Professional Documents
Culture Documents
Volumetric Modulated Arc Therapy: IMRT in A Single Gantry Arc
Volumetric Modulated Arc Therapy: IMRT in A Single Gantry Arc
Karl Ottoa兲
Vancouver Cancer Centre, BC Cancer Agency, Vancouver, British Columbia V5Z 4E6, Canada
共Received 25 June 2007; revised 21 September 2007; accepted for publication 5 November 2007;
published 26 December 2007兲
In this work a novel plan optimization platform is presented where treatment is delivered efficiently
and accurately in a single dynamically modulated arc. Improvements in patient care achieved
through image-guided positioning and plan adaptation have resulted in an increase in overall treat-
ment times. Intensity-modulated radiation therapy 共IMRT兲 has also increased treatment time by
requiring a larger number of beam directions, increased monitor units 共MU兲, and, in the case of
tomotherapy, a slice-by-slice delivery. In order to maintain a similar level of patient throughput it
will be necessary to increase the efficiency of treatment delivery. The solution proposed here is a
novel aperture-based algorithm for treatment plan optimization where dose is delivered during a
single gantry arc of up to 360 deg. The technique is similar to tomotherapy in that a full 360 deg of
beam directions are available for optimization but is fundamentally different in that the entire dose
volume is delivered in a single source rotation. The new technique is referred to as volumetric
modulated arc therapy 共VMAT兲. Multileaf collimator 共MLC兲 leaf motion and number of MU per
degree of gantry rotation is restricted during the optimization so that gantry rotation speed, leaf
translation speed, and dose rate maxima do not excessively limit the delivery efficiency. During
planning, investigators model continuous gantry motion by a coarse sampling of static gantry
positions and fluence maps or MLC aperture shapes. The technique presented here is unique in that
gantry and MLC position sampling is progressively increased throughout the optimization. Using
the full gantry range will theoretically provide increased flexibility in generating highly conformal
treatment plans. In practice, the additional flexibility is somewhat negated by the additional con-
straints placed on the amount of MLC leaf motion between gantry samples. A series of studies are
performed that characterize the relationship between gantry and MLC sampling, dose modeling
accuracy, and optimization time. Results show that gantry angle and MLC sample spacing as low as
1 deg and 0.5 cm, respectively, is desirable for accurate dose modeling. It is also shown that
reducing the sample spacing dramatically reduces the ability of the optimization to arrive at a
solution. The competing benefits of having small and large sample spacing are mutually realized
using the progressive sampling technique described here. Preliminary results show that plans gen-
erated with VMAT optimization exhibit dose distributions equivalent or superior to static gantry
IMRT. Timing studies have shown that the VMAT technique is well suited for on-line verification
and adaptation with delivery times that are reduced to ⬃1.5– 3 min for a 200 cGy
fraction. © 2008 American Association of Physicists in Medicine.
关DOI: 10.1118/1.2818738兴
310 Med. Phys. 35 „1…, January 2008 0094-2405/2008/35„1…/310/8/$23.00 © 2008 Am. Assoc. Phys. Med. 310
311 Karl Otto: Single arc radiation therapy 311
MLC leaf positions. Each sample is effectively a “snapshot” weights are physically impossible and are therefore rejected
of the gantry and MLC at a time point during the delivery. by the optimization. Constraints are also placed on MLC leaf
Continuous gantry and MLC motion is approximated by in- motion and MU variation to preserve continuous delivery.
dividual beams that are optimized at 5–10 deg gantry angle These are referred to as “efficiency constraints.” Defined in
increments with MLC leaf positions varying by up to 2–4 cm terms of gantry rotation angle, they are
between gantry angle samples.9–14 Poor sampling can sub-
stantially degrade plan accuracy, thereby reducing the quality
of the delivered plan and potentially resulting in an unaccept-
⌬x
⌬
ⱕ 冉 冊
dx
d max
,
冉 冊
able dosimetric error.10 Increasing the gantry and MLC sam-
pling frequency will improve accuracy but it will also further ⌬MU dMU
ⱕ , 共1兲
restrict the ability of the optimization and leaf sequencing ⌬ d max
algorithms to derive an acceptable plan. VMAT provides a where x, MU, and are MLC leaf position, MU weight, and
novel solution to several limitations encountered with IMAT. gantry angle, respectively. 共dx / d兲max and 共dMU/ d兲max are
In particular, the VMAT technique is designed so that opti- selected so that there is no particular component of the de-
mized plans may be delivered: livery system compromising a continuous and efficient deliv-
共1兲 Efficiently, in a single gantry arc ery. For example, at maximum speed the gantry rotates
共2兲 With high dose conformality, using a full 360 deg range through a 360 deg arc in 60 s 共共d / dt兲max = 6 deg/ s兲. The
of gantry directions maximum leaf motion speed of the Varian Millennium 120
共3兲 Accurately, with high-resolution sampling of beam di- leaf MLC is 共dx / dt兲max = 3.0 cm/ s. By constraining VMAT
rections during planning. delivery to a maximum leaf displacement of 0.5 cm per de-
gree of gantry rotation,
II. METHODS 冉 冊 冉 冊 冒冉 冊
dx
d max
=
dx
dt max
d
dt max
= 0.5 cm/deg, 共2兲
II.A. Optimization
the maximum total time for MLC motion over a 360 deg arc
The classical approach to IMRT and IMAT planning in- is 60 s, which matches the 60 s gantry rotation period.
volves optimization of fluence maps followed by a MLC leaf Similar constraints defined for 共dMU/ d兲max are based on
sequencing step. An alternative approach is to predefine a the gantry rotation speed and maximum dose rate, where
冉 冊 冉 冊 冒冉 冊
series of beam’s eye view 共BEV兲 aperture shapes that are
Boolean operations of target and healthy tissue structures15 dMU dMU d
= . 共3兲
共e.g., BEV of prostate excluding rectum兲. The number of MU d max dt max dt max
共radiation output兲 for each BEV aperture is then optimized In practice, a MU weight that causes the maximum dose rate
based on a dose-volume cost function. Both of these ap- setting 共dMU/ dt兲max to be exceeded can be delivered by re-
proaches have been applied to IMAT.9–11,14,16–19 Recently, ducing gantry rotation speed 共d / dt兲. Deceleration of the
several investigators have proposed directly optimizing leaf
gantry is undesirable because it will result in increased treat-
positions and segment weights for both static gantry IMRT
ment time as well as potentially less accurate delivery due to
共Refs. 20–22兲 as well as IMAT.11 Generally, this optimization
the substantial angular momentum of the linac apparatus.
technique involves selecting a MLC aperture, then changing
The maximum MU weight is therefore also constrained
either the MU weight or a MLC leaf position for that sample.
throughout the optimization to ensure that the maximum
VMAT dose optimization employs a similar aperture-based
dose rate is rarely exceeded.
method that incorporates MLC leaf positions and MU
Each iteration of the optimization involves randomly se-
weights as optimization parameters. The cost function is
lecting an available gantry sample, then changing either the
based on dose-volume constraints and is defined using the
MU weight or a MLC leaf position for that sample. If a
formulation proposed by Bortfeld et al.23 Minimum and
proposed change does not violate a mechanical or efficiency
maximum dose constraints are specified as a function of vol-
constraint the dose distribution and cost function are calcu-
ume 共e.g., 67 Gy to 95% volume兲 for target and healthy
lated. If the cost is reduced the change is accepted, otherwise
tissue structures individually. For each constraint a relative
it is rejected.
priority or importance value is also assigned. The cost is
calculated for each constraint using a standard quadratic dose
difference function multiplied by the priority value. The total II.C. Progressive sampling of gantry and MLC
cost is the sum total of all individual constraint cost values. positions
Of central importance to the VMAT algorithm is the
II.B. Optimization constraints
mechanism used to sample the dynamic source motion by a
During the optimization MLC leaf positions or MU finite number of static beams. All techniques that involve
weights are constrained such that the aperture shapes and source motion 共e.g., IMAT and tomotherapy兲 model the mov-
MU values are physically achievable in practice. For ex- ing linac source as a series of static source position samples.
ample, overlapping of opposing leaves or negative MU The instantaneous MLC configuration is defined at each
FIG. 1. Continuous source and MLC motion is initially modeled as a series of static source positions in 共a兲. As the optimization progresses new samples are
introduced with the first new sample placed between the first and second existing samples in 共b兲. Further samples are added in 共c兲 to span the full range of
gantry motion. The next sample is added at the beginning of the gantry range in 共d兲. Samples are continuously added in this way until a desired sampling
frequency is reached.
sample, and the MU setting for the sampling interval is as- for sample S + 1 and S − 1 as shown in Eqs. 共4兲–共6兲. Each
signed to that static MLC configuration. For example, a gan- time a new sample is added, the VMAT algorithm continues
try rotation from 0 to 90 deg can be modeled as a series of to optimize both the previous beam samples as well as the
nine evenly spaced samples each covering a 10 deg range. newly added sample. Samples are further introduced using
The source position and instantaneous MLC configuration the schedule depicted in Fig. 1共c兲. After the full gantry range
are defined at 5, 15, 25, 35, 45, 55, 65, 75, and 85 deg of has been resampled 共effectively doubling the sampling fre-
gantry rotation. MU values set at these source positions are quency兲, the process continues by returning to the start of the
for the ranges 0–10, 10–20, 20–30, 30–40, 40–50, 50–60, gantry range as shown in Fig. 1共d兲. In this way the number of
60–70, 70–80, and 80–90 deg, respectively. For the actual samples is continually expanded until a desired sampling fre-
delivery, the MLC leaves and gantry move linearly in be- quency is attained.
tween each sample. Clearly, the accuracy of the delivered Using a computer with 3.0 GHz Xeon processor each it-
plan will depend on the amount MLC leaf positions change eration 共dose and cost calculation兲 takes approximately 1 ms
and how far the gantry rotates in between each sample.10,24 while each sample addition takes approximately 1 s. There is
At the start of VMAT optimization a relatively coarse also a preprocessing step that takes approximately 5 min. All
sampling of the gantry positions is used to model the gantry calculation times depend on the size of the target volume共s兲
rotation range. Samples are included at the beginning and as well as the complexity of the case. Total optimization
end of the range with evenly distributed samples in between times vary from 10 min to 1 h. Algorithms are currently
as depicted in Fig. 1共a兲. After a number of iterations 共MLC implemented in the MATLAB software prototyping environ-
and/or MU weight changes兲 an additional sample is added to ment. There are several noteworthy characteristics to this
the pool of optimizable gantry positions. The new sample is approach. The first is with regard to the relationship between
added midway between two existing samples. MLC posi- the sampling interval and the efficiency constraints. The ef-
tions for the new sample are linearly interpolated from the ficiency constraints defined in Eq. 共1兲 can be rewritten as
冉 冊
MLC positions of the adjacent samples. The MU weight of
dx
the new sample is a function of the MU weight of the adja- ⌬x ⱕ ⌬ ,
cent samples. More precisely, when the new sample is intro- d max
冉 冊
duced the sampling interval covered by the two adjacent
samples is reduced to accommodate the new sample, dMU
⌬MU ⱕ ⌬ . 共7兲
d
MUold共S − 1兲 MUold共S + 1兲 max
MUnew共S兲 = + , 共4兲
3 4 The MLC leaf position change 共⌬x兲 and MU weight 共⌬MU兲
permitted between samples is directly proportional to the
2 MUold共S − 1兲 gantry rotation range 共⌬兲 between samples. When a coarse
MUnew共S − 1兲 = , 共5兲 sampling of gantry positions is used, there is a larger gantry
3
range between samples 共i.e., larger ⌬兲. Therefore, from Eq.
3 MUold共S + 1兲 共7兲, a coarse sampling of gantry positions will allow for a
MUnew共S + 1兲 = , 共6兲 larger maximum MLC leaf displacement and MU weight
4
change at any iteration, thereby providing more flexibility in
where MUnew, MUold, and S are the new MU weight, the old deriving an optimal plan. Conversely, as new samples are
MU weight 共prior to the new sample兲, and the sample index introduced ⌬ will decrease and reduce the flexibility of the
of the new sample, respectively. Nonuniform sampling inter- optimization. The overall effect is that VMAT optimization is
vals arise during the introduction of new samples 关see Figs. effectively not restricted by efficiency constraints at the start
1共b兲 and 1共d兲兴 and result in a nonuniform weighting of adja- but gradually becomes more restricted as new samples are
cent samples where new samples are being added. This effect introduced. A similar characteristic related to the plan accu-
is accounted for by nonuniformly redistributing MU weights racy is also observed. Due to the large gantry range between
samples at the beginning of optimization the accuracy of the cost value. The degree to which optimization flexibility is
optimized plan with respect to the plan that would be deliv- reduced because of higher sampling was characterized by
ered will be relatively poor. As new samples are introduced determining the relationship between the gantry/MLC sam-
the sampling frequency increases, thereby improving the pling frequency and CPU time.
modeling accuracy of dynamic MLC and gantry motion. The
final number of optimized samples is chosen to ensure a II.F. Nasopharynx example
sampling frequency that provides adequate dosimetric accu-
The VMAT technique was applied to a nasopharynx car-
racy.
cinoma patient. During planning, minimum and maximum
Samples are introduced throughout the optimization using
dose-volume constraints were placed on three separate PTVs
a schedule that is a function of the optimization progress.
and multiple critical structures 共brainstem, spinal cord, pa-
The interval 共e.g., CPU time or number of successful itera-
rotids, brain, optical apparatus兲 and are identical to those
tions兲 between the addition of each new sample follows the
used for the patient’s fixed gantry clinical treatment plan
form of a decreasing exponential. When there are fewer
共Eclipse v6.5, Varian Medical Systems Inc.兲. A VMAT plan
samples, i.e., at the beginning of the optimization, the num-
was generated with the optimization time fixed at 60 min
ber of iterations between each sample addition is relatively
共3.2 GHz Xeon processor, 2 GB RAM兲. For comparison, 360
large. As the number of samples increases there are exponen-
deg arc plans using fixed sampling 共constant number of
tially fewer iterations between each new sample. The optimi-
samples兲 were also generated with a range of different sam-
zation terminates once all samples have been introduced.
pling frequencies. The final cost values of these plans were
Each time a sample is added the cost is recalculated with the
then compared to the final cost of the VMAT plan.
new beam configuration. The addition of a new sample will
Finally, plan quality and delivery time were evaluated by
generally perturb the optimization progress, causing a tem-
comparing the VMAT plan to a static gantry 7 beam IMRT
porary increase in the cost. The magnitude of this effect is
plan 共Eclipse––sliding window25兲. Identical constraints were
larger when there are fewer samples but becomes increasing
used in generating the VMAT and static gantry plans. VMAT
less significant as more samples are added. The schedule
plans are delivered on a Varian Cl21EX linac using research
described above is specifically designed to compensate for
control software provided by Varian Medical Systems Inc.
these perturbations by having exponentially more iterations
MLC leaf positions, gantry angles, and fractional MU values
between sample additions when there are fewer samples.
are transferred and monitored by the software throughout the
delivery. A 200 cGy fraction was delivered using 600 MU/
II.D. Dosimetric accuracy and sampling
min maximum dose rate setting.
To maximize dosimetric accuracy when approximating
dynamic source and collimation by multiple static beams, it III. RESULTS
is necessary to determine the minimum sampling frequency
III.A. Dosimetric accuracy and sampling
for gantry and MLC leaf positions. For this purpose a series
of VMAT plans was generated with varying sampling reso- The sampling requirements for VMAT planning were de-
lution. These plans were then compared to the same plan but termined using VMAT plans generated with a range of gantry
resampled at high frequency, which more closely approxi- and maximum MLC leaf sampling. For these plans the gan-
mates true dynamic motion. From these results an estimate try positions were interpolated to high resolution with 0.5
of the required sampling frequency is determined. deg of gantry rotation in between samples. Correspondingly,
the MLC leaves move a maximum of 0.25 cm between
II.E. Optimization time and fixed sampling samples as can be calculated from Eq. 共2兲. Dose calculation
error resulting from different sample spacing was determined
A study was performed to evaluate the effect of different
by comparing each plan with its corresponding highly
sampling frequencies on optimization time. Increasing the
sampled “reference” plan. The results of this study for the
sampling frequency will increase the number of optimizable
nasopharynx carcinoma patient are shown in Fig. 2. Other
gantry positions and restrict the maximum MLC leaf position
treatment sites showed similar results to those presented
changes in between consecutive samples 关Eq. 共6兲兴. The im-
here. From this study it was determined that a sampling of 1
pact of sampling on optimization time was assessed using a
deg and a maximum of 0.5 cm is required for gantry rotation
series of arc plans generated with fixed sampling. These
and MLC leaf motions, respectively, in order to ensure that
plans were created without the progressive sampling method
less than 5% of the volume has a dose error ⬎3%. This
共i.e., not VMAT兲 so that a fixed number of samples are used
threshold was chosen to coincide with typical IMRT quality
throughout the entire optimization. Gantry rotation is re-
assurance criteria. Note that the results of this study are in-
stricted to a single 360 deg arc in order to match the beam
dependent of any errors associated with the delivery device,
delivery geometry of VMAT. Optimization of leaf position
including modeling of MLC characteristics.
and MU weights was performed using a simulated
annealing-based method. For these plans, simulated anneal-
III.B. Optimization time and fixed sampling
ing is required so that the optimization does not become
trapped in a local minimum. The CPU time was recorded for Based on the results from the previous section, VMAT
each plan once the optimization had reached a predetermined plans are generated with gantry angles and MLC positions
FIG. 5. VMAT and Eclipse isodoses 共95% of PTV70, PTV56, and 31.5 Gy兲
are shown on a representative CT slice for the nasopharynx example. Dose
distributions have substantially different shapes due to the use of 360 deg of
gantry rotation with VMAT and seven fixed beam directions with Eclipse.
5
larly well suited to modeling dynamic source and MLC mo- B. C. Davis et al., “Automatic segmentation of intra-treatment CT images
for adaptive radiation therapy of the prostate,” in International Confer-
tion due to the particle-by-particle calculation framework. ence on Medical Image Computing and Computer Assisted Intervention
The calculation time for a Monte Carlo simulation to achieve 共2005兲, Vol. 8, pp. 442–450.
6
a desired level of uncertainty is proportional to the number A. Mestrovic et al., “Direct aperture optimization for online adaptive
of simulated particles and not the number of beams, making radiation therapy,” Med. Phys. 34, 1631–1646 共2007兲.
7
T. R. Mackie et al., “Tomotherapy: A new concept for the delivery of
it potentially more time efficient as well as more accurate dynamic conformal radiotherapy,” Med. Phys. 20, 1709–1719 共1993兲.
than existing algorithms for modeling dynamic source and 8
M. Carol et al., “Initial clinical experience with the Peacock intensity
MLC motion. modulation of a 3-D conformal radiation therapy system,” Stereotact.
Finally, it should be noted that the VMAT technique is not Funct. Neurosurg. 66, 30–34 共1996兲.
9
C. X. Yu, “Intensity-modulated arc therapy with dynamic multileaf colli-
restricted to a 360 deg arc. Reduced arc ranges or non- mation: An alternative to tomotherapy,” Phys. Med. Biol. 40, 1435–1449
coplanar arcs may be desirable in some cases 共e.g., for avoid- 共1995兲.
10
ing beam directions where radiation is transmitted through D. M. Shepard et al., “An arc-sequencing algorithm for intensity modu-
immobilization devices兲. Furthermore, there is no inherent lated arc therapy,” Med. Phys. 34, 464–470 共2007兲.
11
M. A. Earl et al., “Inverse planning for intensity-modulated arc therapy
restriction on the use of overlapping arcs, although prelimi- using direct aperture optimization,” Phys. Med. Biol. 48, 1075–1089
nary results indicate that doing so provides only minimal 共2003兲.
12
improvements in plan quality. A. Gladwish et al., “Segmentation and leaf sequencing for intensity
modulated arc therapy,” Med. Phys. 34, 1779–1788 共2007兲.
Treatment site-specific planning studies must be per- 13
C. X. Yu et al., “Clinical implementation of intensity-modulated arc
formed before any formal conclusions can be made regarding therapy,” Int. J. Radiat. Oncol. Biol. Phys. 53, 453–463 共2002兲.
14
the dose conformality capabilities of VMAT with respect to E. Wong, J. Z. Chen, and J. Greenland, “Intensity-modulated arc therapy
other planning/delivery methods. This is the primary focus of simplified,” Int. J. Radiat. Oncol. Biol. Phys. 53, 222–235 共2002兲.
15
W. De Gersem et al., “An anatomy-based beam segmentation tool for
ongoing development. Thus far, results have shown that intensity-modulated radiation therapy and its application to head-and-
treatment plans generated with VMAT optimization have neck cancer,” Int. J. Radiat. Oncol. Biol. Phys. 51, 849–859 共2001兲.
16
dose distributions that are equivalent to or superior to a com- K. Bratengeier, “2-Step IMAT and 2-Step IMRT in three dimensions,”
mercial static gantry IMRT system. Med. Phys. 32, 3849–3861 共2005兲.
17
C. Cameron, “Sweeping-window arc therapy: An implementation of ro-
tational IMRT with automatic beam-weight calculation,” Phys. Med. Biol.
V. CONCLUSION 50, 4317–4336 共2005兲.
18
S. M. Crooks et al., “Aperture modulated arc therapy,” Phys. Med. Biol.
VMAT is a treatment planning and delivery platform us- 48, 1333–1344 共2003兲.
19
ing a single 360 deg gantry arc. Novel optimization tech- M. A. MacKenzie and D. M. Robinson, “Intensity modulated arc deliv-
niques using progressive gantry and MLC sampling are ca- eries approximated by a large number of fixed gantry position sliding
window dynamic multileaf collimator fields,” Med. Phys. 29, 2359–2365
pable of producing accurate highly conformal dose 共2002兲.
20
distributions with superior delivery efficiency. The introduc- W. De Gersem et al., “Leaf position optimization for step-and-shoot
tion of new methods for on-line imaging and adaptation will IMRT,” Int. J. Radiat. Oncol. Biol. Phys. 51, 1371–1388 共2001兲.
21
D. M. Shepard et al., “Direct aperture optimization: A turnkey solution
inevitably increase radiation therapy treatment times, with
for step-and-shoot IMRT,” Med. Phys. 29, 1007–1018 共2002兲.
accurate motion compensation somewhat limited by the 22
M.-P. Milette and K. Otto, “Maximizing the potential of direct aperture
beam delivery time. Time studies have shown that VMAT optimization through collimator rotation,” Med. Phys. 34, 1431–1438
optimization is well suited for on-line verification and adap- 共2007兲.
23
T. Bortfeld, J. Stein, and K. Preiser, “Clinically relevant intensity modu-
tation with delivery times that are reduced to ⬃1.5– 3 min lation optimization using physical criteria,” in XII International Confer-
for a 200 cGy fraction. ence on the Use of Computers in Radiation Therapy, Salt Lake City, UT,
1997, pp. 1–4.
24
M. W. Kissick, T. R. Mackie, and R. Jeraj, “A delivery transfer function
ACKNOWLEDGMENT 共DTF兲 analysis for helical tomotherapy,” Phys. Med. Biol. 52, 2355–2365
This work was supported in part by Varian Medical Sys- 共2007兲.
25
S. V. Spirou and C. S. Chui, “Generation of arbitrary intensity profiles by
tems. dynamic jaws or multileaf collimators,” Med. Phys. 21, 1031–1041
共1994兲.
a兲 26
Electronic mail: kotto@bccancer.bc.ca J. L. Bedford and S. Webb, “Constrained segment shapes in direct-
1
D. Yan et al., “Adaptive radiation therapy,” Phys. Med. Biol. 42, 123–132 aperture optimization for step-and-shoot IMRT,” Med. Phys. 33, 944–958
共1997兲. 共2006兲.
2 27
P. J. Keall et al., “Motion adaptive x-ray therapy: A feasibility study,” C. Cotrutz and L. Xing, “Segment-based dose optimization using a ge-
Phys. Med. Biol. 46, 1–10 共2001兲. netic algorithm,” Phys. Med. Biol. 48, 2987–2998 共2003兲.
3 28
H. Rehbinder, C. Forsgren, and J. Lof, “Adaptive radiation therapy for P. R. Storchi, L. J. van Battum, and E. Woudstra, “Calculation of a pencil
compensation of errors in patient setup and treatment delivery,” Med. beam kernel from measured photon beam data,” Phys. Med. Biol. 44,
Phys. 31, 3363–3371 共2004兲. 2917–2928 共1999兲.
4 29
L. E. Court et al., “An automatic CT-guided adaptive radiation therapy P. Storchi and E. Woudstra, “Calculation of the absorbed dose distribution
technique by online modification of multileaf collimator leaf positions for due to irregularly shaped photon beams using pencil beam kernels derived
prostate cancer,” Int. J. Radiat. Oncol. Biol. Phys. 62, 154–163 共2005兲. from basic beam data,” Phys. Med. Biol. 41, 637–656 共1996兲.