Int. 1. Radiation Oncology Biol. Phys., Vol. 32, No. 4, pp.

1215- 1225, 1995

Copyright 0 1995 Elsevier Science Ltd
Pergamon Printed in the USA. All rights reserved
0360-3016/95 $9.50 + 00

0360-3016(94)00465-X

l Technical Innovations and Notes

AN ITERATIVE FILTERED BACKPRO JECTION INVERSE TREATMENT

PLANNING ALGORITHM FOR TOMOTHERAPY

TIMOTHY W. HOLMES, PH.D., T. ROCK MACRIE, PH.D. AND PAUL RECKWERDT, B.S.
Departmentsof Medical Physicsand HumanOncology, University of WisconsinSchool of Medicine,
1300University Avenue, Madison, WI

Purpose: An inverse treatment planning algorithm for tomotherapy is described.

Methods and Materials: The algorithm iteratively computes a set of nonnegative beam intensity profiles
that minimizes the least-squares residual dose defined in the target and selected normal tissue regions of
interest. At each iteration the residual dose distribution is transformed into a set of residual beam profiles
using an inversion method derived from filtered backprojection image reconstruction theory. These “resid-
ual” profiles are used to correct the current beam profile estimates resulting in new profile estimates.
Adaptive filtering is incorporated into the inversion model so that the gross structure of the dose distribution
is optimized during initial iterations of the algorithm, and the fine structure corresponding to edges is
obtained at later iterations. A three dimensional, kernel based, convolution/superposition dose model is
used to compute dose during each iteration.
Results: Two clinically relevant treatment planning examples are presented illustrating the use of the
algorithm for planning conformal radiotherapy of the breast and the prostate. Solutions are generally
achieved in lo-20 iterations requiring about 20 h of CPU time using a midrange workstation. The majority
of the calculation time is spent on the three-dimensional dose calculation.
Conclusions: The inverse treatment planning algorithm is a useful research tool for exploring the potential
of tomotherapy for conformal radiotherapy. Further work is needed to (a) achieve clinically acceptable
computation times; (b) verify the algorithm using multileaf collimator technology; and (c) extend the method
to biological objectives.

Radiation therapy, Tomotherapy, Treatment planning, Optimization, Image reconstruction, Filtered back-
projection.

INTRODUCTION a configuration of radiation sources that achieves a high

The introduction of computers into the field of radiation
oncology, about 30 years ago, carried the promise of im-
proved treatment techniques through pretreatment model-
ing and optimization of radiation dose distributions and
computer controlled treatment delivery. During this time
much of the focus has been on the development of more
accurate and efficient dose calculation algorithms. Until
recently, comparatively less emphasishas been placed on
1) providing optimization algorithms that use these dose
calculation models, and 2) treatment delivery technology
that conveniently allows for clinical realization of opti-
mized plans.
Radiotherapy optimization is the process of selecting
degree of tumor control while simultaneously attempting
to minimize the amount of normal tissue damage. This
process involves the application of biological response
data by an oncologist, either implicitly in the form of a
prescribed dose distribution, or explicitly in the form of
dose responsemodels used to compute bioeffect probabil-
ities for comparing rival treatment plans. In the first in-
stance, the dose prescription represents a surrogate for
biological response. The distinction leads to two optimi-
zation approaches.Physical optimization methodsassume
that the prescribed dose distribution is the biologically
optimal dose distribution. The prescription is nearly al-
ways physically unrealizable because zero dose is typi-
cally assigned to normal tissues. These methods attempt

Reprint requeststo: Timothy W. Holmes, Ph.D., Southern
WisconsinRadiotherapyCenter, 1102John Nolen Drive, Madi-
son, Wisconsin53713.
Acknowledgments- We acknowledgeR. H. Huesman,G. T.
Gullberg, W. L. Greenberg,and T. F. Budinger, the authorsof
the “Donner Algorithms for ReconstructionTomography,” for
the use of this software (15). We would like to thank Joseph
1215
0. Deasyfor many stimulatingdiscussionson radiotherapyopti-
mization. This work wassupported,in part, by NC1grant num-
ber CA48902. T. Holmeswas supported,in part, by National
ResearchService Award CA09206 and a grant from the Wis-
consinFoundation.
Accepted for publication 19 August 1994.
1216 I. J. Radiation Oncology l Biology 0 Physics Volume 32. Number 4, 1995

to automate conventional treatment planning normally A. Tomotherapy

performed by a dosimetrist. Biological optimization meth-
ods determine the dose distribution that has the best bio-
logical “figure-of-merit” without any a priori bias regard-
ing what the optimal dose distribution should look like
(18). The underlying biological mechanisms of radiation
damage and repair must be properly modeled. At this
time there is still great uncertainty in the actual form of
the biological response model, owing to an insufficient
amount of reliable clinical response data (7). Conversely,
the physical processes of particle interactions in matter
are well understood and can be accurately modeled, yet
the prescribed dose distribution may not correspond to
the optimal biological effect.
Conventional treatment planning attempts to arrive at
the best approximation of the desired dose distribution
(or the best biological figure-of-merit) by determining the
set of beam profiles and orientations through trial and
error. Historically, this process has been carried out manu-
ally by a dosimetrist. The limitations of equipment, time,
and personnel required to fabricate custom blocks and
compensators and execute treatment delivery has re-
stricted the maximum number of beam portals to be clini-
cally practical (< 4-6). Until recently, the technology of
automatically creating and delivering a large number of B. Tomotherapy Unit
modulated beams was not available to most clinics. This
probably explains why automated optimization methods
have never found popular application even though a num-
ber of approaches were described from the mid-1960s to
the mid-1980s (1, 13, 14, 20, 26, 27, 29, 32, 33).
A renewed interest in automated optimization methods
has occurred since the late 1980s due in part, no doubt, to
the commercial availability of computer-controlled linear
accelerators and multileaf collimators. This new technol-
ogy may make it possible to conveniently treat a configu-
ration of a large number of beams that realize the promise
of conformal radiotherapy techniques initially developed
over 20-30 years ago by Green and colleagues in England
(6, 16), and Takahashi in Japan (34). Several efforts are
underway to achieve conformational dose distributions
using computer-controlled collimation systems. Several
investigators have proposed methods based on modulat-
ing the profile of beam intensity by subdividing the treat-
ment volume into subfields that are differentially modu-
lated, using either asymmetric collimators or two-dimen-
sional (2D) multileaf collimators (5, 19, 25, 37). These Fig. 1. (A) Tomotherapy is carried out by rotating the x-ray
beamwhile the patientis translatedthoughthe planeof irradia-
methods should allow the physical realization of modu- tion. (B) Schematicof the tomotherapyunit proposedby Ma&e
lated beam profiles computed by any of a number of et al. (23).
inverse radiation treatment planning algorithms (2, 4, 12,
18, 21, 30, 35, 36) including the algorithm described in
this work. through the gantry bore (Fig. la). A schematic drawing
We have proposed an integrated approach to treatment showing the main components of a hypothetical tomother-
planning, delivery, and verification of conformal radio- apy unit is given in Fig. lb. The on-board CT scanner is
therapy, called tomotherupy (23). Tomotherapy is analo- used to simultaneously image the anatomy during treat-
gous to spiral computed tomography (CT) scanning where ment. Image detectors are used to detect transit profiles
a megavoltage x-ray fan beam is continuously rotated on of the megavoltage beam and the data used along with the
a ring gantry while the patient is simultaneously translated CT imanes of anatomv, to reconstruct dose distributions as
 Tomotherapy l T. W. HOLMES et al. I225

pub-214, University of California, Berkeley. CA; October 26. Mantel, J.; Perry, H. Automatic variation of field size and
1977. dose rate in rotation therapy. Int. J. Radiat. Oncol. Biol.
16. Jennings, W. A. The tracking cobalt project: From moving- Phys. 2:307-317; 1977.
beam therapy to three-dimensional programmed irradiation. 27. McDonald, S. C.; Rubin, P. Optimization of external beam
In: Orton, C. G., ed., Progress in medical radiation physics, radiation therapy. Int. J. Radiat. Oncol. Biol. Phys. 2:307-
chap. 1-I. New York: Plenum Press; 1985:144. 317; 1977.
17. Kak, A. C.; Slaney, M. Principles of computerized tomo- 28. Reckwerdt, P. Superposition/convolution speed improve-
graphic imaging. New York: IEEE Press; 1988. ments using run-length raytracing. (Abstr.) Med. Phys.
18. Kallman, P., Lind, B. K.; Brahme, A. An algorithm for 19:784; 1992.
maximizing the probability of complication-free tumor con- 29. Redpath, A. T.; Vickery, B. L.; Wright, D. H. A new tech-
trol in radiation therapy. Phys. Med. Biol. 37:871-890; nique for radiotherapy planning using quadratic program-
1992. ming. Phys. Med. Biol. 21:781-791; 1976.
19. Lane, R. G.; Loyd, M. D.; Chow, C. H.; Ekwelundu, E.; 30. Rosen, I. I.; Lane, R. G.; Morrill, S. M.; Belli, J. A. Treat-
Rosen, I. I. Custom beam profiles in computer-controlled ment plan optimization using linear programming. Med.
radiation therapy. Int. J. Radiat. Oncol. Biol. Phys. 22: 167- Phys. 18:141-152; 1991.
174; 1991. 31. Rosenberger, F. U.; Mathews, J. W.; Johns, G. C.; Drzy-
mala, R. E.; Purdy, J. A. Use of transputers for real time
20. Langer M.; Leong, J. Optimization of beam weights under
dose calculation and presentation for three-dimensional ra-
dose-volume restrictions. Int. J. Radiat. Oncol. Biol. Phys.,
13:1255-1260; 1987. diation treatment planning. Int. J. Radiat. Oncol. Biol. Phys.
25:709-719; 1993.
21. Lind, B. Properties of an algorithm for solving the inverse
32. Sonderman, D.; Abrahamson, P. G. Radiotherapy design
problem in radiation therapy. Inverse Problems 6:415 -426; using mathematical programming models. Op. Res.
1990. 33:705-725; 1985.
22. Luenberger, D. G. Linear and nonlinear programming. 2nd 33. Starkshall, G. A constrained least-squares optimization
ed. Reading, MA: Addison-Wesley; 1989. method for external beam radiation therapy treatment plan-
23. Mackie, T. R.; Holmes, T. W.; Swerdloff, S.; Reckwerdt, ning. Med. Phys. 11:659-665; 1984.
P.; Deasy, J. 0.; Yang, J.; Paliwal, B.; Kinsella, T. Tomo- 34. Takahashi, S. Conformation radiotherapy: Rotation tech-
therapy: A new concept for the delivery of conformal radio- niques as applied to radiography and radiotherapy of can-
therapy using dynamic compensation. Med. Phys. 20: 1709- cer. Acta Radiol. Suppl. 242:1-142; 1965.
1719; 1993. 35. Webb, S. Optimization of conformal radiotherapy dose dis-
24. Mackie, T. R.; Reckwerdt, P. J.; Gehring, M. A.; Holmes, tributions by simulated annealing. Phys. Med. Biol.
T. W.; Kubsad, S. S.; Thomadsen, B. R.; Sanders, C. A.; 34:1349-1370; 1989.
Paliwal, B. R.; Kinsella, T. J. Clinical implementation of 36. Webb, S. Optimization by simulated annealing of three-
the convolution/superposition method. In: Proceedings of dimensional, conformal treatment planning for radiation
the 10th International Conference on Computers in Radio- fields defined by a multileaf collimator: II. Inclusion of
therapy, Lucknow, India. Lucknow, India: Alpana Arts; two-dimensional modulation of the x-ray intensity. Phys.
1990:322-325. Med. Biol. 37: 1689- 1704; 1992.
25. Mageras, G. S.; Podmaniczky, K. C.; Mohan. R. A model 37. Zacarias, A. S.; Lane, R. G.; Rosen, I. I. Assessment of a
for computer-controlled delivery of 3-D conformal treat- linear accelerator for segmented conformal radiation ther-
ments. Med. Phys. 19:945-953; 1992. apy. Med. Phys. 20:193-198; 1993.
1218 I. J. Radiation Oncology 0 Biology l Physics
WA(f)
- lIA(9 ---
R(f) ....-
---I
-0.5 -0.4 0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
SpetlalFrequenq
Fig. 2. Frequency spectra of the quantity 1f 1/A(f) for: (A) boCo; (B) 6 MV; (C) 10 MV; and (D) 24 MV x-rays,
Also shown are the spectra for the frequency ramp 1f 1 and the inverse of the filter A(f).
where If 1 is a frequency ramp, A(f) is the frequency
space representation of the 1D dose kernel, and
(,/-)-I is the Butterworth lowpass filter where
$ is the filler cutoff frequency and the exponent 2M con-
trols the filter shape (8). The quantity 1f (/A(f) is plotted
in Fig. 2 for @‘Cogamma rays and 6 MV, 10 MV, and
24 MV x-rays. The frequency spectra of (f I and A(f)-’
are included for comparison. Note that to first order the
shape of the fast inversion filter resembles a frequency
ramp with slope between 2-4 for the range of energies
used clinically. The factor (p/p)-‘exp( + PX) provides ap-
proximate energy conservation. To simplify matters it is
assumed that the patient is water equivalent; therefore
p/p and p are defined for water and f is the average depth
to the center of the tumor for all beams intersecting the
tumor. Typically, the scale factor is in the range 20-40
for pwlp= 0.05 cm*/g and an average attenuation factor
of 1.o-0.5.
Volume 32. Number 4, 1995
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
SpatialFrequency
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
SpatialFrequency
The Butterworth filter is used in the optimization
algorithm to control both the maximum spatial resolu-
tion of the beam profiles and indirectly the maximum
spatial resolution of the optimized dose distribution
during a given iteration (Fig. 3). The cutoff frequency
f0 is adaptively increased from a small starting value
of 0.0625 up to a maximum value less than or equal to
the Nyquist frequency of 0.5. In general, it is desirable
to limit the amount of oscillations in beam profiles
because smooth profiles are easier to realize clinically.
The philosophy adopted here is to optimize the gross
structure of the dose distribution during early iterations
and the penumbra at later iterations by slowly increas-
ing the frequency cutoff when no progress is being
made as measured by the relative change in the objec-
tive function:
(Eq. 9)
 Tomotherapy 0 T. W. HOLMES et al. 1219

A.FastlnwtionFilter B.BulterworthFilter
-I-
1.8 &
1.8 I
1.4 I
!
1.2
I
1
0.8 !
I
0.8
I
0.4
0.2
0
-0.5 Xl.4 0.3 4.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.4 -0.2 0 0.2 0.4
SpatialFrequency Saptbl
Frequency

C.BandlirnitedFastlnrersionFilter D.AdaptiwBandlimiting
! I""'""""'1

h 0.45
e
0.8 ; 0.4
i 0.35
0.8 k9 0.3
< 0.25
t 0.2
5 0.15
z 0.1
0.05
0
-0.4 -0.2 0 0.2 0.4 0 1 2 3 4 5 8 7 8 Q 10 11 12 13 14 15
Saptial
Frequency No.hllons

Fig. 3. Illustration of adaptive bandlimiting of the fast inversion filter. (A) The fast inversion filters. (B) The
Butterworth filters corresponding to M = 4 and f,, given by Eq. 10. (C) The product of the 6 MV fast inversion
filter and the Buttetworth filters for each value of f,,. (D) Plots of f0 and f(d) as a function of iteration number
for a typical treatment planing problem.

The cutoff frequency is updated when the relative
change in the objective function falls below an empiri-
cally determined threshold of 20% unless, of course, the
maximum cutoff frequency has been reached, or if the
solution is diverging as measured by A < 0. The initial
value of 0.0625 corresponds to averaging over eight pen-
cil beams. Subsequent cutoff values are obtained by re-
ducing the number of pencil beams averaged over (n) by
one:
n= (8,7,6,5,4,3,2,1).
f0=$
Small errors in the dose projections will be significantly
amplified by the inversion process because the filtering
operation is a method of deconvolution. These errors are
largely due to inaccurate representation of continuous
edges of regions-of-interest by discrete pixels and show
up as high frequency oscillations in the beam profile (Fig.
4a). One solution is to use antialiased edge representation
at the existing grid resolution. The improvement possible
is shown in Fig. 4b. Better results are also possible using
a higher calculation grid resolution. The major drawback
of this solution is that the computation time is signifi-
(Eq.lO) cantly lengthened if applied to all steps in the algorithm.
We compromise by computing the dose distribution at a
0.5 cm resolution, but perform the inversion at a higher
resolution of 0.125 cm by resampling the dose distribution
 I. J. Radiation Oncology 0 Biology l Physics Volume 32, Number 4, 1995

ues are nonphysical. The profiles can be optionally modi-

!
INITIAL- fied to account for blocking of normal tissue and critical
/ . A. I
! ESTIMATE+-- structures. The beam profiles are subsequently used to
/ compute dose using a 3D superposition calculation (24,
! 28). The current dose estimate is then compared against
!
upper and lower dose constraints defined for the target
region and any regions-of-interest specified in the normal
tissue. The iteration is terminated if a feasible solution is
.-.
reached, wherein the current dose estimate falls between
F the upper and lower dose limits for all regions-of-interest.
If a feasible solution is not achieved, then the algorithm
proceeds to compute the residual dosein the target region-
of-interest and any unblocked normal tissue regions-of-
interest in preparation for updating the beam profiles. The
residual dose is used to compute the objective function,
-20 -15 -10 -5 0 5 10 15 20 which is used to terminate the planning process if it in-
Distance(cm)
creases in successive iterations. If it decreasesthen the
residual dose in the regions of interest is transformed into
residual beam profiles by the inversion method on a slice-
by-slice basis. New beam profile estimates are computed
/ INITIAL-
25 8. ESTIMATE+ by subtracting the residual beam profiles from the current
profile estimatesand a new iteration is begun. Upon termi-
nation the dose calculation can be optionally verified us-
ing an alternative dose calculation model. This is useful

INlllALlZE ENERQY FLUENCE

-20 -15 -10 -5 0 5 10 15 20

Oklance(cm)
Fig. 4. Inversion results for different edge representations of
the external contour. The test consisted of simulating the irradia-
tion of a 30 cm diameter circular uniform phantom by 72 identi-
cal uniform 24 MV beam profiles equally spaced about 360
degrees. (A) Binary edge. (B) Antialiased edge. The calculation BLOCKING CONSTRAINT
grid was 64 x 64 pixels at 0.5 cm resolution.
I

to the higher resolution. The dose residual is computed FORWARD DOSE CALCULAYION c
and extracted from the target and normal tissue using high
resolution region-of-interest masks. The high resolution
projections of the dose residual are obtained and then
resampledto the lower resolution of the multileaf collima-
tor and filtered by I. This modification to the algorithm
is discussedfurther in Holmes and Mackie (11).
The IFB algorithm is implemented as the iterative se-
quence of steps listed in the flowchart shown in Fig. 5.
At the beginning of the iteration the beam profiles are INVERSE DOSE CALCULATION
initialized by inverting the dose prescription. Typically
this is represented as nonzero dose in the target and zero
dose to normal tissues. During each iteration and prior to
computing dose, the beam profiles are verified to seethat
they consist of nonnegative values, becausenegative val- Fig. 5. Flowchart of the iterative optimization algorithm.
 Tomotherapy 0 T. W. HOLMES et al. 1221

if approximations are incorporated in the dose calculation Blocking Strategies

used in the iteration to speed up the inverse planning No Blocks

process.

RESULTS
The definition of beam blocking is one of the initial
steps in the treatment planning process and can have a
significant impact on the results obtained using the IFB
algorithm. Figure 6 illustrates the possible choices of
blocking for a C-shaped target and nearby organ at risk.
I 7
I rl
Note that if the organ at risk is not blocked, it is included Critical Structure Blocked
with the target in the determination of the dose residual; Fig. 6. Diagramshowingthe different blocking strategiesfor a
otherwise if the organ at risk is blocked, the residual is C-shapedtarget with nearby squaresensitivestructure.
computed only over the target region. Dose profiles are
compared in Fig. 7 corresponding to the different types
of blocking for two beam energies. The profiles corre-
spond to a horizontal line through the center of the target

Co!imabdbn,Crilk9SbW%mW0d@d
1.2 1.2 1

I
1

0.6

ij O4

0.4

0.2
/
Od be
0
-15 -10 -5 0 5 IO 15
rmlca (all)

Cdimati,CIkdStWumNotBkdied Wmated,NoBloding
1.2 , I 1.2 >
I I I '1

24MV- - 1 24W- .

0.6
5-t
9

8 O4

0.4

0.2

0
-15 -10 -5 5 10 15
WaA (an)

Fig. 7. Dose profiles for different blocking strategies for the phantom shown in Fig. 6. In all cases the target is
included in the determination of the residual. In panels A-C the normal tissue is blocked by collimating the beam
to the target boundaries. In these examples the critical structure is (A) ignored, (B) blocked, and (C) unblocked,
but includedin the determinationof the residual.(D) Whenboth structureswereunblockedthe algorithmperformed
poorly.
1222 I. J. Radiation Oncology l Biology 0 Physics
Fig.8. Prostate example. Isodose lines shown are 110, 100, 75, 50, and 25%. (A) Central slice. (B) Off-axis slices
1 cm and 3 cm superior and inferior to the central slice.
and critical structure. These results suggest that it is best
to at least collimate the beam to the outline of the target
region to reduce the effect of the beam energy on the
resulting dose distribution. Furthermore, the algorithm
reached an acceptable level of dose uniformity in the
target when the organ at risk was also blocked (Fig. 7b),
although with increased dose to the normal tissue as a
consequence.
The IFB inverse treatment planning algorithm was used
to compute the conformal isodosedistributions for a pros-
tate and a breast shown in Figs. 8 and 9, respectively.
These examples represent two different classesof prob-
lems for conformal radiotherapy. The small size of the
Volume 32. Number 4, 1995
prostate, its location deep inside the body, and the rela-
tively moderate size of adjacent organs at risk (bladder
and rectum) allow a variety of treatment techniques to be
used, from multiple static fields to arcs. In contrast, the
superficial location of the breast and the large size of
adjacent organs at risk (lung, heart, and the contralateral
breast) typically constrain one to a pair of tangential paral-
lel-opposed fields. For the prostate, the goal of conformal
radiotherapy is generally the improvement of local control
via dose escalation; whereas for the breast, especially
early stage disease,it is reduced normal tissue complica-
tions, becauseonly marginal improvement in local control
is felt to be possible using dose escalation.
 Tomotherapy 0 T. W. HOLMES et al 1223

Fig. 9. Breast example. Isodose lines shown are 110, 100, 75, 50, and 25%. (A) Central slice. (B) Off-axis slices
2 cm and 4 cm superior and inferior to the central slice.

To simplify the problem, and test feasibility of the
algorithm, the critical structures were ignored and the
target volume was the only region-of-interest used to
compute the residual dose. The planning objective in both
cases was to minimize the least-squared residual dose in
the target volume to achieve a dose uniformity of 2 5%
in the target volume. The prostate dose distribution was
achieved in 9 iterations using 64 beam directions equally
spaced about 360”. The breast dose distribution was
achieved in 5 iterations using 32 beam directions equally
spaced about 360”. Both examples took over 20 h to com-
pute on a workstation.’ A larger number of iterations was
obtained in the case of the prostate because it is a smaller
structure and requires fewer pencil beams to be optimized
than the breast volume. Due to the long calculation times,
the iterations were terminated prior to convergence to a
minimum of the objective function, and consequently
only a 10% dose uniformity was achieved in the target
for both examples. The dose distributions for the central
axis and four off-axis planes are provided for each exam-
ple. Although + 5% dose uniformity was not achieved,
the dose distributions are nevertheless highly conformal.

’ DEC Station 5000/200, Digital Equipment Corp, Maynard,

MA.
1224 I. J. Radiation Oncology 0 Biology 0 Physics
Furthermore, the dosevolume histograms for these exam-
ples indicate that a sensitive structure receives no more
than about 30-50% of the target dose to one-third of its
volume. This might be reduced further by including the
critical structures in the optimization. Further work is
needed to compare these results against other treatment
delivery methods using coplanar and noncoplanar beam
directions, with and without intensity modulation of the
beam profiles.
DISCUSSION
Long computation times are the major problem facing
optimization of large numbers of beams.The overall time
of the IFB algorithm is dominated by the 3D superposition
dose calculation, which is repeated during each iteration
using the new beam profile estimates. This could be re-
duced somewhat by incorporating useful approximations
that do not compromise accuracy of the dose calculation.
For example, if heterogeneities are not important then
approximate dose calculations that are similar to filtered
backprojection image reconstruction (3, 10) could be
used. Another solution is to use parallel processing com-
puter technology to accelerate the dose computation (31)
by assigning the computation of a subset of one or more
beams to a dedicated central processing unit.
A secondary, but important, issue is the number of
iterations needed to achieve a solution. An advantage of
gradient search algorithms, such as the algorithm de-
scribed here and those described in Bortfeld et al. (2) and
Brahme (4) over simulated annealing, is the small number
of iterations needed to obtain an acceptable solution. For
example, the IFB algorithm typically requires lo-20 iter-
REFERENCES
1. Bahr, G. K.; Kereiakes, J. G.; Horwitz, H.; Finney, R.;
Galvin, J.; Goode, K. The method of linear programming
applied to radiation treatment planning. Radiology 91:686-
693; 1968.
2. Bortfeld, T.; Btirkelbach, J.; Boesecke, R.; Schlegel, W.
Methods of image reconstruction from projections applied
to conformation therapy. Phys. Med. Biol. 35:1423-1434;
1990.
3. Bortfeld, T.; Schlegel, W.; Rein, B. Decomposition of pen-
cil beam kernels for fast dose calculations in three-dimen-
sional treatment planning. Med. Phys. 20:31 l-318; 1993.
4. Brahme, A. Optimization of stationary and moving beam
radiation therapy techniques. Radiother. Oncol. 12: 129-
140; 1988.
5. Convery, D. J.; Rosenbloom, M. E. The generation of inten-
sity-modulated fields for conformal radiotherapy by dy-
namic collimation. Phys. Med. Biol. 37:1359- 1374; 1992.
6. Davy, T. J. Physical aspects of conformation therapy using
computer-controlled tracking units. In: Orton, C. G., ed.
Progress in medical radiation physics, chap. l-11. New
York: Plenum Press: 1985:45-94.
7. Emami, B.; Lyman, J.; Brown, A.; Coia, L.; Goitein, M.;
Munzenrider, J. E.; Shank, B.; Solin, L. J.; Wesson, M.
Tolerance of normal tissue to therapeutic irradiation. Int. J.
Radiat. Oncol. Biol. Phys. 21:109-122; 1991.
Volume 32. Number 4, 1995
ations to achieve a solution for 2D problems, whereas
Webb (35) has reported using several million iterations
using simulated annealing to solve similar problems. The
major limitation of gradient algorithms is that they are
lessgeneral than simulated annealing-they are only ap-
propriate for problems where the objective function is
unimodal; that is, it has only a single minimum value
that is the global minimum. We have found that the IFB
algorithm will likely converge to a feasible solution when
the number of pencil beams (e.g., the “degrees-of-free-
dom”) are comparable to the number of dose constraints
in the target, in analogy with the general principle used
in image reconstruction that the total number of projection
ray samples should be roughly equal to the number of
pixels in the circle of reconstruction (17). This is consis-
tent with requiring that the matrix D be square and its
determinant nonzero so that its inverse D-’ will exist (22).
Consequently, if this criteria is met for a given tomother-
apy inverse treatment planning case, then simulated an-
nealing may not be required, although further research
comparing the two algorithms is needed.
CONCLUSION
We have described an inverse treatment planning algo-
rithm for tomotherapy based on filtered backprojection
image reconstruction. The algorithm is demonstrated us-
ing two clinically relevant treatment planning examples
of the prostate and breast. Future work is necessary to
reduce the computation time, to extend the method to
biological objectives, and to perform experimental verifi-
cation, including comparison with other inverse treatment
planning algorithms.
8. Hamming, R. W. Digital filters. 3rd ed. Englewood Cliffs,
NJ: Prentice Hall; 1989.
9. Holmes, T. W.; Mackie, T. R. A comparison of three in-
verse treatment planning algorithms. Phys. Med. Biol.
39:91-106; 1994.
10. Holmes, T. W.; Mackie, T. R. A filtered backprojection
dose calculation method useful for inverse treatment plan-
ning. Med. Phys. 21:303-313; 1994.
11. Holmes, T. W.; Ma&e, T. R. Suppression of artifacts in
optimized beam profiles determined by the iterative filtered
backprojection inverse treatment planning algorithm. (sub-
mitted to Phys. Med. Biol.); 1993.
12. Holmes, T. W.; Ma&e, T. R.; Simpkin, D.; Reckwerdt, P.
A unified approach to the optimization of brachytherapy
and external beam dosimetry. Int. J. Radiat. Oncol. Biol.
Phys. 20:859-873; 1991.
13. Hope, C. S.; Laurie, J.; Orr, J. S.; Halnan, K. E. Computer
optimization of 4 MeV treatment planning. Phys. Med.
Biol. 10:365-373; 1965.
14. Hope, C. S.; Laurie, J.; Orr, J. S.; Halnan, K. E. Optimiza-
tion of x-ray treatment planning by computer judgment.
Phys. Med. Biol. 12:531-542; 1967.
15. Huesman, R. H.; Gullberg, G. T.; Greenberg, W. L.; Bud-
inger, T. F. Users manual: Donner algorithms for recon-
struction tomography. Lawrence Berkeley Laboratory, LBL
 Tomotherapy l T. W. HOLMES et al. I225

pub-214, University of California, Berkeley. CA; October 26. Mantel, J.; Perry, H. Automatic variation of field size and
1977. dose rate in rotation therapy. Int. J. Radiat. Oncol. Biol.
16. Jennings, W. A. The tracking cobalt project: From moving- Phys. 2:307-317; 1977.
beam therapy to three-dimensional programmed irradiation. 27. McDonald, S. C.; Rubin, P. Optimization of external beam
In: Orton, C. G., ed., Progress in medical radiation physics, radiation therapy. Int. J. Radiat. Oncol. Biol. Phys. 2:307-
chap. 1-I. New York: Plenum Press; 1985:144. 317; 1977.
17. Kak, A. C.; Slaney, M. Principles of computerized tomo- 28. Reckwerdt, P. Superposition/convolution speed improve-
graphic imaging. New York: IEEE Press; 1988. ments using run-length raytracing. (Abstr.) Med. Phys.
18. Kallman, P., Lind, B. K.; Brahme, A. An algorithm for 19:784; 1992.
maximizing the probability of complication-free tumor con- 29. Redpath, A. T.; Vickery, B. L.; Wright, D. H. A new tech-
trol in radiation therapy. Phys. Med. Biol. 37:871-890; nique for radiotherapy planning using quadratic program-
1992. ming. Phys. Med. Biol. 21:781-791; 1976.
19. Lane, R. G.; Loyd, M. D.; Chow, C. H.; Ekwelundu, E.; 30. Rosen, I. I.; Lane, R. G.; Morrill, S. M.; Belli, J. A. Treat-
Rosen, I. I. Custom beam profiles in computer-controlled ment plan optimization using linear programming. Med.
radiation therapy. Int. J. Radiat. Oncol. Biol. Phys. 22: 167- Phys. 18:141-152; 1991.
174; 1991. 31. Rosenberger, F. U.; Mathews, J. W.; Johns, G. C.; Drzy-
mala, R. E.; Purdy, J. A. Use of transputers for real time
20. Langer M.; Leong, J. Optimization of beam weights under
dose calculation and presentation for three-dimensional ra-
dose-volume restrictions. Int. J. Radiat. Oncol. Biol. Phys.,
13:1255-1260; 1987. diation treatment planning. Int. J. Radiat. Oncol. Biol. Phys.
25:709-719; 1993.
21. Lind, B. Properties of an algorithm for solving the inverse
32. Sonderman, D.; Abrahamson, P. G. Radiotherapy design
problem in radiation therapy. Inverse Problems 6:415 -426; using mathematical programming models. Op. Res.
1990. 33:705-725; 1985.
22. Luenberger, D. G. Linear and nonlinear programming. 2nd 33. Starkshall, G. A constrained least-squares optimization
ed. Reading, MA: Addison-Wesley; 1989. method for external beam radiation therapy treatment plan-
23. Mackie, T. R.; Holmes, T. W.; Swerdloff, S.; Reckwerdt, ning. Med. Phys. 11:659-665; 1984.
P.; Deasy, J. 0.; Yang, J.; Paliwal, B.; Kinsella, T. Tomo- 34. Takahashi, S. Conformation radiotherapy: Rotation tech-
therapy: A new concept for the delivery of conformal rdio- niques as applied to radiography and radiotherapy of can-
therapy using dynamic compensation. Med. Phys. 20: 1709- cer. Acta Radiol. Suppl. 242:1-142; 1965.
1719; 1993. 35. Webb, S. Optimization of conformal radiotherapy dose dis-
24. Mackie, T. R.; Reckwerdt, P. J.; Gehring, M. A.; Holmes, tributions by simulated annealing. Phys. Med. Biol.
T. W.; Kubsad, S. S.; Thomadsen, B. R.; Sanders, C. A.; 34:1349-1370; 1989.
Paliwal, B. R.; Kinsella, T. J. Clinical implementation of 36. Webb, S. Optimization by simulated annealing of three-
the convolution/superposition method. In: Proceedings of dimensional, conformal treatment planning for radiation
the 10th International Conference on Computers in Radio- fields defined by a multileaf collimator: II. Inclusion of
therapy, Lucknow, India. Lucknow, India: Alpana Arts; two-dimensional modulation of the x-ray intensity. Phys.
1990:322-325. Med. Biol. 37: 1689- 1704; 1992.
25. Mageras, G. S.; Podmaniczky, K. C.; Mohan. R. A model 37. Zacarias, A. S.; Lane, R. G.; Rosen, I. I. Assessment of a
for computer-controlled delivery of 3-D conformal treat- linear accelerator for segmented conformal radiation ther-
ments. Med. Phys. 19:945-953; 1992. apy. Med. Phys. 20:193-198; 1993.