Home Search Collections Journals About Contact us My IOPscience

A new sparse optimization scheme for simultaneous beam angle and fluence map

optimization in radiotherapy planning

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2017 Phys. Med. Biol. 62 6428

(http://iopscience.iop.org/0031-9155/62/16/6428)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 207.162.240.147
This content was downloaded on 21/07/2017 at 02:45

Please note that terms and conditions apply.

 Institute of Physics and Engineering in Medicine Physics in Medicine & Biology

Phys. Med. Biol. 62 (2017) 6428–6445 https://doi.org/10.1088/1361-6560/aa75c0

A new sparse optimization scheme for

simultaneous beam angle and fluence map
optimization in radiotherapy planning
Hongcheng Liu, Peng Dong and Lei Xing
Department of Radiation Oncology, Stanford University School of Medicine, 875
Blake Wilbur Drive, Stanford, CA 94305-5847, United States of America

E-mail: lei@stanford.edu

Received 22 September 2016, revised 21 April 2017

Accepted for publication 30 May 2017
Published 20 July 2017

Abstract
2,1-minimization-based sparse optimization was employed to solve the beam
angle optimization (BAO) in intensity-modulated radiation therapy (IMRT)
planning. The technique approximates the exact BAO formulation with
efficiently computable convex surrogates, leading to plans that are inferior
to those attainable with recently proposed gradient-based greedy schemes.
In this paper, we alleviate/reduce the nontrivial inconsistencies between the
2,1-based formulations and the exact BAO model by proposing a new sparse
optimization framework based on the most recent developments in group
variable selection.
We propose the incorporation of the group-folded concave penalty (gFCP)
as a substitution to the 2,1-minimization framework. The new formulation is
then solved by a variation of an existing gradient method. The performance of
the proposed scheme is evaluated by both plan quality and the computational
efficiency using three IMRT cases: a coplanar prostate case, a coplanar head-
and-neck case, and a noncoplanar liver case. Involved in the evaluation are
two alternative schemes: the 2,1-minimization approach and the gradient
norm method (GNM).
The gFCP-based scheme outperforms both counterpart approaches.
In particular, gFCP generates better plans than those obtained using the
2,1-minimization for all three cases with a comparable computation time.
As compared to the GNM, the gFCP improves both the plan quality and
computational efficiency.
The proposed gFCP-based scheme provides a promising framework for
BAO and promises to improve both planning time and plan quality.

1361-6560/17/166428+18$33.00  © 2017 Institute of Physics and Engineering in Medicine  Printed in the UK 6428
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Keywords: inverse planning, beam angle selection, IMRT, folded concave

penalty, sparse optimization

(Some figures may appear in colour only in the online journal)

1. Introduction

Beam angle optimization (BAO) in treatment planning is known to be an NP-hard problem

(Azizi Sultan 2006) and is among the hardest numerical problems to which efficient solution
schemes are unknown. Clinically, to generate an acceptable set of beam angles, beams are
placed manually via trial and error. (Meta)-heuristics such as genetic algorithms and simulated
annealing have been proposed for BAO (Ezzell 1996, Pugachev et  al 2001, Li et  al 2005,
Aleman et al 2008, Rocha et al 2013). However, due to the prohibitively large search space,
these heuristics tend to be slow to converge. As an alternative scheme, different ranking crite-
ria were proposed for beam angles to be sequentially added to the plan till an acceptable plan
quality is achieved (Woudstra and Storchi 2000, Pugachev and Xing 2001a, 2001b, Meedt
et al 2003, Meyer et al 2005, Vaitheeswaran et al 2010). Depending on the approximation
involved in the model, this type of scheme may also suffer from suboptimality.
Sparse optimization with total variation regularization was first studied by Zhu and Xing
(2009) to achieve highly conformal dose distribution while reducing the number of segments
in IMRT. Then, Jia et al (2011) proposed a sparse optimization-based scheme called 2,1-mini-
mization, which is an adaptation of the group-lasso method (Yuan and Lin 2005) from the
feature selection schemes in statistical learning and signal processing to the problem of select-
ing beam angles. However, the group-lasso and its variations often result in biasedness in
signal recovery (Huang et al 2012). Mathematically, it is a strictly increasing function, which
penalizes both significantly and insignificantly contributing groups indifferently. Therefore,
its adaptation to beam angle selection probably carries the same problem. While we want
to ensure that the useful beams that contribute significantly to the provision of a conformal
dose distribution are less penalized so that they can remain active (nonzero), the group-lasso
imposes a nontrivial penalty to these beams. Indeed, the very same paper by Jia et al (2011)
that proposes adaptive 2,1-minimization invokes an (unpenalized) fluence map optimization
to generate the final treatment plan, after completion of the 2,1-based beam angle selection.
In this work, we propose a group-folded concave penalty (gFCP) as an alternative scheme
to the 2,1 penalty. The gFCP is shown to be an unbiased scheme for signal recovery and entails
better feature selection properties than 2,1 (Fan and Li 2001, Zhang 2010, Huang et al 2012,
Liu et al 2017). Specifically, Liu et al (2017) showed that gFCP may result in fewer errors in
approximating the true signals as compared to 2,1. Moreover, numerical experiments from the
literature indicate that gFCP outperforms 2,1 in generating significantly contributing (groups
of) signals. Such an advantage is likely to be transferable to treatment planning through the
adaptation of gFCP to BAO. Indeed, our test results for three different cases indicate a sub-
stantial improvement in plan quality and computational efficiency.
The rest of the paper is organized as follows. In the next section, we introduce the gFCP.
We illustrate its unbiasedness and potential advantages over the 2,1 penalty. Despite its non-
convexity, we present a fully polynomial-time approximation scheme to solve for a local solu-
tion. Section 3 then presents three IMRT cases to test the proposed gFCP-based formulations.
Specifically, we compare gFCP with the 2,1-minimization model and gradient norm method
(GNM), proposed by Bangert and Unkelbach (2016), for a coplanar prostate case, a coplanar
head-and-neck case, and a noncoplanar liver case. In all three cases, the gFCP is significantly

6429
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

better in terms of the resultant plan quality than the 2,1-minimization model, and is substanti­
ally faster than the GNM, even though the latter is claimed to be the ‘accelerated scheme’
(Bangert and Unkelbach 2016). We attribute the success to (i) the proposed gFCP scheme
being a better approximation to the exact BAO formulation than the 2,1-minimization model,
and (ii) gFCP allowing the simultaneous determination of both the beam angles and the flu-
ence map by (locally) solving the formulation once, while the GNM involves the repetitive
computation of a fluence map optimization (FMO) model. The computational overhead to
achieve a local solution is much less than the global optimization.

2. Methods

2.1.  Folded concave penalty-based scheme for beam angle selection

Consider a generic modeling framework for radiation therapy inverse planning of the form
min F (w)
w
(1)
s.t. w ∈ W,
where F is the optimization criteria to be minimized and W ⊆ N+ denotes the feasible region
defined by any set of constraints for the beamlet intensities, denoted by w, with  and N
denoting the real space and the number of beamlets, respectively. This formulation is gener-
ally enough to cover most, if not all, of the inverse planning formulations discussed in the
literature. In this paper, we focus on a special case of F(w) in the form of the quadratic one-
sided penalties as follows:
 2 2
F (w)
(2) := αn (max {0, Tn − dn (w)}) + βn (max {0, dn (w) − Tn }) ,
n

where n is the voxel index, Tn denotes the target dose level, αn and βn represent the penalty
weights for underdosing and overdosing costs, and dn (w) represents the dose level at the nth
voxel with the beamlet intensity equal to w. Nonetheless, we would like to emphasize that our
proposed schemes are flexible in choosing the objective function F .
Suppose that prior to BAO, a pool of candidate beam angles B := {Bj : j ∈ Θ} is given,
where Bj and Θ denote the set of beamlets at the jth angle and the index set of beam angles,
respectively. The BAO here proceeds by removing unfavorable beams. An exact BAO form­
ulation, which is more general than the existing mixed integer formulations for exact BAO, as
presented by Bangert and Unkelbach (2016) and Yang et al (2006), has the following form:
min F (w)
w
(3)
s.t. w ∈ W
Card(D (w))  C,
where D (w) := { j ∈ Θ : wi > 0 for some i ∈ Bj } is the index set of the beam angles that
has at least one strictly positive beamlet intensity, i.e. the index set of beam angles in use.
Card(D (w)) denotes the cardinality of D(w), and C is the allowed number of active beam
angles. Due to the presence of the cardinality constraint, this is enforced to select the beams
that contribute the most to the plan quality so that the number of active beams satisfies the
prescribed upper bound C . The presence of the beam angle constraint, as in the last constraint
of (3), introduces nonconvexity to the model.
 
  
 2 
Noticing that Card(D (w)) =  wi : j ∈ Θ  , an equivalent formulation to (3) is
 i∈Bj 
0
6430
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Figure 1.  The exact penalty (a), the 2,1 penalty (b) and the gFCP (c).

  
  
 
min sup F (w) + γ ·  w2i : j ∈ Θ  − C
(4)
w γ0  i∈Bj 
0
s.t. w ∈ W.
Problems (3) and (4) have been widely studied in applications of signal processing and sparse
recovery. Due to the nonconvexity of both formulations, a common tractable solution approach
is to approximate (4) with an 2,1-based formulation
  2
min F (w) + γ · wi
w
(5) j∈Θ i∈Bj

s.t. w ∈ W,
for some parameter γ subject to tuning. There are two differences between (4) and (5): (i)
in model (4), γ is a decision variable in the lower-level supreme problem, while in model
(5), γ is a prefixed parameter; (ii) model (5) approximates the cardinality-based penalty
 
    
 2  2
 wi : j ∈ Θ  with a group 2,1 penalty j∈Θ i∈Bj wi . Based on (5), Jia et al
 i∈Bj 
0
(2011) first introduced the 2,1-minimization approach to select the beam angles.
Similar to (5), for efficient computation, we propose to approximate the exact beam angle
optimization formulation using the group-folded concave penalty (gFCP). The gFCP concave
penalty is a recent variable selection tool used in statistical learning to screen out insignifi-
cantly contributing groups of variables. Analogously, in the beam angle optimization problem,
beamlets at the same angle naturally form a group, and the use of the gFCP can screen out the
angles that are not contributing enough to the treatment quality.
Specifically, the proposed model for the simultaneous optimization of the inverse plan and
beam angle is written as
min F (w) + γ · P (w, B) ,
(6)
w∈W

where, again, parameter γ > 0 is subject to tuning and converts the units. Instead of minimiz-
ing the number of beam angles, a surrogate formulation  is used to penalize the use of each
beam angle. The penalty is formulated as P (w, B) := p(w, Bj ), where
j∈Θ
6431
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Figure 2.  The pseudocode for algorithm 1 and the bisection algorithm for BAO.


 w2i [aλ−θ]
i∈Bj
p (w, Bj ) : = 0 a
+
dθ
   2
 i∈B wi 
 
 λ w2i − 2aj , if w2i  aλ
i∈Bj i∈Bj
= 

 aλ2
w2i > aλ
 2 , if (7)
i∈Bj

for any w  0, where [ · ]+ denotes the larger number between 0 and the argument, and a
and λ are two tuning parameters for the penalty function. The intuition of this penalty is to
approximate the stepwise function of penalizing the number of beamlets by a continuously

6432
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Table 1.  Experiment settings and parameters.

Cases Prostate Head and neck Liver

Modality Coplanar IMRT Coplanar IMRT Noncoplanar IMRT
Beamlet size (cm  ×  cm) 1  ×  1 1  ×  1 1  ×  1
Voxel size 3.00  ×  3.00  ×  2.50 1.86  ×  1.86  ×  2.50 3.00  ×  3.00  ×  2.50
(mm  ×  mm  ×  mm)
Initial candidate beam 18 equispaced beams 18 equispaced beams 56 prefixed candidate
angles combinations of gantry
and couch angles

differentiable function, as illustrated in figure 1. This penalty can be used in most inverse plan-
ning formulations as an additional term in the objective function.
The proposed gFCP formulation results in a better approximation to (4) than the 2,1-mini-
mization formulation, as illustrated by figure 1, where three different penalty functions are
plotted. In this figure, we consider only the value of the penalty function for one group of
beamlets, that is Θ = {1}, and the beamlets share the same gantry angle. Then, the cardinal-
ity-based penalty is equivalent to a stepwise function that takes a value of 1 when any of the
beamlets have a nonzero intensity, and a value of 0, otherwise. Figure 1(a) presents such a
stepwise function. In particular, the 2,1 penalty is strictly increasing, while the gFCP becomes

2
constant once the magnitude of the beamlet intensity from the same angle, i∈Bj wi ,
becomes larger than a certain threshold. As also seen in the figure, this property makes it a
tighter approximation to the cardinality-based penalty than the 2,1 penalty. It has also been
reported in the literature that local solutions to FCP-penalized formulations often achieve suf-
ficient quality in approximating the cardinality-based function (Liu et al 2017).
The parameters a and λ determine the shape of gFCP. Specifically, the partial derivative of

2
the penalty dP(w,B)
dwi =  λwi
2
− wi
a
if 0 < i ∈{i :i ∈Bj , for j: i∈Bj } (wi ) < aλ
i ∈{i :i ∈Bj , for j: i∈Bj } (wi )

and dP(w,B)
dwi = 0, otherwise. Since the first derivative (namely the slope) of P strictly increases
in a, this then determines how quickly the slope of the penalty function decreases to zero for a
fixed λ. Parameter λ determines the trade-off between the penalty and the fidelity term, as (6)
can be viewed as a bi-objective optimization that minimizes these two functions simultane-
ously. Moreover, aλ is the threshold for the penalty function to become a constant.

2.2.  The solution algorithm

We elect to compute a local solution to (6) using a gradient-based algorithm, which does
not incur a significant increase in the computational burden to the original inverse planning
model. Specifically, we adopt a gradient method adapted from Wang et al (2014) and Nesterov
(2013). The idea of the gradient method stems from an observation of the following equality:
P (w, B) = G1 (w, B) + G2 (w, B)
(8)

where G1 (w, B) := j∈Θ g1,j (w, B) and
  2

 i∈Bj wi  2

 − 2a , if wi  aλ
i∈Bj
g1,j (w, B) =
(9)   2 ,
 aλ2

 2 −λ w2i , if wi > aλ
 i∈Bj i∈Bj

6433
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

 
as well as G2 (w, B) := j∈Θ λ i∈Bj w2i . The purpose of separating P into two terms is
that G1 (w, B) is continuously differentiable in w, while G2 (w, B) is not differentiable wher-

2
ever i∈Bj wi = 0 for any j. The gradient method then handles the two terms differently in
the computation. The pseudocode of the algorithm is given in figure 2.
In particular, when the feasible region W consists of only bound constraints, i.e. when
N
W := {w : 0  w  U} for some U ∈ (0, ∞) , formulation (10) admits a closed-form equiv-
alent reformulation as
   

min T k
 −sλ     
i
· wki − s · Hi wk , B , Ui , Tik   s · λ
wk+1
i :=  Tik 
+


0, otherwise

k
 k  k   
 ka short-hand notation Ti := wi − s · Hi w , B : i ∈ Bj for j : i ∈ Bj , where
with
Hi w , B and Ui denote the ith dimension of H and U . A derivation of this closed form is
given in the appendix. In algorithm 1, w0 is set to be the solution to the FMO problem with all
candidate beams allowed to be active.

2.3. Evaluation

We implemented algorithm 1 using MatLab and tested the approach on a coplanar prostate
case, a coplanar head-and-neck case and a noncoplanar liver case. For the first two cases, we
employed CERR (Deasy et al 2003), which is a computational environment for radiotherapy
research, with default parameters (beam energy is 6 MV) to generate the dose-influence matri-
ces. The third case was from Craft et al (2014), with a beam energy of 6 MV and ready-to-
use dose-influence matrices in CERR IMRTP GUI, which is an IMRT planning graphic user
interface within CERR. In this case, we followed Craft et al (2014) to use a normalized dose,
so as to be consistent with the original example in their paper. Some other experiment settings
are summarized in table 1. Specifically, the pools of candidate beam angles are reported in the
row of ‘initial candidate beam angles’. The active beam angles are then selected from these
pools by the tested schemes.
For each of the cases, we first solved an inverse planning problem as in model (6) with
all candidate beams after careful tuning of the parameters, αn and βn, through trial and error,
to obtain a clinically acceptable plan. The same set of parameters was then used to test all
the BAO schemes. Some straightforward tuning of parameter λ was involved to ensure that
the number of active beam angles equaled exactly the prescribed number. In particular, we
elected to use only eight active beams for all three cases. All experiments were conducted
using MatLab in a single thread on a PC with Intel Core i7 3.40 GHz CPU and 16 GB RAM.

3. Results

3.1.  Prostate coplanar IMRT case

In figure 3, we compare the dose distributions generated by algorithm 1 with a different num-
ber of beam angles from both the axial and sagittal orientations. It can be seen that the pro-
posed scheme is capable of selecting the beam angles that satisfy the prescribed number of
beams. Although we anticipate a decay in the plan quality with the reduction in the allowed
number of beams, the output plan quality is maintained at a reasonable level. In particular,

6434
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Figure 3.  The sagittally ((a), (c), (e) and (g)) and axially ((b), (d), (f) and (h)) oriented
dose distribution with different beam numbers generated by gFCP.
6435
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Figure 4.  The DVHs of (a) the PTV, (b) bladder, (c) rectum, and (d) body in the plans
generated by algorithm 1 with a different number of beams.

from subplots (a), (c), (e) and (g) of the figure, the critical structures including the bladder and
rectum are well spared from radiation, even when only four beams are allowed in the treat-
ment. The same pattern is observed in figure 4, which presents the DVHs of the PTV, bladder,
rectum and body in the plans generated by the proposed approach with a different prescription
of beam numbers. As can be seen from this figure, although it may be possible to improve dose
distribution in the low-dose regions by reducing the number of beam angles, increasing the
number makes it easier to improve the DVHs of high or moderate doses.
We further compare the proposed scheme with two alternative schemes, the
2,1-­minimization approach (Jia et al 2011) and the gradient norm method (GNM) proposed
by Bangert and Unkelbach (2016), with the number of beam angles prescribed to be eight.
The comparison is presented in figure  5, where it is seen that all three schemes lead to
similar PTV coverage (the maximal difference is less than 0.01 or 1%). Meanwhile, gFCP
results in much better sparing of the critical structures. For example, the percentage of the
rectum volume receiving a dose above 10 Gy is 27%, 38% and 60% for gFCP, 2,1-minimi-
zation, and GNM, respectively. Thus the plan generated by gFCP significantly improves the

6436
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Figure 5.  The DVHs (of prostate and rectum in (a), body in (b), and bladder in (c)) of
the outputs from gFCP (solid lines), 2,1 minimization (dot-dashed lines), GNM (dotted
lines), and FMO all with candidate beam angles (dashed lines).

6437
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Figure 6.  Dose distributions corresponding to GNM (a), 2,1 minimization (b), and
gFCP (c).

Table 2.  Computation time and selected beam or gantry/couch angles.

Scheme gFCP 2,1 GNM λ a

Prostate coplanar Gantry angles 40, 60 0, 20 0, 20 160 0.1
80, 100 40, 60 40, 140
240, 260 180, 200 180, 300
280, 300 32, 340 320, 340
CPU time 16.3 min 15.9 min 28.5 min
Head and neck coplanar Gantry angles 0, 80 0, 60 20, 40 100 0.1
100, 240 80, 100 80, 120
260, 280 240, 260 240, 280
300, 340 280, 300 320, 340
CPU time 31.0 min 30.6 min 89.4 min
Liver noncoplanar Gantry/couch 286/0 279/15 18/−  32 140 0.1
angles 296/17 296/17 198/32
306/32 306/32 270/0
314/13 314/13 296/17
324/32 324/32 306/32
32/−  90 32/−  90 314/13
330/59 330/59 324/32
342/32 342/32 330/59
CPU time 10.7 min 11.4 min 23.1 min

dose distribution obtained using the alternative schemes. Similar patterns can be observed in
the dose distribution presented in figure 6, where gFCP selects the beam angles of the most
reasonable orientations. In contrast, the resulting choices of beam angles by the 2,1-mini-
mization approach and the GNM are insufficient to avoid unnecessary doses on the critical
structures, the rectum and bladder.
Table 2 reports the output beam angles from different schemes and the total computa-
tion time. For gFCP, since the initial solution for algorithm 1 is chosen to be the output
from FMO with all candidate beam angles, the reported CPU time includes both the ini-
tialization and the runtime of algorithm 1. Likewise, for 2,1-minimization, the reported
CPU time is the sum of the runtime for the first step of the beam angle selection and that
for the second step of FMO with the selected beams. As seen in this table, the GNM is the
slowest, while gFCP and 2,1 incur similar computational costs. In summary, the proposed

6438
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Figure 7.  DVHs (of PTVs and body in (a), optic nerves in (b), parotids and mandible
in (c), and brainstem and spinal cord in (d)) of the outputs from gFCP (solid lines), 2,1
minimization (dot-dashed lines), GNM (dotted lines) and FMO with all the candidate
beam angles (dashed lines).

scheme generates the best balance between the plan quality and the computational cost
than the existing schemes involved in the comparison. The computation took approxi-
mately 3.5 GB memory.

3.2.  Coplanar head and neck IMRT case

Figure 7 compares the DVHs generated by four different schemes: gFCP, 2,1, GNM and
FMO with all 18 beam angles. From the figure, we observe that the gFCP generates a very
competitive plan against the GNM and the FMO with all the candidate beam angles, as
all three schemes produce closely similar PTV coverage and critical structure sparing.
Specifically, for all three PTVs, the difference between the three approaches is smaller than
0.04 (or 3%), and the maximal differences for the DVHs of the brainstem and spinal cord are
no more than 0.05 (or 5%). In contrast, 2,1-minimization generates the lowest plan quality in
terms of the DVH. Despite this, the weighting factors for all four schemes are kept the same,
and 2,1-minimization results in undesirably high maximal doses to the spinal cord (50.9 Gy)
and brainstem (65.9 Gy).
Figure 8 further presents the dose distributions generated by gFCP, 2,1 and GNM. As illus-
trated in this figure, the dose distributions by gFCP and for GNM are closely similar; they
both appear to be superior to the dose distributions by 2,1 in sparing the nontarget regions. For
example, the nontarget area around the mandible receives a noticeably higher dose in the plan
generated by 2,1 than the other two plans, as we compare subplots (a), (d) and (g).
6439
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Figure 8. Dose distributions of the outputs from GNM ((a)–(c)), 2,1 minimization
((d)–(f)) and gFCP ((g)–(i)).

Again, table 2 reports the output beam angles from different schemes and the total time.
The GNM is significantly slower compared to both gFCP and 2,1, while the latter two incur
comparable computational costs. The computation took approximately 5 GB memory.

3.3.  Noncoplanar liver IMRT case

Figure 9 compares the DVHs of the plans generated by gFCP, 2,1 and GNM. As can be seen
from the figure, gFCP provides better PTV coverage than the other two alternative schemes
(significantly better than 2,1 and slightly more homogeneous than the GNM). For example, on
the PTV, the percentage volumes with a normalized dose of 0.9 units and above are 0.98, 0.96
and 0.83, respectively, for gFCP, GNM and 2,1-minimization. In terms of sparing the critical
structure, gFCP performs competitively against 2,1, and significantly better than the GNM,
especially in sparing the heart. For example, the percentage heart volumes with a normalized

6440
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Figure 9.  The DVHs of the outputs from gFCP (solid lines), 2,1 minimization (dot-
dashed lines), and GNM (dashed lines). The x-axis is the normalized dose (all dose
levels are divided by the maximum dose of the gFCP output). In subplot (b), there is no
data plotted beyond the normalized dose of 0.16 units.

dose of 0.23 units and above are 0.049, 0.074 and 0.041, respectively, for gFCP, GNM and
2,1-minimization.
In terms of computational efficiency, as reported in table 2, the GNM generates the highest
computational cost, while both gFCP and 2,1 yield much smaller costs. The computation took
approximately 7 GB memory.

4. Discussion

As pointed out by Bangert and Unkelbach (2016), in spite of the wide popularity of rotation
therapy such as VMAT, the problem of BAO remains essential in many treatment sites, espe-
cially when using noncoplanar beams and VMAT treatment using noncoplanar trajectories (Li
and Xing 2013, Xing and Li 2013). Computation time has been a persisting concern in solving
BAO, as traditionally many repetitions of FMO problems are solved in order to generate rea-
sonable choices of gantry angles. In addressing this issue, the 2,1-minimization scheme by Jia
et al (2011) and the GNM by Bangert and Unkelbach (2016) are state-of-the-art approaches.
In particular, the 2,1-minimization scheme employs techniques in sparse optimization for a
variable selection (Zhu et al 2008, Zhu and Xing 2009), which approximates the exact BAO

6441
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

formulation by a convex surrogate. As a result, the BAO admits tractable (approximate) solu-
tion algorithms.
This paper studies a new sparse optimization scheme that benefits from new advances in
sparse optimization. It has been shown that a recent group-sparsity-inducing penalty called
gFCP is superior to 2,1-based formulations for variable selection. Such superiority is trans-
ferrable to the BAO problem according to our experiments in three different IMRT treatment
cases, a coplanar prostate case, a coplanar head and neck case, and a noncoplanar liver case.
For all three cases, the gFCP-based formulation results in a significant improvement in terms
of DVHs, presumably because the gFCP is innately a better approximation to the exact BAO
formulation than the 2,1-based penalty.
Also involved in our numerical comparisons is the most recent BAO algorithm called
GNM (Bangert and Unkelbach 2016), which was shown to be capable of reducing the BAO
runtime by one to two orders of magnitude over the naïve implementation of iterative BAO.
Nonetheless, our algorithm turned out to be substantially faster. Meanwhile, in terms of the
resulting plan quality, the gFCP-based formulation is significantly better than GNM in the
prostate case and very competitive compared to GNM—if not better. The comparative advan-
tage of gFCP over GNM in plan quality is likely to be generalizable, as the GNM is intrinsi-
cally a greedy scheme that heuristically takes into consideration the beam angles that are
locally most promising at the current stage. Meanwhile, the gFCP provides a better global
balance between FMO and beam angle selection and seeks to minimize FMO while penal-
izing the use of more beam angles. Algorithm 1 entails less computational cost than GNM
because the latter requires the beam angles to be added iteratively—one beam is added to the
active set of beams after the computation of an FMO subproblem. Though each subproblem
is potentially much smaller than the FMO with full consideration of all beam angles, the fact
that many repetitions of such computation are needed leads to a considerable amount of com-
putation. In contrast, in our proposed scheme, it is only necessary to run algorithm 1 once.
The one-time calculation of algorithm 1 is comparable to the gradient descent algorithm that
is invoked after several times in GNM. Moreover, the computational efficiency of the gFCP
scheme is also generalizable, because algorithm 1 is a fully polynomial-time approximation
scheme as per the analysis by both Wang et al (2014) and Nesterov (2013).
Since the proposed scheme achieves only a local solution to the nonconvex optimization
formulation, changing the initial solution may change the solution quality. Nonetheless, our
numerical tests show that starting algorithm 1 with the solution to the FMO problem provides
a reliable choice of initialization.
Finally, we emphasize that our proposed algorithm 1 is capable of simultaneously selecting
the beam angles and the beamlet intensities at the active beam angles. This is different from
the 2,1-minimization scheme, in which a pure FMO computation with prefixed beam angles
is needed to generate the final fluence map. Such a second step seems very necessary, as indi-
cated by our prostate case, where ∗2,1 —which is the solution to the 2,1-minimization model
without the follow-up FMO—results in an unacceptable plan quality. It is expected that the
computational cost of gFCP will be further reduced by better initialization.

5. Conclusion

A new sparse optimization-based BAO framework is proposed. Traditionally, heuristics and

2,1-minimization approaches have been proposed to quickly generate promising beam angles.
This paper demonstrates that by incorporating a new sparse optimization framework based on
gFCP, a significant increment in both plan quality and computational efficiency is possible.

6442
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Our results may be further extended to facilitate VMAT and station parameter optimized radi-
ation treatment (SPORT, Li and Xing 2013, Xing and Li 2013).

Acknowledgments

This work is partially supported by NIH (5R01 CA176553) and a Faculty Research Award
from Google Inc. We wish to acknowledge helpful discussions with Drs Y Ye and S Boyd
during the course of this study.

Appendix 

This appendix presents a derivation of the closed form solution to (10) when
W := {w : 0  w  U}, for any U > λ . We first notice that problem (10) is decomposable
into a sequence of smaller optimization problems in the form: for each j ∈ Θ

1 k
 k  2

min (wi : i ∈ Bj ) − (wi : i ∈ Bj ) + s · (Hi w , B : i ∈ Bj ) + λ w2i . (A.1)
w∈W 2s
i∈B j


Therefore, to find the closed-form solution to (10) it suffices to study (A.1).
In the following, we consider four different cases: case 1, when
T k −sλ  k      
0  j T k · wi − s · Hi wk , B  Ui for all i, where Tjk := wki − s · Hi wk , B : i ∈ Bj .
 j
T k −sλ  k 
Then, one can easily check that wk+1i := j T k · wi − s · Hi (wk , B) satisfies the KKT con-
 j
dition to (A.1), which is written as
1  k+1   wk+1
wi − wki + s · Hi wk , B + λ  i
= 0.
s
(A.2)   k+1 2
wi
i∈Bj

k+1
 then (wi : i ∈ Bj ) is a solution
Since (A.1) is convex,
 k 
to (A.1).
k  k
Case 2, when Tj  − s · λ  0 and wi − s · Hi w , B  0 for all i ∈ D for some index
set D ⊆ Bj . Suppose that there is a dimension τ ∈ D such that wk+1
 k+1 τ = 0 . Let y = (yi : i ∈ Bj )
  
with yi = wi k+1
for all i = τ and yτ = 0. Then wi : j ∈ Bj  > yk+1

i : j ∈ Bj  and
 k+1       2 
 w : i ∈ Bj − wki : i ∈ Bj + s · Hi wk , B : i ∈ Bj  > (yi : i ∈ Bj ) − (wki : i ∈ Bj ) +
i
  2  
s · (Hi wk , B : i ∈ Bj ) , which then means that wk+1 i : j ∈ Bj is not a solution to (A.1) and
k
  which then indicates that wτ = 0.
results in a contradiction,
k 
Case 3, when Tj  − s · λ  0 and wki − s · Hi wk , B  0 for some i. In this case,
   k  
we let wk+1 i = 0 for all i ∈ B j. Then 0 ∈ 1
s w k+1
j − w k
i + s · H w , B : i ∈ Bj −
     
   k+1   k+1 
λ∂ (wk+1 j : j ∈ Bj ) , where ∂ (w j : j ∈ B j )  is the subdifferential of  w j : j ∈ Bj 
 
in wk+1j : j ∈ Bj . This indicates that the first-order optimality condition of the subproblem
(A.1) is satisfied. Since (A.1) is convex, then (wk+1
i : i ∈ Bj ) is a solution to (A.1).
Tjk −sλ  k  k 
Case 4, when · wi − s · Hi w , B > Ui , for all i ∈ D for some index set
Tjk 
D ⊆ Bj . Suppose that 0  wk+1 τ < Uτ for some τ ∈ D. Then the KKT condition yields that

6443
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

   wk+1
for some ρi,2  0: 1
s wk+1
i − wki + s · H wk , B + λ  i
2
− ρi,2 = 0 , which yields
i∈Bj (wk+1
i )
Tjk −sλ  k   
that wk+1
τ = · wτ − s · Hτ wk , B + ρi,2 > Uτ and leads to a contradiction. This
Tjk 
implies that wk+1
τ = Uτ .
Combining the four cases above, we know that
   
  
min Tjk −sλ k k 
· T , U , Tj   s · λ
(wk+1i : i ∈ Bj ) =  Tjk 

+ ,


0, otherwise,
which applied to all j ∈ Θ immediately leads to the desired result.

References

Aleman  D  M, Kumar  A, Ahuja  R  K, Romeijn  H  E and Dempsey  J  F 2008 Neighborhood search

approaches to beam orientation optimization in intensity modulated radiation therapy treatment
planning J. Global Optim. 42 587–607
Azizi Sultan  A  S 2006 Optimization of beam orientations in intensity modulated radiation therapy
planning PhD Thesis Technische Universität Kaiserslautern
Bangert  M and Unkelbach  J 2016 Accelerated iterative beam angle selection in IMRT Med. Phys.
43 1073–82
Craft D, Bangert M, Long T, Papp D and Unkelbach J 2014 Shared data for intensity modulated radiation
therapy (IMRT) optimization research: the CORT dataset Gigascience 3 37
Deasy  J  O, Blanco  A  I and Clark  V  H 2003 CERR: a computational environment for radiotherapy
research Med. Phys. 30 979–85
Ezzell G A 1996 Genetic and geometric optimization of three-dimensional radiation therapy treatment
planning: II. Beam’s eye-view, back projection, and projection through CT sections Med. Phys.
23 293–305
Fan J and Li R 2001 Variable selection via nonconcave penalized likelihood and its oracle properties
J. Am. Stat. Assoc. 96 1348–60
Huang J, Breheny P and Ma S 2012 A selective review of group selection in high-dimensional models
Stat. Sci. 27 481–99
Jia X, Men C, Lou Y and Jiang S B 2011 Beam orientation optimization for intensity modulated radiation
therapy using adaptive l2,1-minimization Phys. Med. Biol. 56 6205–22
Li R and Xing L 2013 An adaptive planning strategy for station parameter optimized radiation therapy
(SPORT): segmentally bosted VMAT Med. Phys. 40 050701
Li Y, Yao D, Yao J and Chen W 2005 A particle swarm optimization algorithm for beam angle selection
in intensity modulated radiotherapy planning Phys. Med. Biol. 50 3491–514
Liu  H, Yao  T, Li  R and Ye  Y 2017 Folded concave penalized sparse linear regression: sparsity,
statistical performance, and algorithmic theory for local solutions Math. Prog. (doi:10.1007/
s10107-017-1114-y)
Meedt  G, Alber  M and Nusslin  F 2003 Non-coplanar beam direction optimization for intensity-
modulated radiotherapy Phys. Med. Biol. 48 2999–3019
Meyer  J, Hummel  S  M, Cho  P  S, Austin-Seymour  M  M and Philips  M  H 2005 Automatic selection
of non-coplanar beam directions for three-dimensional conformal radiotherapy Br. J. Radiol.
78 316–27
Nesterov Y 2013 Gradient methods for minimizing composite functions Math. Prog. 140 125–61
Pugachev A, Li J, Boyer A L, Hancock S, Le Q T, Donaldson S and Xing L 2001 Role of beam orientation
optimization in IMRT Int. J. Radiat. Oncol. Biol. Phys. 50 551–60
Pugachev A and Xing L 2001a Computer-assisted selection of coplanar beam orientations in intensity-
modulated radiation therapy Int. J. Radiat. Oncol. Biol. Phys. 51 1361–70
Pugachev A and Xing L 2001b Pseudo beam’s eye view as applied to beam orientation selection in IMRT
Int. J. Radiat. Oncol. Biol. Phys. 51 1361–70

6444
Phys. Med. Biol. 62 (2017) 6428 H Liu et al

Rocha H, Dias J M, Ferreira B C and Lopes M C 2013 Beam angle optimization for intensity-modulated
radiation therapy using a guided pattern search method Phys. Med. Biol. 58 2939–53
Vaitheeswaran R, Narayanan V K, Bhangle J R, Nirhali A, Kumar N, Basu S and Maiya V 2010 An
algorithm for fast beam angle selection in intensity modulated radiotherapy Med. Phys. 37 6443–52
Wang Z, Liu H and Zhang T 2014 Optimal computational and statistical rates of convergence for sparse
nonconvex learning problems Ann. Stat. 42 2164–201
Woudstra E and Storchi P R M 2000 Constrained treament planning using sequential beam selection
Phys. Med. Biol. 45 2133–49
Xing  L and Li  R 2013 Inverse planning in the age of digital LINACs: station parameter optimized
radiation therapy (SPORT) J. Phys.: Conf. Ser. 489 012065
Yang R, Dai J, Yang Y and Hu Y 2006 Beam orientation optimization for intensity-modulated radiation
therapy using mixed integer programming Phys. Med. Biol. 51 3653–66
Yuan M and Lin Y 2005 Model selection and estimation in regression with grouped variables J. R. Stat.
Soc. B 68 49–67
Zhang C 2010 Nearly unbiased variable selection under minimax concave penalty Ann. Stat. 28 894–942
Zhu L, lee L, Ma Y, Ye Y, Mazzeo R and Xing L 2008 Using total-variation regularization for inverse
planning with field specific numbers of segments Phys. Med. Biol. 53 6653–72
Zhu L and Xing L 2009 Search for IMRT inverse plans with piecewise constant fluence maps using
compressed sensing techniques Med. Phys. 36 1895

6445