Treatment planning for volumetric modulated arc therapy

James L. Bedford

Citation: Medical Physics 36, 5128 (2009); doi: 10.1118/1.3240488

View online: http://dx.doi.org/10.1118/1.3240488
View Table of Contents: http://scitation.aip.org/content/aapm/journal/medphys/36/11?ver=pdfcov
Published by the American Association of Physicists in Medicine

Articles you may be interested in

Direct leaf trajectory optimization for volumetric modulated arc therapy planning with sliding window delivery
Med. Phys. 41, 011701 (2014); 10.1118/1.4835435

A method for optimizing LINAC treatment geometry for volumetric modulated arc therapy of multiple brain
metastases
Med. Phys. 37, 4146 (2010); 10.1118/1.3455286

Development and evaluation of an efficient approach to volumetric arc therapy planning

Med. Phys. 36, 2328 (2009); 10.1118/1.3132234

Volumetric modulated arc therapy: IMRT in a single gantry arc

Med. Phys. 35, 310 (2008); 10.1118/1.2818738

Design and dosimetry of a few leaf electron collimator for energy modulated electron therapy
Med. Phys. 34, 4782 (2007); 10.1118/1.2795827
Treatment planning for volumetric modulated arc therapy
James L. Bedforda兲
Joint Department of Physics, The Institute of Cancer Research and The Royal Marsden
NHS Foundation Trust, Downs Road, Sutton, Surrey SM2 5PT, United Kingdom
共Received 26 March 2009; revised 14 July 2009; accepted for publication 10 September 2009;
published 9 October 2009兲
Purpose: Volumetric modulated arc therapy 共VMAT兲 is a specific type of intensity-modulated
radiation therapy 共IMRT兲 in which the gantry speed, multileaf collimator 共MLC兲 leaf position, and
dose rate vary continuously during delivery. A treatment planning system for VMAT is presented.
Methods: Arc control points are created uniformly throughout one or more arcs. An iterative
least-squares algorithm is used to generate a fluence profile at every control point. The control
points are then grouped and all of the control points in a given group are used to approximate the
fluence profiles. A direct-aperture optimization is then used to improve the solution, taking into
account the allowed range of leaf motion of the MLC. Dose is calculated using a fast convolution
algorithm and the motion between control points is approximated by 100 interpolated dose calcu-
lation points. The method has been applied to five cases, consisting of lung, rectum, prostate and
seminal vesicles, prostate and pelvic lymph nodes, and head and neck. The resulting plans have
been compared with segmental 共step-and-shoot兲 IMRT and delivered and verified on an Elekta
Synergy to ensure practicality.
Results: For the lung, prostate and seminal vesicles, and rectum cases, VMAT provides a plan of
similar quality to segmental IMRT but with faster delivery by up to a factor of 4. For the prostate
and pelvic nodes and head-and-neck cases, the critical structure doses are reduced with VMAT, both
of these cases having a longer delivery time than IMRT. The plans in general verify successfully,
although the agreement between planned and measured doses is not very close for the more com-
plex cases, particularly the head-and-neck case.
Conclusions: Depending upon the emphasis in the treatment planning, VMAT provides treatment
plans which are higher in quality and/or faster to deliver than IMRT. The scheme described has
been successfully introduced into clinical use. © 2009 American Association of Physicists in
Medicine. 关DOI: 10.1118/1.3240488兴

Key words: intensity-modulated arc therapy, IMAT, segmental, step and shoot, IMRT

I. INTRODUCTION and various planning studies have been conducted to deter-

Intensity-modulated arc therapy 共IMAT兲 as proposed by Yu1
has been used for several years to treat patients in a short
treatment time using high-quality dose distributions.2,3 This
technique makes various gantry arcs with dynamic multileaf
collimation to build up fluence modulation. A related tech-
nique using a single gantry arc and rather more vigorous leaf
motion has also been proposed by Crooks et al.4 The possi-
bilities of IMAT have given rise to several planning
algorithms5,6 and related theoretical concepts.7–9 However,
all of these methods involve the delivery of dose with con-
stant dose rate. Recently, linear accelerators have become
available which are able to vary gantry speed, multileaf col-
limator 共MLC兲 leaf position, and dose rate simultaneously, a
technique known as volumetric modulated arc therapy
共VMAT兲.
Treatment planning for VMAT is challenging due to the
requirement to sequence the MLC leaves in such a way that
the leaf motion is minimal and within the allowed maximum
leaf speed of the linear accelerator. Furthermore, the beam
weights should be such that the dose rate of the accelerator is
not required to drop to a value lower than the accelerator can
reliably deliver. Several schemes have been proposed,10–16
mine how VMAT compares in dosimetric quality to segmen-
tal intensity-modulated radiation therapy 共IMRT兲 and
tomotherapy.17–23 However, there are still some open
questions.24
This paper therefore describes a treatment planning
scheme for VMAT with particular emphasis on providing
deliverable plans which can be executed and verified with
confidence. General methods are described, but the scheme is
illustrated with reference to the Elekta Synergy 共Crawley,
U.K.兲 accelerator. This implementation of VMAT has vari-
able gantry speed, variable collimator angle, variable colli-
mator position, variable MLC leaf position, and variable
dose rate. A number of control points are prescribed to the
accelerator, each defining the gantry angle, the collimator
angle, the position of the collimators, the position of the
MLC leaves, and the cumulative monitor units delivered at a
particular instance in the treatment delivery. The delivery
control system then moves the gantry, collimators, and MLC
leaves dynamically between the positions specified at the
control points, while the dose rate is chosen so that at each
control point, the correct cumulative monitor units have been
delivered. A feedback system repeatedly monitors the actual

5128 Med. Phys. 36 „11…, November 2009 0094-2405/2009/36„11…/5128/11/$25.00 © 2009 Am. Assoc. Phys. Med. 5128
5129 James L. Bedford: VMAT treatment planning
accelerator conditions and adjusts the motion and dose rate
to ensure that the prescription is accurately followed. The
planning algorithm is described with reference to this para-
digm and then illustrated using a variety of clinical cases of
varying complexity. In particular, three novel features are
presented: 共a兲 An accurate means of handling clinical objec-
tives and constraints in an iterative least-squares optimiza-
tion of fluence, which allows this fast but seldom used opti-
mization method to be used for the many beam orientations
which are encountered in VMAT, 共b兲 a control point group-
ing and sequencing scheme for VMAT, and 共c兲 an efficient
method of calculating dose for the dynamic aperture motion
in VMAT.
II. METHODS
II.A. Background
The VMAT planning algorithm has been implemented in
the AUTOBEAM 共v4.9兲 in-house software developed at the
Royal Marsden and described previously.25–27 CT scans and
outlines are read from PINNACLE3 共Philips Radiation Oncol-
ogy Systems, Madison, WI兲 and inverse planning is then
performed in a separate software. The inverse planning algo-
rithm consists of three stages: Fluence optimization, segmen-
tation, and direct-aperture optimization.
II.B. Arc definition
The algorithm begins with the definition of arcs. The arcs
are defined by uniformly distributing the desired number of
control points throughout the specified range of gantry,
couch, and collimator angles. There can be any number of
arcs, noncoplanar if desired. At this point, having defined the
control points, it is possible to proceed immediately to
direct-aperture optimization to determine the optimum aper-
ture shapes and monitor units for the control points. How-
ever, this is very inefficient, so we use the well-established
three-step process of fluence optimization, segmentation, and
then direct-aperture optimization, the latter being used for
fine-tuning. The concept of a control point group is intro-
duced with this in view. A single control point cannot, on its
own, deliver intensity-modulated radiation to the patient: A
number of control points are required to be collected together
so that, working together, albeit at slightly different gantry or
collimator angles, they deliver a modulated fluence profile to
the patient. The control point group provides this collection
of control points. Beginning with the first control point of the
first beam, a set of control point groups is defined whose
gantry, couch, and collimator angles differ by less than or
equal to a specified tolerance 共Fig. 1兲. For example, if there
are two arcs with 5° spacing of control points, one a clock-
wise arc starting at 190° and ending at 170° and the other an
anticlockwise arc starting at 170° and ending at 190° 共a typi-
cal arrangement兲, and the tolerance angle is 10°, the control
points at 190°, 195°, and 200° in each of the two beams, a
total of six control points, form the first segment group, and
so on. From these six control points, an optimal fluence map
can be reproduced with deliverable aperture shapes. Note
Medical Physics, Vol. 36, No. 11, November 2009
5129
FIG. 1. Segment grouping in the segmentation step of the inverse planning
algorithm.
that although the above illustration uses two arcs, this algo-
rithm is general and can be used for any number of arcs.
By setting the tolerance angle for the control point group-
ing, the user can therefore specify how the control points are
to be used. Compared to segmental delivery, a large toler-
ance angle corresponds to few static beams with many seg-
ments per beam and a small tolerance angle corresponds to
many static beams with few segments per beam.
II.C. Fluence optimization
In principle, one fluence profile is required to be deter-
mined for each control point group. At the segmentation
stage, the control points of the group are then set to represent
the fluence profile. The fluence profile could be determined
for the gantry angle at the center of the gantry angle range
represented by the control point group 共e.g., Ref. 16兲. How-
ever, this is not optimal, since the control points of the group
are at different gantry angles, and the positions of the organs
at risk in the beam’s eye view vary with gantry angle. Cre-
ating a single fluence profile at a single gantry angle, seg-
menting this fluence profile, and then distributing the seg-
ments to other gantry angles in the control point group
results in beam apertures which do not properly match the
organs at risk. An alternative strategy is therefore used. A
fluence profile is generated for each discrete beam orienta-
tion, as defined by gantry, couch, and collimator angles, for
which there is a control point. 共Collimator angle is included
here as rotation of the collimator implies rotation of the cor-
responding fluence map, which requires explicit reoptimiza-
tion of fluence.兲 Thus, the example in Fig. 1 would have
three fluence profiles per control point group. Consequently,
when the segmentation step is reached, each control point is
associated with a fluence profile which belongs properly to
its own gantry, couch, and collimator angles. 共See Sec. II D
below for details of exactly how the segmentation proceeds.兲
Fluence is optimized using the iterative least-squares al-
gorithm as originally implemented in image reconstruction
and applied to radiation therapy by Xing and Chen.28 The
method iterates through the pixels of the fluence map, com-
paring the dose actually delivered by each pencil beam of
5130 James L. Bedford: VMAT treatment planning
radiation and the prescribed dose and correcting the fluence
accordingly. Let the fluence delivered by the jth element of
the entire set of fluence profiles around the patient be de-
noted by w j and the dose delivered to voxel i of the patient
by unit fluence element j be denoted by dij. If the prescribed
dose at voxel i of the patient is denoted by dip and the corre-
sponding actual dose is di = 兺 jdijw j, then the correction ap-
plied to w j at each iteration is
兺i␣idij共dip − di兲
⌬w j = R , 共1兲
兺i␣id2ij
where R = 0.001 is a relaxation parameter included to pro-
mote divergence and ␣i is the importance factor associated
with each individual voxel. The presence of the first dij in the
numerator is to weight the correction according to the con-
tribution that a fluence element makes to each voxel in the
patient. In other words, deviations from the prescribed dose
at voxels which receive a high dose from w j influence the
correction more than deviations at voxels which receive a
low dose from w j: There is no point in making a large change
to a fluence element because of a large deviation from the
prescribed dose at a particular voxel if that fluence element
does not contribute significantly to that voxel. The denomi-
nator is a normalization factor.
So far, the method follows that of Xing and Chen.28 How-
ever, the method28 has the same limitation as all of the in-
verse planning algorithms which rely on a prescribed dose
distribution, which is that the prescribed dose is not actually
known. Rather, a series of dose, dose-volume, and biological
constraints is available, and the prescribed dose can be any-
thing appropriate to these limits. Hence, the method is used
in an adapted form, in which dip is simply equal to the pre-
scribed dose for target structures and zero elsewhere. The
importance factors ␣i are then used to control the algorithm
by evaluating an objective function. The objective function is
constructed from treatment plan statistics, which can be 共a兲
root-mean-square dose deviation from a specified dose level,
used for controlling PTV dose inhomogeneity, 共b兲 minimum
dose, 共c兲 maximum dose, 共d兲 mean dose, 共e兲 irradiated vol-
ume, 共f兲 dose to the hottest fraction of a structure, 共g兲 tumor
control probability, or 共h兲 normal tissue complication prob-
ability. These quantities can be maximized or minimized but
for simplicity, minimization is considered here. The treat-
ment plan statistics are represented as f 1共di兲 , f 2共di兲 , f 3共di兲 , . . .
in the form of M objectives and N constraints. The objectives
relate to volumes of interest Vm 共m = 1 , . . . , M兲 and the con-
straints relate to volumes of interest Vn 共n = 1 , . . . , N兲, with
multiple objectives and constraints allowed for each volume,
so that some of the Vm and Vn are the same volume. In
addition, the volumes may overlap in space. Each objective
and constraint implies a corresponding spatial importance
factor:
再 冎
␤ m f m, i 苸 V m ,
␣im = 共2兲
0 otherwise,
Medical Physics, Vol. 36, No. 11, November 2009
5130
␣in = 再 H共f n − cn兲A共f n − cn兲,
0 otherwise,
i 苸 Vn ,
冎 共3兲
where ␤m are user-defined importance factors, cn are limits
on the N constraints, and A is a large constant importance
factor.27 H is the Heaviside unit step function. Then:
M N
␣i = 兺 ␣im + n=1
m=1
兺 ␣in . 共4兲
This is used in Eq. 共1兲. The effect is that if the dose to a
structure is far from the dose-volume constraint, for ex-
ample, the importance factor is large, giving emphasis to that
structure in the optimization. Once a constraint is met, the
objective function for that structure becomes zero, and the
importance factor consequently zero as well. Overall, the al-
gorithm proceeds by calculating ⌬w j for all j, then imple-
menting the corrected w j, giving the new improved dose dis-
tribution, calculating the new set of spatial importance
factors ␣i, then repeating the process for 100 iterations. At
the completion of the fluence optimization, the fluence pro-
files are lightly filtered using a cosine-shaped low-pass filter
to remove any unrealistic high-frequency components which
cannot be sensibly segmented.
II.D. Segmentation
Segmentation is achieved by stratification of the fluence
profiles. The individual fluence profile for each control point
in the group is divided into as many intensity levels as there
are control points in the group, the first level being specified
by the user and the remainder being equally spaced in inten-
sity. Over the range of angles in the group, the fluence pro-
files are very similar due to the iterative least-squares algo-
rithm used for the fluence optimization but vary slightly due
to the difference in angle over the width of the segment
group. The first control point in the group is then set to
deliver the first level of its intensity map, the second segment
is then set to deliver the second level of its own respective
intensity map, and so on 共Fig. 2兲. This is approximately
equivalent to segmenting one single fluence profile, except
that each control point aperture aligns itself with the fluence
profile intended for its own particular gantry, couch, and col-
limator angles. However, it is not an exact process due to the
changes in the fluence profiles between beam angles. For
example, the segment marked with an asterisk in Fig. 2 is not
delivered by any of the control points. This is because at the
gantry angle of control points 3 and 4, a second peak in the
fluence profile has arisen, so that control point 4 addresses
the second peak rather than the main peak. Effects such as
this are not dealt with at this stage as they can be corrected
for by the direct-aperture optimization later on.
The aperture shape at each level in the intensity map is
selected by bracketing the leftmost 共in an arbitrary sense兲
peak in the fluence profile to rise above the specified inten-
sity level. Computationally, this is achieved as follows. If
there are s apertures to be identified, the first segment having
weight a and the remainder having weight b, then a + 共s
− 1兲b = 1 for a single-peaked fluence map with unit maximum
5131 James L. Bedford: VMAT treatment planning 5131

FIG. 2. Segmentation step of the inverse planning algorithm showing how

segments are extracted from a segment group. The colors show which ap-
erture is delivered by each segment. The first segment of each intensity map
may have a different intensity specified by the user.

intensity. The segmentation proceeds by drawing all pixels of

the fluence map which are nonzero to a binary bitmap, set-
ting to zero the rightmost pixels in any rows of the bitmap
having multiple disconnected islands 共so that the leftmost
islands remain兲, then finding the boundary of the bitmap.
This defines an aperture shape, and the weight of this seg-
ment is then subtracted from the overall fluence map as it has
been accounted for. This process is repeated until all of the s
segments have been identified. If there are multiple peaks in
the bitmap, some of the fluence map is still unaccounted for
because only the leftmost peaks have been bracketed. Thus,
the weights of the s segments are all increased by a factor
共s + 1兲 / s and the whole process is repeated until the s seg-
ments account for the entire fluence map.
Once the segmentation is completed, the allowed range of
leaf motion is enforced by moving each leaf to a position
suitably close to its position at the previous control point.
This enforcement of allowed range of leaf motion is gov-
erned by the user’s intention for the treatment delivery. There
is actually no limit on the allowable range of leaf positions in
the delivery because if the required leaf motion is too great
to move the leaf at maximum gantry speed and dose rate, the FIG. 3. Schematic representation of control point interpolation for dose
accelerator lowers the gantry speed and dose rate so that the calculation.
leaf motion can be achieved within the permissible maxi-
mum speed. However, this may not be the user’s intention.
For example, for a lung treatment with Active breathing co- segmentally 共i.e., with step-and-shoot delivery兲. This is due
ordinator 共Elekta, Crawley, U.K.兲, it is important to complete to the nature of the segmentation algorithm and can be seen
the treatment in the minimum length of time so as to mini- clearly from Fig. 2. However, during dynamic delivery, the
mize the number of breath holds that the patient is required dose is deposited between control points, so that the first
to perform. The user therefore supplies the minimum gantry control point has zero weight by definition, and the weights
speed for the delivery. From this, the maximum leaf motion of the subsequent control points refer to the monitor units
between segments is simply calculated. delivered between the previous control point and the present
one. At the start of the direct-aperture optimization, the
weight of the first control point in each arc is therefore set to
II.E. Direct-aperture optimization
zero. From this point onwards, the dose calculation also
At the end of the segmentation stage, the aperture shapes takes into account the motion of the aperture shapes between
and segment weights are set as if the plan is to be delivered control points. Both of these changes cause a temporary re-

Medical Physics, Vol. 36, No. 11, November 2009

5132 James L. Bedford: VMAT treatment planning
duction in plan quality, but the optimization subsequently
corrects for this perturbation.
The direct-aperture optimization phase takes the deliver-
able but suboptimal plan produced by the segmentation step
and improves it for delivery. The optimization is a con-
strained optimization, with the collimator positions and the
positions of several key MLC leaves being optimized and the
remaining MLC leaves being fitted by a polynomial function,
so as to ensure rational aperture shapes. There are also con-
straints on minimum leaf gap, minimum aperture width and
length, and minimum equivalent square. The method uses a
multidimensional downhill search, relying on the initial seg-
mentation to ensure that it begins in the approximate region
of the global minimum of the objective function. The down-
hill search cycles through the collimators and MLC leaves to
冦 冧
M N
兺
m=1
␤m f m共di兲 if 兺 H共f n共di兲 − cn兲 = 0
n=1
f共di兲 = N M
A + 兺 H共f n共di兲 − cn兲共f n共di兲 − cn兲 + B 兺 ␤m f m共di兲
n=1 m=1
where B is a small constant importance factor. Since this has
a similar form to Eqs. 共2兲 and 共3兲, the direct-aperture optimi-
zation proceeds in a similar direction in parameter space to
the fluence optimization. The remaining details of the opti-
mization are very similar to those published previously26,27
and so are not repeated here.
II.F. Dose calculation
Dose calculations are performed using a fast convolution
algorithm with inhomogeneity correction.30 The method can
therefore be used for lung planning and other dosimetrically
challenging sites. In order to accurately model the motion of
the apertures between the control points, 100 linear interpo-
lations of the apertures are made between control points and
dose is summed over these interpolated apertures. To clarify
the explanation, we refer to control points as the actual states
supplied to the accelerator and segments as static beams with
a specific orientation, aperture shape, and weight, used for
computation purposes only. Let the P control points be in-
dexed by p 共p = 1 , . . . , P兲 and let the pth control point have
beam orientation A p, aperture shape S p, and segment weight
w p. The weight of the first control point has been set to zero
and remains zero, i.e., w0 = 0. For each of control points
2 , . . . , P, 100 interpolated segments are created, each having
orientation A pq, aperture shape S pq, and weight w pq 共q
= 1 , . . . , 100兲, such that
A pq = 共1 − 0.01q兲A p−1 + 0.01qA p , 共6兲
S pq = 共1 − 0.01q兲S p−1 + 0.01qS p , 共7兲
Medical Physics, Vol. 36, No. 11, November 2009
5132
be optimized, selects each in turn, and systematically finds
the optimum leaf position of the selected leaf, before moving
on to the next leaf. The allowed range of leaf positions is
taken into account when systematically searching through
the leaf/jaw positions. Segment weights are also optimized
during this process. Ten complete cycles through all of the
variables are completed, as the optimum value of one vari-
able depends upon the current values of the other variables.
This method has been shown to converge to the minimum of
an optimization space, so long as there are no local
minima.29 In this case, there are local minima, but the initial
segmentation is used to ensure that the optimization starts in
the vicinity of the global minimum.
Using the notation introduced above in Sec. II C, the ob-
jective function is
共5兲
otherwise,
w pq = 0.01w p . 共8兲
In principle, dose is now calculated for each of these inter-
polated segments to model the motion of the aperture be-
tween control points. However, this would in practice require
a prohibitively long time. We therefore make the approxima-
tion
A pq = A p, p = 2, . . . , P, q = 1, . . . ,100. 共9兲
So long as the P control points are reasonably closely
spaced, this is sufficiently accurate in practice. The value of
this is that we now have only P − 1 segment orientations to
calculate, as opposed to 100⫻ 共P − 1兲 orientations. Each con-
trol point now has 100 segments, all at that orientation. Fig-
ure 3 illustrates the situation.
The form of the dose calculation is as follows.30 Let the
beam space be spanned by a coordinate system whose x and
y axes are orthogonal to the beam axis and whose z axis is
parallel to the beam axis and let the equivalent square field
size of aperture S pq be s pq. Then the total energy released per
unit mass 共TERMA兲 is characterized by an output factor
f pq共s pq兲, an incident fluence profile r pq共x , y , S pq兲, and a depth
component t pq共z , s pq兲. The TERMA is convolved with a con-
volution kernel k p共x , y , z兲, which is the same for all segments
at a given orientation. The dose due to each set of 100 seg-
ments, all at the same orientation, can be calculated as
5133 James L. Bedford: VMAT treatment planning 5133

TABLE I. Details of the clinical planning cases.

Segment Nomina Segmental

Number Control points Gantry range group size l gantry speed gantry angles Segments
Site Prescribed dose of arcs per arc 共deg兲 共deg兲 共deg/s兲 共deg兲 per segmental beam

0, 70, 145,
Lung 64 Gy in 32 # 1 51 185–175 30 6.0 215, 290 5
Prostate
and seminal 0, 70, 145,
vesicles 74 Gy, 71 Gy, 59 Gy in 37 # 1 71 185–175 20 3.0 215, 290 8
185–175, 180, 250,
Rectum 54 Gy, 45 Gy in 25 # 2 51 175–185 30 3.0 325, 35, 110 10
Prostate 185–175, 180, 250, 325,
and pelvic nodes 70 Gy, 60 Gy in 35 # 2 71 175–185 30 1.5 35, 110 10
185–175, 0, 50, 105,
175–185, 155, 205,
Head and neck 65 Gy, 54 Gy in 30 # 3 51 185–175 40 1.5 255, 310 10

100 of the direct-aperture optimization is that it can be terminated

d p共x,y,z兲 = 兺 兵关f pq共s pq兲r pq共x,y,S pq兲t pq共z,s pq兲兴 early if a good quality solution has been obtained before the
q=1 共generous兲 normal number of iterations has been completed.
丢 k p共x,y,z兲其. 共10兲 The final plan is exported from AUTOBEAM to the accel-
erator 共or record and verify system兲 in DICOM format. The

再 冎
The convolution can be brought outside of the summation:
100
d p共x,y,z兲 = 兺 关f pq共spq兲rpq共x,y,Spq兲tpq共z,spq兲兴
q=1
丢 k p共x,y,z兲.
Moreover, if the equivalent square field size does not change
too significantly between the 100 segments, it is sufficient to
use the mean depth dose component for all segments:
100
t̄ p共z兲 = 0.01 兺 t pq共z,s pq兲.
q=1
plan is delivered dynamically, with gantry speed, collimator
position, MLC leaf position, and dose rate varying between
control points. In addition, a PINNACLE3 script is written
which can then be run within PINNACLE3 to set up the plan
共11兲 for visualization purposes as pseudoarcs consisting of static
beams. Similarly to the process described above, each con-
trol point is represented as a step-and-shoot beam of ten seg-
ments. The ten segments are equally weighted and contain
linear interpolations of the apertures at the control points.
Only ten interpolations are used at this stage for purposes of
共12兲 practicality. This approach allows the well-established
PINNACLE3 collapsed cone convolution dose calculation to be

The depth component can then be removed from the summa- used for the final dose calculation. Note, however, that no
modification can be made to the plan in PINNACLE3 as the

再 冎
tion:
100
d p共x,y,z兲 = t̄ p共z兲 兺 r pq共x,y,S pq兲 丢 k p共x,y,z兲.
q=1
Hence, the complete calculation for all 100 interpolated seg-
ments can be achieved by a straightforward two-dimensional
summation of the individual aperture shapes into an effective
fluence, followed by the same ray tracing and convolution as
a single segment. Overall, the calculation of dose for an arc
with P control points takes a similar time to the calculation
of dose for P − 1 simple static beams. The complete inverse
planning process, including fluence optimization, segmenta-
tion, and direct-aperture optimization, takes around 3–24 h
depending upon the complexity of the plan, running in JAVA
on a single core of a 2.4 GHz AMD Opteron processor in a
Sun Ultra 40 M2 workstation 共Sun Microsystems, Santa
Clara, CA兲. Further work is in progress to optimize the speed
of execution and to introduce multithreaded operation so as
to reduce this time. At present, several plans can be calcu-
lated simultaneously on different processor cores, so the long
calculation time is not clinically prohibitive. Also, the nature
prescription is sent to the accelerator directly from AUTO-
BEAM. In summary, the true VMAT prescription is sent di-
共13兲 rectly from AUTOBEAM to the accelerator, while the static
approximation is sent from AUTOBEAM to PINNACLE3 for vi-
sualization purposes only. The software has been introduced
into clinical use.31
II.G. Planning study
Five cases were retrospectively planned with this scheme
as illustrations of the performance for plans of varying com-
plexity 共Table I兲. The VMAT plans were compared against
segmental 共step-and-shoot兲 treatment plans. In order to en-
sure consistency in the planning comparison, AUTOBEAM was
also used for the segmental plans, so that the same objec-
tives, constraints, and segment size limitations could be ap-
plied. The cases were 共a兲 a lung case, intended to be simple
and fast as a precursor to hypofractionated stereotactic body
irradiation, 共b兲 a treatment of prostate and seminal vesicles
according to the CHHIP trial, consisting of three target vol-
umes, 共c兲 a rectal tumor, consisting of two treatment vol-

Medical Physics, Vol. 36, No. 11, November 2009

5134 James L. Bedford: VMAT treatment planning
umes, the main volume containing the primary tumor, lo-
coregional lymph nodes plus a 1.5 cm margin, and a boost
volume consisting of the primary tumor alone plus a 1.5 cm
margin, 共d兲 prostate and pelvic lymph nodes, giving an ex-
ample of a more challenging concave target volume, and 共e兲
a head-and-neck case, consisting of primary and nodal vol-
umes, which was the most complex of the plans considered.
The emphasis in the choice of inverse planning parameters
for these cases was from fast delivery 共a兲 to high-quality
dose distribution 共e兲. The nominal gantry speed and the num-
ber of arcs were the two main factors in influencing this
emphasis 共see Table I兲.
All cases were delivered on Elekta Synergy linear accel-
erators 共Elekta Ltd., Crawley, U.K.兲 running RTDESKTOP
v7.01. The accelerators were fully commissioned for VMAT
delivery.32 The lung case and the prostate and seminal
vesicles case were planned for, and delivered using, a unit
with a Beam Modulator head. This had 4 mm leaf width at
the isocenter, fixed jaws, and interdigitation. The remaining
cases were planned for, and delivered using, an MLCi head,
which had 10 mm leaf width at the isocenter, variable jaws,
and no interdigitation. The segmental IMRT plans were de-
livered at a dose rate of 600 MU/min. The delivery times,
from the start of the first beam to the end of the last beam,
were recorded for both the VMAT plans and the correspond-
ing segmental IMRT plans. A DELTA4 phantom 共Scandidos,
Uppsala, Sweden兲 was also used to verify the dose distribu-
tions calculated in PINNACLE3.33 The percentage of the points
of measurement agreeing to within 3% of the dose to the
primary PTV and 3 mm was recorded for all of the plans.
Detectors recording less than 20% of the dose to the primary
PTV were excluded from the gamma analysis. The plans
were assessed against the departmental acceptance criterion
that 90% of points of measurement should agree with the
plan to within 3% and 3 mm.
III. RESULTS
Typical control point beam’s eye views as seen in AUTO-
BEAM are shown in Fig. 4 for the prostate and seminal
vesicles 共Beam Modulator兲 and prostate and pelvic nodes
共MLCi兲 cases. The type of MLC 共i.e., with or without inter-
digitation, fixed or moving collimators兲 influences the shape
of the segments considerably. Transaxial, sagittal, and coro-
nal dose distributions are shown in Fig. 5 for the VMAT
plans for all five clinical cases. It can be seen that VMAT
TABLE II. Irradiated volumes of normal tissue.
VMAT V20
Treatment site 共cm3兲 a
Lung 3528
Prostate and seminal vesicles 4598
Rectum 6055
Prostate and pelvic nodes 8526
Head and neck 4708
a
V20 is the volume of the patient receiving 20% of the highest prescribed
dose, excluding the planning target volume but including critical structures.
Medical Physics, Vol. 36, No. 11, November 2009
5134
FIG. 4. Representative beam’s eye views of segments in 共a兲 the prostate and
seminal vesicles case and 共b兲 the prostate and pelvic nodes case showing the
grayscale fluence map and the segmented leaf positions for the Beam Modu-
lator head and the MLCi head, respectively.
produces very conformal plans with well-controlled dose dis-
tributions. Dose-volume histograms comparing VMAT with
segmental IMRT for all cases are shown in Fig. 6. For the
lung case, prostate and seminal vesicles, and rectum cases,
there is generally 共with a few exceptions兲 a small improve-
ment in target coverage and a small reduction in critical
structure dose. For the prostate and pelvic nodes and head-
and-neck cases, VMAT gives considerably reduced critical
structure doses compared to IMRT. The irradiated volume of
normal tissue is generally higher with VMAT, with the ex-
ception of the head-and-neck case 共Table II兲.
The total monitor units and delivery times are given in
Table III. For the simpler cases, the monitor units with
VMAT are similar to those for IMRT, and the delivery times
are rather shorter. For the more complex cases, where the
emphasis is on plan quality rather than delivery speed, more
monitor units are required with VMAT than for IMRT and
the delivery time is longer.
The VMAT plan for the prostate and pelvic nodes case is
shown in Fig. 7 as it appears in the DELTA4 software. Similar
results are obtained for both the IMRT and VMAT plans for
all of the cases. The percentage of measurements agreeing
with the plan to within 3% and 3 mm drops gradually as the
IMRT V20
cases become more complex 共Table III兲. Assessed against the
共cm3兲 a criterion that 90% of the measurements should agree with the
plan to within 3% and 3 mm, all of the verification results are
3334 acceptable for both VMAT and IMRT plans, with the excep-
3385
tion of the head-and-neck VMAT case.
5472
7847
4982 IV. DISCUSSION
This paper takes the previous scheme of IMRT with con-
strained aperture shapes and applies it to planning for
5135 James L. Bedford: VMAT treatment planning
FIG. 5. Transaxial, sagittal, and coronal views of the VMAT plans for the five cases.
VMAT. The concept of simple apertures is very suited to
VMAT because with VMAT, the apertures smoothly vary
with time. Complex “flag-pole” segments where leaves are
positioned behind the opposing jaw so as to overcome con-
straints in collimator position are not sensible in VMAT be-
cause the aperture shapes produced as the leaves move
smoothly toward and away from this position are unusual
and likely to give poor dosimetric agreement with the treat-
ment planning system.
In order to improve the agreement of the delivered plan
with the treatment planning system, the inverse planning has
been designed to include modeling of the leaf motion. For
simple cases, where the leaves do not move much between
segments, this may not be necessary: It may be sufficient to
sum the contributions of the apertures at the control points
only. However, for complex cases where the leaves move
several centimeters between control points, some form of
Medical Physics, Vol. 36, No. 11, November 2009
5135
modeling is necessary to correctly predict the delivered dose.
A classic example of this is for apertures irradiating either
side of the spinal cord in a head-and-neck case. In a segmen-
tal treatment plan, one aperture may irradiate one side of the
cord and another aperture may irradiate the other side of the
cord. The same two apertures, if used as two adjacent control
points in a VMAT plan, will give rise to a distribution of
dose across the cord as the leaves move from one side of the
cord to the other. However, if the motion of the leaves is not
modeled, the treatment planning system will predict near-
perfect sparing of the cord, because only the apertures at the
control points, i.e., either side of the cord, are taken into
account. It is for this reason that leaf motion modeling has
been included in the inverse planning scheme.
The cases have been chosen to show the range of possible
modulation complexities, from a single-arc lung plan with
fairly conformal apertures to a three-arc head-and-neck plan
5136 James L. Bedford: VMAT treatment planning
FIG. 6. Dose-volume histograms for each of the five clinical cases. 共a兲 Lung, 共b兲 prostate and seminal vesicles, 共c兲 rectum, 共d兲 prostate and pelvic nodes, and
共e兲 head and neck. Solid lines: VMAT. Broken lines: Segmental IMRT.
with apertures which differ very much from the conformal
beam’s eye view of the PTV. VMAT can be used for confor-
mal arcs as well as for complex aperture shapes, the three
basic accelerator capabilities used in VMAT, variation of
gantry speed, leaf position, and dose rate, being used to pro-
vide a well-balanced conformal plan. As well as choosing an
appropriate degree of modulation for the case to be treated, it
is also sensible to choose the arcs to suit the target arrange-
ment. A single arc may have benefits in terms of delivery
efficiency, but multiple arcs are necessary for accurate deliv-
ery of complex plans. The present algorithm can deal with
both of these situations. For example, the lung and prostate
and seminal vesicles cases can be accurately planned and
delivered with a single arc, whereas in the rectum case, a
clockwise arc with nonconformal aperture shape ranging
Medical Physics, Vol. 36, No. 11, November 2009
5136
over the full extent of the phase 1 volume, followed by an
anticlockwise conformal arc directed to the phase 2 volume,
works well. The prostate and pelvic nodes and head-and-
neck cases require several modulated arcs for reliable plan-
ning and delivery. As a general rule, it has been found that
cases with multiple target volumes requiring multiple pre-
scription doses are best dealt with using several arcs, one
directed to each of the target volumes. Hence, the head-and-
neck case has been planned using an arc to the high-dose
volume plus an arc to each of the nodal PTVs.
The main factor influencing the complexity of the modu-
lation is the nominal gantry speed. If the inverse planning is
performed with the gantry intended to move at maximum
speed, only limited motion of the collimators and MLC
leaves is allowed. Conversely, if considerable motion of the
5137 James L. Bedford: VMAT treatment planning 5137

TABLE III. Delivery and verification statistics.

VMAT IMRT VMAT IMRT

Fraction dose VMAT IMRT treatment time treatment time verification agreement verification agreement
Treatment site 共Gy兲a 共MU兲 共MU兲 共min:s兲 共min:s兲 共%兲b 共%兲b

Lung 2.0 350 333 1:25 6:40 99.8 100.0

Prostate and seminal vesicles 2.0 396 370 2:30 7:30 99.5 98.9
Rectum 2.16 290 285 5:15 9:00 96.7 95.6
Prostate and pelvic nodes 2.0 627 462 13:05 9:40 90.6 99.7
Head and neck 2.167 938 467 15:45 13:15 88.2 95.9
a
In the cases with multiple prescription doses, the fraction size is given for the planning target volume with the highest dose.
b
Percentage of measurements agreeing with the planned dose to within 3% and 3 mm.

collimators and MLC leaves is allowed at the planning stage,
a more exquisite treatment plan will result but the delivery
time will be much longer. The parameters used for planning
the cases in this paper have been chosen to illustrate the
complete range of plan quality and delivery speed. Other
choices are possible: For example, the prostate and pelvic
nodes and head-and-neck cases could have been planned for
improved speed compared to segmental delivery. In this case,
the plan quality would not be so high, but the delivery would
be faster.
The complexity of the plan also influences the verification
results. If the nominal gantry speed is high, there is less
collimator and MLC leaf motion, which means that the dose
calculation model with interpolated collimator and MLC leaf
positions is more accurate. It is also easier for the accelerator
to deliver the plan as the delivery speeds are more uniform.
Finally, the dose measurement device, in this case the
DELTA4, has itself a finite uncertainty in measurement. It is
thought that all of these factors contribute to the failure of
the head-and-neck case to satisfy the gamma criterion. This
case has 51 control points per arc, which equates to 7° con-
trol point spacing, which is fairly large, together with a low
nominal gantry speed, giving rise to considerable motion be-
tween these control points, and these factors combine to cre-
ate a challenge for the dose calculation to model the dynamic
delivery. The motion between the control points is also chal-
lenging for the accelerator to deliver smoothly. Improving
the accuracy of planning and delivery for such cases is the
subject of an ongoing study.
The improved speed of delivery with VMAT for the
simple cases should lead to an overall improvement in effi-
ciency of treatment. VMAT lends itself naturally to simulta-
neous cone-beam imaging and treatment due to the gantry
rotation. For imaging schedules where cone-beam scans are
acquired on the first days of treatment, followed by a shift,
these initial scans could be obtained simultaneously with the
VMAT delivery,34 leading to a short treatment time. This
would be beneficial for the departmental workflow and
would also lead to improved accuracy of delivery and a more
favorable experience for the patient.
V. CONCLUSIONS
An inverse planning scheme for VMAT, using constrained
aperture shapes and modeling of MLC leaf motion between
control points, has been developed and clinically imple-
mented. Comparison studies show that this inverse planning
scheme used in conjunction with an Elekta accelerator pro-
duces treatments which are slightly improved in efficacy to
segmental IMRT, with a treatment time which is up to four
times faster.

ACKNOWLEDGMENTS
The author would like to thank Mr. Jim Warrington for his
supervision of the VMAT project and Elekta Ltd. for their
collaboration, particularly Mr. Kevin Brown and Mr. Paul
Boxall. AUTOBEAM has been developed under a grant from
The Institute of Cancer Research and the present project has
also received support from Cancer Research U.K. 共Program
Grant No. C46/A2131兲. The authors acknowledge NHS
funding to the NIHR Biomedical Research Centre.
a兲
Author to whom correspondence should be addressed. Electronic mail:
4
FIG. 7. DELTA results for the VMAT prostate and pelvic nodes plan. The james.bedford@icr.ac.uk; Telephone: ⫹44 20 8661 3477; Fax: ⫹44 20
planned dose distribution is shown in grayscale and with isodoses, the mea- 8643 3812.
1
sured dose is shown in colored points, and the histograms of dose difference, C. X. Yu, “Intensity-modulated arc therapy with dynamic multileaf colli-
distance to agreement, and gamma 共3% and 3 mm兲 are shown from left to mation: An alternative to tomotherapy,” Phys. Med. Biol. 40, 1435–1449
right below. 共1995兲.

Medical Physics, Vol. 36, No. 11, November 2009

5138 James L. Bedford: VMAT treatment planning
2
W. Duthoy, W. De Gersem, K. Vergote, M. Coghe, T. Boterberg, Y. De
Deene, C. De Wagter, S. Van Belle, and W. De Neve, “Whole abdomi-
nopelvic radiotherapy 共WAPRT兲 using intensity-modulated arc therapy
共IMAT兲: First clinical experience,” Int. J. Radiat. Oncol., Biol., Phys. 57,
1019–1032 共2003兲.
3
W. Duthoy, W. De Gersem, K. Vergote, T. Boterberg, C. Derie, P. Smeets,
C. De Wagter, and W. De Neve, “Clinical implementation of intensity-
modulated arc therapy 共IMAT兲 for rectal cancer,” Int. J. Radiat. Oncol.,
Biol., Phys. 60, 794–806 共2004兲.
4
S. M. Crooks, X. Wu, C. Takita, M. Watzich, and L. Xing, “Aperture
modulated arc therapy,” Phys. Med. Biol. 48, 1333–1344 共2003兲.
5
M. A. Earl, D. M. Shepard, S. Naqvi, X. A. Li, and C. X. Yu, “Inverse
planning for intensity-modulated arc therapy using direct aperture optimi-
zation,” Phys. Med. Biol. 48, 1075–1089 共2003兲.
6
C. Cameron, “Sweeping-window arc therapy: An implementation of ro-
tational IMRT with automatic beam-weight calculation,” Phys. Med. Biol.
50, 4317–4336 共2005兲.
7
K. Bratengeier, “2-step IMAT and 2-step IMRT: A geometrical approach,”
Med. Phys. 32, 777–785 共2005兲.
8
K. Bratengeier, “2-step IMAT and 2-step IMRT in three dimensions,”
Med. Phys. 32, 3849–3861 共2005兲.
9
S. Webb, Contemporary IMRT: Developing Physics and Clinical Imple-
mentation 共Institute of Physics Publishing, Bristol, 2005兲.
10
D. M. Shepard, D. Cao, M. K. N. Afghan, and M. A. Earl, “An arc-
sequencing algorithm for intensity modulated arc therapy,” Med. Phys.
34, 464–470 共2007兲.
11
A. Gladwish, M. Oliver, J. Craig, J. Chen, G. Bauman, B. Fisher, and E.
Wong, “Segmentation and leaf sequencing for intensity modulated arc
therapy,” Med. Phys. 34, 1779–1788 共2007兲.
12
S. Ulrich, S. Nill, and U. Oelfke, “Development of an optimization con-
cept for arc-modulated cone beam therapy,” Phys. Med. Biol. 52, 4099–
4119 共2007兲.
13
K. Otto, “Volumetric modulated arc therapy: IMRT in a single gantry
arc,” Med. Phys. 35, 310–317 共2008兲.
14
C. Wang, S. Luan, G. Tang, D. Z. Chen, M. A. Earl, and C. X. Yu,
“Arc-modulated radiation therapy 共AMRT兲: A single-arc form of
intensity-modulated arc therapy,” Phys. Med. Biol. 53, 6291–6303
共2008兲.
15
S. Luan, C. Wang, D. Cao, D. Z. Chen, D. M. Shepard, and C. X. Yu,
“Leaf-sequencing for intensity-modulated arc therapy using graph algo-
rithms,” Med. Phys. 35, 61–69 共2008兲.
16
K. Bzdusek, H. Friberger, K. Eriksson, B. Hårdemark, D. Robinson, and
M. Kaus, “Development and evaluation of an efficient approach to volu-
metric arc therapy planning,” Med. Phys. 36, 2328–2339 共2009兲.
17
D. Cao, T. W. Holmes, M. K. N. Afghan, and D. M. Shepard, “Compari-
son of plan quality provided by intensity-modulated arc therapy and he-
lical tomotherapy,” Int. J. Radiat. Oncol., Biol., Phys. 69, 240–250
共2007兲.
18
L. Cozzi, A. Clivio, G. Bauman, S. Cora, G. Nicolini, R. Pellegrini, E.
Vanetti, S. Yartsev, and A. Fogliata, “Comparison of advanced irradiation
techniques with photons for benign intracranial tumours,” Radiother. On-
col. 80, 268–273 共2006兲.
19
A. Fogliata, A. Clivio, G. Nicolini, E. Vanetti, and L. Cozzi, “A treatment
Medical Physics, Vol. 36, No. 11, November 2009
5138
planning study using non-coplanar static fields and coplanar arcs for
whole breast radiotherapy of patients with concave geometry,” Radiother.
Oncol. 85, 346–354 共2007兲.
20
D. Palma, E. Vollans, K. James, S. Nakano, V. Moiseenko, R. Shaffer, M.
McKenzie, J. Morris, and K. Otto, “Volumetric modulated arc therapy for
delivery of prostate radiotherapy: Comparison with intensity-modulated
radiotherapy and three-dimensional conformal radiotherapy,” Int. J. Ra-
diat. Oncol., Biol., Phys. 72, 996–1001 共2008兲.
21
M. Iori, G. M. Cattaneo, E. Cagni, C. Fiorino, G. Borasi, C. Riccardo, C.
Iotti, F. Fazio, and A. E. Nahum, “Dose-volume and biological-model
based comparison between helical tomotherapy and 共inverse-planned兲
IMAT for prostate tumours,” Radiother. Oncol. 88, 34–45 共2008兲.
22
K. Yoda, K. Nakagawa, K. Shiraishi, Y. Okano, K. Ohtomo, and R. G.
Pellegrini, “Dose verification of intensity-modulated arc therapy using an
ERGO+ + treatment planning system and Elekta internal multileaf colli-
mators for prostate cancer treatment,” Br. J. Radiol. 82, 328–331 共2009兲.
23
P. Zygmanski, W. Högele, R. Cormack, L. Chin, and R. Löschel, “A
volumetric-modulated arc therapy using sub-conformal dynamic arc with
a monotonic dynamic multileaf collimator modulation,” Phys. Med. Biol.
53, 6395–6417 共2008兲.
24
T. Bortfeld and S. Webb, “Single-arc IMRT?,” Phys. Med. Biol. 54, N9–
N20 共2009兲.
25
J. L. Bedford and S. Webb, “Elimination of importance factors for clini-
cally accurate selection of beam orientations, beam weights and wedge
angles in conformal radiation therapy,” Med. Phys. 30, 1788–1804
共2003兲.
26
J. L. Bedford and S. Webb, “Constrained segment shapes in direct-
aperture optimization for step-and-shoot IMRT,” Med. Phys. 33, 944–958
共2006兲.
27
J. L. Bedford and S. Webb, “Direct-aperture optimization applied to se-
lection of beam orientations in intensity-modulated radiation therapy,”
Phys. Med. Biol. 52, 479–498 共2007兲.
28
L. Xing and G. T. Y. Chen, “Iterative methods for inverse treatment plan-
ning,” Phys. Med. Biol. 41, 2107–2123 共1996兲.
29
W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Nu-
merical Recipes in C 共Cambridge University Press, Cambridge, 1992兲.
30
J. L. Bedford, “Speed versus accuracy in a fast convolution photon dose
calculation for conformal radiotherapy,” Phys. Med. Biol. 47, 3475–3484
共2002兲.
31
J. L. Bedford, V. N. Hansen, H. A. McNair, A. H. Aitken, J. E. C. Brock,
A. P. Warrington, and M. Brada, “Treatment of lung cancer using volu-
metric modulated arc therapy and image guidance: A case study,” Acta
Oncol. 47, 1438–1443 共2008兲.
32
J. L. Bedford and A. P. Warrington, “Commissioning of volumetric modu-
lated arc therapy 共VMAT兲,” Int. J. Radiat. Oncol., Biol., Phys. 73, 537–
545 共2009兲.
33
J. L. Bedford, Y. K. Lee, P. Wai, C. P. South, and A. P. Warrington,
“Evaluation of the Delta4 phantom for IMRT and VMAT verification,”
Phys. Med. Biol. 54, N167–N176 共2009兲.
34
K. Nakagawa, A. Haga, K. Shiraishi, H. Yamashita, H. Igaki, A. Terahara,
K. Ohtomo, S. Saegusa, T. Shiraki, T. Oritate, and K. Yoda, “First clinical
cone-beam CT imaging during volumetric modulated arc therapy,” Ra-
diother. Oncol. 90, 422–423 共2009兲.