Journal of Network Intelligence

2019
c ISSN 2414-8105 (Online)
Taiwan Ubiquitous Information Volume 4, Number 2, May 2019

Aperture Generation for Intensity-modulated

Radiotherapy Scheme based on Gradient Information

Jie Yang1,2 , Zhi-Guo Gui 1,3

, Li-Yuan Zhang1 , Peng-Cheng Zhang1∗
1
National Key Laboratory for Electronic Measurement Technology,
North University of China, Taiyuan Shanxi 030051, China
2
School of Medicine Management,
Shanxi University of TCM, Taiyuan 030619, China
3
Key Laboratory of Instrumentation Science and Dynamic Measurement of Ministry of Education,
North University of China, Taiyuan Shanxi 030051, China
∗
Corresponding Author: zhangpc198456@163.com

Received March 2018; revised December 2018

Abstract. We propose an algorithm for aperture generation (AG) for step and shoot de-
livery of intensity-modulated radiotherapy. The method is an approach to direct aperture
optimization (DAO) that exploits gradient information to initialize the aperture shapes.
Because the relationship between dose distribution and leaf positions of multileaf collima-
tors is not linear, new apertures are generated by using a region growing segmentation
algorithm, rather than solving a shortest-path problem. This yields a well-behaved opti-
mization problem for the aperture shapes, which can be solved efficiently. In this paper,
the AG algorithm is embedded into a DAO approach. After a new aperture is generated
based on the region growing segmentation algorithm and added to the treatment plan,
we optimize aperture weights and leaf positions for the new set of apertures. We tested
the proposed method on four prostate cases using the Computational Environment for
Radiotherapy Research algorithm for our dose calculation. The computational results
indicate that the proposed AG algorithm can produce highly conformal treatment plans
and decrease optimization time.
Keywords: Intensity-modulated radiotherapy; Direct aperture optimization; Gradient;
Region growing segmentation.

1. Introduction. Treatment planning for intensity-modulated radiotherapy (IMRT) has

conventionally employed a two-step process. The first stage is called the fluence map op-
timization (FMO) problem, which involves finding an optimal intensity profile for each
beam. In the second stage, the fluence maps are decomposed into a set of apertures
that are deliverable using a multileaf collimator (MLC). The main disadvantage of the
two-step process is that a loss in treatment quality can occur. To address this drawback,
direct aperture optimization (DAO)—which directly optimizes the aperture shapes and
weights—was developed [1-14]. There are basically two types of DAO approaches, and
each one usually starts with different initial aperture shapes. One type optimizes the
shape and weight of a predetermined number of apertures using a deterministic or sto-
chastic search (also called aperture modulation). In the works of Shepard et al. [1], Earl
et al. [2], and Cao et al. [5], the apertures are initialized with beam-eye-view (BEV)
projections of the target shapes. The original methods of Li et al. [3] and Cotrutz and
Xing [4] are initialized with leaf sequencing after FMO. Methods based on deterministic or
stochastic search methods rely on a good initial set of apertures, which can enable them

58
 Aperture Generation for Intensity-modulated Radiotherapy Scheme 59

to escape local minima. Nonetheless, they often cannot guarantee that all deliverable
apertures are considered. The second approach involves column generation, which starts
with an empty set of apertures and iteratively adds apertures to the treatment plan [6-14].
In the generic column-generation approach described by Romeijn et al. [7], the treatment
plan iteratively adds apertures by solving a shortest-path problem without modifying the
existing apertures. However, Carlsson [13] proved that the plan improvement rate is much
higher with the adjustable aperture approach than with the fixed aperture approach early
in the solution process. Meanwhile, Cassioli and Unkelbach [12] proposed an algorithm
that uses a trust-region-based local search approach combined with column generation to
optimize the aperture shapes. However, because the relationship between dose distribu-
tion and the leaf positions of multileaf collimators is not linear, the shortest-path problem
is complex, with a variable corresponding to each possible aperture shape. Consequently,
this approach tends to generate large apertures, and is not sufficiently fast. On the basis
of the above comparison and considerations, to improve the DAO algorithm and reduce
computation time, an algorithm for aperture generation (AG) based on gradient informa-
tion is herein proposed. The proposed method is different from those in previous studies
insofar as new apertures are generated using a region growing segmentation algorithm,
rather than solving a shortest-path problem. The AG algorithm is embedded into a DAO
approach; the optimization scheme starts with an empty set of apertures, and then gen-
erates a new aperture based on the AG algorithm. After the aperture is added to the
treatment plan, we optimize aperture weights and leaf positions for the new set of aper-
tures. The above procedure is then iteratively executed until the termination condition
is reached. In addition, the constraints imposed by the MLC are incorporated during
optimization. In this paper, results for four prostate cases are presented. The remainder
of this paper is organized as follows. In Section 2, the mathematical model and solution
for the proposed AG method are introduced. In Section 3, the computational results are
presented. Finally, the conclusions are summarized in Section 4.

2. Material and Methods.

2.1. Cost Function. Traditionally, some criteria are used to construct the cost function,
whereas others constitute constraints. In the proposed approach, on the other hand,
the weighted sum of a sub-objective function is used. The cost function can then be
formulated:
Traditionally, some criteria are used to construct the cost function, whereas others
constitute constraints. In the proposed approach, on the other hand, the weighted sum
of a sub-objective function is used. The cost function can then be formulated:
I
P
minimize F = δi fi (D) (1)
i=1

where I is the number of sub-objective functions, weight δi reflects the importance of the
ith sub-objective function, and fi (D) is the ith sub-objective function. In addition, D is
the dose distribution vector defined on the patient voxels. There are many different types
of sub-objectives [15-23]. They correspond to specifications for a variety of factors ranging
from the dose to the region of interest (ROI). In this work, four different specification types
are used. The sub-objective function for the specification Dmin ≥ yGy , representing the
stipulation that no ROI voxel should receive a dose less than yGy, is given by
V
1 X
fmin (D) = H(y − dv ) · (y − dv )2 (2)
V v=1
60 J. Yang, Z. G. Gui , L. Y. Zhang, and P. C. Zhang

1 dv < y
H(y − dv ) = (3)
0 dv ≥ y
where V specifies the voxels included in the ROI; and dv is the dose in the voxel v. For
the specification Dmin ≤ yGy ,, representing the stipulation that the uniform dose to the
ROI should not exceed yGy, the sub-objective function is given by
V
1 X
fmean (D) = (dv − y)2 (4)
V v=1
For the specificationVd1 ≤ y%, representing the stipulation that the volume of the ROI
receiving doses greater than d1 should not exceed y%, the sub-objective function is given
by
V
1 X
fDVH (D) = H(dv − d1 ) · H(d2 − dv ) · (dv − d1 )2 (5)
V v=1

1 dv > d1
H(dv − d1 ) = (6)
0 dv ≤ d1

1 d2 > dv
H(d2 − dv ) = (7)
0 d2 ≤ dv
where d2 is in the current dose volume histogram (DVH) and V(d2) = y%. For the spec-
ification N T CP ≤ y%, representing the stipulation that the normal tissue complication
probability (NTCP) does not exceed y%, the sub-objective function is given by
fNTCP (D) = H(ln(1 − y)) − ln(1 − N T CP (D)) · (ln(1 − y)) − ln(1 − N T CP (D)) (8)

1 N T CP (D) > y
H(ln(1 − y)) − ln(1 − N T CP (D)) = (9)
0 N T CP (D) ≤ y
where
Z gEU D(D)−D50
1 mD50 −t2
N T CP (D) = √ e 2 dt (10)
2π −∞
in which !1 /a
1X a
gEU D (D) = d (11)
V v∈V v
D50 is the whole organ dose for which the NTCP is 50%. In addition, m is obtained
by fitting the tolerance doses for the uniform whole and partial organ irradiation to the
function [24], and a is a control parameter [25]. Choi and Deasy [26] proved that a
convex function will be minimized when minimizing gEUD(di) to normal tissues (a ≥ 1).
Alternatively, when maximizing the gEU D(di ) of the target (a < 1), a concave function
will be maximized.
In FMO, the sub-objective function can usually be formulated by the beamlet inten-
sities. By contrast, for DAO, the objective function cannot be easily formulated in the
closed form as a function of the leaf positions. For the proposed AG method, a mathe-
matical model was developed to formulate dose D as a linear function of aperture shapes
and aperture weights on the basis of the dose-influence matrix. An advantage of this
mathematical model is that the highly nonlinear and non-convex problem of aperture
optimization is decomposed into a sequence of well-behaved optimization problems that
can be efficiently solved using standard algorithms. For the sake of convenience, the po-
sitions of leaves contained within the MLC replace the aperture shapes. We assume that
the aperture set is denoted as K, the set of weights is denoted as W, and each aperture
k ∈ Khas a weight wk ∈ W (wk > 0). The MLC consists of M identical leaf pairs, which
 Aperture Generation for Intensity-modulated Radiotherapy Scheme 61

Figure 1. Illustration of aperture generation (M = 9, L = 8)

we also refer to as rows. Each row has length L, and the positions of the left and right
leaves, respectively, of the mth row in the k th aperture are denoted by xlmk , xrmk ∈ [−1, L].
Dose D can be mathematically expressed as
X
D=P Bk wk (12)
k∈K
[
bjmk


Bk = (13)
m∈M,j∈L

1, xlmk < j < xrmk
bjmk = Ψmj (xlmk , xrmk ) = (14)
0, j ≤ xlmk or xrmk ≤ j
where P is the dose-influence matrix that contains the dose to each patient voxel from
each pencil beam, and Bk is the set of binary vectors specifying the beamlets that are
exposed by aperture k. Meanwhile, bjmk denotes the binary variable to bixel j from row m in
aperture k as a step function of the leaf position. To improve the quality of the deliverable
apertures, we consider the following two common sets of hardware constraints. (i) The
consecutiveness constraint simply corresponds to the fact that apertures are shaped by
pairs of leaves; i.e., in each given row m, each voxel should be exposed consecutively. (ii)
The interdigitation constraint holds that, in addition to (i), the left leaf of a row cannot
overlap with the right leaf of an adjacent row.

2.2. AG method. In the generic column-generation approach, the shape of the new
aperture is determined by solving a shortest-path problem as described by Carlsson [13].
However, the relationship between the dose distribution and the leaf positions of a multi-
leaf collimator is not linear, making the shortest-path problem overly complex. Unlike the
generic column-generation approach, the proposed AG method generates new apertures
using a simple method—namely, the region growing segmentation method. In the aper-
ture generation module, one aperture per beam can be generated with the AG method;
i.e., the leaf is assigned to a bixel. This is conducted in three serial steps to obtain the
aperture shape. The first step is to calculate the gradient matrix of the cost function.
The second step is to directly assign the leaf of each row to a bixel using the region grow-
ing segmentation method. The third step is to use the matrix’s minimum gradient as a
starting point and impose constraints on the leaf positions. A detailed description of the
procedure is summarized in what follows.
First, the gradient matrix of the cost function is computed. Let gmk (j) denote the
gradient of cost function F with respect to the beamlet j in row m in aperture k, which
62 J. Yang, Z. G. Gui , L. Y. Zhang, and P. C. Zhang

is evaluated for the current aperture set (K̄, W̄ ). Meanwhile, gmk (j) is given by

∂F
gmk (j) = Pmkj wk (15)
∂D

Second, to solve the problem of assigning the leaf of each row to a bixel, we apply region
growing to open or close the bixel. We define some seed points and then calculate the
connectivity between each seed bixel and its neighborhoods. The main steps for aperture
generation based on region growing are as follows:
Step 1: Selection of initial seed points. The matrix’s minimum gradient and each row’s
minimum gradient is found quickly. Let gminmk = minj∈L gmk (j) denote the minimum
gradient of row m in aperture k, and let the position of gminmk be(m,j 0 m ). We select
the initial seed points for region growing according to the minimum gradient of row m in
aperture k. The set S of initial seed points is defined as

S = {gminmk |m ∈ M } (16)

Step 2: Region growing. If any bixel belonging to the seed point’s 2-connected neigh-
borhood (those to the left and right of the seed points) satisfies the similarity criterion
gmk (j) < 0, we add it to the seed point set.
Step 3: Termination. The left leaf is assigned to bixel j + 1 when the termination
condition is met (i.e., if moving to the left gmk (j) is positive for the first time). Similarly,
the right leaf is assigned to bixel j–1 (i.e., if moving to the right gmk (j) is positive for the
first time). Note that we set gmk (−1) and gmk (L) to large values.
In the third stage, we must ensure that the new aperture remains deliverable. Thus, we
have to ensure that the selected positions xlmk , xrmk for all m ∈ M will lead to deliverable
solutions. We therefore use the Gmink as a starting point, and going line-by-line, up or
down, have satisfy

xlmk < xrmk ∀k ∈ K, m ∈ M, xlmk ∈ [−1, L], xrmk ∈ [−1, L]

xmk < xm−1k ∀k ∈ K, m ∈ M, xlmk ∈ [−1, L], xrmk ∈ [−1, L]
l r
(17)
xrmk > xlm−1k ∀k ∈ K, m ∈ M, xlmk ∈ [−1, L], xrmk ∈ [−1, L]

where Gmink = minm∈M,j∈L gmk (j) denotes the minimum gradient of aperture k, and
the position of Gmink is(m0 ,j 0 m ). Finally, constraints on the minimum aperture size are
imposed. If the number of non-zero bmj values for the new aperture (m ∈ M, j ∈ L)
exceeds the lower limit, it is added to K̄.
 Aperture Generation for Intensity-modulated Radiotherapy Scheme 63

2.3. Integration into A DAO Approach. The proposed AG algorithm is next em-
bedded into a DAO approach. This approach is similar to the one used by Cassioli et
al. [12] and Wang et al. [14], but our approach replaces the shortest-path problem with
the AG method. Simultaneous optimization of the aperture shape and weight is indeed
possible, but because there is a coupling, optimization is difficult. Thus, it is preferable
to handle them separately. To optimize the weights W̄ of the current aperture set, the
weight optimization module is entered after a new aperture is added to the treatment
plan. This procedure is equivalent to Eq. (12) with bk fixed. The extension of the L-
BFGS algorithm for bound-constrained problems (L-BFGS-B) [27] is used to optimize all
weights for the set of apertures. This is because this restricted problem is much easier to
solve owing to the number of variables being smaller and D being linear in W̄ . This is
in contrast to the non-convexity introduced by the leaf position variables, and it ensures
that the weights are optimal or close to optimal so the plan can be saved. The aperture
64 J. Yang, Z. G. Gui , L. Y. Zhang, and P. C. Zhang

Figure 2. Flow chart of the AG method

weights optimization problem is given by

I
P
minimizeW F = δi fi (D) (18)
i=1
subject to wk > 0, Bk fixed. Then, the aperture modification module is entered. In
the approach described by Carlsson [13], apertures with weights at the lower limit will be
removed in the aperture removal module to improve the plan, but this makes computation
extremely complex. Cassioli and Unkelbach [12] exploited gradient information to locally
optimize the positions of the leaves of a multileaf collimator. Thus, we exploit the aperture
shape optimization (ASO) algorithm to modify the existing apertures K̄. We repeat
the procedure until the relative improvement is less than 10−1 between two consecutive
iterations. The overall scheme for the complete DAO algorithm is shown in Fig. 2.

3. Experiments and Analysis.

3.1. Algorithm Details. In this section, we investigate the feasibility and performance
of the proposed method in four clinical prostate tumor cases. We investigate the integra-
tion of the AG algorithm into a DAO approach. Specifically, we compare the following
two algorithms: ag − our DAO algorithm in which a new aperture is generated with the
AG algorithm (Algorithm 1); sp − the DAO proposed by Cassioli et al. [12] in which a
new aperture is generated with shortest-path problem algorithm.
In these cases, the planning target volume (PTV) was the prostate, and the organs at
risk (OARs) were the bladder and rectum. Normal tissue was 5 cm outside the PTV. For
all four cases, we designed plans using five 6 − M V photon beams. The five beams were
distributed around the patient with angles of 36o , 100o , 180o , 260o , and 324o , respectively.
Figure 3 shows an example of the target delineation. The matrix of the beamlet dose
deposition coefficients was computed using the pencil-beam dose calculation algorithm
through the Computational Environment for Radiotherapy Research (CERR) interface
[28]. The proposed method was implemented in a Microsoft Visual C + + environment on
 Aperture Generation for Intensity-modulated Radiotherapy Scheme 65

Figure 3. Typical CT slice illustrating the target and the critical structure
delineation

Table 1. Total volumes (cubic cm) for the four prostate tumor cases

Table 2. Weights of cost function for the four cases

a personal computer equipped with an Intel Core i3 CPU@ 3.50 GHz and 4 GB memory.
The termination criterion was defined as a relative improvement of less than 10−1 between
two consecutive iterations. Table 1 summarizes the model dimensions for the four prostate
tumor cases.
We considered a biological indices (NTCP-based) cost function:

δ1 fNTCP (DBladder ) + δ2 fNTCP (DRectum )

(19)
+δ3 fmean (DPTV ) + δ4 fmin (DPTV ) + δ5 fDVH (DNormal )
66 J. Yang, Z. G. Gui , L. Y. Zhang, and P. C. Zhang

Figure 4. Objective function values for case 1 for NTCP-based cost function

Table 3. Parameters for the NTCP

(a) a (b) b

Figure 5. DVHs of case 1: (a) all DVHs, and (b) DVHs of PTV

Table 2 shows the weights of the NTCP-based cost function (Eq. (19)) for the four
cases. The parameters for NTCP [29, 30] are tabulated in Table 3. The mean dose for
the PTV was 78 Gy, and the minimum dose for the PTV was 74 Gy.
 Aperture Generation for Intensity-modulated Radiotherapy Scheme 67

Table 4. Dose for case 1

Table 5. Final value of cost function, number of apertures, and runtime

Table 6. ag: DVH criteria

Table 7. sp: DVH criteria

3.2. Evaluate Criteria. We used clinical DVH criteria to objectively verify the ability
of our models to create clinically acceptable treatment plans. The criteria employed are
based on the current clinical guidelines formulated by Marks et al. [31]:
PTV
– At least 99% of the PTV should receive 93% of the prescribed dose (0.93 × 75Gy).
68 J. Yang, Z. G. Gui , L. Y. Zhang, and P. C. Zhang

– At least 95% should receive the prescribed dose (75 Gy).

– No more than 10% should be overdosed by more than 10%
of the prescribed dose (1.1 × 75 Gy).
– No more than 1% should be overdosed by more than 20%
of the prescribed dose (1.2 × 75 Gy).
Rectum
– No more than 50% should be overdosed by more than 50 Gy.
– No more than 35% should be overdosed by more than 60 Gy.
– No more than 25% should be overdosed by more than 65 Gy.
– No more than 20% should be overdosed by more than 70 Gy.
– No more than 15% should be overdosed by more than 75 Gy.
Bladder
– No more than 50% should be overdosed by more than 65 Gy.
– No more than 35% should be overdosed by more than 70 Gy.
– No more than 25% should be overdosed by more than 75 Gy.
– No more than 15% should be overdosed by more than 80 Gy.

3.3. Convergence Properties. In Fig. 4, we plot the objective function value as a

function of the iteration. The ag and sp algorithms stop after at most 100 iterations.
The dotted line indicates the objective function obtained for fmo (the FMO algorithm
as proposed by Zhang et al. [32], used as an ideal solution benchmark. The details of
the progress of FMO have been published elsewhere [32-34].). Note that this objective
function refers to the ideal FMO solution without a sequencing step.
The test results show that both ag (solid line) and sp (dashed line) come substantially
close to the fmo solution (dotted line). They also show that the ag algorithm is convergent
and stable. This indicates that, with an increasing number of iterations, at some point
the incremental improvement in the treatment plan quality may no longer be clinically
significant. Hence, rather than allowing the ag algorithm to formally converge, we propose
terminating the algorithm when it is observed that the treatment plan quality with respect
to a particular criterion has not markedly improved in recent iterations, i.e., when the
relative improvement between successive iterations is less than 10−1 .

3.4. Comparison of Plans. Figure 5 compares the DVHs of the final plan for the two
algorithms. The solid lines indicate the DVHs obtained for ag; those for sp are indicated
by dashed lines. From the DVH comparison in Fig. 6, it can be confirmed that ag yields
treatment plans that are close to the sp solution based on the NTCP cost function. The
proposed method is thus efficient and effective. We observe the same dose coverage on
the PTV for these two optimization results. The mean doses for PTV were 76.36 Gy and
76.38 Gy for ag and sp, respectively. For the OARs, both results met the current clinical
guidelines. However, ag’s overall protection performance for OARs, shown in Table 4, was
slightly better than sp’s. Table 4 shows the mean dose, maximum dose, and minimum
dose for the OARs and PTV.
We computed the final cost function value, number of apertures, and runtime (in min-
utes), all of which can be found in Table 5. Tables 6 and 7 show the values of the DVH
criteria for the targets and main critical structures for all four cases of ag and sp, respec-
tively. For example, the data in the column labeled PTV@93% represent the fraction of
volume (in %) of PTV receiving 0.93 × 75 Gy. The labels of the other columns follow a
similar format.
In comparing Tables 5, 6, and 7, it is apparent that treatment plans of a very similar
quality can be obtained. We observe in the results that there is the same dose coverage for
 Aperture Generation for Intensity-modulated Radiotherapy Scheme 69

the PTV using the sg method, the ag method, and objective function values. However,
the ag approach can reduce the programs’ execution time by over 15.90% on average.

4. Conclusions. We demonstrated the feasibility of a direct aperture optimization method

based on gradient information for IMRT. The proposed AG algorithm generates new
apertures using the region growing segmentation method, leading to the fewest possible
apertures and the simplest possible aperture shapes. All of the cases investigated demon-
strate that the proposed AG algorithm provides simple, effective IMRT solutions, and
results that approximate the ideal FMO solution. In particular, for the examples con-
sidered here, the AG algorithm obtained slightly better treatment plan quality and less
optimizing time compared to the column-generation algorithm. On average, the proposed
algorithm reduced the programs’ execution time by 15.90%. The proposed approach facil-
itates a tradeoff between treatment plan quality and delivery efficiency. Future research
can extend this work by explicitly incorporating a treatment plan efficiency measure, such
as reducing the number of apertures per beam.

Acknowledgment. This work was supported in part by the National Nature Science
Foundation of China under Grant 61271357, in part by the National Nature Science
Foundation of China under Grant 11605160, in part by the National Key Scientific In-
strument and Equipment Development Project under Grant 2014YQ24044508, in part by
the opening project of State Key Laboratory of Explosion Science and Technology (Bei-
jing Institute of Technology) under Grant KFJJ13-11M, and in part by Research Project
Supported by Shanxi Scholarship Council of China under Grant 2016-089.

REFERENCES

[1] D. M. Shepard, M. A. Earl, X. A. Li, S. Naqvi, and C. X. Yu, Direct aperture optimization: A
turnkey solution for step-and-shoot IMRT, Med. Phys., vol. 29, pp. 1007-1018, 2002.
[2] M. A. Earl, M. K. N. Afghan, C. X. Yu, Z. Jiang, and D. M. Shepard, Jaws-only IMRT using direct
aperture optimization, Med. Phys., vol. 34, pp. 307-314, 2007.
[3] Y. Li, J. Yao, and D. Yao, Genetic algorithm based deliverable segments optimization for static
intensity-modulated radiotherapy, Phys. Med. Biol., vol. 48, pp. 3353-3374, 2003.
[4] C. Cotrutz and L. Xing, Segment-based dose optimization using a genetic algorithm, Phys. Med.
Biol., vol. 48, p. 2987, 2003.
[5] R. Cao, X. Pei, H. Zheng, L. Hu and Y. Wu, Direct aperture optimization based on genetic algorithm
and conjugate gradient in intensity modulated radiation therapy, Chin. Med. J., vol. 127, pp. 4152-
4153, 2014.
[6] F. Preciado-Walters, M. P. Langer, R. L. Rardin, and V. Thai, Column generation for IMRT cancer
therapy optimization with implementable segments, Ann. Operat. Res., vol. 148, pp. 65-79, 2006.
[7] H. E. Romeijn, R. K. Ahuja, J. F. Dempsey, and A. Kumar, A column generation approach to
radiation therapy treatment planning using aperture modulation, SIAM J. Optim., vol. 15, pp.
838-862, 2005.
[8] C. Men, H. E. Romeijn, Z. C. Taskin, and J. F. Dempsey, An exact approach to direct aperture
optimization in IMRT treatment planning, Phys. Med. Biol., vol. 52, pp. 7333-7352, 2007.
[9] E. Salari, C. Men, and H. E. Romeijn, Accounting for the tongue-and-groove effect using a robust
direct aperture optimization approach, Med. Phys., vol. 38, pp. 1266-1279, 2011.
[10] G. Bednarz, D. Michalski, C. Houser, M. S. Huq, Y. Xiao, P. R. Anne, and J. M. Galvin, The use
of mixed-integer programming for inverse treatment planning with pre-defined field segments, Phys.
Med. Biol., vol. 47, pp. 2235-2245, 2002.
[11] E. Salari and J. Unkelbach, A column-generation-based method for multi-criteria direct aperture
optimization, Phys. Med. Biol., vol. 58, pp. 621-639, 2013.
[12] A. Cassioli and J. Unkelbach, Aperture shape optimization for IMRT treatment planning, Phys.
Med. Biol., vol. 58, pp. 301-318, 2012.
[13] F. Carlsson, Combining segment generation with direct step-and-shoot optimization in intensity-
modulated radiation therapy, Med. Phys., vol. 35, pp. 3828-3838, 2008.
70 J. Yang, Z. G. Gui , L. Y. Zhang, and P. C. Zhang

[14] J. Wang, X. Pei, R. Cao, L. Hu, and Y. Wu, A multiphase direct aperture optimization for inverse
planning in radiotherapy, Nucl. Sci. Tech., vol. 26, pp. 48-53, 2015.
[15] J. F. Fowler, The linear-quadratic formula and progress in fractionated radiotherapy, Brit. J. Radiol.,
vol. 62, pp. 679-694, 1989.
[16] A. Brahme, Treatment Optimization Using Physical and Radiobiological Objective Functions in
Smith (ed.), Radiation Therapy Physics, Springer, Berlin, pp. 209-246, 1995.
[17] G. J. Kutcher and C. Burman, Calculation of complication probability factors for non-uniform
normal tissue irradiation: The effective volume method, Int. J. Radiat. Oncol. Biol. Phys., vol. 16,
pp. 1623-1630, 1989.
[18] M. Alber and R. Reemtsen, Intensity modulated radiotherapy treatment planning by use of a barrier-
penalty multiplier method, Optim. Method. Softw., vol. 22, pp. 391-411, 2007.
[19] J. O. Deasy, Multiple local minima in radiotherapy optimization problems with dose–volume con-
straints. Med. Phys., vol. 24, pp. 1157-1161, 1997.
[20] J. Llacer, J. O. Deasy, T. R. Portfeld, T. D. Solberg, and C. Promberger, Absence of multiple local
minima effects in intensity modulated optimization with dose-volume constraints, Phys. Med. Biol.,
vol. 48, pp. 183-210, 2003.
[21] H. E. Romeijn, J. F. Dempsey, and J. G. Li, A unifying framework for multi-criteria fluence map
optimization models, Phys. Med. Biol., vol. 49, pp. 1991-2013, 2004.
[22] I. E. Naqa, G. Suneja, P. E. Lindsay, A. J. Hope, J. R. Alaly, and V. Vicic, Dose response explorer:
an integrated open-source tool for exploring and modelling radiotherapy dose–volume outcome rela-
tionships, Phys. Med. Biol., vol. 51, pp. 5719-5735, 2006.
[23] J. Zhu, B. S. Li, H. Z. Shu, T. Bai,Y. Yin, and L. M. Luo, Review of normal tissues complication
probability models, Chin. J. Biomed. Eng., vol. 33, pp. 233-240, 2014.
[24] J. T. Lyman, ”Complication probability as assessed from dose-volume histograms,” Radiat. Res.
Suppl., vol. 8, pp. S13-S19, 1985.
[25] A. Niemierko, A generalized concept of equivalent uniform dose (EUD), Med. Phys., vol. 26, p. 1100,
1999.
[26] B. Choi and J. O. Deasy, The generalized equivalent uniform dose function as a basis for intensity-
modulated treatment planning, Phys. Med. Biol., vol. 47, pp. 3579-3589, 2002.
[27] R. H. Byrd, P. Lu, and J. Nocedal, A limited memory algorithm for bound constrained optimization,
SIAM J. Sci. Comp., vol. 16, pp. 1190-1208, 1995.
[28] J. O. Deasy, A. I. Blanco, and V. H. Clark, CERR: a computational environment for radiotherapy
research, Med. Phys., vol. 30, pp. 979-985, 2003.
[29] E. Dale, T. P. Hellebust, A. Skjønsberg, T. Høgberg, and D. R. Olsen, Modeling normal tissue
complication probability from repetitive computed tomography scans during fractionated high-dose-
rate brachytherapy and external beam radiotherapy of the uterine cervix, Int. J. Radiat. Oncol. Biol.
Phys., vol. 47, pp. 963-971, 2000.
[30] S. T. H. Peeters, M. S. Hoogeman, W. D. Heemsbergen, A. A. M. Hart, P. C. M. Koper, and J. V.
Lebesque, Rectal bleeding, fecal incontinence, and high stool frequency after conformal radiotherapy
for prostate cancer: Normal tissue complication probability modeling. Int. J. Radiat. Oncol. Biol.
Phys., vol. 66, pp. 11-19, 2006.
[31] L. B. Marks, E. D. Yorke, and A. Jackson, Use of normal tissue complication probability models in
the clinic, Int. J. Radiat. Oncol. Biol. Phys., vol. 76, pp. 10-19, 2010.
[32] L. Y. Zhang, P. C. Zhang, Z. G. Gui, H. Z. Shu, and J. Yang, Plan optimization method of IMRT
based on biological indices, Appl. Res. of Comp., vol. 34, no. 5, pp. 1303-1307, 2017.
[33] J. Yang, and Z. G. Gui, ”Base on Study Of L-BFGS and NSGA-II Hybrid Calculation to IMRT
Inverse Planning Optimization, J. North Uni. Chin., vol. 6, pp. 706-712, 2015.
[34] J. Yang, P. C. Zhang, L. Y. Zhang, H. Z. Shu, B. S Li, and Z. G. Gui, Particle swarm optimizer for
weighting factor selection in intensity-modulated radiation therapy optimization algorithms, Phys.
Med., vol. 33, pp. 136-145, 2017.