CHAPTER 11

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

The Inverse Problem in Radiation
Therapy Treatment Planning

In this chapter we introduce the problem of therapy treatment planning

in radiation oncology. There are some good reasons to construct a fully
discretized model of this problem, to which we can apply the various row­
action algorithms and block-iterative methods discussed in earlier chapters.
We will explain our reasons and also present some computational results.
Radiation therapy maintains a prominent place in the treatment of
neoplastic lesions (tumors). Over the past ninety years, medical experience
and laboratory studies have established the principles whereby radiation
can be used as an effective modality for oncological treatment.
Radiation dose is a measure of the actual energy absorbed in a biologi­
cal (or other) absorber from a given x-ray beam. It is measured in units of
energy per mass. The dose unit is called gray and is abbreviated by Gy, and
1 Gy = 1 Joule/Kg (where Joule is the standard unit of energy). It is very
difficult to define quantitatively the actual biological damage produced by
the dose but since the dose is closely related to the amount of ionization
produced we can roughly estimate the degree of biological damage. Al­
though all viable tissues are affected to some degree by ionizing radiation,
a dose below a certain empirical limit, or tolerance value, does not cause
irreparable damage to normal tissue. However, a dose of radiation that ex­
ceeds the tolerance can be expected to cause permanent tissue damage, and
the tolerance value varies with tissue type. For any given organ, the risk
of injury depends on the total and incremental dose, and on the fraction
of the organ irradiated. Similarly, empirical values have been determined
for doses lethal to different tumors. Generally the range of dose that will
destroy a tumor without an unacceptable risk of injury to normal tissue is
very narrow.
The radiation dose delivered to any point within a region is determined
by two components, the primary and the scatter radiation. The primary
radiation causes that portion of the dose which results from interactions
lying directly along the path of the undeviated radiation beam. All other
contributions to the dose at a specific point come from radiation deflected,
i.e., scattered, to that point from within the radiation field.

Parallel Optimization, Yair Censor, Oxford University Press (1997), © 1997 by

Oxford University Press, Inc., DOI: 10.1093/9780195100624.003.0011
 307

After the lesion is diagnosed, a decision must be made on treatment

modality, e.g., surgery, chemotherapy, radiation therapy, or some combi­
nation of these. If a tumor is of the size and type likely to respond to
radiation therapy, the radiation therapist (physician) delineates the vol­

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

ume to be treated and prescribes the dose he believes most effective to
the lesion. Together with the radiation therapist, the radiation dosimetrist
(technician) must ensure that this physician-specified dose is correctly de­
livered to the site of the lesion, without endangering the function of critical
organs or normal tissue.
Radiation can be delivered in two ways: by brachytherapy, i.e., the di­
rect implantation of radioactive sources into the lesion, and by teletherapy,
i.e., the use of beams of penetrating radiation directed at the lesion from
an external source. When radioactive sources are implanted into tumors
the dose is determined by the strength of the sources, the geometric ar­
rangement of these sources within the tumor region, and the duration of
insertion. Brachytherapy calculations are relatively straightforward and
show a predictable dose gradient from the high-dose regions around the
sources to the small-dose regions at increasing distances from the sources.
When radiation is delivered by beams from an external source (e.g.,
cobalt-60 or a linear accelerator), the situation becomes quite complex.
Beams from the external radiation source are shaped and directed at a
specific target region, namely, the volume of the neoplastic lesion and the
immediately surrounding volume that is likely to contain microscopic dis­
eases. The beams must be aimed and contoured so that when their dose
contributions are added, the total dose is lethal to the neoplastic cells in
the target volume but not to the healthy tissue. The clinical simplicity,
reproducibility, and ability to plan and deliver a uniform dose throughout
the tumor volume makes teletherapy the more widely used modality of ra­
diation therapy. In this chapter we are concerned only with teletherapy
treatment planning.
The teletherapy radiation source is mounted in a gantry (or arm) that
rotates in a circular arc in some plane at a specified distance from the axis
of rotation. A patient is positioned on the therapy table so that the center
of rotation is placed in the target volume. The target can be irradiated
by positioning the gantry at a number of fixed angles around the arc of
rotation or by rotating the gantry continuously along segments of the arc.
For teletherapy planning, the contours of the patient boundary, the tar­
get, and the critical organs are delineated in one or more contiguous trans­
verse sections of the patient. Only where the anatomy changes rapidly from
section to section (as between the head and neck or the neck and shoulders)
are several contiguous sections required for accurate calculation of delivered
dose, called dosimetry. We restrict ourselves to two-dimensional treatment
planning, meaning that one patient section is coincident with the plane of
gantry motion and all radiation rays are confined to that plane. Extensions
308 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

to three-dimensional radiation therapy treatment are possible.

The radiation therapist chooses a treatment dose with a high probability
of eradicating the tumor without permanent normal tissue injury. The
dosimetrist then arranges the treatment fields to cover the target volume

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

while trying to minimize the amount of normal tissue within the irradiated
field. Radiation beam modifiers, called wedges or compensators, are added
to increase the dose homogeneity across the target volume. The dosimetrist
tries to achieve: a uniform dose distribution across the tumor; a minimum
dose to the critical organs and other nondiseased tissues; and a minimum of
complexity in the technical setup. Most often the experience and judgment
of the dosimetrist quickly lead to an efficacious setup.
There are a number of cases, however, that present a challenge, for
example when tumors are juxtaposed against critical organs such as the
spinal cord, optic nerve, or kidney. The dosimetrist must then consider
many possibilities, including fixed fields, arc rotation, beam modifiers, or a
combination of these. Achieving a satisfactory treatment plan can therefore
become time consuming and tedious. How can we determine if a feasible
treatment plan that satisfies all of the radiation therapist’s specifications
even exists? Should we accept a plan after a preset number of tries if there
is still the possibility of a significantly better plan? How do we know when
we are close to an optimal plan in these complex cases?
In this chapter we describe a mathematical modeling approach to radi­
ation therapy treatment planning (RTTP) that aims toward automating
the selection of an acceptable setup once the physician’s specifications are
given. The mathematical formulation of the RTTP problem leads to a
mathematical problem of operator inversion, which can be approached by
either continuous analytical methods or by fully discretizing the model at
the outset and resorting to linear algebraic or optimization techniques in
a finite-dimensional space. This dichotomy between solution approaches is
reminiscent of the situation in image reconstruction from projections (see
Chapter 10). The reasons that we favor the full discretization approach
will become clear as the discussion unfolds.

11.1 Problem Definition and the Continuous Model

Let the plane of the gantry motion be parametrized by polar coordinates r
and 0 where r is the distance of a point from the center of rotation of the
gantry and 0 is an angle measured clockwise from the positive y-axis (see
Figure 11.1).
Let Z)(r, 0} be a real-valued nonnegative function of the polar coordi­
nates whose value describes the dose absorbed at a point in the patient
section coincident with the plane of gantry motion. D(r, 0) is referred to
as the dose distribution. We define a ray to be a directed line along which
radiation travels away from the source, where source is a short term for
teletherapy source position. A ray is specified by the gantry angle u of its
Sect. 11.1 Problem Definition and the Continuous Model 309

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.1 Patient space in the plane of the gantry circle,
(r, 0) denotes a point in patient space. Rays are defined by the
gantry angle u of the source and the distance w in the “beam
window plane” (shown perpendicular to the central ray at the
isocenter). Here ray (u, w) is shown to pass through the point
at (r, 0).

source on the gantry circle and by the length w determined by w = dtanA,

where d is the radius of the gantry circle and A is the angle of the ray mea­
sured clockwise from the central ray. Each ray is outgoing from a source.
Thus rays (u, 0) and (u 4- 7T, 0) represent coincident lines but distinct rays
since they have opposite directions. The domain of u is 0 < u < 2tt, and
the domain of w is — W < w < W where 2W is the beam width.
The real-valued nonnegative function p(u, w) for 0 < u < 2tt and — W <
w < W represents the radiation intensity along the ray (zz, w) due to a point
source on the gantry circle. We refer to p(u, w) as the radiation intensity
distribution. The continuous version of the forward and inverse problems
of radiation therapy treatment planning can now be formulated.

11.1.1 The continuous forward problem

The continuous forward problem of RTTP is the following. Assume that the
cross section Q of the patient and its radiation absorption characteristics
are known. Given a radiation intensity distribution p(u, w) for 0 < u < 2tt
and — W < w < W, find the dose distribution D(r,0) for all (r, 0) e Q
from the formula

D(r, 0) = A[p(u, w)](r, 0), (11.1)

where A is the dose operator. This operator relates the dose distribution
to the radiation intensity distribution.
310 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

In other words, the forward problem amounts to the calculation of the

total dose absorbed at each point of a patient section when all parame­
ters of each radiation beam are specified and the description of the patient
section is known. The difficulties associated with the forward problem

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

(i.e., dose calculation, or dosimetry) stem from the fact that there exists
no closed-form analytic representation of the dose operator A that will
enable us to use equation (11.1) for the calculation of D(r,0). Although
the interaction between radiation and tissue is measured and understood
at the atomic level, the situation is so complex that, to solve the forward
problem in practice, a good state-of-the-art computer program, which rep­
resents a computational approximation of the operator A and which enables
reasonably good dose calculations, must be used.
Let us elaborate on what we intend by stating “there exists no closed-
form analytic representation of the dose operator A.” We actually mean the
following: If certain simplifying assumptions are made about the physics of
the model as well as the particulars of the desired dose distribution, then it
is sometimes possible to express the dose operator in a closed-form analytic
formula. This has been done by various authors and some references are
given in Section 11.7.
In current practice of RTTP, when dose calculations are performed to
verify the dose distribution that will result from a proposed treatment plan,
the goal is to obtain results that are as accurate as possible. To achieve
this, various empirical data, which are often condensed in look-up tables,
are incorporated into the forward calculation. Thus, the true forward cal­
culation, or true dose operator, is not represented by a closed-form ana­
lytic relation between the radiation intensity distribution p(u, w) and the
dose distribution D(r, 0), but by a software package that calculates D(r, 0)
from p(u,w). These arguments are well expressed in the words of Dr.
M. Goitein, from Massachusetts General Hospital in Boston. Describing
the 3D-CATP (Three-Dimensional Computer-Aided Treatment Planning)
program he says:
Dose calculations are the most critical and most CPU-intensive
calculations the system produces. The system must interrogate
every pixel in the path of the beam. In order to derive the dose
at any point, the system has to take into account the beam’s
energy profile, and the density of the tissue (gray scale value
of the pixels) preceding that point. The system can then look
up the dose in a depth-dose curve table. The system also takes
into account a scattering factor, through another table look-up.
None of the individual calculations are particularly complex,
but there are a tremendous number of calculations required.
Once the dose distribution calculations have been completed
for individual treatments, they can be combined into an overall
dose distribution.

Thus, what we really mean by saying that there is no closed-form ana­

Sect. 11.1 Problem Definition and the Continuous Model 311

lytic expression for A is that we choose to adhere to the software represen­

tation rather than compromise on the accuracy of the forward calculations
by allowing simplifying assumptions that might lead to a closed-form ana­
lytic mathematical formula.

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

11.1.2 The continuous inverse problem
The inverse problem of radiation therapy is the treatment planning prob­
lem:
Given a description of the patient section, the dose prescribed
for the target, and the maximum permissible doses to the tar­
get, critical organs, and other tissues, calculate the external
configuration and relative intensities of radiation sources (i.e.,
the radiation field) that will deliver the specified radiation
doses (or some acceptable approximation thereof).

Assuming that the cross section Q of the patient and its radiation absorp­
tion characteristics are known, and given a prescribed dose distribution
Z)(r, 0), the problem is to find a radiation intensity distribution p(zz,w)
such that equation (11.1) holds, or p(u,w) = A-1[Z)(r, #)] where A-1 is
the inverse operator of A. This is the inversion problem that we want
to solve, in a computationally tractable way, although no closed-form an­
alytic mathematical representation is available for the dose operation A.
The dose at (r, 0) is the sum of the dose contributions from the sources at
all the different gantry angles. Thus

s
D(r,e) = ^yiDdr,e), (11.2)
i=l

where, for each i = 1, 2,..., S, the value Di(r, 3) is the dose deposited at
point (r, 0) by a beam of unit intensity from the zth source, and yi is the
total intensity of the zth source. It is also possible to interpret Z?2(r, 0) as
the dose per unit time deposited at (r, 0) by the zth beam, and y2 as the
time the zth beam is kept on.
The dose can be further partitioned into two components that are due
to primary and scattered radiation. Thus

A(r, 0) = D^pr\r, 0) + D\sc\r, ff). (11.3)

The primary dose f)^pr> is physically due to the first interaction of beam
photons with the tissue medium, and is delivered along the fan of rays of
the zth beam profile. The value of the primary dose at a point in the patient
depends on the distance from the source to the patient surface, the depth of
the point, the electron density distribution, the angle between the central
ray (i.e., the ray through the center of the gantry circle) and the ray to
312 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

the point, and the value of the beam profile for the latter ray. The scatter
dose D\ is due to radiation scattered to a point after beam photons first
interact with other points in the medium. That is, the scatter dose is due
to secondary interactions. The scatter dose at any position depends on the

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

depth of this location in the tissue and the area of irradiated tissue. Both
primary and scatter dose also depend on the beam energy, the incident
photon spectrum, and the beam modifiers.
For the purposes of our study, it will be assumed here that the dose
Di(r, 0), and its components 0) and D-sc\r, 0) for each source z,
can be calculated accurately once the beam parameters and patient section
information are specified. That is, we assume that we can solve the forward
problem and calculate D(r, 0) accurately from (11.2) and (11.3). This
assumption is confirmed by innumerable direct measurements in water and
tissue-equivalent phantoms.
Whereas a dose distribution that solves the forward problem is always
obtained for a specified radiation intensity field, the inverse problem may
have no solution at all, since some prescribed dose distributions may be
unobtainable from any radiation field. Therefore, we do not aim at a solu­
tion of the continuous inversion problem, but look instead at a feasibility
formulation that relaxes the equality in (11.1) in the following manner.
Let D = D(r,0) and D = be two dose distribution functions
whose values represent upper and lower bounds, respectively, on the per­
mitted and required dose inside the patient’s crosssection. A radiation
therapist defines D and D for each given case and will accept as a solu­
tion to the RTTP problem any radiation intensity distribution p(zz, w) that
satisfies

D(ry0) < A[p(u, w)](r,0) < D(r, 0), for all (r, 0) e Q. (11.4)

In target regions (tumors) the lower bound D is usually the important

factor because the dose there should exceed some given value. In critical
organs and other healthy tissues Dfr, #) — 0, so that J9(r, 0) is the dose
that cannot be exceeded. Any solution p(u, w) that fulfills (11.4), for given
D and D, is a feasible solution to the RTTP problem.

11.2 Discretization of the Feasibility Problem

In the approach presented here, we adhere to the computerized calculation
of the dose operator A. Full discretization of the problem at the outset is
used to circumvent the difficulties associated with the inversion of A. We
also neglect the effect of scatter. The patient’s cross section Q is discretized
into a grid of points represented by {(r7, 0;) | j — 1,2,..., J}. Define Ajp
by
(U-5)
Sect. 11.2 Discretization of the Feasibility Problem 313

and call Aj a dose functional, for every j = 1,2,..., J. Acting on a radiation

intensity distribution p(u, w), the functional Aj provides Ajp, which is the
dose absorbed at the Jth grid point of the patient’s cross section Q due to
the radiation intensity field p. The dose distributions D and D are specified

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

at the grid points by giving, for all j = 1, 2,..., J,

=T)j, (11.6)

thus converting (11.4) into a finite system of interval inequalities

Dj <^jP<Dj, j = l,2,...,J. (11.7)

Denoting hereafter by D (D) the J-dimensional column vector whose /th

element is Dj (Dj), the inverse problem of RTTP is restated as follows:
Given vectors D and D of permitted and required doses at J
grid points in the patient’s cross section Q, find a radiation
intensity distribution p = p(u,w) such that (11.7) holds.

In continuing the discretization process of the problem it is assumed

that a set of I basis radiation intensity fields is fixed and that their non­
negative linear combinations can give adequate approximations to any ra­
diation intensity field we wish to specify. This is done by discretizing the
region 0 < u < 2tt, — W < w < W in the (u, w)-plane into a grid of points
given by {(uj, wf) | i = 1,2,...,!}. A radiation intensity distribution

I.1.’ if (U.S)
v 7 [0, otherwise, v 7

is a unit intensity ray and serves as a member of the set of basis intensity
fields, i ~ 1, 2,..., I, A desired radiation intensity distribution p that solves
(11.7) is approximated by

i
p(u,w) = X?CTZ(11, w), (11.9)
2=1

where X{ is the intensity of the zth ray, and it is required that xr > 0, for
all i = 1,2,... ,1. Once the grid points are fixed, any radiation intensity
distribution p that can be presented as a nonnegative linear combination
of the rays is uniquely determined by the coefficients Xi, 1 < i < I. The
vector x = (rrj, in the /-dimensional Euclidean space IR7 is referred to as
the radiation vector or basic solution.
Further, assume that the dose functionals Aj are linear and continuous.
This assumption cannot be mathematically verified due to the absence of
an analytic representation of A or Aj, but it is a reasonable assumption
314 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

based on the empirical knowledge of Aj. Using linearity and continuity of

all A/s, we can write Ajp ~ Ajp = ZZLi Xi&j&i- For j = 1,2,..., J, and
2 = 1,2,...,/, denote by
dij = Ajai (11.10)

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

the dose deposited at the jth point (rj, 0j) in the patient’s cross section Q
due to a unit intensity ray w). The fully discretized feasibility inverse
problem of RTTP then becomes the linear interval feasibility problem of
finding a vector x 6 IR7 such that

i
Dj < Xidij < Dj, j = 1,2,..., J,
2=1

Xi >0, z = 1,2,...,/. (11-11)

Let the set of pixels in the discretized patient cross section be denoted
by N = {1,2,..., J}. Organs within the patient section are then defined as
subsets of N. The subsets Bk C N, where k = 1,2,..., K denote K critical
organs to be spared from excessive radiation. Let the values bk denote the
corresponding upper bounds on the dose permitted in each critical organ.
The subsets Tq C N where q = 1,2,... ,Q denote Q target regions. Let
the values tq denote the corresponding prescribed lower bounds for the
absorbed dose in each. All the Bk and Tq are pairwise disjoint. The set
of pixels inside the patient section that are not in any Bk or Tq are called
the complement, denoted as the subset C C N, and c is the upper bound
for the total permitted dose there. It is assumed that the definition of all
subsets BkyTq, and C and the prescription of all bk,tq, and c are given
by the radiotherapist as input data for the discretized treatment planning
problem.
Problem (11.11) then becomes the following system of linear inequali­
ties, which we call the basic model'.

i
for all j € Bk, k = 1, 2,.. .,K, (11.12)
VI
A

for all j ETq, q = 1,2,.. ■ ,Q, (11.13)

VI

I
dijXi < c, for all j E C, (11.14)
2=1

%i > 0, for all 2 = 1,2,..., I. (11.15)

With bk, tq, and c given and the dij's calculated from (11.10), the math­
ematical question represented by the basic model (11.12)—(11.15) is to find
Sect. 11.2 Discretization of the Feasibility Problem 315

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.2 Discretization of patient section into pixels, and
discretization of the radiation field into sources and rays.

a nonnegative solution vector x = (xi) for a system of linear inequalities.

This is a linear feasibility problem. Any vector x whose components Xi
satisfy the basic model is acceptable in terms of the dose specifications set
forth. Consult Figure 11.2 in which the individual ray intensities Xi are
double-indexed to identify the number of the source from which they are
derived in the discretization process. With the discretization of source po­
sitions on the gantry arc as described above, we may consider a discretized
ray space, namely, a rectangular grid of I points (in the nonnegative or-
thant of the ray space), with S equally spaced locations in the interval
0 < u < 2tt along the iz-axis and M equally spaced locations placed in the
interval — W < w < W along the w-axis. We enclose the points in the ray
space in pixels and assign intensity x^ to the zth pixel, where i = i(s,rri)
according to the lexicographic ordering formula i — m -F {s ~ 1)M (see
Figure 11.3).
For this problem to be well defined the quantities dij must be precal­
culated, i.e., by apportioning the calculated dose per unit beam intensity
between the rays representing that beam in the discretization. A forward
problem solver in the form of a computer program is assumed available to
calculate each specific dose ds(j), which is the dose per unit time, absorbed
at pixel j of the patient section when the radiation field is due to a radi-
316 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.3 Discretized ray space wherein each column rep­
resents a single beam.

ation beam from source s. For our discretized radiation field, the gantry
angle u assumes only the values

US = s = 1,2, (11.16)
o
We must now attribute the specific dose at pixel j to a contribution from
each of the rays in our discretized ray space. We thus need to calculate the
numbers dij, i = 1,2,... ,1, j = 1,2,..., J, which represent the specific
dose absorbed in the Jth pixel due to radiation from the zth ray alone.
We have already decided to consider only the primary component of the
radiation dose; therefore we use a dose apportionment scheme which is now
defined and explained (see Figure 11.4).
Assume that a source of radiation is placed at gantry angle us, and that
the value ds(j) has been calculated for each pixel j by means of a forward
problem solver. This value is apportioned among a chosen set of rays, and
at present, only rays that pass through or immediately straddle the pixel
are used for apportionment. A possible apportionment scheme results from
the formula

dzj = (11,17)
Sect. 11.3 Computational Inversion of the Data 317

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.4 Specific dose absorbed in pixel j due to a source
is apportioned among rays labeled 1 to 4 (in this case) to obtain
the quantities dij.

where crij is the distance from the center of the Jth pixel to the zth ray.
The summation over p goes up to the number Ms(j) of rays from source s
at gantry angle us that pass through or immediately straddle the jth pixel.
If the pixel center lies on a ray whose index is t so that crtj = 0, then we
take dtj = ds(j) and set d^ = 0 for rays i t. In computational trials,
however, the apportionment scheme of (11.17) was found to be too sensitive
to small changes in the planar discretization and ray-sampling schema. A
more satisfactory apportioning scheme was found to be

(11.18)
Ms(jy

This simple scheme is the one used to obtain the numerical results demon­
strated in Section 11.5.

11.3 Computational Inversion of the Data

A procedure that solves the fully discretized inversion problem formulated
above can be constructed in the following manner:
Step 1. Read as input the parameters necessary for the forward problem
solver (dose calculation program). These are the beam width, beam
energy, calculation tables, etc. Read as input the patient boundary
318 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.5 Derivation of the basic solution of ray intensities
from physician requirements, patient and machine data, and
discretization.

contour and the contours of the internal structures (targets and crit­
ical organs). Read as input the physician’s prescription for the dose
distribution.
Step 2. Choose a discretization grid for the cross section, and determine
whether each pixel lies within a tumor region, within a critical or­
gan, or elsewhere. Assign to each pixel the appropriate values of the
physician-chosen bounds on radiation dose.
Step 3. Choose the number S of possible gantry (source) positions. Calcu­
late and save the specific dose distribution (ds(j))j=1 for each gantry
position of the beam. No blocks, wedges, or filters are used to modify
the beam profile. Since scatter radiation is omitted here, this forward
dose calculation is quite fast.
Step 4. Choose the number of rays per beam for discretization.
Step 5. For each pixel, the specific dose due to the source at gantry angle
us, s = 1,2,..., S is apportioned among those rays passing through or
straddling the pixel. This dose apportionment is applied to the dose
distributions calculated in Step 3 for all the sources. In this way the
numbers dij are calculated.
Step 6. Use an iterative algorithm to solve the resulting linear feasibility
problem. Initially all rays are assigned zero intensity. During the
iterative process, the algorithm gradually changes these intensities as
it tries to satisfy the prescribed physician constraints. Rays that do
not cross any tumor region are permanently assigned zero intensity.
Sect. 11.4 Consequences and Limitations 319

/ INPUT: Correction FINAL OUTPUT: 1

Beam
Reduction For Scattered Clinical
'Ray Intensities —>.
—> Scheme —► Radiation Treatment

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Effects Plan f
10 11 12 vsf

Figure 11.6 Derivation of the clinical treatment plan from a

basic solution of ray intensities.

Figures 11.5 and 11.6 describe the overall process. All boxes in these
diagrams are numbered for easy reference. Box 3 represents and assumes
the availability of a state-of-the-art computer program for forward calcu­
lation. Given discretization data for pixels, beams, and rays (box 1), and
data on the patient’s cross section and the treatment machine parameters
(box 2), it calculates the numbers ds(j), s = 1,2,..., S, j = 1,2,..., J.
Each ds(j) is the dose absorbed at the jth location of the cross section (i.e.,
at (rj,0j)) due to a unit intensity of radiation from source s on the gantry
circle. The dose apportionment scheme of box 4 distributes these values
among individual rays. The resulting dv’s are the coefficients of the sys­
tem (11.11), which also needs the physician requirements from box 5. An
iterative algorithm (box 7) then produces an approximate basic solution of
ray intensities x = (xi)-==1.
If a treatment machine that could deliver pencil-thin single rays of con­
trolled intensity existed, then the basic solution could have been imple­
mented clinically. Since this is not yet the case, we use this solution as
input to the process described in Figure 11.6, which first employs a beam
reduction scheme that extracts from the basic solution a clinically accept­
able treatment plan. After reducing the number of beams, one would need
to correct the plan to incorporate the effect of scattered radiation because
initially the calculations in the system of Figure 11.5 were for primary dose
only.

11.4 Consequences and Limitations

If a mathematically feasible solution to the radiation therapy treatment
planning problem exists, i.e., if the system (11.12)—(11.15) is feasible, then
the final set of ray intensities, obtained after stopping the algorithm, pro­
vides an approximation to a mathematically feasible radiation field. On
the other hand, if a feasible solution to the treatment planning problem
does not exist, i.e., if the system (11.12)—(11.15) is not consistent, then the
final set of ray intensities could provide a possible compromise assuming
that the nonnegativity constraints (11.15) are satisfied. With the final set
of ray intensities denoted by the vector x * = (x*
), the question is how to
implement this derived radiation field as a treatment plan.
320 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

A straightforward implementation of radiation field x* calls for a ma­

chine that will be able to emit controlled amounts of radiation in pencil-
thin rays and repeat this in a multitude of directions. The number of
directions, determined by the fineness of the grid for radiation field dis­

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

cretization, may range into the thousands. Available radiation treatment
machines, however, use finite-width beam sources. Although a beam pro­
file may be modified to some extent by using certain accessories such as
blocks, wedges, and compensators, we cannot expect present day machines
to implement accurately our finely discretized radiation fields.
While the technology may sufficiently improve to meet such a challenge,
we are studying the mathematical problem of how to extract from the ray
intensities solution vector x* a set of physical parameters such as: number
of sources (beams), gantry angles, specification of beam modifiers, and
intensities of sources (beams) that will constitute a clinically acceptable
and implementable treatment plan. Figure 11.3 shows that the discretized
radiation field x* can be considered a distribution of values x’f (or z* )
over the pixels (s, m) that represent the rays in this figure. To be clinically
implementable, a derived radiation field should have nonzero values in only
a few (say up to 10) columns in Figure 11.3 (each representing a single
source or gantry position), and only a linear or piecewise linear change of
ray intensities over a succession of rays within a column (to allow wedge
or block beam modifications). To meet these requirements, a method must
be devised to extract a blockwise varying radiation field from the finely
discretized radiation field.
At present we examine the distribution of ray intensities (or .r
*)
and choose those source positions us that provide the greatest ray intensity
contributions. We then repeat the feasibility calculation described above
for a limited number (not necessarily uniformly spaced) of source positions
around the gantry arc.

11.5 Experimental Results

The system (11.11) constitutes the mathematical model (box 6 in Fig­
ure 11.5) that we adopt for computing a solution of the inverse problem
in RTTP. The physician’s requirements are given by the lower and upper
bounds, Dj and Dj, j = 1,2,..., J. The coefficients dij are calculated
with the aid of a state-of-the-art forward problem solver, according to the
principles stated above. We apply to this problem the fully simultaneous
Cimmino algorithm. This is the block-AMS algorithm (Algorithm 5.6.2)
in which there is only one block (M = 1) that contains all inequality con­
straints. The Cimmino algorithm (for box 7 in Figure 11.5) was tested for
the two-dimensional case of a single transverse section of a patient with
beams confined to the plane of the section. No algorithmic changes are
required to extend the applicability to three-dimensional problems in ra­
diation therapy, although numbering of volume elements (voxels) and rays
Sect. 11.5 Experimental Results 321

(the input to the implementation), and the graphic presentation of the

calculated dose in a three-dimensional volume (the output of the imple­
mentation) would differ.
To use the algorithm the clinician must specify: the contours for the

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

patient, target, and critical organs; the dose prescription for the target,
together with a minimum and a maximum dose over it; the maximum dose
allowed for specified critical organs; and the maximum dose allowed for the
remaining noncritical tissue. In addition, the clinician may assign relative
weights (of importance) to each region within the cross section, such that
the algorithm generates a solution that least violates the constraints of the
heavily weighted regions and distributes proportionately more error to the
least weighted regions.
Figure 11.7 shows a computed tomography (CT) section of a patient’s
skull. In Figure 11.8 the picture is discretized into squares with dimensions
0.85x0.85 cm (not drawn to scale). The region marked by x is the target
containing the tumor. The dose to be delivered to the target is between 50
and 53 Gy. The region marked with 2 corresponds to the brain stem, which
is a critical organ. The dose there is not to exceed 40 Gy. Regions marked
with 3 and 4 are the left and right eyes respectively, and are not to receive
any radiation. Finally, the dose delivered to the remaining unlabelled tissue
should not exceed 45 Gy (see Figure 11.9).
The constraint (or regional) weights for the Cimmino algorithm are
chosen to be fixed, i.e., in Algorithm 5.6.2 the weights wu(i) = w(i) are
independent of the iteration index v and have the same value for all pixels
i that belong to a specific organ region in the patient’s cross section. Note
that in the fully discretized model of RTTP each inequality represents
a pixel in the cross section, whereas in image reconstruction discretized
models (Chapter 10) each constraint (equality or inequality) comes from
an individual ray traversing the object. The weights w(z) chosen here are:
ten for the target region and for the eyes, three for the brain stem, and
one for the unlabelled tissue. We start the discretization of the model
with 24 uniformly spaced beam positions, calculate the intensities of 50
rays for each beam (a total of 50 x 24 rays), and then plot the resulting
dose distribution. The 50 rays of each beam are set to span the target,
which means that rays which do not intersect the target are set, and kept
throughout, at zero intensity.
In clinical situations, complex multibeam treatment plans are generally
infeasible and the number of beam positions must be limited to fewer than
five or six. To this end, we describe our beam reduction scheme (box 11 in
Figure 11.6), which leads to our clinical treatment plan (box 13 in Figure
11.6). Correction for scattered radiation effects (box 12 in Figure 11.6)
has not been incorporated in the experiments described here. Ideally, one
would hope that a reduced number of beams, suitable for a practical clinical
treatment plan, could be read directly from the basic solution vector of ray
322 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Right eye Left eye

Brainstem Target
Figure 11.7 A CT section of the patient to be treated. Our
experimental results are demonstrated on this case. (Repro­
duced from Censor, Altschuler, and Powlis 1988).

intensities x* = (x).
* Since we are unaware of a method that will directly
reduce beams, we repeatedly eliminate certain beams from the original 24
beams, used for the initial discretization of the full radiation field, and
then repeat the computational process (described in Figure 11.5) until a
satisfactory plan is achieved. Thus, we use the Cimmino algorithm to select
systematically and iteratively a small number of beam positions that can
deliver a clinically acceptable dose distribution to the patient. This is done
in the following manner. Once we have at hand a basic solution x* = (z )
*
with the 24 x 50 ray intensities #*, we calculate for each beam the sum
Sect. 11.5 Experimental Results 323

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.8 The discretized section derived from Figure 11.7
(with the pixels compressed, for convenience of display, in the
vertical direction).

of the individual ray intensities (this number is proportionate to the dose

contribution of the whole beam) and the sum of the absolute values of
the differences of adjacent ray intensities (this number is indicative of the
beam smoothness across its profile). Because the ray intensity sums of
the 24 beams vary considerably, we separate the beams into two groups.
The high ray-intensity group has 11 beams; thus the other 13 beams can
be eliminated. Applying the Cimmino algorithm a second time to the 11
remaining beams, on the basis of ray intensity sums alone we can determine
six beams with higher such sums. When the Cimmino algorithm is applied
a third time—to the remaining six beams—the resulting ray intensity sums
are all very similar and moreover, the beams are adjacent along an arc
from 45° to 120° (see Table 11.1). We then use the criterion of smoothness
across the target, and eliminate two beams with significantly large adjacent
difference sums. A fourth application of Cimmino’s algorithm to the four
remaining beams shows that one beam (the one at 45°) can be eliminated
on the basis of ray sum and smoothness. The fifth application of Cimmino’s
algorithm to three beams shows virtually the same ray sums and adjacent
difference sums for each beam; and a sixth application of the algorithm
gives the ray intensities for a two-beam treatment plan.
In this way we have used two empirical criteria to reduce the number
324 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

0 0 0 0 0 0 0 0 0 |45 45
11 0 0 0 00 0 0 0 0
0 0 0 0 0 0 F0 10olp50 145 45 0 0 pr10o|| 00 0 0 0 0
0 0 0 0 0 F 45 45 45 F 0 0 0 0 0
nr 45]10

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

0 0 0 0 | 45"101p5 45 45 45 45 0 1|45"|| 0 0 0 0
0 0 0 |45 45 45 45 45 45 45 45 45 45 45 45 45 45 0 0 0
0 01 45 45 45 45 45 45 45 45 45 45 45 45 45 45 0 0 0
0 0 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 0 0
0 o|I 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 0 0
0 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 0 0
0 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 0 0
0 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 0 0
0 45 45 45 45 45 45 45 45 45 45 45 1 1
45 ["50] 45 45 45 45 0 0
0 0 0
45 45 45
0 145 45
45 45
45
45
45
45
T455 0
0
0
0

45 45 45 45
45 45 45
45 45
45 45
45
45
45
45
|45
0
45
0
ZD
0
0 0 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 0 0 0
0 0 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 0 0 0
0 0 0 45 45 45 45 45 45 45 45 45 45 45 45 45 0 0 0 0
0 0 0 45 45 45 45 45 45 45 45 45 45 45 45 0 0 0 0 0
0 0 0 45 45 45 45 45 45 45 45 45 45 0 0 0 0 0 0
0 0 0 45 45 45 45 45 45 45 0 0 0 0 0 0 0
0 0 0 o o 0 0 0 0 0 0 0 0 0 0 0 0

Figure 11.9 Physician-prescribed treatment requirements.

(We use a minus sign to indicate the target pixels containing
the tumor.)

of beams in the final treatment plan. The more important criterion is a

threshold on the sum of ray intensities of the beams and it can be used
if the ray intensity sums are sufficiently different. The secondary criterion
is the choice of beams with least sum of absolute values of differences of
intensities of adjacent rays. This criterion helps to differentiate between
beams with similar sums of ray intensities.
When employing the full set (24 in our case) of beams there are no
violations of the dose prescription, indicating that a feasible solution exists.
As more beams are eliminated with our empirical criteria, errors begin
to occur, indicating that the dose prescription cannot be satisfied with
few beams (see Table 11.2). Figure 11.10 shows the dose distribution and
the dose prescription violations obtained with the six-beam configuration
of our empirical method; and Figure 11.11 shows the same information
for the two-beam configuration. A trained dosimetrist, who did not use
our model, chose a two-beam configuration of 60° and 120°—which are
exactly the same angles derived by our empirical method for the two-beam
configuration.
Sect. 11.5 Experimental Results 325

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.10 Dose distribution (top figure) and dose pre­

scription violations (bottom figure) for the six-beam configu­
ration.
326 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.11 Dose distribution (top figure) and dose pre­

scription violations (bottom figure) for the two-beam configu­
ration.
Sect. 11.5 Experimental Results 327

Table 11.1 Repeated application of Cimmino’s algorithm to

choose beams. (The two leftmost columns in the table display
the beam numbers and angles respectively.)

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Absolute sum of
Sum of ray intensities adjacent ray
intensity differences
Number of beams Number of beams
24 11 6 4 3 2 24 11 6 4 3 __ 2
1 0 10 - - - - 4 - - - -
2 15 126 - - - - 28 - - - -
3 30 165 133 - - - 22 24 - - -
4 45 179 147 143 137 - 34 28 28 34 -
5 60 184 151 147 146 145 150 28 26 24 26 32 40
6 75 184 155 151 - - 38 34 36 - -
7 90 184 150 152 - - 46 36 38 - -
8 105 187 157 150 152 147 34 30 28 32 34
9 120 186 154 148 143 149 149 30 26 28 24 30 34
10 135 153 112 36 28
11 150 65 - 32 -
12 165 145 104 38 32
13 180 48 - 28 -
14 195 133 - 24 -
15 210 148 124 24 26
16 225 147 125 28 24
17 240 139 22
18 255 138 28
19 270 134 32
20 285 134 22
21 300 129 20
22 315 126 22
23 330 21 12
24 345 109 28 -

Table 11.2 Increasing error with decreasing number of

beams—for the procedure described in text to eliminate beams
from a treatment plan.

Number of Number of Absolute

beams pixels with errors sum of errors
24 2 2
6 3 3
4 3 5
3 4 5
2 10 22
328 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

11.6 Combination of Plans in Radiotherapy

We conclude our chapter on radiation therapy treatment planning with a
discussion of how to systematically combine several given treatment plans,
which we term plan combination in RTTP. Plan combination gives rise to

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

a problem that is neither of the forward nor of the inverse type discussed
earlier, but one that addresses a situation when for a specific clinical case,
a set of several treatment plans is proposed, for which both the forward
and the inverse problems have already been solved with whatever avail­
able methods. This means that for each plan both a description of the
dose distribution and a detailed scheme of machine setup parameters for
dose delivery are available. Additionally, it is assumed that although all
proposed plans approximate the desired and prescribed dose distribution,
they all violate the prescribed dose in at least one significant region of the
volume to be treated.
Such situations arise in clinically complicated cases, so that even an ex­
perienced dosimetrist will have difficulty in finding an acceptable treatment
plan. Alternatively, it is conceivable that a set of such inaccurate treat­
ment plans will be generated intentionally in some time- or effort-saving
crude planning method. Both possibilities are especially plausible for truly
three-dimensional planning, and present a dilemma to the radiotherapist.
He must either choose a single plan from the set of plans, or attempt to
create new plans, probably based on relaxed requirements. We will now
discuss treatment plans as vectors in the Euclidean space, and define their
equivalence, acceptability and realizability. A simple linear algebraic model
for combining them is utilized in order to derive, from the given set of ap­
proximate plans, a combined treatment plan that will be both acceptable
and technically realizable. In the event that such a combined plan does
not exist, the alternatives for relaxing the treatment requirements will be
systematically considered.
11.6.1 Basic definitions and mathematical modeling
Let Qj, j = 1,2,... ,n, be n mutually disjoint subsets of a cross section Q
such that Q = Uj=1Qj. These could be exactly identified with structures
such as tumor regions and critical organs, but could also be (as will be
understood later) any other regions in Q for which the dose distributions
in any of the proposed plans do not conform with the requirements.
Definition 11.6.1 (Treatment Plan) A vector x — (xj)^ E IRn is called
a treatment plan vector relative to a given cross section Q, and its partition
{Qj}, j = l,2,...,n, if Xj is nonnegative and it represents the total dose
absorbed everywhere inside the region ilj for all j = 1,2,..., n.

Definition 11.6.2 (Radio-equivalence) Two radiation therapy treatment

plans that give rise to two treatment plan vectors x\x2 are called radio­
equivalent if x1 = x2 (component wise) regardless of the actual treatment
Sect. 11.6 Combination of Plans in Radiotherapy 329

machine setups and the types of machines that yield them.

Let I = (lj) and u = (uj) be two given treatment plan vectors in which
lj and Uj are respectively, the lower and upper bounds on the dose at region
Clj. These vectors are assumed to be prescribed by the radiotherapist.

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Definition 11.6.3 (Acceptable Treatment Plan) Given Q, subsets Qy,
j = 1,2,... ,n, I and u, a treatment plan vector x is called acceptable if
regardless of how x is delivered, lj < Xj < Uj, j = 1,2,... , n.
Definition 11.6.4 (Realizable Treatment Plan) A treatment plan vector
x is called realizable if there exists (in-house) a clinical machine setup that
can deliver it.
Realizability is a subjective and user-dependent concept that may vary
according to equipment availability and treatment planning capabilities.
The zero vector 0 € IRn is always realizable, but not acceptable unless
1 = 0. It is henceforth assumed that by utilization of whatever planning
procedure (or procedures) and by using whatever equipment, a finite set
of treatment plan vectors {x1^2,... ,xm} has been generated for a given
clinical case such that every xl = (x\, x^ ..., xln) is realizable in the given
clinical environment, but none of the xl is acceptable. This means that for
each i, 1 < i < m there exists at least one j, 1 < j < n, such that either
x^ < lj or xlj > Uj.
A nonnegative combination of treatment plan vectors is a treatment
plan vector y = aixl suc^ that the linear combination coefficients
ai, 02,..., are nonnegative real numbers. From the nature of RTTP, a
nonnegative combination of realizable vectors is also realizable.
The treatment plan vectors I and o, prescribed for a given clinical case,
determine a box B (henceforth called treatment box) , in IRn defined as

B = {x € IRn | lj < Xj < Uj, j = 1,2,... ,n}. (11.19)

Obviously, every treatment plan vector that belongs to B is acceptable.

The set {x1, x2,..., xm} of realizable but unacceptable treatment plan
vectors, together with the always realizable zero vector generate a polyhe­
dral convex cone C with apex at the origin, given by the convex hull of all
nonnegative linear combinations of the vectors {x2}^, i.e.,
m
C = conv{?/ e IRn | y = ^^aix'1, a* > 0, i = 1,2,... ,m}. (11.20)
i=l

Following these definitions, the linear algebraic model for plan combination
in RTTP can be formulated as follows:
Given Q, Qj, j = 1,2, ...,n, I and u, and given a set of
realizable treatment plan vectors denoted by {x, x2,..., xm},
find a treatment plan vector y such that y G B n C.
330 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.12 A plan combination problem with n = 2
regions and a set {x1, x2, x3} of realizable but unacceptable
treatment plans. In this drawing, B n C 0 and the problem
is feasible.

Since the plan combination problem is a linear feasibility problem (see

Chapter 5), we call any treatment plan vector y that solves it a feasible
solution. Figure 11.12 depicts a typical feasible plan combination problem
with m = 3 and n — 2. The overall methodology considered in this section
is summarized in the flow chart presented in Figure 11.13 whose meaning
becomes clear after reading the next subsections.

11.6.2 The feasible case

The formulation of the plan combination problem makes a thorough anal­
ysis possible and suggests several options for radiation therapy treatment.
The problem poses first the question of feasibility, i.e., does at least one
acceptable linear combination of realizable plans exist? This question can
be answered using linear programming methods.
First, we give a dual formulation of the plan combination problem.
Let us organize the given set {x1, x2,..., xm} of realizable but probably
unacceptable treatment plan vectors in an n x m matrix X whose zth
column is x\ The transpose of the Jth row of X is denoted by rJ, i.e.,
= (xpXp ... , x™)T. The dual formulation of the plan combination
problem of finding a vector y C B A C is the following linear feasibility
problem:
Sect. 11.6 Combination of Plans in Radiotherapy 331

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.13 Methodological flow chart of plan combination.

Given Z, u and r-7, j — 1, 2,..., n, find a vector a = (ai) e lRm such that

lj < (a, r7) < Uj, j = 1,2,..., n,

i — 1,2,..., m. (U-21)
di > 0,

The duality between the plan combination problem and (11.21) is ob­
vious, and it is also clear that the plan combination problem is feasible if
and only if (11.21) is.
The simplex method of linear programming is a practical tool for solving
(11.21). Phase 1 of the simplex method is concerned in particular with
332 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

finding a feasible solution to a linear programming problem and it answers
the question whether (11.21) is feasible or not. For the remainder of this
section we assume feasibility, i.e., that B Cl C 0.
Therefore, in this case more information can be extracted, and probably
several feasible treatment plan vectors can be suggested. The physician
may be interested in a treatment plan that is not only feasible, but is
also optimal in some sense. For example, in addition to feasibility we
may wish to minimize the total radiation dose /(a) given to the patient,
where /(a) = 522L1 or alternatively, to maximize the total dose
to the target #(a), where g(o) = and is a subset of
{1,2,..., n} that consists of all the indices corresponding to the target.
Another criterion for selecting one realizable and acceptable treatment plan
vector over another may be that a combined treatment plan that uses fewer
beams, i.e., = 0, for some or several indices i C {l,2,...,m}is preferable,
perhaps for clinical reasons.
These are all linear programming problems and the simplex method or
an interior point algorithm (see Chapter 8) can be applied. Since in our case
the polyhedral set described by (11.21) is bounded, it is impossible to have
an unbounded solution. The following example illustrates this approach.
Example 11.6.1 (Plan combination) Suppose that the target for the ra­
diation treatment is in the chest, surrounded by the two lungs, the heart
and the spinal cord, as illustrated in Figure 11.14- Let the vectors of lower
and upper bounds be I = (0,0,0,0,0,54)T and u = (20,30,44,44,60, 100)T,
Sect. 11.6 Combination of Plans in Radiotherapy 333

respectively. The entries in each vector are doses measured in Gy units

and correspond to the organs as follows: the first entry corresponds to the
right lung, the second entry to the left lung, the third to the heart, the
fourth to the spinal cord, the fifth to the complementary tissue and the

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

sixth entry corresponds to the target. Let z1,#2,#3 be three given vectors
of realizable but unacceptable treatment plans: x1 = (15,25,40,43,45,50)T,
x2 = (30,40,50,40,60,60)T, x^ = (40,40,60,55,60,55)T. Let aua2, and
be the corresponding coefficients that we seek for the linear combination.
Then the linear feasibility problem for this example is to find nonnegative
numbers ai,a2, and as such that

0 < 15ai -b 30ti2 "b 40^3 < 20,

0 < 25&1 4- 40(^2 40^3 < 30,
0 < 40ai 50(22 4" 60(23 _ 44, , .
0 < 43(21 + 40a2 4- 55a3 < 44, U }
0 < 45ai 4- 60(22 4- 60(23 < 60,
54 < 50ai + 60(22 4- 55a3 < 100.

We impose an objective function f(a) to minimize the total radiation

given to the organs surrounding the target tissue. By way of example we
take f(a) = 168ai 4-220a2 4-255a3, which has to be minimized subject to a >
0 and to all inequalities in (11.22). A solution with a linear programming
algorithm yields the following coefficients (all numbers are rounded): a± =
0.83, 02 = 0.21, (23 = 0.0. The treatment vector corresponding to the
amount of radiation each organ receives, if this combined plan is used,
is y = (18.72, 29.10, 43.62, 44, 49.86, 54)T. The objective function value
is 185.31 Gy. Moreover, elementary sensitivity analysis shows that if we
decrease the lower bounds of the radiation given to the target by one unit
(i.e., from 54 Gy to 53 Gy), then the total radiation to all but the target
organ will decrease by 4-72 Gy.
A different objective function, which might at times be more desir­
able, is one that maximizes the total radiation g(a) = 50(21 4- 60&2 + 55^3
given to the target tissue, subject to the constraints. The solution ob­
tained for this problem is: a± = 0.8, 02 = 0.24, <23 = 0, with total ra­
diation to the target 54-4 Gy, and with a combined treatment plan vector
y = (19.2, 29.6, 44, 44, 50.4, 54.5)T.

11.6.3 The infeasible case

When no feasible solution exists (see, for example, Figure 11.15), it is not
necessarily a hopeless situation, and there are several possible alternatives.
The analysis of Phase 1 of the simplex method provides some insight to the
infeasibility issue, as it yields important information about the model, such
as which constraints are violated if the current (infeasible) solution is used,
and which constraints are tight for this solution, namely constraints that
334 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.15 BAC = 0, an infeasible plan combination prob­
lem.

are met with equality. Using this information, the physician may decide to
reformulate the dose requirements I and u only in so far as necessary to ar­
rive at a feasible combined treatment plan. Such a possibility is illustrated
in the next example.
Example 11.6.2 (An infeasible case) Suppose that for the same patient
as in Example 11.6.1 we have the following realizable but unacceptable
plans x1 = (15,25,40,45,45,55)T, x2 = (30,40,50,50,60,50)T, and x3 =
(40,40,60,55,60,40)T. Application of Phase 1 of the simplex method yields
an infeasible solution with a± = 0.9777, <22 = 0.0, = 0.0, giving rise to
the treatment vector: y = (14.67,24.44,39.11,44,44,53. 78)T. The violated
constraint is the sixth one, corresponding to the target region. It is violated
(underdosed) by 0.2222 Gy. In such a situation, the physician may decide
that a relatively small decrease in the lower bound of the radiation require­
ment for the target can—and should—be done. Thus to obtain a feasible
solution, the lower bound should be reduced from 54 Gy to, say, 53.7 Gy.
An alternative change might be an increase in the upper bound of the con­
straint that contributes to the infeasibility, namely the fourth constraint. A
simple computation indicates that the problem becomes feasible if the up­
per bound (44 Gy) is replaced by 44Gy, which results in the combined
treatment vector y = (14.73,24.55,39.27,44.18,54)T.
There is an alternative approach for the infeasible case. Rather than
relaxing the vectors I and/or u as dictated by the sensitivity analysis of the
Sect. 11.6 Combination of Plans in Radiotherapy 335

simplex method, the physician is called upon to provide additional input.

He is required to point out the relative importance (priorities) of delivering
the prescribed doses to the various regions. An iterative algorithm is then
applied to the dual formulation of the plan combination problem (11.21)

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

and the result is called an adjusted solution to the original infeasible plan
combination problem.
Various iterative algorithms are applicable to (11.21), many of which
can be derived as special cases of the block-iterative projections (BIP)
method (Algorithm 5.6.1). Some, such as ART3 (Algorithm 5.10.2) or
ARM (Algorithm 5.10.1) are specifically tailored to handle an interval
linear feasibility problem. However since many of these algorithms rely
on feasibility of (11.21) for their convergence, we construct here an ad
hoc fully simultaneous version of ART3 and demonstrate its capability to
(asymptotically) generate adjusted solutions to the original infeasible plan
combination problem.
We use the notation of Section 5.10. For each j, an interval linear
inequality of (11.21) represents in lRm a hyperslab, whose bounding hyper­
planes are Lj = {a e IRm | (a,r7) = lj} and Uj = {aE IRm | (a, rj) — Uj}.
The median hyperplane is, for each J, Hj = {a G IRm | (a, r7) = |(Z7 4-u7)},
and the half-width of the Jth hyperslab is Wj = (uj — Ij)/2 || r7 ||. The
signed distance of a point z G lRm from Hj is given by

Let be a sequence of iterates. We denote the distances of an

element of the sequence from the median hyperplane by

dj ~ d(a",HjY (11.24)

The sign 4-1 or -1 of dj is denoted by tj = sign (dj). With these definitions

a simultaneous ART3 algorithm that aims at solving the feasibility problem
(11.21) can be constructed. Assume that a set of real numbers tt7 for all
j = 1, 2,..., n is given, such that 0 < 7r7 < 1 for all J, and 71j ~ I-
By assigning such a set of priorities tf7 the therapist indicates the relative
importance attributed to each face of the treatment box 5, defined by
(11.19), thus to each interval of the system (11.21). Analogous to a typical
step of ART3 we define, for a given iterate G IRm and with respect to
each hyperslab Qj — {a G IRm | lj < (a,ri) < Uj}, the step sizes

0, if | dj |< Wj,
sv,j Oil/ (dj tjWj), if Wj < | dj | < 2w7, (11.25)
dj^ if 2wj < | dj |,
336 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

where dj and tj are determined from the formulae given above. Factors
that determinine the value of the constants are explained later.

Algorithm 11.6.1 Simultaneous ART3

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Step 0: (Initialization.) a0 e IRm is arbitrary.

Step 1: (Iterative step.) Given the current iterate ay, calculate for each
index j, 1 < j < n,

au+1 — av - sv,j (11.26)

then calculate
(11.27)

It is straightforward to verify that each intermediate iterate ai'+lj is a

relaxed orthogonal projection of ay onto Qj. Thus if Ap denotes relaxation
parameters, then equation (11.26) can be rewritten as

a-'+iJ = + (Pq.^) - a1') , (11.28)

where,

(11.29)

For = a = 2 the iterative step (11.26) coincides with that of the

original ART3 algorithm. However, to ensure applicability of the theory to
the present simultaneous ART3 algorithm, the relaxation parameters in
(11.28) must fulfill the condition that for all v > 0,

< Ap < 2 — €2, (11.30)

with any fixed 6i,62 > 0.

If this condition is adhered to, then the simultaneous ART3 algorithm
becomes a special case of the nonlinear fully simultaneous Cimmino al­
gorithm, i.e., the BIP algorithm (Algorithm 5.6.1) with all indices of all
constraints lumped into a single block. Then the theory guarantees con­
vergence of the simultaneous ART3 algorithm. The algorithm converges
globally to a solution of (11.21) if (11.21) is feasible. If (11.21) is infeasible,
then the iterates {ap} generated by the simultaneous ART3 algorithm can
be shown to converge locally to a point that minimizes the convex com­
bination with {ttj} as coefficients of the squares of the distances to the
hyperslabs Qj, j = 1,2,..., n.
Sect. 11.6 Combination of Plans in Radiotherapy 337

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

Figure 11.16 Dual space description of an infeasible plan
combination problem.

In this sense we call such a limit point an adjusted solution to the infea­
sible original plan combination problem. It violates some or all constraints
which define the treatment box B, but at the same time it is a weighted
least squares solution to the linear feasibility problem - the weights being
the prescribed priorities {7Tj}y=1. Nonnegativity of the iterates {a17} gen­
erated by the simultaneous ART3 algorithm is not guaranteed even if the
initial point a° E JRm is chosen so that aQ > 0. To overcome this difficulty
we perform orthogonal projections onto the nonnegative orthant between
every two consecutive iterations of the simultaneous ART3 algorithm. This
is done by replacing a"
* 1 in (11.27) by dI/+1 and then adding to the iterative
step the operation

a^+1,if^+1 >o,
(11.31)
0, if^+1<0.

Algorithm 11.6.1 with (11.30)—(11.31) converges to a nonnegative weight­

ed least squares solution of the system (11.21), i.e., to a point in the region
marked A in Figure 11.16.
Example 11.6.3 Using the data in Example 11.6.2, we run a computer
program that executes the simultaneous ART3 algorithm. We stopped
after p iterations, where p is the smallest integer such that the inequality
338 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

|| || / || || < IO 5 holds. If equal weights are given to all organs,

then the solution is a = (0.98, 0.0013, 0.001 )T, and the resulting plan is
y — (14.75, 24.54, 39.25, 44.13, 44.15, 53.9)T. We see that the problematic
constraints, namely the fourth (which is associated with the spinal cord)

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

and the sixth (the target) are violated. Nevertheless, y is a least-squares
point as described above. If different weights are given, the algorithm con­
verges to a weighted least-squares point. For example, if the weights given
in this example are (0.05, 0.05, 0.05, 0.05, 0.05, 0.75) then the correspond­
ing treatment plan is y = (14.8,24.62,39.35,44.23,44.27,53.99)T, resulting
from the solution a = (0.98, 0.002, 0.0018)T.
It should be noted that the two alternatives suggested for the infeasible
case, may be mathematically and practically equivalent for some cases.
The same feasible solution obtained by moving an upper or lower bound,
as described in the first option for the infeasible case, can be obtained using
the simultaneous ART3 with appropriate weights.

11.7 Notes and References

For a treatise on the physics of radiology see, e.g., Johns and Cunningham
(1974) or De Vita, Hellman, and Rosenberg (1989). Some condensed details
about radiation quantities related to this chapter can be found in Barrett
and Swindell (1981, Appendix D). See also the computationally oriented
book of Wood (1981), or a conference proceedings such as Bruinvis et
al. (1987). An up-to-date review of the field (including 130 references) is
found in Brahme (1995b). This work contains also a concise description of
the biological objective function optimization approach, recently pursued
with promising results by Brahme and his coworkers. Much can be learned
about recent progress from a special issue of the International Journal of
Imaging Systems and Technology, Brahme (1995a).
11.1 We presented the continuous model and formulated the continuous
forward and inverse problems in RTTP in general terms only. Work
in this direction (as opposed to the fully discretized approach that
we have adopted here) originates in the paper by Brahme, Roos and
Lax (1982), followed by the papers of Cormack (1987), Cormack and
Cormack (1987), and Cormack and Quinto (1989, 1990). The for­
ward problem, represented by (11.1), i.e., the process commonly called
dose calculations, has received great attention for many years now and
its complexity brought about many methods and formulae. Several
relevant references are: Geijn (1972), Cunningham, Shrivastava, and
Wilkinson (1972), Sontag and Cunningham (1978), Lane, Bloch f‘d
Davis (1974), Cunningham (1972), and the review by Goitein
Dose calculations continue to form a substantial research issue, see,
e.g., Altschuler, Sontag and Bloch (1987), and Nafstadius, Brahme
and Nordell (1984). The quotation of Dr. M. Goitein is taken from
Sect. 11.7 Notes and References 339

a report by Hindus (1988); see also Barrett and Swindell (1981, Ap­
pendix D.5.3). The feasibility formulation (11.4) was proposed by
Altschuler and Censor (1984).
11.2 Mathematical methods (mostly optimization theory techniques) are

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

used in radiation therapy in ways that differ conceptually from the
approach presented here. These other approaches can be generally
identified as follows.
Comparison among rival plans. Here several treatment plans are com­
pared, based on their score with respect to some predetermined quality
index. The treatment plans are all fixed, and the selection of a plan
depends largely on the definition of the quality index. Various quality­
index functions were proposed and advocated on different grounds.
These include, by way of example, the complication probability fac­
tor (CPF) of Dritschilo et al. (1978) and Wolbarst et al. (1980), the
normal tissue dose (NTD), the optimum target dose (OTD), and the
critical organ dose (COD), in Kartha et al. (1982); see also Wolbarst,
Chin, and Svensson (1982).
Optimization with respect to a few setup parameters. Here the com­
pared treatment plans are not fixed, but most setup parameters are
kept unchanged while a few (one, two, or more) are allowed to take
several possible values which will result in different treatment plans.
A criterion function is set up, and a mathematical programming (i.e.,
optimization) problem is solved to yield an optimal choice of the un­
fixed variables. Usually linear programming is used. Although such
an approach is logically also a comparison among rival plans, it is ap­
parent that because of the greater freedom in the selection process
(i.e., the process of mathematical optimization) many more options
are compared. Into this category fall the work of Gallagher (1967),
Bahr et al. (1968), Redpath, Vickery and Wright (1976), McDonald
and Rubin (1977), and Starkschall (1984). The implementation of
some of these methods, such as in McDonald and Rubin (1977) and
in Starkschall (1984) constrains the dose distribution at only a small
number of points (typically less than 20) chosen by the dosimetrist.
Additional work with optimization techniques includes Legrass et al.
(1986), Fymat et al. (1988), the method of Webb (1989) with simulated
annealing, Swan (1981), Burkard et al. (1994), and many others. For
general reviews of the status of involvement of mathematical modeling
see, e.g., Aird (1989) and Goitein (1990).
Another attempt at approaching the problem mathematically is given
in the papers of Ebert (1977a, 1977b). His models are different from
those studied here. The S-model assumes a source of radiation that
moves around the patient section in orbits, while his F-model allows
340 The Inverse Problem in Radiation Therapy Treatment Planning Chap. 11

some parameters to vary as others remain fixed and therefore falls into
the second class of methods described above.
Among more recent advances we refer the reader to the works of
Brahme, Lind, and Kailman (1990), Lind (1990), and Lind and Kailman

Downloaded from https://academic.oup.com/book/53915/chapter/422193784 by OUP site access user on 12 May 2024

(1990). Their approach is to apply various iterative techniques to a
discretized version of a Fredholm integral equation of the first kind.
See also Soderstrom (1995). Raphael (1991, 1992a, 1992b) in his dis­
sertation and subsequent publications formulated the RTTP problem
as constrained optimization in appropriate C2 function space and ap­
plied to it iterative algorithms. His work and the work of Brahme
and co-workers mentioned above, along with the full discretization ap­
proach presented here, show in more than one way the usefulness of
iterative techniques for the solution of the inverse problem in RTTP.
The approach that we presented in this chapter differs fundamentally
from all these approaches. We do not compare fixed rival plans. We
do not select a plan by allowing one or few parameters to change se­
quentially. We rather address the inverse problem within a general
framework where both the patient section and the sources are fully
discretized and the full discretization is the tool for handling the mod­
eling difficulties inherent in the continuous model.
11.3-11.5 The presentation in these sections, as well as in the previous
one, is based on works of Altschuler, Censor, and Powlis (1988, 1992),
Censor, Altschuler, and Powlis (1988), Censor, Powlis, and Altschuler
(1987), and Powlis et al. (1989). A review appears also in Altschuler,
Censor, and Powlis (1992). A description of a new generation of radia­
tion therapy machines is given by Brahme (1987). Experiments with a
similar approach are reported in Altschuler, Powlis, and Censor (1985)
and Powlis, Altschuler, and Censor (1985) where the method is applied
to uniform beams which are not further discretized into rays.
11.6 The material of this section comes from Censor et al. (1988) and
Censor and Schwarz (1989). It should be noted that the plan combi­
nation approach is closely related to the full discretization philosophy
given earlier. The inverse problem of RTTP, in its fully discretized
formulation is actually a high-resolution plan combination problem.