Professional Documents
Culture Documents
net/publication/357898855
Article in IEEE transactions on neural networks / a publication of the IEEE Neural Networks Council · January 2022
DOI: 10.1109/TEVC.2022.3144675
CITATION READS
1 308
8 authors, including:
Some of the authors of this publication are also working on these related projects:
Real-time control and optimisation in a Free Piston Engine Generator (FPEG) View project
All content following this page was uploaded by Ye Tian on 18 January 2022.
Multileaf
collimator
(MLC) Radiation
beam
Multileaf
collimator
(MLC)
Fluence map Aperture 1 and metaheuristics, just to name a few [17], [18]. So far,
× ××××××× × × × ×××× × direct aperture optimization has been a mature technique
× ×× ××× × × × ×××× ×
Grid × × × × × ×××× × and adopted in commercial systems [19].
× × × × × × ×× × Nevertheless, IMRT planning in clinical practice still
× ×× ××× × × × ×× ×
× ×× ×× × × × ×××× × requires much human expertise and takes hours or
Therapeutic
field
× ×× × × × × ×××× × even days for one clinical case [20]. Once the doctor
× ××××××× × × × ×××× × decides the treatment protocol, the medical physicist first
builds an optimization model and obtains its optimal
Aperture 3 Aperture 2 solution (i.e., treatment plan) via a planning method,
× ××××××× × × ×× ××× × then discusses with the doctor the quality of the plan
× ×× ××× × × ×× ××× ×
× × × × × ×× × according to the dose-volume histogram. If the plan
× × × × × × × × is not clinically acceptable, the medical physicist must
× ×× ××× × × × ×× × build a new model and re-run the planning method.
× ×× ×× × × × × ×
× ×× × × × ×× ××× × This procedure will probably be repeated many times
× ××××××× × × ×× ××× × as shown in Fig. 1, which is relatively time-consuming
since only one plan is obtained at a time.
Fig. 2. An illustrative example of MLC sequencing.
On the other hand, IMRT planning is naturally a
multi-objective optimization problem since the radiation
deteriorate the effectiveness since complex fluence maps damages both tumor and healthy cells [20]. To ensure the
cannot be accurately divided into few apertures [12]. By minimum doses of tumor cells and limit the maximum
contrast, some other IMRT planning methods directly doses of healthy cells, multiple conflicting objectives
optimize the shapes and beam weight of each aperture, should be optimized simultaneously. While most opti-
which is known as direct aperture optimization [13]. mizers adopted in existing planning methods are only
While the optimization problem contains a large number for single-objective optimization, the multiple objectives
of discrete variables (i.e., MLC leaf positions) and con- have to be aggregated by different weights. To determine
tinuous variables (i.e., beam weights), it is ineffective to the optimal weights for the objectives, an optimizer
directly optimize all of them by using the optimizers at should be employed many times to obtain the optimal
hand. To address this dilemma, some methods limit the solution for each setting of weights [21], [22]. In short,
shapes of apertures to certain ones [12], some methods the optimization process is also inefficient since only one
optimize the MLC leaf positions by random search [13] solution is obtained at a time.
or enumeration [14], some methods iteratively add new Therefore, the efficiency of IMRT planning can be
apertures and optimize only the beam weights [15], [16], considerably improved by generating multiple diverse
and some methods optimize the beam weights and MLC plans at a time, where ample alternatives are provided
leaf positions by mathematical programming methods to alleviate the uncertainty in both modeling and weight-
IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. , NO. , MONTH YEAR 3
ing. Although multi-objective evolutionary algorithms of voxel v in terms of the i-th grid in the therapeutic
(MOEAs) have been applied to a variety of applications field of a beam. Besides, the intensity yi of the grid is
with large numbers of decision variables [23], [24], few determined by the beam weights and MLC leaf positions
of them have been adopted in IMRT planning. This of all the apertures related to a beam:
is because existing MOEAs mainly focus on continu- ∑
ous optimization, whereas little attention has been paid yi = wj pij , (2)
to large-scale combinatorial optimization problems like j
IMRT planning. The discrete variables in IMRT planning where wj denotes the beam weight of the j-th aperture,
pose stiff challenges for existing MOEAs to find well- pij = 0 indicates that the i-th grid is shielded by the
converged and diversified solutions [25]. Hence, this MLC in the j-th aperture, and pij = 1 otherwise.
paper aims to improve the efficiency of IMRT planning Based on the doses of voxels in PTV and OAR, a
by customizing an MOEA, which can obtain a set of solu- variety of objective functions have been formulated to
tions that present diverse trade-offs between the multiple describe the treatment protocol [7], [14], [26]–[29]. This
conflicting objectives in IMRT planning. Specifically, this paper adopts the following four types of objective func-
paper includes the following three contributions: tions to be minimized, which are easy to calculate and
1) The IMRT planning is formulated into a large-scale contain just a few parameters that should be predefined
combinatorial many-objective optimization prob- by medical physicists.
lem. The problem definition is applicable to dif- Minimum Dose: This objective function is to ensure the
ferent clinical cases, where medical physicists only minimum dose of each voxel, which is defined as:
need to vary the parameters in the problems. 1 ∑
2) A coevolutionary algorithm is proposed for solving M inDose(x, V, dmin ) = (max{0, dmin − dv })2 ,
|V |
the above problem, termed MOEA/IMRT. In or- v∈V
(3)
der to handle the high-dimensional discrete search
where x is a solution determining the values of wj and
space, the proposed algorithm evolves one popula-
pij in (2). Besides, V denotes a region of PTV or OAR and
tion with fine encoding and evolves another popu-
dmin denotes the predefined minimum dose. Obviously,
lation with rough encoding, where the balance be-
the value of M inDose is zero only when all doses are
tween exploitation and exploration can be achieved
greater than or equal to dmin .
by the cooperation of the two populations. Besides,
Maximum Dose: This objective function is to limit the
two local search strategies are suggested to repair
maximum dose of each voxel, which is defined as:
invalid solutions and accelerate the convergence.
3) The effectiveness of the proposed MOEA/IMRT 1 ∑
M axDose(x, V, dmax ) = (max{0, dv − dmax })2 ,
is tested on six clinical cases of different can- |V |
v∈V
cers. According to the experimental results, the (4)
proposed algorithm has better convergence and where dmax denotes the predefined maximum dose.
diversity performance than state-of-the-art MOEAs Similarly, the value of M axDose is zero only when all
and IMRT planning methods. doses are less than or equal to dmax .
The rest of this paper is organized as follows. Section II Prescribed Dose: This objective function is to make the
gives the proposed optimization model, and reviews ex- doses of all voxels as close to the prescribed dose as
isting multi-objective IMRT planning methods and large- possible, which is defined as:
scale MOEAs. Section III describes the proposed MOEA 1 ∑
for IMRT planning. Section IV analyzes the experimental P reDose(x, V, dpre ) = (dv − dpre )2 , (5)
|V |
results. Lastly, Section V draws the conclusion. v∈V
case. While existing work aggregates all the objectives single-objective optimization problems [34], [35], and
into one objective [7], [14], [26], [29] or two objectives some work searches for the optimal weights via particle
[27], [30], this paper builds the following large-scale swarm optimization [28].
combinatorial many-objective optimization problem by Although the above methods involve the concepts of
aggregating only similar objectives: “multi-objective optimization” and “Pareto optimality”,
they can only find a single solution at a time. On the
Minimize f (x) = (f1 (x), . . . , f5 (x))
∑ contrary, a few MOEAs have also been adopted in IMRT
f1 (x) = M inDose(x, V, dmin ) planning to find multiple solutions simultaneously. In
V ∈P T V [30], a multi-objective particle swarm optimization algo-
∑
f2 (x) = M axDose(x, V, dmax ) rithm was employed for the IMRT planning with two
V ∈P T V objectives. In [27], a classical evolutionary algorithm
∑ NSGA-II [36] was directly applied to the IMRT planning
+ DV C(x, V, ddvc , r)
V ∈P T V , (7) with multiple objectives and constraints. In [25], the
∑ performance of NSGA-II was enhanced by a tailored
f3 (x) = M axDose(x, V, dmax ) initialization strategy, where the initial solutions are
V ∈OAR
∑ generated by a gradient based algorithm.
f4 (x) = DV C(x, V, ddvc , r) As a consequence, existing IMRT planning methods
V ∈OAR can hardly strike a balance between efficiency and effec-
∑
f5 (x) = P reDose(x, V, dpre ) tiveness. On the one hand, the methods based on single-
V ∈P T V objective optimization should be performed many times
where x is a solution containing all the beam weights to generate a set of diverse plans, which is very inef-
and MLC leaf positions, P T V is the set of all regions ficient in clinical practice. On the other hand, although
in PTV, and OAR is the set of all regions in OAR. The a few methods adopt classical MOEAs to find multiple
first objective f1 is the core objective, which defines the plans in a single run, their effectiveness has not been
minimum dose of PTV and ensures the effectiveness of fully verified via comparative studies. More importantly,
radiotherapy. The second objective f2 contains the maxi- no state-of-the-art MOEA has been adopted and no
mum dose and dose-volume constraint of PTV, which search strategy has been tailored for IMRT planning.
are aggregated into a single objective since they are To take full advantage of evolutionary multi-objective
positively correlated with each other. The third objective optimization in IMRT planning, this paper proposes a
f3 defines the maximum dose of OAR and the fourth new MOEA with several customized strategies, includ-
objective f4 defines the dose-volume constraint of OAR, ing a coevolutionary framework and two local search
and the last objective f5 defines the prescribed dose of strategies. In Section IV, some state-of-the-art MOEAs
PTV, which makes the doses of voxels uniform so that will be tested in the experiments, which are verified to
the plan is more practical. be underperformed by the proposed algorithm.
Given a treatment protocol, the medical physicist
should first define the parameters dmin , dmax , dpre , ddvc , C. Existing Large-Scale MOEAs
and r for each region V , then optimize the above prob-
The proposed optimization problem is characterized
lem via an IMRT planning method or other optimizer.
by many objectives, large number of decision variables,
In the next two subsections, some state-of-the-art IMRT
and discrete search space. After decades of development,
planning methods and MOEAs are reviewed.
MOEAs are able to handle many objectives via enhanced
dominance relations [37], diversity based selection [38],
B. Existing Methods for Multi-Objective IMRT Planning objective decomposition [39], or performance indicators
In spite of the multiple conflicting objectives in IMRT [40]. However, due to the black-box search paradigms of
planning, existing planning methods are mostly based MOEAs, most of them encounter difficulties in handling
on single-objective optimization algorithms due to their a large number of decision variables. To tackle the curse
fast convergence speed. While most planning methods of dimensionality of decision space, some MOEAs based
aggregate the multiple objectives into a single one via the on decision variable grouping, decision space reduction,
weights set by hand, some other methods focus on the or novel search strategies have been proposed in recent
adaptive detection of the optimal weights. For example, years [24]. The decision variable grouping based MOEAs
some work only searches for the optimal solutions for adopt the divide-and-conquer strategy that divides the
each single objective [31], some work samples several decision variables via different grouping strategies, such
sets of weights and obtains the optimal solution for each as random grouping [41], [42], differential grouping
of the aggregated objectives [21], some work aggregates [43], and variable clustering [44]. The decision space
the objectives via principal component analysis [32], reduction based MOEAs aim to directly reduce the num-
some work navigates over the Pareto front by interpola- ber of decision variables via different dimensionality
tion between pre-found solutions [22], [33], some work reduction methods, such as problem transformation [45],
converts the original problem into several constrained problem reformulation [46], and unsupervised neural
IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. , NO. , MONTH YEAR 5
networks [47]. The novel search strategy based MOEAs Fine encoding
Rough encoding
x2 = 3 x4 = 5
suggest new reproduction operators to accelerate the
x1 = 4 × × × × × × × × x2 = 4 × × × × × × × ×
convergence, such as competitive swarm optimizer [48],
x3 = 3 × × × × × × x4 = 3 × × × × × × × ×
self-evaluation evolution [49], and generative adversarial x5 = 1 × × x 6 = 1 x1 = 3 × × × × ×
networks [50]. These MOEAs have shown obvious supe- x7 = 2 × × × x8 = 1 × × × × ×
riority over classical ones in solving a variety of large- x9 = 3 × × × × × × x10 = 3 × × × × ×
scale multi-objective optimization problems [51]. x11 = 3 × × × × × x12 = 2 × × × × ×
Nevertheless, they cannot be directly applied to IMRT x13 = 3 × × × × x14 = 1 x3 = 3 × × × × ×
planning since all the three categories of search strate- x15 = 4 × × × × × × × × x16 = 4 × × × × × × × ×
gies are tailored for large-scale continuous optimization, Beam weight: x17 = 0.8 Beam weight: x5 = 0.8
10 0 10 000 evaluations
100 000 evaluations
Input: N (population size), K (number of exchanged
200 000 evaluations solutions)
Output: P1 (final population)
10-2
1 P1 ← Randomly generate N solutions of fine encoding;
2 P2 ← Randomly generate N solutions of rough encoding;
f2
3 [P1 , P2 ] ← LocalSearchA(P1 , P2 );
-4
4 while termination criterion is not fulfilled do
10
5 P1′ ← Select N parents from P1 via binary tournament
selection;
6 O1 ← Generate offsprings based on P1′ ;
10-6
10
-4
10
-3
10
-2
10
-1 7 P2′ ← Select N parents from P2 via binary tournament
f1 selection;
MOEA/IMRT with fine encoding and rough encoding 8 O2 ← Generate offsprings based on P2′ ;
100 evaluations 9 [O1 , O2 ] ← LocalSearchA(O1 , O2 );
100
10 000 evaluations 10 [O1′ , O2′ ] ← Exchange(O1 , O2 , K);
P1 ← P1 ∪ O1 ∪ O1′ ;
100 000 evaluations
200 000 evaluations 11
12 P2 ← P2 ∪ O2 ∪ O2′ ;
-2 13 P1 ← Truncate P1 via environmental selection;
10
14 P2 ← Truncate P2 via environmental selection;
f2
15 P1 ← LocalSearchB(P1 );
10
-4 16 return P1 ;
10-6
-4 -3 -2 -1
having N solutions with fine encoding and a popula-
10 10 10 10
f1 tion P2 having N solutions with rough encoding are
randomly initialized, and the first local search strat-
Fig. 4. Comparison between the first two objectives of the populations egy LocalSearchA() is conducted on the two popula-
on the Esophagus dataset, which are collected at different generations tions to repair invalid solutions. At each generation of
of the proposed algorithm with fine encoding and both the two
encoding schemes. MOEA/IMRT, two mating pools P1′ and P2′ are first
constructed by selecting N solutions from P1 and P2 ,
1 respectively. Then, the mating pools are used to gener-
ate two offspring population O1 and O2 via simulated
0.95
binary crossover and polynomial mutation, and the off-
0.9
springs are also repaired by the first local search strat-
egy. Afterwards, in the operation of offspring exchange
0.85 Fine encoding Exchange(), some offsprings with fine encoding in O1
Rough encoding
Rough + fine encoding are converted into solutions with rough encoding and
stored in O2′ , and some offsprings with rough encoding
0.8
1 2 3 4 5
Number of evaluations 105
in O2 are converted into solutions with fine encoding and
Fig. 5. Convergence profiles of mean HV obtained by MOEA/IMRT stored in O1′ . Lastly, the population P1 is combined with
with a single and two encoding schemes on the Esophagus dataset. O1 ∪ O1′ and truncated by environmental selection, and
the population P2 is combined with O2 ∪ O2′ and also
runs. According to the figure, the algorithm with rough truncated by environmental selection. Besides, the sec-
encoding converges much faster than the algorithm with ond local search strategy LocalSearchB() is conducted
fine encoding since the rough encoding has a better on P1 for further enhancement of the convergence.
exploration ability. However, the algorithm with rough
encoding converges prematurely and finally obtains the For the selection strategies in the proposed
worst HV value, since it cannot represent irregular aper- MOEA/IMRT, the mating selection and environmental
ture shapes to further reduce the objective values. By selection in SPEA2 [55] with shift-based density
contrast, the algorithm with fine encoding obtains better estimation [38] are adopted for simplicity, which are
result due to its good exploitation ability. Therefore, by flexible to be embedded in different MOEAs and
adopting both the rough encoding and fine encoding, have demonstrated high effectiveness in handling many
the proposed MOEA/IMRT has good exploration ability objectives. For the exchange of offsprings, MOEA/IMRT
and good exploitation ability at the same time, leading converts between the solutions with fine encoding and
to the fastest convergence speed and the best HV value. rough encoding by the steps given in Algorithm 2.
Firstly, some solutions are separately selected from the
two offspring populations to be exchanged, which is
B. General Procedure of MOEA/IMRT achieved by the same strategy as the environmental
The procedure of the proposed MOEA/IMRT is pre- selection of MOEA/IMRT. Note that only K solutions
sented in Algorithm 1. To begin with, a population P1 rather than all the solutions are selected here, since
IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. , NO. , MONTH YEAR 7
TABLE I
D ETAILED I NFORMATION A BOUT S IX D ATASETS U SED IN THE E XPERIMENTS
1 1 1
Dose-volume constraint
Relative volume (%)
(a) Maximum and minimum doses of PTV. (b) Maximum dose of OAR. (Ideal case: the (c) Dose-volume constraint of OAR. (Ideal
(Ideal case: the vertical part of the curve is curve is to the left of the vertical line) case: the curve is to the lower left of the
between the two vertical lines) two points)
Fig. 8. Illustration of different objectives in a dose volume histogram (DVH). A point (x, y) on the curve indicates that the ratio of doses greater
than x in the PTV or OAR is y.
distribution index is set to 20; CCGDE3 uses differential are reported. Besides, the Friedman test with Bonferroni
evolution and polynomial mutation, where both the correction [63] at a significance level of 0.05 is adopted
learning rate F and the crossover rate CR are set to for statistical test, where ‘+’, ‘−’, and ‘≈’ indicate that
0.5; WOF-SMPSO uses particle swarm optimization and the HV values obtained by an algorithm are significantly
polynomial mutation; LMOCSO uses competitive swarm better, significantly worse, and statistically similar to the
optimizer and polynomial mutation; Cao’s algorithm proposed MOEA/IMRT, respectively.
and Wang’s algorithm uses conjugate gradient method. 6) Dose-Volume Histogram (DVH): The DVH is also
3) Problem-Dependent Strategies: In spite of the differ- used to visually assess the quality of plans, which is the
ent numbers of objective functions in the datasets, they main assessment criterion used by medical physicists. A
are aggregated into five objectives by (7) and optimized DVH depicts the cumulative distribution of the doses in
by all the compared MOEAs. As for Cotrutz’s algorithm, PTV and OAR, where a point (x, y) on a curve indicates
Cao’s algorithm, and Wang’s algorithm, they aggregate that the ratio of doses greater than x in the PTV or
all the objective functions into a single one. Except OAR is y. Fig. 8 illustrates a cumulative distribution
for the proposed MOEA/IMRT adopting two encoding curve together with different objectives in a DVH, where
schemes, all the other compared algorithms represent the vertical part of the curve should be between the
solutions with fine encoding. To ensure the validity of two vertical lines of maximum and minimum doses in
the solutions, all the compared MOEAs perform the Fig. 8(a), the curve should be to the left of the vertical line
LocalSearchA() described in Algorithm 3 once an off- of maximum dose in Fig. 8(b), and the curve should be
spring is generated. Since Cotrutz’s algorithm, Cao’s to the lower left of the points of dose-volume constraint
algorithm, and Wang’s algorithm can only generate one in Fig. 8(c). If the curve does not meet any of the above
solution at a time, a set of uniformly distributed weight conditions, it implies that the plan is not optimal on the
vectors are sampled for the aggregation of objective corresponding objective.
functions, where one solution is obtained by optimizing
the single objective aggregated by each weight vector
and all the solutions constitute the final population. B. Comparisons Between MOEA/IMRT and State-of-the-Art
4) Population Size and Termination Criterion: For fair- Table II lists the HV results obtained by the nine
ness, the population size of all the compared MOEAs is compared algorithms on six datasets. It is clear that the
set to 100 and the maximum number of function evalu- proposed MOEA/IMRT obtains the best performance on
ations of all the compared MOEAs is set to 500000. Be- all the datasets, whose HV values are always signifi-
sides, Cotrutz’s algorithm, Cao’s algorithm, and Wang’s cantly better than all the other compared algorithms.
algorithm are performed 100 times and the maximum For a visual comparison, Fig. 9 compares the objective
number of function evaluations is set to 5000 each time. values of the populations with median HV obtained
5) Performance Indicator: Since the true Pareto fronts of by the proposed MOEA/IMRT and each of the other
real-world problems are unknown, the HV indicator [54] algorithms on the Esophagus dataset. It can be observed
is adopted to assess the performance of the compared that the population obtained by MOEA/IMRT has bet-
algorithms. To determine the reference point for HV ter convergence than those obtained by Two Arch2,
calculation, the non-dominated solutions obtained by CCGDE3, WOF-SMPSO, Cotrutz’s algorithm, Cao’s al-
all the algorithms in all the runs are collected, and the gorithm, and Wang’s algorithm, and has better diver-
reference point consists of the maximum objective values sity than those obtained by RVEA and LMOCSO. For
of the non-dominated solutions on all dimensions. To Two Arch2 and RVEA tailored for many-objective opti-
eliminate the influence of randomness, each algorithm mization, the convergence performance of RVEA is bet-
is performed on each dataset for 30 independent runs, ter than Two Arch2, which means that the coevolution-
where the mean and standard deviation of the HV values ary framework of Two Arch2 is ineffective for the many
IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. , NO. , MONTH YEAR 11
TABLE II
HV R ESULTS O BTAINED BY N INE A LGORITHMS ON S IX D ATASETS OF IMRT P LANNING . T HE B EST R ESULT IN E ACH R OW IS H IGHLIGHTED
Problem Two Arch2 RVEA CCGDE3 WOF-SMPSO LMOCSO Cotrutz’s algorithm Cao’s algorithm Wang’s algorithm MOEA/IMRT
Esophagus 9.6064e-1 (8.85e-3) − 9.6384e-1 (3.75e-3) − 0.0000e+0 (0.00e+0) − 9.5540e-1 (9.52e-3) − 9.5822e-1 (1.31e-2) − 3.4490e-2 (3.22e-2) − 0.0000e+0 (0.00e+0) − 9.1906e-1 (1.00e-2) − 9.8045e-1 (2.61e-3)
Renal 9.9330e-1 (1.72e-3) − 9.8747e-1 (1.95e-3) − 0.0000e+0 (0.00e+0) − 9.7799e-1 (8.29e-3) − 9.8099e-1 (1.46e-3) − 9.0421e-1 (2.20e-2) − 0.0000e+0 (0.00e+0) − 9.8749e-1 (6.67e-4) − 9.9768e-1 (3.60e-4)
RightLung 8.8601e-1 (1.13e-2) − 7.5203e-1 (2.43e-1) − 2.4031e-2 (5.33e-2) − 7.2896e-1 (2.64e-1) − 8.4637e-1 (1.72e-2) − 4.8269e-1 (2.72e-1) − 0.0000e+0 (0.00e+0) − 7.2737e-1 (4.29e-3) − 9.0784e-1 (6.61e-3)
Sneezer 9.8459e-1 (1.41e-3) − 9.7737e-1 (1.89e-3) − 0.0000e+0 (0.00e+0) − 8.9300e-1 (1.27e-1) − 9.8122e-1 (1.06e-3) − 9.6587e-1 (5.76e-3) − 1.0271e-1 (1.41e-2) − 9.7198e-1 (2.43e-3) − 9.8680e-1 (1.10e-3)
Throat 8.4575e-1 (1.77e-2) − 5.9789e-1 (3.27e-2) − 0.0000e+0 (0.00e+0) − 5.8518e-1 (1.18e-1) − 2.2960e-1 (2.14e-1) − 0.0000e+0 (0.00e+0) − 0.0000e+0 (0.00e+0) − 0.0000e+0 (0.00e+0) − 8.5912e-1 (9.62e-3)
Cervical 9.6661e-1 (3.46e-3) − 9.6897e-1 (5.40e-3) − 1.1003e-1 (5.95e-2) − 9.2699e-1 (3.31e-2) − 9.7579e-1 (1.67e-3) − 8.4165e-1 (1.33e-2) − 0.0000e+0 (0.00e+0) − 9.4716e-1 (3.58e-3) − 9.8394e-1 (1.61e-3)
Fig. 9. Parallel coordinates of the objective values of the populations with median HV obtained by the compared algorithms on the Esophagus
dataset. MOEA/IMRT has better convergence performance than Two Arch2, CCGDE3, WOF-SMPSO, Cotrutz’s algorithm, Cao’s algorithm, and
Wang’s algorithm, and better diversity performance than RVEA and LMOCSO.
objectives of IMRT; besides, the reference vector based and OAR2, it can be found that all the algorithms can
selection strategy of RVEA is imperfect since it has worse almost meet the requirement of dose-volume constraints.
diversity performance than MOEA/IMRT. For CCGDE3, For the curves of OAR3, all the algorithms can mostly
WOF-SMPSO, and LMOCSO tailored for large-scale op- meet the requirement of maximum dose. As for the most
timization, the novel search strategy of LMOCSO is important curves of PTV, only RVEA, LMOCSO, and
more effective than the coevolutionary framework and the proposed MOEA/IMRT can meet the requirement of
variable grouping strategy of CCGDE3 and the two both maximum and minimum doses. The band of PTV
encoding schemes and decision space reduction strategy obtained by MOEA/IMRT between the vertical lines of
of WOF-SMPSO, since the former exhibits much better maximum and minimum doses is obviously wider than
convergence performance; besides, the diversity perfor- those obtained by RVEA and LMOCSO, which indi-
mance of LMOCSO is also worse than MOEA/IMRT. cates a better diversity performance of MOEA/IMRT.
For Cotrutz’s algorithm, Cao’s algorithm, and Wang’s In addition, Figs. S1–S5 in the Supplementary Materials
algorithm tailored for IMRT planning, they have worse show the DVH of the populations obtained on the other
convergence performance than MOEA/IMRT since the five datasets, where MOEA/IMRT can also obtain a set
aperture shapes cannot be well optimized by using ge- of well-converged and diversified solutions on all the
netic algorithm (without local search), MLC sequencing, datasets.
and enumeration, respectively.
Moreover, Fig. 10 shows the DVH of the populations
with median HV obtained by the compared algorithms C. Verification of the Core Components in MOEA/IMRT
on the Esophagus dataset. To better reflect the popula- Firstly, MOEA/IMRT is compared to a variant without
tion diversity, the range of all the curves in a population the local search strategy LocalSearchB() to verify the
rather than a single curve is plotted, where a curve effectiveness of the local search strategy for enhancing
corresponds to the cumulative distribution of the doses the population with fine encoding. According to the
in PTV or OAR of one solution. For the curves of OAR1 convergence profiles of mean HV shown in Fig. 11,
IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. , NO. , MONTH YEAR 12
Fig. 10. DVH of the populations with median HV obtained by the compared algorithms on the Esophagus dataset. The meanings of the elements
in the figures are referred to Fig. 8.
0.95
of the solutions obtained by MOEA/IMRT with and
without local search, where the curves corresponding
0.9
to the solution with the smallest sum of all objectives
0.85 in the population with median HV are shown. It can
0.8
be found that the two algorithms obtain the same good
With LocalSearch
Without LocalSearch curves of OAR1–OAR3, while the algorithm without
1 2 3 4 5 local search obtains a curve of PTV obviously exceeding
Number of evaluations 105
1 the maximum dose. Therefore, performing local search
With local search (PTV) Without local search (OAR2)
0.9 With local search (OAR1)
With local search (OAR2)
Without local search (OAR3)
Maximum/Minimum dose (PTV)
on the population with fine encoding can improve the
0.8 With local search (OAR3)
Without local search (PTV)
Dose-volume cons (OAR1)
Dose-volume cons (OAR2)
performance of MOEA/IMRT indeed.
0.7 Without local search (OAR1) Maximum dose (OAR3)
Then, MOEA/IMRT is compared to three variants
Relative volume (%)
0.6
0.5
with different values of K, i.e., the number of exchanged
0.4
solutions in Algorithm 2. According to the results shown
0.3
in Fig. 12, the algorithms with different values of K
0.2 gain very similar HV values, which means that the
0.1 performance of MOEA/IMRT is not sensitive to the
0
0 1000 2000 3000 4000 5000 6000 7000
setting of K. Besides, K = 10 leads to a relatively
Dose (/cGy) better performance than the other settings, hence K is
suggested to be 10 in all cases.
Fig. 11. Convergence profiles of mean HV and DVH of the solutions
obtained by MOEA/IMRT with and without local search on the Lastly, MOEA/IMRT is compared to two variants,
Esophagus dataset.
where one variant evolves only a population with rough
encoding and the other variant evolves only a popula-
the elimination of the local search strategy decreases tion with fine encoding. According to the HV results
the convergence speed of MOEA/IMRT and leads to a obtained on all the datasets given in Table SI in the
worse HV value. Besides, the figure also shows the DVH Supplementary Materials, the MOEA/IMRT with two
IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. , NO. , MONTH YEAR 13
0.8
of apertures in the decision variables [11].
0.7
0.6
R EFERENCES
0.5
[22] D. Craft and C. Richter, “Deliverable navigation for multicriteria [42] L. M. Antonio, C. A. Coello Coello, M. A. R. Morales, S. G. Bram-
step and shoot IMRT treatment planning,” Physics in Medicine & bila, J. F. González, and G. C. Tapia, “Coevolutionary operations
Biology, vol. 58, p. 87, 2013. for large scale multi-objective optimization,” in Proceedings of the
[23] W. Hong, P. Yang, and K. Tang, “Evolutionary computation for 2020 IEEE Congress on Evolutionary Computation, 2020.
large-scale multi-objective optimization: A decade of progresses,” [43] M. Li and J. Wei, “A cooperative co-evolutionary algorithm for
International Journal of Automation and Computing, vol. 18, pp. 155– large-scale multi-objective optimization problems,” in Proceedings
169, 2021. of the 2018 Annual Conference on Genetic and Evolutionary Compu-
[24] Y. Tian, L. Si, X. Zhang, R. Cheng, C. He, K. C. Tan, and Y. Jin, tation Conference, 2018, pp. 1716–1721.
“Evolutionary large-scale multi-objective optimization: A survey,” [44] X. Zhang, Y. Tian, R. Cheng, and Y. Jin, “A decision variable
ACM Computing Surveys, vol. 54, no. 8, p. 174, 2022. clustering-based evolutionary algorithm for large-scale many-
[25] M. Lahanas, D. Baltas, and N. Zamboglou, “A hybrid evolu- objective optimization,” IEEE Transactions on Evolutionary Compu-
tionary algorithm for multi-objective anatomy-based dose opti- tation, vol. 22, no. 1, pp. 97–112, 2018.
mization in high-dose-rate brachytherapy,” Physics in Medicine & [45] H. Zille, H. Ishibuchi, S. Mostaghim, and Y. Nojima, “A frame-
Biology, vol. 48, no. 3, pp. 399–415, 2003. work for large-scale multiobjective optimization based on prob-
[26] Y. Chen, D. Michalski, C. Houser, and J. M. Galvin, “A determinis- lem transformation,” IEEE Transactions on Evolutionary Computa-
tic iterative least-squares algorithm for beam weight optimization tion, vol. 22, no. 2, pp. 260–275, 2018.
in conformal radiotherapy,” Physics in Medicine & Biology, vol. 47, [46] C. He, L. Li, Y. Tian, X. Zhang, R. Cheng, Y. Jin, and X. Yao,
no. 10, pp. 1647–1658, 2002. “Accelerating large-scale multiobjective optimization via problem
[27] R. Cao, Y. Wu, X. Pei, J. Jing, G. Li, M. Cheng, G. Li, and L. Hu, reformulation,” IEEE Transactions on Evolutionary Computation,
“Multi-objective optimization of inverse planning for accurate vol. 23, no. 6, pp. 949–961, 2019.
radiotherapy,” Chinese Physics C, vol. 35, no. 3, pp. 313–317, 2011. [47] Y. Tian, C. Lu, X. Zhang, K. C. Tan, and Y. Jin, “Solving large-
[28] J. Yang, P. Zhang, L. Zhang, H. Shu, B. Li, and Z. Gui, “Parti- scale multiobjective optimization problems with sparse optimal
cle swarm optimizer for weighting factor selection in intensity- solutions via unsupervised neural networks,” IEEE Transactions
modulated radiation therapy optimization algorithms,” Physica on Cybernetics, vol. 51, no. 6, pp. 3115–3128, 2021.
Medica, vol. 33, pp. 136–145, 2017. [48] Y. Tian, X. Zheng, X. Zhang, and Y. Jin, “Efficient large-scale multi-
[29] C. Cotrutz, M. Lahanas, C. Kappas, and D. Baltas, “A multiob- objective optimization based on a competitive swarm optimizer,”
jective gradient-based dose optimization algorithm for external IEEE Transactions on Cybernetics, vol. 50, no. 8, pp. 3696–3708, 2020.
beam conformal radiotherapy,” Physics in Medicine & Biology, [49] W. Hong, K. Tang, A. Zhou, H. Ishibuchi, and X. Yao, “A scalable
vol. 46, no. 8, pp. 2161–2175, 2001. indicator-based evolutionary algorithm for large-scale multiobjec-
[30] G. Li and D. Cao, “A multi-objective particle swarm algorithm tive optimization,” IEEE Transactions on Evolutionary Computation,
for the optimization of IMRT inverse planning,” in Proceedings vol. 23, no. 3, pp. 525–537, 2018.
of the 3rd International Conference on Biomedical Engineering and [50] C. He, S. Huang, R. Cheng, K. C. Tan, and Y. Jin, “Evolutionary
Informatics, 2010. multiobjective optimization driven by generative adversarial net-
[31] A. Jalalimanesh, H. S. Haghighi, A. Ahmadi, H. Hejazian, and works (GANs),” IEEE Transactions on Cybernetics, vol. 51, no. 6,
M. Soltani, “Multi-objective optimization of radiotherapy: dis- pp. 3129–3142, 2021.
tributed Q-learning and agent-based simulation,” Journal of Ex-
[51] H. Zille and S. Mostaghim, “Comparison study of large-scale
perimental & Theoretical Artificial Intelligence, vol. 29, no. 5, pp.
optimisation techniques on the LSMOP benchmark functions,”
1071–1086, 2016.
in Proceedings of the 2017 IEEE Symposium Series on Computational
[32] T. Spalke, D. Craft, and T. Bortfeld, “Analyzing the main
Intelligence, 2017.
trade-offs in multiobjective radiation therapy treatment planning
[52] K. Deb and R. B. Agrawal, “Simulated binary crossover for
databases,” Physics in Medicine & Biology, vol. 54, no. 12, p. 3741,
continuous search space,” Complex Systems, vol. 9, no. 4, pp. 115–
2009.
148, 1995.
[33] R. Bokrantz, “Distributed approximation of Pareto surfaces in
multicriteria radiation therapy treatment planning,” Physics in [53] K. Deb and M. Goyal, “A combined genetic adaptive search
Medicine & Biology, vol. 58, no. 11, p. 3501, 2013. (GeneAS) for engineering design,” vol. 26, no. 4, 1996, pp. 30–
[34] S. Breedveld, P. R. M. Storchi, and B. J. M. Heijmen, “The 45.
equivalence of multi-criteria methods for radiotherapy plan opti- [54] L. While, P. Hingston, L. Barone, and S. Huband, “A faster
mization,” Physics in Medicine & Biology, vol. 54, no. 23, p. 7199, algorithm for calculating hypervolume,” IEEE Transactions on
2009. Evolutionary Computation, vol. 10, no. 1, pp. 29–38, 2006.
[35] R. van Haveren, S. Breedveld, M. Keijzer, P. Voet, B. Heijmen, [55] E. Zitzler, M. Laumanns, and L. Thiele, “SPEA2: Improving the
and W. Ogryczak, “Lexicographic extension of the reference point strength Pareto evolutionary algorithm for multiobjective opti-
method applied in radiation therapy treatment planning,” Euro- mization,” in Proceedings of the Fifth Conference on Evolutionary
pean Journal of Operational Research, vol. 263, no. 1, pp. 247–257, Methods for Design, Optimization and Control with Applications to
2017. Industrial Problems, 2001, pp. 95–100.
[36] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and eli- [56] R. Wang, R. C. Purshouse, and P. J. Fleming, “Preference-
tist multiobjective genetic algorithm: NSGA-II,” IEEE Transactions inspired coevolutionary algorithms for many-objective optimiza-
on Evolutionary Computation, vol. 6, no. 2, pp. 182–197, 2002. tion,” IEEE Transactions on Evolutionary Computation, vol. 17, no. 4,
[37] Y. Tian, R. Cheng, X. Zhang, Y. Su, and Y. Jin, “A strengthened pp. 474–494, 2013.
dominance relation considering convergence and diversity for [57] H. Wang, L. Jiao, and X. Yao, “Two Arch2: An improved two-
evolutionary many-objective optimization,” IEEE Transactions on archive algorithm for many-objective optimization,” IEEE Trans-
Evolutionary Computation, vol. 23, no. 2, pp. 331–345, 2019. actions on Evolutionary Computation, vol. 19, no. 4, pp. 524–541,
[38] M. Li, S. Yang, and X. Liu, “Shift-based density estimation for 2015.
Pareto-based algorithms in many-objective optimization,” IEEE [58] Y. Tian, T. Zhang, J. Xiao, X. Zhang, and Y. Jin, “A coevolu-
Transactions on Evolutionary Computation, vol. 18, no. 3, pp. 348– tionary framework for constrained multi-objective optimization
365, 2014. problems,” IEEE Transactions on Evolutionary Computation, vol. 25,
[39] K. Li, K. Deb, Q. Zhang, and S. Kwong, “Combining dominance no. 1, pp. 102–116, 2021.
and decomposition in evolutionary many-objective optimization,” [59] H. Zille and S. Mostaghim, “Linear search mechanism for multi-
IEEE Transactions on Evolutionary Computation, vol. 19, no. 5, pp. and many-objective optimisation,” in Proceedings of the 2019 In-
694–716, 2015. ternational Conference on Evolutionary Multi-Criterion Optimization,
[40] Y. Tian, R. Cheng, X. Zhang, F. Cheng, and Y. Jin, “An indicator 2019, pp. 399–410.
based multiobjective evolutionary algorithm with reference point [60] R. Cheng, Y. Jin, M. Olhofer, and B. Sendhoff, “A reference vector
adaptation for better versatility,” IEEE Transactions on Evolutionary guided evolutionary algorithm for many-objective optimization,”
Computation, vol. 22, no. 4, pp. 609–622, 2018. IEEE Transactions on Evolutionary Computation, vol. 20, no. 5, pp.
[41] L. M. Antonio and C. A. Coello Coello, “Use of cooperative 773–791, 2016.
coevolution for solving large scale multiobjective optimization [61] C. Cotrutz and L. Xing, “Segment-based dose optimization using
problems,” in Proceedings of the 2013 IEEE Congress on Evolutionary a genetic algorithm,” Physics in Medicine & Biology, vol. 48, no. 18,
Computation, 2013, pp. 2758–2765. pp. 2987–2998, 2003.
IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. , NO. , MONTH YEAR 15
[62] R. Cao, X. Pei, H. Zheng, J. Jing, M. Cheng, G. Li, and Y. Wu, Xingyi Zhang (SM’18) received the B.Sc. degree
“Beam intensity map optimization based on conjugate gradient in from Fuyang Normal College, Fuyang, China,
intensity modulated radiation treatment,” Nuclear Physics Review, in 2003, and the M.Sc. and Ph.D. degrees from
vol. 27, no. 2, pp. 212–217, 2010. Huazhong University of Science and Technol-
[63] J. Derrac, S. Garcia, D. Molina, and F. Herrera, “A practical tutorial ogy, Wuhan, China, in 2006 and 2009, respec-
on the use of nonparametric statistical tests as a methodology tively.
for comparing evolutionary and swarm intelligence algorithms,” He is currently a Professor with the School
Swarm and Evolutionary Computation, vol. 1, no. 1, pp. 3–18, 2011. of Artificial Intelligence, Anhui University,
[64] S. Breedveld, P. Storchi, P. Voet, and B. Heijmen, “iCycle: Inte- Hefei, China. His current research interests in-
grated, multicriterial beam angle, and profile optimization for clude unconventional models and algorithms of
generation of coplanar and noncoplanar IMRT plans,” Medical computation, evolutionary multi-objective op-
Physics, vol. 39, no. 2, pp. 951–963, 2012. timization, and logistic scheduling. He is the recipient of the 2018
and 2021 IEEE Transactions on Evolutionary Computation Outstanding
Paper Award and the 2020 IEEE Computational Intelligence Magazine
Outstanding Paper Award.
Xi Pei received the B.Sc. degree from Nan-
jing University of Aeronautics and Astronau-
Ye Tian received the B.Sc., M.Sc., and Ph.D.
tics, Nanjing, China, in 2006, and the Ph.D.
degrees from Anhui University, Hefei, China,
degrees from University of Chinese Academy
in 2012, 2015, and 2018, respectively.
of Sciences, Hefei, China, in 2012.
He is currently an Associate Professor with
He is currently an Associate Professor with
the Institutes of Physical Science and Infor-
the School of Nuclear Science and Technol-
mation Technology, Anhui University, Hefei,
ogy, University of Science and Technology of
China. His current research interests include
China, Hefei, China. His main research interests
evolutionary computation and its applications.
include key technologies of radiotherapy and
He is the recipient of the 2018 and 2021
their applications.
IEEE Transactions on Evolutionary Computa-
tion Outstanding Paper Award, the 2020 IEEE Kay Chen Tan (SM’08-F’14) received the B.Eng.
Computational Intelligence Magazine Outstanding Paper Award, and (First Class Hons.) degree in electronics and
the 2022 IEEE Computational Intelligence Society Outstanding PhD electrical engineering and the Ph.D. degree
Dissertation Award. from the University of Glasgow, Glasgow, U.K.
He is currently a Chair Professor of the Depart-
ment of Computing at the Hong Kong Poly-
technic University, Hong Kong. He has pub-
lished over 200 journal papers and six books,
and holds one U.S. patent on surface defect
Yuandong Feng received the B.Sc. degree from detection.
Hunan Normal University, Changsha, China, Prof. Tan is currently the Vice-President (Pub-
in 2018, where he is currently pursuing the lications) of IEEE Computational Intelligence Society, USA. He has
M.Sc. degree with the School of Computer Sci- served as the Editor-in-Chief of IEEE Transactions on Evolutionary
ence and Technology, Anhui University, Hefei, Computation from 2015-2020 and IEEE Computational Intelligence
China. Magazine from 2010-2013, and currently serves as the Editorial Board
His current research interests include large Member of over 10 journals. He was an IEEE Distinguished Lecturer
scale multi-objective optimization and its appli- Program (DLP) speaker and currently serves as the Chief Co-Editor of
cations. Springer Book Series on Machine Learning: Foundations, Methodolo-
gies, and Applications.
Yaochu Jin (SM’02-F’16) received the B.Sc.,
M.Sc., and Ph.D. degrees from Zhejiang Univer-
sity, Hangzhou, China, in 1988, 1991, and 1996,
respectively, and the Dr.-Ing. degree from Ruhr
Chao Wang received the B.Sc. degree from University Bochum, Germany, in 2001.
Suihua College, Suihua, China, in 2012, and the He is an Alexander von Humboldt Profes-
M.Sc. and Ph.D. degrees from Harbin Engineer- sor for Artificial Intelligence, Chair of Nature
ing University, Harbin, China, in 2015 and 2018, Inspired Computing and Engineering, Faculty
respectively. of Technology, Bielefeld University, Germany.
He is currently a Lecturer with the School of He is also a Distinguished Chair, Professor
Artificial Intelligence, Anhui University, Hefei, in Computational Intelligence, Department of
China. His main research interests include evo- Computer Science, University of Surrey, Guildford, U.K. He was a
lutionary computation, combinatorial optimiza- “Finland Distinguished Professor” of University of Jyväskylä, Fin-
tion, and intelligent transportation. land, “Changjiang Distinguished Visiting Professor”, Northeastern
University, China, and “Distinguished Visiting Scholar”, University
of Technology Sydney, Australia. His main research interests include
evolutionary optimization, evolutionary learning, trustworthy machine
learning, and evolutionary developmental systems.
Prof. Jin is presently the Editor-in-Chief of Complex & Intelligent
Systems. He was the Editor-in-Chief of the IEEE TRANSACTIONS
Ruifen Cao received the B.Sc. degree from ON COGNITIVE AND DEVELOPMENTAL SYSTEMS in 2016-2021
Henan University in 2004, and the Ph.D. de- and an IEEE Distinguished Lecturer in 2013–2015 and 2017–2019, the
gree from Hefei Institutes of Physical Science, Vice President for Technical Activities of the IEEE Computational
Chinese Academy of Sciences, Hefei, China, in Intelligence Society (2015–2016). He is the recipient of the 2018 and 2021
2010. IEEE Transactions on Evolutionary Computation Outstanding Paper
From 2010 to 2015, she was a Research Assis- Award, and the 2015, 2017, and 2020 IEEE Computational Intelligence
tant and Associate Researcher with the infor- Magazine Outstanding Paper Award. He was named by the Web of
mation science and technology. Since 2019, she Science as a Highly Cited Researcher from 2019 to 2021 consecutively.
has been an Associate Professor with the School He is a Member of Academia Europaea.
of Computer Science and Technology, Anhui
University. Her current research interest focuses
on medical image processing, artificial intelligence and its application
in radiotherapy.
View publication stats