Physica Medica xxx (2016) xxxxxx

Contents lists available at ScienceDirect

Physica Medica
journal homepage: http://www.physicamedica.com

Intelligence-guided beam angle optimization in treatment planning of

intensity-modulated radiation therapy
Hui Yan , Jian-Rong Dai
Department of Radiation Oncology, Cancer Hospital Chinese Academy of Medical Sciences, PO Box 2258, Beijing 100021, China

a r t i c l e i n f o a b s t r a c t

Article history: An intelligence guided approach based on fuzzy inference system (FIS) was proposed to automate beam
Received 22 January 2016 angle optimization in treatment planning of intensity-modulated radiation therapy (IMRT). The model of
Received in Revised form 12 April 2016 FIS is built on inference rules in describing the relationship between dose quality of IMRT plan and irra-
Accepted 14 June 2016
diated region of anatomical structure. Dose quality of IMRT plan is quantified by the difference between
Available online xxxx
calculated and constraint doses of the anatomical structures in an IMRT plan. Irradiated region of
anatomical structure is characterized by the metric, covered region of interest, which is the region of
Keywords:
an anatomical structure under radiation field while beams eye-view is conform to target volume.
Fuzzy inference system
Beam angle optimization
Initially, an IMRT plan is created with a single beam. The dose difference is calculated for the input of
Intensity-modulated radiation therapy FIS and the output of FIS is obtained with processing of fuzzy inference. Later, a set of candidate beams
is generated for replacing the current beam. This process continues until no candidate beams is found.
Then the next beam is added to the IMRT plan and optimized in the same way as the previous beam.
The new beam keeps adding to the IMRT plan until the allowed beam number is reached. Two spinal
cases were investigated in this study. The preliminary results show that dose quality of IMRT plans
achieved by this approach is better than those achieved by the default approach with equally spaced
beam setting. It is effective to find the optimal beam combination of IMRT plan with the intelligence-
guided approach.
2016 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.

1. Introduction may be less effective. Moreover, an increased number of beams

Intensity-modulated radiation therapy (IMRT) was developed in
order to deliver a highly conformal dose to tumor while bearing
surrounding normal tissue, especially critical organs, under clini-
cally acceptable dose level. For achieving this goal, multiple beams
with modulated fluence maps from different angles were required.
In current state-of-the-art treatment planning systems of IMRT,
once beam angles are determined fluence maps can be automati-
cally computed by inverse planning optimization algorithms. Usu-
ally, the optimal combination of beam angles is unknown
aforehand and has to be guessed by trial-and-error approach.
Several researchers attempt to resolve this issue with different
way [15]. It was found that the optimal combination of beam
angles for IMRT plan tends to be an even distribution over an angu-
lar range of 02p [1]. It was also demonstrated that with increasing
number of beams the quality of plan dose could be improved [2].
However, for the cases with irregular spatial distribution of inter-
ested anatomical structures equally spaced beam configuration
Corresponding author.
E-mail address: hui.yan@cicams.ac.cn (H. Yan).
results in prolonged treatment time, which may increase delivery
error caused by involuntary patient movement and intra-
fractional organ motion. Therefore, the optimization of beam
angles should have taken spatial distribution of interested anatom-
ical structures into account for an improved dose quality of IMRT
plan [69].
In general, the approaches of Beam angle optimization (BAO)
can be divided into two categories. The first category treats BAO
as an independent part in addition to fluence map optimization
(FMO). An optimal beam set is selected based on scores of beam
angles using prior knowledge, then forwarded to FMO for plan dose
distribution. It is efficient to determine a suitable beam set for FMO
but the optimality of resulting beam set is not guaranteed. The sec-
ond category treats BAO as an integrated part of FMO, and jointly
optimizes beam angles and fluence map. With the inclusion of
beam angles, the solution space of FMO is expanded and could
be non-convex [1]. To avoid trapping in local minima, it is ideal
to thoroughly search the solution space in order to achieve global
minima with exhaustive search strategy [10,11]. The stochastic
approach, such as simulated annealing and genetic algorithm,
which have been previously used in inverse planning optimization,

http://dx.doi.org/10.1016/j.ejmp.2016.06.005
1120-1797/ 2016 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.

Please cite this article in press as: Yan H, Dai J-R. Intelligence-guided beam angle optimization in treatment planning of intensity-modulated radiation
therapy. Phys. Med. (2016), http://dx.doi.org/10.1016/j.ejmp.2016.06.005
2 H. Yan, J.-R. Dai / Physica Medica xxx (2016) xxxxxx

are also effective for the problem of BAO [1216]. It is also feasible 2. Materials and methods
to find a global minima by mixed integer linear programming
[17,18]. It should note that as a general flaw of stochastic 2.1. Covered region of interest
approaches it is usually time-consuming. Besides stochastic
approaches, heuristic approaches are able to find an optimal com- Coverage region of interest is a metric in measuring the region
bination of beam angles for an IMRT plan. In addition, several met- of an interested structure (PTV, OAR, etc.) under radiation field of a
rics in characterizing the effect of beam on those interested beam at a specific angle while beam aperture is shaped to target
structures, such as the entropy and Fourier transform measures volume. In this study, it is assumed that there are three types of
[19], the maximum beam separation and minimized non-target anatomical structures: target volume (TV), critical organ (CO),
irradiation [20], and the score of beam angle based on pseudo and normal tissue (NT). As shown in Fig. 1 one beam passes
beams-eye-view technique [21,22] were proposed. These metrics through the CT volume of a patient and intersects with the regions
could keep those favorable beams while remove those unfavor- of TV, CO, and NT along its path. The overlapping regions between
able beams. radiation path and interested structures are labeled with different
It is believed the experience of human planner is critical in textures in Fig. 1. Since beam aperture is conformal only to TV, par-
guiding the process of BAO for an IMRT plan. However, most of tial volume of CO and NT are outside of radiation field. As shown in
prior knowledge of human planners is vague and hardly quantified Fig. 1 two third of CO is outside of radiation field. The whole vol-
for computer applications. So far, there are few studies in applying ume of TV is covered by radiation filed but its distance from radi-
intelligence technique in BAO of IMRT treatment planning. Novel ation source is varied. The fraction of CO (and NT) inside radiation
approach was proposed to use a pattern search method framework filed is changed by beam angle, and its distance from radiation
in the optimization of BAO problem [23]. Beams-eye-view dose source is also varied.
(BEVD) metric furnishes a prior knowledge in guiding the search- For computation of CRI, both fraction of interested structures
ing of optimal beam ensembles. Although BEVD represents an intu- inside radiation field and its distance to radiation source are con-
itive consideration of the deliverable dose capability to the target sidered. At first, the voxels of interested anatomical structure
of a single beam direction, the whole BAO procedure is virtually inside radiation field are counted. Then the inverses of distances
dominated by optimization algorithm and less controlled by between the voxel and radiation source are summed. Finally, the
human knowledge. Machine learning technique, such as artificial value of sum is normalized by the total number of voxels inside
neural network (ANN), was used to map anatomical features to radiation field. This final value is CRI for the interested anatomical
beam angle scores and learn the relationship using clinically structure at a given beam angle. Its value represents the effect of a
approved plans [24]. As beam scores obtained by ANN the follow- beam on an interested structure. Note that the distance was calcu-
ing optimization algorithm will use them to avoid the unnecessary lated as the center of voxel to the source and the partial volume
search directions. Recently the relationship between beam angles effect was not considered. As radiation source closer to an organ,
and anatomical features was modeled by advanced machine learn- the value of CRI is increased, otherwise decreased. As the number
ing technique forest regression algorithm [25]. It is capable of of voxels inside radiation field increased, the value of CRI is
mapping a multitude of anatomical features into an individual increased. For TV, the beam with larger value of CRI is preferred
beam score. An optimization scheme is then built to select beam since higher dose to tumor could be achieved. However, for CO
while considering the inter-beam dependencies. The results are and NT, the beam with smaller value of CRI is preferred since
promising and would be helpful in reducing the manual planning low dose could be achieved. As there are several CRI for different
workload. interested anatomical structures, it is better to represent them
Among those prior knowledge guide BAO algorithms, the scores
of beams are derived from the clinically approved plan or provided
by empirical function, and then used by optimization algorithms. It
is rare that the prior knowledge of human reasoning process in
trial-and-error procedure of beam selection is learned and then
applied to guide the searching of the best beam ensembles. It is
expected that there is a way to bridge the gap between vague
knowledge in trial-and-error procedure and precise quantities
used by computer software. As fuzzy inference system is dedicated
to simulate the human reasoning process based on the vague
knowledge, it is ideal for this problem. In this study we proposed
an intelligence guided approach in assisting the process of BAO.
It incorporates both metric of beam geometry and prior knowledge
of human planner into the optimization process of beam angles.
Similar as our previous studies of applying fuzzy inference system
into the parameter optimization of inverse planning [2629], the
beam angles are another set of planning parameters outside of
FMO and with different characteristics. The remainder of this paper
was organized as follow. In materials and methods section the con-
cept of CRI and the principle of FIS were introduced. The imple-
mentation of inverse planning algorithm for FMO employed was
described, and the process of intelligence guided approach was
explained. For testing purpose the detail of clinical case study
was introduced. In results section, dose volume histogram (DVH),
dose distribution, and CRI values of plans resulted by the intelli-
gence approach and default approach were presented and ana-
lyzed. Finally the benefit and limitation of this approach was Fig. 1. Demonstration of single beam plan used for calculation of CRIs for TV, CO,
discussed. and NT.

Please cite this article in press as: Yan H, Dai J-R. Intelligence-guided beam angle optimization in treatment planning of intensity-modulated radiation
therapy. Phys. Med. (2016), http://dx.doi.org/10.1016/j.ejmp.2016.06.005
 H. Yan, J.-R. Dai / Physica Medica xxx (2016) xxxxxx
with a single value. For this purpose, a General CRI (GCRI) is
defined as below.
CRITV n  CRICO n  CRINT n
GCRIn
CRITV n CRICO n CRINT n
Here, GCRIn is the value of GCRI at nth beam. CRITV n, CRICO n;
and CRINT n are the values of CRI for TV, CO, and NT at nth beam.
The larger value of GCRIn indicates the beam is close to TV and
away from CO (NT). On the contrary, the smaller value of GCRIn
indicates the beam is away from TV and close to CO (NT).
When values of CRI for all beam angles (varied from 0 to 360)
are calculated for an interested anatomical structure, they are
ranked in ascending order and stored in a lookup table for next
step. The lookup tables are created before the process of BAO.
Fig. 2 illustrates the plot of CRI values as a function of beam angle
varied from 0 to 360 for TV, CO, and NT, respectively. This plot is
calculated from the voxels of CT images as shown in Fig. 1. There
are three CRI curves corresponding to three interest anatomical
structures. For different treatment sites and beam angles, the CRI
plot would be different. It is observed that the curves of CRI are
similar to those of mean organ-at-risk data reported by Shi et al.
[15]. In their study, the correlation relationship between the mag-
nitude of mean organ-at-risk data and dose sparing of OAR was
found. The beam angles preferred for target volume may be not
desirable for critical organ (normal tissue). The trade-off between
dose uniformity and conformity of TV and dose sparing of CO has
to be determined by human planner.
2.2. Fuzzy inference system
A fuzzy inference system is an intelligent model, which can per-
form human reasoning based on linguistic rules provided by expert
[26,27]. Generally, FIS consists of three components: Fuzzifier,
Inference Engine, and Defuzzifier as shown in Fig. 3. The inputs
of FIS DDTV ; DDCO ; DDNT are dose differences between calculated
and prescription/constraint doses, and the outputs of FIS
DCRITV ; DCRICO ; DCRINT are adjustments of CRI values for TV, CO,
and NT. The detailed description of FIS and the process of fuzzy
inference can be found in our previous publications [21,22]. In gen-
eral, input was first converted to values in the intermediate form,
and then processed by inference engine. The result in the interme-
diate form was finally converted to real numbers for output pur-
pose. As shown in Fig. 3, a set of functions, called membership
functions, are specified in Fuzzifier and used to convert input val-
ues from real numbers to membership values. These membership
values are then processed by inference engine to derive another
set of membership values according to logical rules. In Defuzzifier
another sets of membership functions are specified and used to
Fig. 2. The illustration of CRI values for TV, CO, and NT as a function of beam angles varied from 0 to 360.
Please cite this article in press as: Yan H, Dai J-R. Intelligence-guided beam angle optimization in treatment planning of intensity-modulated radiation
therapy. Phys. Med. (2016), http://dx.doi.org/10.1016/j.ejmp.2016.06.005
3
convert the resulting membership values to real numbers. In this
study a simple rule set consisting of three if-then rules are pro-
vided. The content of the rule set is listed below.
1
(1) If calculated dose of TV is less than prescribed dose, then
CRITV is increased.
(2) If calculated dose of CO is more than constraint dose, then
CRICO is decreased.
(3) If calculated dose of NT is more than constraint dose, then
CRINT is decreased.
The first rule is assigned the highest priority and the second rule
is assigned less priority. The third rule is assigned the least priority.
These rules can be rewritten as:
(1) If DCTV < DPTV ; then DCRITV > 0.
(2) If DCCO > DPCO ; then DCRICO < 0.
(3) If DCNT > DPNT ; then DCRINT < 0.
Here DCTV and DPTV are calculated dose and prescribed dose of TV. DCCO
and DPCO are calculated dose and constraint dose of CO, and DCNT and
DPNT are calculated dose and constraint dose of NT. DCRITV , DCRICO ,
and DCRINT are CRI values of TV, CO, and NT, respectively. If dose
differences are defined as DDTV DCTV  DPTV ; DDCO DCCO  DPCO ,
and DDNT DCNT  DPNT , for TV, CO, and NT, respectively, then these
rules can be further simplified as:
(1) If DDTV < 0, then DCRITV > 0.
(2) If DDCO > 0, then DCRICO < 0.
(3) If DDNT > 0, then DCRINT < 0.
With the simplified form of if-then rules, the experience of
human planners in trial-and-error procedure of manual beam
angle optimization is digitized.
2.3. Fluence map optimization
The objective function used in this study is defined as:
XXX  2
f x wijk Dijk  P ijk 2
i j k
P
Dijk Nn1 An;ijk xn is the calculated dose at voxelijk. An;ijk is the dose
deposition coefficient, i.e., the dose per unit intensity of nth pencil
beam at voxelijk. For simplicity, primary-only dose of 6 MV photon
beam at each ray is used for dose calculation and lateral scattering
is omitted. Pijk and wijk are prescription (or constraint) dose and
weighting factor at voxelijk which are defined as:
4 H. Yan, J.-R. Dai / Physica Medica xxx (2016) xxxxxx
Fig. 3. The illustration of a fuzzy inference system.
8 8
< PTV ; if i; j; k 2 XTV
> < W TV ; if i; j; k 2 XTV
> 2.4. Beam angle optimization
pijk PCO ; if i; j; k 2 XCO ; wijk W CO ; if i; j; k 2 XCO
>
: >
: With FIS model established beam configuration of an IMRT plan
PNT ; if i; j; k 2 XNT W NT ; if i; j; k 2 XNT
Here, XTV , XCO and XNT denote the volumes of TV, CO, and NT,
respectively. PTV denotes the prescription doses for TV, and its value
is set to 100. PCO and PNT denote constraint doses for CO and NT, and
their values are usually set to 0. W TV , W CO ; and W NT denote the
weighting factors for TV, CO, and NT, respectively. In this study,
they are 100, 100, and 20 for TV, CO, and NT, respectively. The min-
imization of the objective function under the constraint of xn P 0 is
written as:
minff xg
x 3
subject to xn P 0; 8n:
Eq. (3) can be solved by the fast-monotonic-descent (FMD)
method developed by Li, which is an iterative gradient technique
for the quadratic objective function [30]. It should note that those
traditional gradient-based inverse planning optimization algo-
rithms could be available in this study.
Fig. 4. Flowchart of intelligence guided approach for beam angle optimization.
Please cite this article in press as: Yan H, Dai J-R. Intelligence-guided beam angle optimization in treatment planning of intensity-modulated radiation
therapy. Phys. Med. (2016), http://dx.doi.org/10.1016/j.ejmp.2016.06.005
can be automatically determined by the intelligence guided
approach as illustrated in Fig. 4. At first, the initial plan is created
 
with a single beam. The dose differences DDiTV ; DDiCO ; DDiNT
between actual and prescription/constraint doses of the interested
anatomical structures at ith iteration are calculated for the input of
FIS. With process of fuzzy inference the output of FIS, the adjust-
 
ment of CRI DCRIiTV ; DCRIiCO ; DCRIiNT , at ith iteration is obtained
for the interested anatomical structures. According to the adjust-
ment of CRI, the ranges of CRI for the candidate beams are deter-
h i h i
mined by CRIiTV ; CRIiTV DCRIiTV , CRIiCO ; CRIiCO DCRIiCO , and
h i
CRIiNT ; CRIiNT DRIiNT for the interest anatomical structures. Here,
CRIiTV ; CRIiCO ; and CRIiNT are CRI values at ith iteration. With the
ranges of CRI obtained, the beams within these ranges are found
according to the lookup tables of CRI previously established. For
example, the candidate beams for TV can be found by checking
the beams with their CRI values within range of
h i
CRIiTV ; CRIiTV DCRIiTV . For three types of interested anatomical
structures, multiple sets of candidate beams are obtained and their
intersection set is the final set of candidate beams. If there is no
intersection sets found, the searching range of CRI values could
be extended properly. Among these beams the one satisfying the
common constraints (such as the minimal interval between beams,
no opposite beam, etc.) and with the largest value of GCRI is chose
to replace the current beam. Such process continues until no can-
didate beam is found for the first beam. Then, the next beam is
added to the IMRT plan and optimized in the same way as the first
beam. The new beam is added to the IMRT plan until the maximum
number of beams is reached.
2.5. Clinical case study
Two clinical cases were examined by the intelligence guided
approach. Both cases were spine cases. In the first case, TV (tumor)
is surrounded by three COs (spinal cord, left kidney, and right kid-
ney). Spinal cord partially touches tumor. The default beam angles
are 30, 60, 90, 120, 150, 210, 240, 270, 300, 330, and 360
for the 11-beam IMRT plan. The prescription and important factor
for TV are 100 and 100. The tolerance dose and weighting factor for
critical organ are 30 and 80, while the tolerance dose and weight-
ing factor for normal tissue are 20 and 30. In the second case, a
concaved TV (tumor) is surrounded by three COs (spinal cord, left
kidney, and right kidney). Spinal cord was almost completely
 H. Yan, J.-R. Dai / Physica Medica xxx (2016) xxxxxx
encapsulated by tumor. The default beam angles are 40, 80, 120,
160, 200, 240, 280, 300, and 360 for the 9-beam IMRT plan.
The setting of prescription doses and important factors for TV,
COs, and NT is the same as that of the first case. With the applica-
tion of FIS, the beam angles were automatically determined and
compared with the default beam setting. Note that in intelligence
guided approach the opposite beam is prohibited and the minimal
beam intervals for 11-beam plan and 9-beam plan are 30 and 40,
respectively. The dose distribution and DVH are main tools in eval-
uating dose quality of IMRT plans.
3. Results
For the 11-beam IMRT plan, the beam settings chosen by
default and automated methods are shown in Fig. 5a and b. The
plot of CRI values as a function of beam angle varied from 0 to
360 is shown in Fig. 6. The values of CRI for TV, CO, and NT are rep-
resented by solid diamond, solid square, and solid triangle, respec-
tively. The beam settings resulted by default and automated
methods are labeled by plan A and plan B, respectively. For the
selected beams in plan A, the mean values of CRI are 0.28, 0.26,
and 0.66 for TV, CO, and NT, respectively. While for the selected
beams in plan B the mean values of CRI are 0.34, 0.18, and 0.70
for TV, CO, and NT, respectively. For TV, the mean value of CRI
resulted by automated method is more than that by default
method. For CO, the mean value of CRI resulted by automated
method is less than that by default method. For NT, the mean value
of CRI by automated method is more than that by default method
slightly. The dose distributions resulted by both methods are com-
pared as shown in Fig. 7. The dose distribution is displayed with
the dose level of 10 Gy. It is observed the dose distributions of both
plans are comparable. The DVHs of target volume and normal tis-
sue of both plans are shown in Fig. 8a and b. They are almost the
same for both plans. The DVHs of three COs resulted by both meth-
ods are compared as shown in Fig. 8ce. The DVH of CO1 (spinal
cord) is improved by automated method comparing with that by
default method. The DVH of CO2 (left kidney) and CO3 (right kid-
ney) resulted by both methods are comparable.
For 9-beam IMRT plan, two beam settings resulted by default
and automated method are compared as shown in Fig. 9. The plot
of CRI values as a function of beam angle varied from 0 to 360 is
shown in Fig. 10. For the selected beams in plan A the mean values
of CRI are 0.33, 0.39, and 0.41 for TV, CO, and NT. While for the
selected beams in plan B the mean values of CRI are 0.42, 0.25,
Fig. 5. The beam setting resulted by default method (a) and automated method (b) in clinical case 1.
Please cite this article in press as: Yan H, Dai J-R. Intelligence-guided beam angle optimization in treatment planning of intensity-modulated radiation
therapy. Phys. Med. (2016), http://dx.doi.org/10.1016/j.ejmp.2016.06.005
5
and 0.36 for TV, CO, and NT, respectively. For TV the mean value
of CRI by automated method is more than that by default method.
For CO and NT, the mean value of CRI result by automated method
is less than that by default method. The dose distributions of both
plans are compared as shown in Fig. 11. It is observed there is less
doses of CO2 (left kidney) and CO3 (right kidney) resulted by auto-
mated method than those resulted by default method. The DVHs of
target volume and normal tissue resulted by both methods are
shown in Fig. 12a and b, and they are almost the same. The DVHs
of three COs resulted by both methods are compared as shown in
Fig. 12ce. The DVH of CO1 (spinal cord) is improved by automated
method than that by default method. The DVHs of CO2 (left kidney)
and CO3 (right kidney) are both improved by automated method
than that by default method.
4. Discussions
An intelligence guided approach was developed for BAO of IMRT
treatment planning. This method employed CRI as an ranking func-
tion for those beams which could be potentially selected for the
final plan. It was observed that beam angle with smaller CRI values
of critical organs could result in better dose sparing of critical
organ. In our study, the improvement of dose sparing of critical
organs was directly related to the decrease of mean values of CRI
of the selected beams. In another word, dose of critical organ could
be reduced when beams with smaller values of CRI for critical
organs are selected. Other researchers also reported similar find-
ings using different characteristic data, such as mean organ-at-
risk data [17], score of beam orientation using pseudo beams-
eye-view [21,22]. Since dose sparing for critical organs and dose
conformity to tumor are both important, it is necessary to select
those beams with larger values of CRI for TV. Our study demon-
strates that the mean values of CRI for TV resulted by automated
method are more than that resulted by default method.
In clinical practice the fine-tuning process of beam angles is
dominated by human planners. In our study, to mimic this process
the logic rules are provided and their priorities are specified for dif-
ferent types of interested structures. In addition, the distance
between radiation source and interest organs is counted in the
computation of CRI. Both measures assure that the priority knowl-
edge of human planner can be conveyed in the intelligence guided
approach. As expected, the automated method resulted in an IMRT
plan with improved dose sparing for critical organs while compa-
rable dose conformity for TV. The mean values of CRI are larger
6 H. Yan, J.-R. Dai / Physica Medica xxx (2016) xxxxxx

Fig. 6. The CRI plot for TV, CO, and NT in clinical case 1 and the corresponding beam settings resulted by default method (a) and automated method (b).

Fig. 7. The dose distributions resulted by default method (a) and automated method (b) in clinical case 1.

for TV and smaller for COs (and NT) in the plan resulted by the
automated method. It agrees with the assumption that the best
beams are those ones with more influence on tumor and less influ-
ence on critical organs. The current inference mechanism is rela-
tive simple which only consists of few rules and less flexible for
different treatment scenarios. The capability of self-learning
method was demonstrated in our previous work which planning
parameter of inverse planning was automatically adjusted by a
neuro-fuzzy system [29]. It will be our future work that the config-
uration of fuzzy inference system could be learned from clinical
practicing data.
It should note that the beam setting found by the intelligence
guided approach may be not globally optimal. This is a drawback
of intelligence-based approach. The fuzzy inference system is
designed to mimic the process of human reasoning but less effec-
tive in term of mathematical optimality. Similar as trial-and-error
process performed by human planners, the final beam combination
obtained by the intelligence guided approach may be sub-optimal
but its dose distribution is superior to those resulted by default
equally spaced beam setting accepted for clinical treatment. In
another word, the beam set obtained by intelligence guided
approach may be not globally optimal but closer to those accepted
by physicians. With the introduction of FIS, the prior knowledge of
human planner can be easily processed by computer applications.
In dealing with different treatment sites, different sets of logic
rules can be provided for the best performance of FIS. In the future,
these rule sets can be automatically learned from clinical
practicing data. From the clinical point of view, the intelligence
guided approach is closest to the way which human planners rou-
tinely practiced in beam angle selection of treatment planning.
The intelligence guided approach is different from the other
existing BAO algorithms. Instead of decoupling beam angle selec-
tion from FMO as mentioned in the first category of BAO [9,21],
the intelligence guided approach presents a way that beam angles
and fluence map are optimized in an alternative manner. Beam
angles are set first then FMO is performed. In turn, beam angles
are adjusted based on the dose response of the resulting plan. Dif-
fering from the joint optimization of beam angle and fluence map
as mentioned in the second category of BAO [4,14,16,22], the
adjustment of beam angles in intelligence guided approach is dom-
inated by FIS instead of optimization algorithms. Beam score or the
other prior information such as BEVD is helpful in identifying bad
beam angles and avoiding spending valuable time on unnecessary
searching direction, but is less useful in guiding the searching pro-
cess along the most promising directions. With prior knowledge,
the efficiency of BAO algorithms could be improved but a large
number of plan optimizations are still unavoidable.
In contrast, FIS functions as a decision-maker and navigates
high-dimensional searching space skillfully with the embedded
prior knowledge. With FIS guided searching strategy, only few
searching iterations are needed and a clinically accepted dose plan
would be achieved within a few minutes. To achieve the plans with

Please cite this article in press as: Yan H, Dai J-R. Intelligence-guided beam angle optimization in treatment planning of intensity-modulated radiation
therapy. Phys. Med. (2016), http://dx.doi.org/10.1016/j.ejmp.2016.06.005
 H. Yan, J.-R. Dai / Physica Medica xxx (2016) xxxxxx 7

Fig. 8. The DVHs of (a) target volume, (b) normal tissue, (c) critical organ 1 (spinal cord), (d) critical organ 2 (left kidney), and (e) critical organ 3 (right kidney) resulted by
default method (Plan A) and automated method (Plan B) in clinical case 1.

Fig. 9. The beam setting resulted by default method (a) and automated method (b) in clinical case 2.

decent dose distributions as shown in Figs. 7 and 11, the total num-
bers of iterations are 150 and 90 for spinal case 1 (11 beams) and
case 2 (9 beams). The computation time would be approximately
10 min on a general purpose computer. The comparison between
intelligence-guide approach with the other existing approaches is
difficult because they employed different searching strategies,
patient cases, dose calculation algorithms, and computing plat-
forms. However, it is clear that our approach would be more effec-
tive if they are implemented in the same environment (with the
same hardware, patient case, dose calculation algorithm, and

Please cite this article in press as: Yan H, Dai J-R. Intelligence-guided beam angle optimization in treatment planning of intensity-modulated radiation
therapy. Phys. Med. (2016), http://dx.doi.org/10.1016/j.ejmp.2016.06.005
8 H. Yan, J.-R. Dai / Physica Medica xxx (2016) xxxxxx

Fig. 10. The CRI plot for TV, CO, and NT in clinical case 2 and the corresponding beam settings resulted by default method (a) and automated method (b).

Fig. 11. The dose distributions resulted by default method (a) and automated method (b) in clinical case 2.

computing platform) due to the application of human knowledge-
based inference system.
Current research focused on the feasibility of applying FIS in
BAO of treatment planning and two spinal cases are relatively sim-
ple for demonstration purpose. For those realistic cases with more
complex spatial distribution of interested anatomical structures,
more work should be done in the future. At first, the objective
function will be improved to incorporate dose-volume based con-
straint instead of dose based constraint since it is more effective to
limit high dose on OAR and normal tissues. Certain non-dosimetric
constraints will also be considered to limit beam angle which may
cause collisions (between gantry and couch/patient) and go
through prosthesis region. Second, more realistic patient cases will
be tested. For real case, more anatomical structures with irregular
shape and spatial distributions will be dealt with by FIS. Since
more anatomical structures included, the parameters related to
the configuration of FIS such as, membership function, inference
rules, etc. will be extended considerably, which will result in pro-
longed time in setting up the configuration of FIS. Therefore, a
more efficient mechanism to learn these parameters is crucial. As
demonstrated in our previous study, the neuro-fuzzy system will
be a good solution for this issue. Third, the intelligence guided
approach will be applied to non-coplanar IMRT. Since couch rota-
tion is included in the plans with non-coplanar beams, the solution
space would be expended exponentially. For example, if the beam
interval of gantry and couch rotation is 1, the total number of can-
didate beams is 64,800. It would be difficult for the intelligence
guided approach to resolve such issue within a few minutes. How-
ever, it could be resolved by multi-resolution search strategy [10].
In this method, the problem was initially resolved by beams with
larger spacing. Then beam spacing was reduced and candidate
beams were sough within the selected beams in the previous res-
olution. This process will continue until the finest beam spacing is
reached.
5. Conclusion
An intelligence guided approach for BAO of IMRT treatment
planning was developed and tested by two clinical related cases.
Based on our results, the dose sparing of COs and NT are improved
with beam configuration resulted by the automated method while
the dose conformity of TV are comparable to those resulted by the
default method with equally spaced beam setting. The mean values
of CRI for TV are increased and the mean values of CRI for COs (and
NT) are decreased by the automated method while comparing with
those resulted by the default method. The introduction of FIS pro-
vided a tool to convert prior knowledge of human planner to logic

Please cite this article in press as: Yan H, Dai J-R. Intelligence-guided beam angle optimization in treatment planning of intensity-modulated radiation
therapy. Phys. Med. (2016), http://dx.doi.org/10.1016/j.ejmp.2016.06.005
 H. Yan, J.-R. Dai / Physica Medica xxx (2016) xxxxxx
Fig. 12. The DVHs of (a) target volume, (b) normal tissue, (c) critical organ 1 (spinal cord), (d) critical organ 2 (left kidney), and (e) critical organ 3 (right kidney) by default
method (Plan A) and automated method (Plan B) in clinical case 2.
rules which can be utilized by computer program. It is promising to
apply the intelligence-guided method in the automation of BAO in
IMRT treatment planning.
Acknowledgements
The authors gratefully acknowledge Ye-Xiong Li and Yi-Ming
Hu at Cancer Hospital of Chinese Academy of Medical Sciences
for their generous support. The Author also wants to thank Fang-
Fang Yin (Duke University Medical Center) and Huai-Qun Guan
(Saint Vincent Hospital) for their helps.
References
[1] Bortfeld T, Schlegel W. Optimization of beam orientation in radiation therapy:
some theoretical considerations. Phys Med Biol 1993;38:291304.
[2] Crooks SM, Pugachev A, King C, Xing L. Examination of the effect of increasing
the number of radiation beams on a radiation treatment plan. Phys Med Biol
2002;47:3485501.
[3] Pugachev A, Li JG, Boyer AL, Hancock SL, Le Q-T, Donaldson SS, Xing L. Role of
beam orientation optimization in intensity-modulated radiation therapy. Int J
Radiat Oncol Biol Phys 2001;50:55160.
[4] Djajaputra D, Wu Q, Wu Y, Mohan R. Algorithm and performance of a clinical
IMRT beam-angle optimization system. Phys Med Biol 2003;48:3191212.
[5] Gokhale P, Hussein EMA, Kulkarni N. Determination of beam orientations in
radiotherapy planning. Med Phys 1994;21:393400.
Please cite this article in press as: Yan H, Dai J-R. Intelligence-guided beam angle optimization in treatment planning of intensity-modulated radiation
therapy. Phys. Med. (2016), http://dx.doi.org/10.1016/j.ejmp.2016.06.005
9
[6] Das IJ, Cheng EC, Anderson PR, Movsas B. Optimum beam angles for the
conformal treatment of lung cancer: a CT simulation study. Int J Cancer
2000;90:35965.
[7] Rowbottom CG, Webb S, Oldham M. Improvements in prostate radiotherapy
from the customization of beam directions. Med Phys 1998;25:11719.
[8] Das S, Cullip T, Tracton G, Chang S, Marks L, Anscher M, Rosenman J. Beam
orientation selection for intensity-modulated radiation therapy based on
target equivalent uniform dose maximization. Int J Radiat Oncol Biol Phys
2003;55:21524.
[9] Das SK, Marks LB. Selection of coplanar or non coplanar beams using three-
dimensional optimization based on maximum beam separation and
minimized nontarget irradiation. Int J Radiat Oncol Biol Phys 1997;38:64355.
[10] Wang X, Zhang X, Dong L, Liu H, Wu Q, Mohan R. Development of methods for
beam angle optimization for IMRT using an accelerated exhaustive search
strategy. Int J Radiat Oncol Biol Phys 2004;60(4):132537.
[11] Bangert M, Ziegenhein P, Oelfke U. Comparison of beam angle selection
strategies for intracranial IMRT. Med Phys 2013;40(1):011716.
[12] Pugachev AB, Boyer AL, Xing L. Beam orientation optimization in intensity-
modulated radiation treatment. Med Phys 2000;27:123845.
[13] Rowbottom CG, Khoo VS, Webb S. Simultaneous optimization of beam
orientations and beam weights in conformal radiotherapy. Med Phys
2001;28:1696702.
[14] Stein J, Mohan R, Wang X, Bortfeld T, Wu Q, Preiser K, Ling C, Schlegel W.
Number and orientations of beams in intensity modulated radiation
treatments. Med Phys 1997;24:14960.
[15] Haas OCL, Burnham KJ, Mills JA. Optimization of beam orientation in
radiotherapy using planar geometry. Phys Med Biol 1998;43:217993.
[16] Li Y, Yao J, Yao D. Automatic beam angle selection in IMRT planning using
genetic algorithm. Phys Med Biol 2004;49:191532.
[17] Souza WDD, Meyer RR, Shi L. Selection of beam orientations in intensity-
modulated radiation therapy using single-beam indices and integer
programming. Phys Med Biol 2004;49:346581.
10 H. Yan, J.-R. Dai / Physica Medica xxx (2016) xxxxxx
[18] Wang C, Dai J, Hu Y. Optimization of beam orientations and beam weights for
conformal radiotherapy using mixed integer programming. Phys Med Biol
2003;48:406576.
[19] Soderstrom S, Brahme A. Selection of suitable beam orientations in radiation
therapy using entropy and Fourier transform measures. Phys Med Biol
1992;37:91124.
[20] Das SK, Bell M, Markers LB, Rosenman J. A preliminary study of the role of
modulated electron beams in intensity modulated radiotherapy, using
automated beam orientation and modality selection. Int J Radiat Oncol Biol
Phys 2004;59:60217.
[21] Pugachev A, Xing L. Pseudo beam eye-view as applied to beam orientation
selection in intensity-modulated radiation therapy. Int J Radiat Oncol Biol Phys
2001;51:136170.
[22] Pugachev A, Xing L. Incorporating prior knowledge into beam orientation
optimization in IMRT. Int J Radiat Oncol Biol Phys 2002;54:156574.
[23] Rocha H, Dias J, Ferreira B, Lopes M. Beam angle optimization for intensity-
modulated radiation therapy using a guided pattern search method. Phys Med
Biol 2013;58(9):293953.
Please cite this article in press as: Yan H, Dai J-R. Intelligence-guided beam angle optimization in treatment planning of intensity-modulated radiation
therapy. Phys. Med. (2016), http://dx.doi.org/10.1016/j.ejmp.2016.06.005
[24] Rowbottom C, Webb S, Oldham M. Beam-orientation customization using an
artificial neural network. Phys Med Biol 1999;44(9):225162.
[25] Amit G, Purdie T, Levinshtein A, Hope A, Lindsay P, Marshall A, Jaffray D, Pekar
V. Automatic learning-based beam angle selection for thoracic IMRT. Med Phys
2015;42(4):19922005.
[26] Yan H, Yin F, Guan H, Kim JH. Fuzzy logic guided inverse treatment planning.
Med Phys 2003;30:305260.
[27] Yan H, Yin F, Guan H, Kim JH. AI-guided parameter optimization in inverse
treatment planning. Phys Med Biol 2003;48(21):356580.
[28] Yan H, Yin F, Willett C. Evaluation of an artificial intelligence guided inverse
planning system: clinical case study. Radiother Oncol 2007;83(1):7685.
[29] Stieler F, Yan H, Lohr F, Wenz F, Yin F. Development of a neuro-fuzzy technique
for automated parameter optimization of inverse treatment planning. Radiat
Oncol 2009;4:39.
[30] Li RP, Yin FF. Optimization of inverse treatment planning using a fuzzy weight
function. Med Phys 2000;27:691700.