). The
Clinical Example median dose to the GTV was normalized to 66 Gy,
and the target volume received a median dose of
In this section an IMRT treatment of a head and 54 Gy. Treatment was delivered in 30 fractions, re-
neck tumor is demonstrated. The patient, a 63-year- sulting in a single fraction dose of 2.2 Gy to the GTV
old woman, was diagnosed with lymphoepithelioma and 1.8 Gy to the Target. This technique is also called
originating from the left eustachian tube. The mac- “simultaneously integrated boost (SIB).” The SIB is a
roscopic tumor spread, as seen in the MRI image, unique feature of IMRT treatments which simplifies
was 3×1.5×2 cm3. No positive lymph nodes or distant treatment and potentially leads to higher local control
metastases were found. Concomitant to the radio- rates (Mohan et al. 2000; Lauve et al. 2004). Another
therapy, the patient received chemotherapy consist- outstanding feature of IMRT treatments for head and
ing of 5-FU and Cisplatin. neck tumors is the sparing of salivary glands. In the
On the left side of Fig. 17.4, the structures of this example given here, the mean dose to the right pa-
case are shown as they appear in the observer’s view rotid could be restricted to 22 Gy (the left parotid
of the TPS. Two different target structures were de- gland, located near the macroscopic tumor, receives
fined: the uninvolved lymph-node-bearing tissue a mean dose of 37.4 Gy). One year after treatment,
(“Target”) that was electively treated because the the patient still does not suffer from xerostomia, an
tumor is known to have a high risk of lymphatic inevitable side effect of conventional head-and-neck
216 T. Bortfeld and C. Thieke

Eyes

Chiasm
Optic Brain
nerves stem
Left
parotid
Right
parotid

Target

Right Lung
Left
Lung
Spinal
Cord

Fig. 17.4. Organ contours and dose distribution of the IMRT head and neck case

radiotherapy that can greatly reduce the quality of life. A new optimization paradigm, the multicriteria
Especially in cases with a very good long-term prog- optimization, is a promising concept to overcome
nosis, like in this example, such considerations are these problems. In the following we present the
important. Also the doses to all other organs at risk in joint effort of the Massachusetts General Hospital
this case could be kept well below the tolerance level. (Boston, Mass, USA), the Fraunhofer Institute for
Industrial Mathematics (Kaiserslautern, Germany),
and the German Cancer Research Center (Heidelberg,
Germany; Küfer et al. 2000; Bortfeld et al. 2002;
17.4 Küfer et al. 2003; Thieke 2003).
Multicriteria Optimization in Inverse Planning To motivate multicriteria optimization, it is help-
ful to step back and look at the original optimization
In the previous sections we saw that inverse planned problem: There is a target structure which should be
IMRT is already in use in clinical practice; however, treated with full dose, and there are organs at risk
it also became clear that “inverse planning” as done nearby that should receive as little dose as possible.
today leaves a lot to be desired, and a lot of research But even with IMRT, due to the physical properties of
effort is still put into this field. Presently, inverse plan- photon beams (namely, the depth-dose curve with an
ning is still a recursive approach that might include exponential tail and the penumbra), it is impossible
several optimization runs and can therefore be quite to deliver the ideal dose distribution, which is 100%
time-consuming. Setting the optimization param- in the target and 0% everywhere else; therefore, ev-
eters is non-intuitive. The penalty factors are artifi- ery treatment plan is a certain compromise between
cial parameters that do not have a clinical meaning, these conflicting goals. Mathematically speaking, the
and even the constraints, which should stand for the optimization problem does not have a single solution
tolerance doses of the particular organ at risk, are where all criteria are at their individual optimum at
often used as mere steering parameters when they the same time. This is exactly the definition of a mul-
cannot be fulfilled in every voxel of the structure any- ticriteria optimization problem. Instead of a single
way. The impact of changes to the parameters cannot optimum, there is a whole solution space of optimal
be seen a priori. Instead, a new optimization run is compromises, also called Pareto optimal solutions.
needed, making changes to the plan non-interactive. A compromise is Pareto optimal when it cannot be
The planning might stop when a certain time limit improved in one criterion without worsening at least
has been reached, not necessarily when the best treat- on other criterion. The whole set of Pareto optimal
ment plan for the patient has been found. solutions is called the Pareto front.
Optimization of Treatment Plans, Inverse Planning 217

The result of multicriteria optimization for inverse well as the equivalent uniform dose (EUD) discussed
planning is therefore no longer a single treatment above, or models of tumor control probability (TCP)
plan. Instead, a whole database of Pareto optimal and normal tissue complication probability (NTCP).
treatment plans is generated. Each plan represents a In multicriteria optimization, these costlets are evalu-
certain clinical compromise, e.g., plan A of the data- ated separately; they are not simply added up to yield
base might have a lower dose in an organ at risk than a single score. We propose the use of EUD because it
plan B. But since every plan is Pareto optimal, plan A is clinically meaningful, is still in the dose domain
will also deliver more dose to another organ at risk and therefore familiar to the treatment planner, and
and/or less dose to the target than plan B. Which com- is differentiable and convex which allows for efficient
promise is the best for the patient is a clinical decision optimization algorithms.
that cannot be made by the computer. Different clini- Inverse planning for radiotherapy using the mul-
cians might even choose different plans because they ticriteria optimization program differs significantly
prefer different balances between tumor dose and from the current trial-and-error approach. In a first
normal tissue sparing. The decision for a particular step, the database of Pareto optimal plans, based on the
plan is made by interactively browsing the database individual patient geometry and contours, is generated.
of plans with an intuitive graphical interface. This can take up to several hours, which is not a big is-
In order to rank different dose distributions in sue since no human interaction is required during the
the target and organs at risk, a costlet function has to calculation. In the second step, the graphical user inter-
be defined separately for each structure. Options are face (Fig. 17.5) comes up. The planner browses through
the dose-based and dose-volume-based functions, as the database interactively using this interface.

Fig. 17.5. User interface of the interactive navigator (developed by ITWM, Kaiserslautern)
218 T. Bortfeld and C. Thieke

The user interface is divided into two main win- Multicriteria optimization for radiotherapy plan-
dows: the left panel shows the part used for the navi- ning as shown in this chapter is still under develop-
gation. Here the complete database is visualized by ment, but when it becomes available in clinical rou-
a star with one axis for each structure of the treat- tine it may change the way radiotherapy planning is
ment plan. The gray area shows all EUD values that done in the future.
can be reached for each organ. This way the complete
planning horizon becomes apparent at first sight. The
planner does not have to find out by trial and error
what dose might be achievable in a certain critical 17.5
structure, instead he just looks at the extreme values Conclusion
on the EUD navigation axes. In addition, the current
treatment plan is represented by a polygon connect- Optimized inverse planning can yield superior treat-
ing the EUD values of all structures. The right panel ment plans, especially in complex situations with
visualizes the current plan as dose volume histogram convex–concave target volumes and nearby critical
and as 3D dose distribution in frontal, sagittal, and structures; however, the optimization criteria must be
transverse projection. All plans are normalized to the carefully chosen. Determining appropriate optimiza-
same target EUD, and the planning horizon can be tion criteria is not straightforward and requires some
changed by the global normalization factor. trial and error in a “human iteration loop.” Using
At the beginning, the system suggests a starting current commercial inverse planning systems this
solution. If one or more aspects of the plan are not process can be quite time-consuming. Experienced
desirable, e.g., the EUD in one critical structure is too treatment planners know how to steer an IMRT plan
high, the planner can immediately go to another so- in the desired direction by appropriately changing
lution in which that specific criterion is fulfilled. This the optimization criteria. Also, class solutions can
is done by grabbing the marker at the EUD bar of help to avoid or reduce the “human iteration loop” in
the particular organ and dragging it in the desired cases that do not vary too much between individu-
direction. Instantaneously, the system finds the cor- als, such as prostate treatments, because optimization
responding solution in its database and updates the criteria can be re-used. Nevertheless, plan optimiza-
information for the other structures and in the DVH tion leaves something to be desired. The main prob-
and isodose windows. lem is that it may not be possible to come up with a
In our opinion the advantages of this concept are quantitative, complete optimization formulation for
threefold: radiotherapy planning in the near future; however, an
• Artificial weight factors, which have no clinical achievable alternative is to design optimization sys-
meaning, are avoided. The whole concept is based tems that let the physicians exercise their experienced
on dose-like values, which are amenable to a clini- clinical judgment or intuition in the most direct in-
cal interpretation. teractive way. Therefore, some future developments
• Unnecessarily high doses in some of the critical aim at a more interactive approach towards inverse
structures, which can occur in constrained opti- planning. Multicriteria optimization and navigating
mization, are avoided by definition of the Pareto a treatment plan database have been described as
optimal solution. promising approaches in this context.
• Plan tuning can be done interactively using
“knobs” that have a clinical meaning. It is easy to
do a sensitivity analysis and determine the depen-
dency of, for example, the target EUD on any of References
the critical-structure EUDs.
Alber M, Nüsslin F (2000) Intensity modulated photon beams
subject to a minimal surface smoothing constraint. Phys
In summary, the planner gains a complete overview Med Biol 45:N49–N52
over the individual solution space and can choose Alber M, Nüsslin F (2001) Optimization of intensity modulated
the compromise that is really the best for the patient, radiotherapy under constraints for static and dynamic
whereby the time required for human intervention is MLC delivery. Phys Med Biol 46:3229–3339
Bahr GK, Kereiakes JG, Horwitz H, Finney R, Galvin J, Goode
only seconds to minutes. Compared with inverse plan-
K (1968) The method of linear programming applied to
ning with scalar objective functions, the workload of radiation treatment planning. Radiology 91:686–693
the treatment planner can be reduced and the quality Birkhoff G (1940) On drawings composed of uniform straight
of the resulting treatment plan can be improved. lines. J Math 19:3
Optimization of Treatment Plans, Inverse Planning 219

Bortfeld T (1999) Optimized planning using physical objec- mization problems with dose-volume constraints. Med
tives and constraints. Semin Radiat Oncol 9:20–34 Phys 24:1157–1161
Bortfeld T (2003) Physical optimization. In: Palta JR, Mackie DeNeve W, DeWagter C, DeJaeger K, Thienpont M, Colle C,
TR (eds) Intensity-modulated radiation therapy: the state Derycke S, Schelfhout J (1996) Planning and delivering
of the art. Medical Physics Publishing, Madison, Wisconsin, high doses to targets surrounding the spinal cord atthe
pp 51–76 lower neck and upper mediastinal levels: static beam-seg-
Bortfeld T, Schlegel W (1993) Optimization of beam orienta- mentation technique executed with a multileaf collimator.
tions in radiation therapy: some theoretical considerations. Radiother Oncol 40:271–279
Phys Med Biol 38:291–304 Fraass B (2002) Differences between plan evaluation and
Bortfeld T, Bürkelbach J, Boesecke R, Schlegel W (1990) Meth- the optimization problem statement, and the difference
ods of image reconstruction from projections applied to it makes. http://www.isye.gatech.edu/nci-nsf.orart.2002/
conformation radiotherapy. Phys Med Biol 35:1423–1434 talks.php.
Bortfeld T, Boyer AL, Schlegel W, Kahler DL, Waldron TJ (1994) Gustafsson A, Langer M (2000) Dose-volume constrained
Realization and verification of three-dimensional confor- radiotherapy optimization for a clinical treatment plan-
mal radiotherapy with modulated fields. Int J Radiat Oncol ning system. Thirteenth International Conference on the
Biol Phys 30:899–908 use of Computers in Radiation therapy. Springer, Berlin
Bortfeld T, Stein J, Preiser K (1997) Clinically relevant intensity Heidelberg New York
modulation optimization using physical criteria. Twelfth Gustafsson A, Lind BK, Brahme A (1994) A generalized pencil
International Conference on the Use of Computers in beam algorithm for optimization of radiation therapy. Med
Radiation Therapy, Salt Lake City, Utah Phys 21:343–356
Bortfeld T, Thieke C, Küfer K.-H., Trinkaus H (2002) New Holmes T, Mackie TR, Simpkin D, Reckwerdt P (1991) A uni-
approaches in intensity-modulated radiotherapy: a new fied approach to the optimization of brachytherapy and
optimization paradigm. In: Kogelnik HD, Lukas P, Sedl- external beam dosimetry. Int J Radiat Oncol Biol Phys
mayer F (eds) Progress in radio-oncology, vol 7. Monduzzi 20:859–873
Editore, Bologna, pp 251–258 Hope CS, Laurie J, Orr JS, Halnan KE (1967) Optimization of
Brahme A (1984) Dosimetric precision requirements in radia- X-ray treatment planning by computer judgement. Phys
tion therapy. Acta Radiol Oncol 23:379–391 Med Biol 12:531–542
Brahme A (1993) Optimization of radiation therapy and the Hunt MA, Hsiung CY, Spirou SV, Chui CS, Amols HI, Ling CC
development of multileaf collimation. Int J Radiat Oncol (2002) Evaluation of concave dose distributions created
Biol Phys 25:373–375 using an inverse planning system. Int J Radiat Oncol Biol
Brahme A (1994) Optimization of radiation therapy. Int J Phys 54:953–962
Radiat Oncol Biol Phys 28:785–787 ICRU (1993) Prescribing, recording, and reporting photon
Brahme A, Roos JE, Lax I (1982) Solution of an integral equa- beam therapy. International Commission on Radiation
tion in rotation therapy. Phys Med Biol 27:1221–1229 Units and Measurements, report 50. Bethesda, Maryland
Carol MP, Nash RV, Campbell RC, Huber R, Sternick E (1997) IMRT (2001) Intensity-modulated radiotherapy: current
The development of a clinically intuitive approach to status and issues of interest. Int J Radiat Oncol Biol Phys
inverse treatment planning: partial volume prescription 51:880–914
and area cost function. Twelfth International Conference Keller-Reichenbecher M-A, Bortfeld T, Levegrün S, Stein J,
on the Use of Computers in Radiation Therapy. Salt Lake Preiser K, Schlegel W (1998) Intensity modulation with the
City, Utah step and shoot technique using a commercial MLC: a plan-
Censor Y (2003) Mathematical optimization for the inverse ning study. Int J Radiat Oncol Biol Phys 45:1315–1324
problem of intensity-modulated radiation therapy. Inten- Kessen A, Grosser K-H, Bortfeld T (2000) Simplification of
sity-modulated radiation therapy: the state of the art. In: IMRT intensity maps by means of 1-D and 2-D median-
Palta JR, Mackie TR (eds) Medical Physics Publishing, pp filtering during the iterative calculation. Thirteenth Inter-
25–49 national Conference on the Use of Computers in Radiation
Censor Y, Altschuler MD, Powlis WD (1988) A computational therapy. Springer, Berlin Heidelberg New York
solution of the inverse problem in radiation-therapy treat- Küfer K-H, Hamacher HW, Bortfeld TR (2000) A multicriteria
ment planning. Appl Math Comput 25:57–87 optimization approach for inverse radiotherapy planning.
Cho PS, Marks RJ II (2000) Hardware-sensitive optimiza- Proc XIIIth ICCR, Heidelberg 2000. Springer, Berlin Heidel-
tion for intensity modulated radiotherapy. Phys Med Biol berg New York, pp 26–29
45:429–440 Küfer K-H, Scherrer A, Monz M, Alonso F, Trinkaus H, Bort-
Cho PS, Lee S, Marks RJ II, Oh S, Sutlief SG, Phillips MH (1998) feld T, Thieke C (2003) Intensity modulated radiotherapy: a
Optimization of intensity modulated beams with volume large scale multi-criteria programming problem. OR Spec-
constraints using two methods: cost function minimization trum 25:223–249
and projections onto convex sets. Med Phys 25:435–443 Kwa SL, Lebesque JV, Theuws JC, Marks LB, Munley MT, Bentel
Chui CS, Chan MF, Yorke E, Spirou S, Ling CC (2001) Delivery G, Oetzel D, Spahn U, Graham MV, Drzymala RE, Purdy
of intensity-modulated radiation therapy with a conven- JA, Lichter AS, Martel MK, Ten Haken RK (1998) Radiation
tional multileaf collimator: comparison of dynamic and pneumonitis as a function of mean lung dose: an analysis
segmental methods. Med Phys 28:2441–2449 of pooled data of 540 patients. Int J Radiat Oncol Biol Phys
Cormack AM, Cormack RA (1987) A problem in rotation ther- 42:1–9
apy with X-rays: dose distributions with an axisof symme- Langer M, Leong J (1989) Optimization of beam weights under
try. Int J Radiat Oncol Biol Phys 13:1921–1925 dose-volume restrictions. Int J Radiat Oncol Biol Phys
Deasy JO (1997) Multiple local minima in radiotherapy opti- 13:1255–1259
220 T. Bortfeld and C. Thieke

Langer M, Brown R, Urie M, Leong J, Stracher M, Shapiro J Redpath AT, Vickery BL, Wright DH (1975) A set of fortran
(1990) Large scale optimization of beam weights under subroutines for optimizing radiotherapy plans. Comput
dose-volume restrictions. Int J Radiat Oncol Biol Phys Programs Biomed 5:158–164
18:887–893 Romeijn HE, Ahuja RK, Dempsey JF, Kumar A, Li JG (2003) A
Langer M, Morrill S, Brown R, Lee O, Lane R (1996) A com- novel linear programming approach to fluence map opti-
parison of mixed integer programming and fast simulated mization for intensity modulated radiation therapy treat-
annealing for optimizing beam weights in radiation ther- ment planning. Phys Med Biol 48:3521–3542
apy. Med Phys 23:957–964 Shepard DM, Earl MA, Li XA, Naqvi S, C Yu (2002) Direct aper-
Langer M, Lee EK, Deasy JO, Rardin RL, Deye JA (2003) Opera- ture optimization: a turn key solution for step-and-shoot
tions research applied to radiotherapy, an NCI-NSF-spon- IMRT. Med Phys 29:1007–1018
sored workshop 7–9 February, 2002. Int J Radiat Oncol Biol Shepard DM, Ferris MC, Olivera GH, Mackie TR (1999) Opti-
Phys 57:762–768 mizing the delivery of radiation therapy to cancer patients.
Lauve A, Morris M, Schmidt-Ullrich R, Wu Q, Mohan R, SIAM Rev 41:721–733
Abayomi O, Buck D, Holdford D, Dawson K, Dinardo L, Söderström S, Brahme A (1995) Which is the most suitable
Reiter E (2004) Simultaneous integrated boost intensity- number of photon beam portals in coplanar radiation
modulated radiotherapy for locally advanced head-and- therapy. Int J Radiat Oncol Biol Phys 33:151–159
neck squamous cell carcinomas. Part II: clinical results. Int Söderström S, Brahme A (1996) Small is beautiful: and often
J Radiat Oncol Biol Phys 60:374–387 enough. Int J Radiat Oncol Biol Phys 34:757–758
Levegrun S, Jackson A, Zelefsky MJ, Skwarchuk MW, Venkatra- Spirou SV, Chui C-S (1998) A gradient inverse planning algo-
man ES, Schlegel W, Fuks Z, Leibel SA, Ling CC (2001) Fit- rithm with dose-volume constraints. Med Phys 25:321–
ting tumor control probability models to biopsy outcome 333
after three-dimensional conformal radiation therapy of Starkschall G, Pollack A, Stevens CW (2000) Using a dose-
prostate cancer: pitfalls in deducing radiobiologic param- volume feasibility search algorithm for radiation treatment
eters for tumors from clinical data. Int J Radiat Oncol Biol planning. Thirteenth International Conference on the Use
Phys 51:1064–1080 of Computers in Radiation therapy. Springer, Berlin Hei-
Llacer J, Deasy JO, Portfeld TR, Solberg TD, Promberger C delberg New York
(2003) Absence of multiple local minima effects in inten- Stein J, Mohan R, Wang XH, Bortfeld T, Wu Q, Preiser K, Ling
sity modulated optimization with dose-volume constraints. CC, Schlegel W (1997) Number and orientation of beams
Phys Med Biol 48:183–210 in intensity-modulated radiation treatments. Med Phys
Mackie R, Deasy J, Holmes T, Fowler J (1994) Optimization of 24:149–160
radiation therapy and the development of multileaf col- Thieke C (2003) Multicriteria optimization in inverse radio-
limation. Int J Radiat Oncol Biol Phys 28:784–785 therapy planning. Doctor of Natural Sciences dissertation.
McDonald SC, Rubin P (1977) Optimization of external beam Combined Faculties for the Natural Sciences and for Math-
radiation therapy. Int J Radiat Oncol Biol Phys 2:307–317 ematics, Ruperto-Carola University of Heidelberg
Mohan R, Ling CC (1995) When becometh less more? Int J Thieke C, Bortfeld T, Niemierko A (2002) Direct Consider-
Radiat Oncol Biol Phys 33:235–237 ation of EUD Constraints in IMRT Optimization. Med Phys
Mohan R, Wang XH (1996) Physical vs biological objectives for 29:1283
treatment plan optimization. Radiother Oncol 40:186–187 Webb S (1989) Optimisation of conformal radiotherapy
Mohan R, Wu Q, Manning M, Schmidt-Ullrich R (2000) dose distributions by simulated annealing. Phys Med Biol
Radiobiological considerations in the design of fraction- 34:1349–1370
ation strategies for intensity-modulated radiation therapy Webb S (1992) Optimization by simulated annealing of three-
of head and neck cancers. Int J Radiat Oncol Biol Phys dimensional, conformaltreatment planning for radiation
46:619–630 fields defined by a multileaf collimator:II. Inclusion of the
Natterer F (1986) The mathematics of computerized tomogra- two-dimensional modulation of the X-ray intensity. Phys
phy. Teubner, Stuttgart Med Biol 37:1689–1704
Niemierko A (1992) Random search algorithm (RONSC) for Webb S, Convery DJ, Evans PM (1998) Inverse planning with
the optimization of radiation therapy with both physi- constraints to generate smoothed intensity-modulated
cal and biological endpoints and constraints. Int J Radiat beams. Phys Med Biol 43:2785–2794
Oncol Biol Phys 23:89–98 Wu Q, Mohan R (2002) Multiple local minima in IMRT optimiza-
Niemierko A (1997) Reporting and analyzing dose distribu- tion based on dose-volume criteria. Med Phys 29:1514–1527
tions: a concept of equivalent uniform dose. Med Phys Wu Q, Mohan R, Niemierko A, Schmidt-Ullrich R (2002) Opti-
24:103–110 mization of intensity-modulated radiotherapy plans based
Niemierko A (1999) A generalized concept of equivalent uni- on the equivalent uniform dose. Int J Radiat Oncol Biol
form dose (EUD). Med Phys 26:1100 Phys 52:224–235
Pugachev A, Li JG, Boyer AL, Hancock SL, Le QT, Donaldson Yajnik S, Rosenzweig KE, Mychalczak B, Krug L, Flores R, Hong
SS, Xing L (2001) Role of beam orientation optimization in L, Rusch VW (2003) Hemithoracic radiation after extra-
intensity-modulated radiation therapy. Int J Radiat Oncol pleural pneumonectomy for malignant pleural mesothe-
Biol Phys 50:551–560 lioma. Int J Radiat Oncol Biol Phys 56:1319–1326