Organ sample generator for expected treatment dose construction

and adaptive inverse planning optimization

Xiaobo Nie, Jian Liang, and Di Yan
Department of Radiation Oncology, Beaumont Health System, Royal Oak, Michigan 48073

(Received 9 February 2012; revised 5 October 2012; accepted for publication 17 October 2012;
published 26 November 2012)
Purpose: To create an organ sample generator (OSG) for expected treatment dose construction and
adaptive inverse planning optimization. The OSG generates random samples of organs of interest
from a distribution obeying the patient specific organ variation probability density function (PDF)
during the course of adaptive radiotherapy.
Methods: Principle component analysis (PCA) and a time-varying least-squares regression (LSR)
method were used on patient specific geometric variations of organs of interest manifested on multiple
daily volumetric images obtained during the treatment course. The construction of the OSG includes
the determination of eigenvectors of the organ variation using PCA, and the determination of the
corresponding coefficients using time-varying LSR. The coefficients can be either random variables
or random functions of the elapsed treatment days depending on the characteristics of organ variation
as a stationary or a nonstationary random process. The LSR method with time-varying weighting
parameters was applied to the precollected daily volumetric images to determine the function form of
the coefficients. Eleven h&n cancer patients with 30 daily cone beam CT images each were included
in the evaluation of the OSG. The evaluation was performed using a total of 18 organs of interest,
including 15 organs at risk and 3 targets.
Results: Geometric variations of organs of interest during h&n cancer radiotherapy can be repre-
sented using the first 3 ∼ 4 eigenvectors. These eigenvectors were variable during treatment, and
need to be updated using new daily images obtained during the treatment course. The OSG generates
random samples of organs of interest from the estimated organ variation PDF of the individual. The
accuracy of the estimated PDF can be improved recursively using extra daily image feedback during
the treatment course. The average deviations in the estimation of the mean and standard deviation of
the organ variation PDF for h&n cancer radiotherapy were less than 2 and 1 mm, respectively, for
most organs after the second week of treatment. After the first three weeks of treatment, the mean
discrepancy of the dose estimation accuracy was within 1% for most of organs, the corresponding
standard deviation was within 2.5% for parotids, the brain stem and the cochleae, and within 1% for
other organs.
Conclusions: A patient specific OSG is feasible and can be used to generate random samples of or-
gans of interest for the expected treatment dose construction and adaptive inverse planning. The accu-
racy of the OSG can be improved continuously and recursively during the adaptive treatment course
using daily volumetric image feedback. © 2012 American Association of Physicists in Medicine.
[http://dx.doi.org/10.1118/1.4765457]

Key words: organ geometric variation, patient specific organ sample generator, expected treatment
dose, adaptive inverse planning optimization

I. INTRODUCTION and online patient surface or diaphragm motion.3, 4 In addi-

Geometric variations of a patient’s anatomy in shape, size,
and position during the course of radiotherapy have been de-
scribed using a random process. This random process can
be characterized by a position variation probability density
function (PDF) of tissue elements in organs of interest.1 The
organ variation/motion PDF can be used to construct ex-
pected treatment dose and perform adaptive inverse planning
optimization.2 However, studies on organ variation PDF con-
struction or estimation during the treatment course have been
very limited.
A few modeling studies on lung tumor motion have been
performed using pretreatment 4D computed tomography (CT)
tion, if an organ geometric variation can be described us-
ing a stationary random process with the normal distribution,
the estimation of the mean and the standard deviation of the
tissue element position will be sufficient1, 5 to describe the
process. However, tumors and certain normal organs, such
as parotid glands, shrink significantly during the treatment
course.6 The geometric variation of these organs is therefore
nonstationary.1 Lung tumor volume shrinkage has been stud-
ied and estimated using a locally weighted regression model,7
but a general estimation model on tissue element position dis-
tribution has not been explored.
Volumetric images, such as daily helical CTs, onboard
cone beam (CB) CTs, as well as MRIs, obtained during the

7329 Med. Phys. 39 (12), December 2012 0094-2405/2012/39(12)/7329/9/$30.00 © 2012 Am. Assoc. Phys. Med. 7329
7330 Nie, Liang, and Yan: Organ sample generator 7330

Daily Volumetric Images in the Pre-collected

first τ days of the treatment Patient Image
Data

Determine & Select Eigenvectors Estimate the meean Determine Model Parameters
Vj(τ): Equation (1) Coefficients Cj(t): Equuation (5) L; λ: Equations (6b) and (7)

Determine the coeffficients

Cj(t): Equations (8) and (9)

The Organ
g Samplep Generator: {Random Samples}
Equation (10) O
Organs off IInterest

1. Expected Treatment Dose

2 Ad
2. Adaptive
ti IInverse Pl
Planning
i

F IG . 1. Flowchart for the OSG construction.

treatment course can be used to determine organ geometric
variation/motion PDF. However, all these measurements share
the same caveat in PDF construction due to the small sample
size of the measurements. Therefore, using a limited number
of samples to construct the organ variation/motion PDF is not
only of academic interest, but also benefit in clinical practice
of adaptive radiotherapy.
In this study, principle component analysis (PCA),8 and
a time-varying least-squares regression (LSR) method9 were
applied to create a recursive organ sample generator (OSG).
The OSG reads daily volumetric images, and generates ran-
dom samples of organs of interest from a distribution obeying
patient specific organ variation PDF. A retrospective study
was performed to evaluate the OSG, using daily CBCT im-
ages obtained during h&n cancer treatment.
II. METHODS AND MATERIALS
Planning and daily treatment volumetric images of each
patient, obtained during the treatment course, are used in the
construction of the OSG. Position variations of tissue ele-
ments represented by the planning CT image voxels within
the patient’s body are determined by deformable image regis-
tration of each daily image to the planning image. The center
of each image voxel is defined as the tissue element position.
During treatment, the patient specific OSG is constructed
in two steps. First, the PCA method is applied to the posi-
tions of all tissue elements in organs of interest manifested on
all daily images obtained, to date, during the treatment course
to determine the corresponding eigenvectors. These eigenvec-
tors represent the principle components of the organs’ geo-
metric variation contained in the daily image measurements.
In the second step, the time-varying LSR method is applied
to determine the best expansion coefficient for each eigenvec-
tor. The expansion coefficient, in this study, is defined to be a
normally distributed random variable or function of time. The
model parameters of the coefficient are predetermined using
a precollected patient image dataset. Using both the eigenvec-
tors and the coefficients, the OSG is formed to generate ran-
dom samples of organs of interest during the treatment course
for the expected treatment dose construction and adaptive in-
verse planning optimization. Figure 1 provides a flowchart
of the OSG construction process which will be discussed in
Secs. II.A.1–II.A.5.
II.A. Constructing patient specific organ
sample generator
II.A.1. Organ variation eigenvectors from daily images
Utilizing the PCA method to describe organ geometric
variation has been previously introduced by Sohn et al.10 in
a prostate variation study. A similar method and description
will be used here.
Let a patient have τ daily volumetric images obtained dur-
ing the treatment course. Each image includes Ne numbers of
tissue elements that are represented by image voxels in or-
gans of interest. The total number of tissue elements varies by
patient and treatment site, and can be up to 106 . The organs’
geometry in the tth daily image can be described by a position
vector defined as
ptT = [x1 , x2 , . . . , xm ] ∈ R m , m = 3Ne ; 1 ≤ t ≤ τ,
where the components of the vector pt represent all element
positions in all organs of interest observed at a time instant
during the tth treatment day.
The organs’ mean position vector calculated using the
daily imaging observations in the first τ treatment days is
therefore
1
τ
p̄ = pt .
τ t=1

Medical Physics, Vol. 39, No. 12, December 2012

7331 Nie, Liang, and Yan: Organ sample generator
Considering a matrix which consists of the position vari-
ations from its mean position observed in the τ treatment
days, P = [p̃1 , p̃2 , . . . , p̃τ ]m×τ and p̃t = pt − p̄, the covari-
ance matrix shall be Cm×m = τ −1 1
P P T . Although the covari-
ance matrix is large in size, its number of nonzero eigenval-
ues nλ is less than the number of variation observations,8 i.e.,
τ . Each eigenvector obtained from this covariance matrix de-
fines a spatial random variation vector of all tissue element
positions. The eigenvectors, Vj ; j = 1, . . . , nλ , with nonzero
eigenvalues form a complete spatial coordinate base for the
observed position variations used in the covariance matrix.8
Therefore, the observed variations can be expressed using a
linear combination of these eigenvectors, such as,
nλ
pt = p̄ + Cj (t) · Vj , t = 1, 2, . . . , τ, (1)
j =1
where the coefficient Cj (t) is the projection of the position
vector pt in the direction of eigenvector Vj and nλ ≤ τ − 1.
The eigenvectors were normalized such that
1  T  1 
m
Vj · Vj = xi · xi = 1,
Ne Ne i=1
where {xi |i = 1, . . . , m} are the components of Vj . Therefore,
once the eigenvectors are determined, the known organ ge-
ometry can be described using the linear combination of the
nλ eigenvectors and the coefficients, {Vj , Cj |j = 1, . . . , nλ }.
Based on the theory of the PCA,8 the organ geometric vari-
ation space spanned by these eigenvectors is identical to the
one formed by all observed organ geometry variations. The
eigenvectors can then be considered to be the new or the prin-
ciple axes of the organ spatial variation. The expansion co-
efficient of each eigenvector is the coordinate along the axis
defined by the eigenvector in the new spatial reference frame.
The coefficient value along the eigenvector direction is di-
rectly related to the corresponding eigenvalue of the covari-
ance matrix8 such that
1 
τ C̄ˆ j (t) = T (t) · (τ ) =
λj = [Cj (t)]2 , j = 1, . . . , nλ , (2)
τ − 1 t=1
where λj is the jth eigenvalue. This implies that if an eigen-
value λj is small, the corresponding coefficient Cj (t) will be
small for all t = 1, . . . , τ . Thus, the corresponding eigenvec-
tor Vj will have less contribution to the organ variation.
Equation (1) can be used to generate extra organ samples
with different variation patterns by selecting different coeffi-
cients such that
nλ
p̂t = p̄ + Ĉj (t) · Vj , t > τ. (3)
j =1
These organ samples may be used to simulate the potential or-
gan variations for the remaining treatment days, that is, for the
treatment day t > τ . To achieve this, however, random organ
samples, which have the spatial variation distribution obeying
the actual patient organ variation PDF, need to be generated.
Utilizing the observations of a patient’s organ variation ob-
tained during the early treatment days, one may provide the
best estimation of the coefficients in Eq. (3) with respect to
Medical Physics, Vol. 39, No. 12, December 2012
7331
the actual organ variation PDF for the remaining treatment.
The coefficients can be either random variables or random
functions of treatment time depending on whether the organ
variation process is stationary or nonstationary.1 The meth-
ods used to determine the coefficients, as well as the corre-
sponding parameters, will be discussed in Secs. II.A.2–II.A.4.
Equation (3) with the optimal selection of eigenvectors and
coefficients can be used to generate random organ samples
from the patient specific organ variation PDF. Therefore, it is
named the organ sample generator, OSG.
II.A.2. Estimation of the mean coefficients
from daily images
The LSR method9 with time-varying weighting parameters
was employed to estimate the mean of the coefficients. Let τ
be the last treatment observation day, the loss function in the
time-varying LSR can be
1
τ
Vj (θ, τ ) = W (t, τ ) · (C̄j (t) − T (t) · (τ ))2 ,
2 t=1
j = 1, 2, . . . , nλ . (4)
The vector (t) consists of the regressors, and the vector (τ )
consists of the parameters of the regression.
By minimizing the loss function in Eq. (4), we get

τ
(τ ) = (τ ) · (t) · W (t, τ ) · C̄ˆ j (t),
t=1
 τ −1

(τ ) = (t) · W (t, τ ) · T (t) .
t=1
Once the optimized parameters are obtained, the mean of the
coefficients can be constructed as

L
ϕl (t) · θl (τ ), t > τ ;
l=1
j = 1, 2, . . . , nλ . (5)
In this work, the power function was used for the regressors,
(t), and an exponential weighting function with a forgetting
factor λ was employed for W (t, τ ), such that
ϕl (t) = t l−1 , l = 1, 2, . . . , L,
W (t, τ ) = λτ −t
where L is the largest order of the power function. Therefore,
the regression function is a polynomial of order L. When the
exponential weight functions are used, the solution of LSR
can be calculated recursively.9 The forgetting factor λ ∈ [0,
1] is used to assign different weights to different treatment
days. Together with L, this factor also plays a role in the esti-
mation of the mean for a nonstationary variation. It equals to
one for the stationary variation. In this study, the two model
parameters, L and λ, used in describing the coefficients were
predetermined using precollected patient image data.
7332 Nie, Liang, and Yan: Organ sample generator 7332

II.A.3. Determine L, λ from the precollected patient images

daily image data
The two model parameters for the coefficients can be deter-
mined using precollected daily patient images by minimizing
the objective function of the deviation between the estimation
and the observation. The estimation deviation of patient k in
the database is defined as
 1/2
1 T Ne
Eτ k (L, λ) = δ 2
, (6a)
Ne · (T − τ ) t=τ +1 i=1 it
where δ it is the difference between the estimated and the ac-
tual mean positions of the tissue element i on the treatment
day t. The final treatment day is denoted by T. For a station-
ary random process, the actual mean is constant and can be
obtained by averaging all position samples. For a nonstation-
ary process, the mean varies with time and the form of varia-
tion depends on the problem.
If the position vectors are expanded on the nf (≤T − 1)
eigenvectors with nonzero eigenvalue formed by the organ
variations obtained from the T treatment days, the estimation
deviation in Eq. (6a) can then be calculated using the expan-
sion coefficients as
⎛ ⎞1/2
1 T nf
Eτ k (L, λ) = ⎝ (C̄ˆ j (t) − C̄j (t))2 ⎠ .
(T − τ ) t=τ +1 j =1
(6b)
Here, C̄j (t) is the expansion coefficient of the jth eigenvector
of the actual mean position, and C̄ˆ j (t) is the expansion coeffi-
cient of the jth eigenvector of the estimated mean position in
Eq. (5).
The average estimation deviation calculated using treat-
ment observations up to the treatment time τ is defined as
the root mean square (RMS) of all patients, such as
⎛ ⎞1/2
1 
Npat
Eτ (L, λ) = ⎝ E 2 (L, λ)⎠ , (7)
Npat k=1 τ k
where Npat is the total number of patients. The two parame-
ters, the order of the polynomial L and the forgetting factor λ,
can then be determined by minimizing the average deviation
of all patients for a given τ .
II.A.4. Estimated coefficients
Once the two model parameters are determined, the best
estimation of the mean coefficient C̄ˆ j (t) can be obtained us-
ing the Eq. (5). The final coefficient will be the mean plus a
random component, i.e.,
Ĉj (t) = C̄ˆ j (t) + ξj , j = 1, 2, . . . , nλ , (8)
where ξ j is a random number drawn from the normal distribu-
tion with mean = 0 and a standard deviation, σ̂j . The standard
deviation is determined in this study using the observed daily
Medical Physics, Vol. 39, No. 12, December 2012
 1/2
1  τ
σ̂j (τ ) = (Cj (t) − C̄ˆ j (t))2 ,
τ − L − 1 t=1
j = 1, 2, . . . , nλ . (9)
II.A.5. Patient specific organ sample generator
Based on Eq. (3), the OSG constructed after the first τ
treatment days is, therefore,

nλ
p̂(t) = p̄(τ ) + Ĉj (t) · Vj (τ ), t > τ. (10)
j =1
Figure 1 shows the flowchart of the patient specific OSG
construction. During the patient’s treatment course, the daily
volumetric images are obtained. At the treatment date τ ,
the eigenvectors are calculated using Eq. (1), meanwhile the
estimated coefficients are determined from Eqs. (8) and (9)
based on the mean coefficients estimated using Eq. (5)
with the two predetermined model parameters obtained from
Eqs. (6b) and (7). The OSG as indicated by Eq. (10) is then
formed by substituting the eigenvectors and the coefficients
into Eq. (3). The OSG is used to generate random organ sam-
ples for the remaining treatment days t > τ , and updated re-
cursively once new daily patient images are available.
II.B. Evaluation
One planning CT image and T (=30) daily treatment
CBCT images for each of 11 h&n cancer patients, who have
been treated using CBCT guided IMRT, were used in the
evaluation. Eighteen organs of interest were manually con-
toured on each planning CT image. The 15 organs at risk
(OAR) are left and right parotids, mandible, spinal cord, brain
stem, left and right cochleae, inferior, middle and superior
constrictors, supraglottic and glottis larynx, oral cavity, left
and right submandible glands (SMG), as shown in Fig. 2.
The three targets are GTV, primary CTV and CTV for elec-
tive nodal regions (CTV2). Tissue element position variations
in each organ of interest were determined by performing de-
formable image registration between the planning helical CT
and the daily CBCT images using inhouse developed regis-
tration software.11, 12 The deformable image registration was
performed on each CBCT after excluding the rigid body dis-
placement of patient setup.
PCA was applied to the position variations of all tissue ele-
ments for each patient. The OSG was constructed at the end of
the first week of the treatment, and then updated at the end of
each following week during the first five weeks of treatment.
For each patient, the OSG updated on the day τ generated the
rest of T − τ samples of organs of interest for each statistics
test. The sampling estimation accuracy was evaluated using
the average deviation of the sample means and the sample de-
viations between the estimate and the actual observation of all
the individuals.
7333 Nie, Liang, and Yan: Organ sample generator 7333

F IG . 3. First three eigenvectors of the right parotid gland of a patient. It

shows the eigenvectors (arrows) on the mean parotid surface. The first and
third plots are the patient anterior view. The second is the right view.

(EUD) of each organ of interest was then calculated13 with

F IG . 2. Organs at risk in h&n cancer radiotherapy, displayed in order from volume effect factors = 0.5, 0.5, 1, 20, and 4.6 for parotid
the front to back and up to down, are the mandible, the oral cavity, the con- glands, submandible glands, mandible, spinal cord, and brain
strictors, the pharynxes, the parotid and submandible glands, the brain stem,
the cord, and the cochleae.
stem, respectively. For GTV, CTV, and CTV2, the D95 (the
minimum dose to 95% of the volume) was used in the eval-
uation. For the other organs, the minimum dose to 1% of the
The average deviation of the sample means, Emean,τ ,k , for organ volume (D1) was used.
the patient k is defined as
1
2
Emean,τ,k = p̂(t) − p(t) · p̂(t) − p(t) , (11) III. RESULTS
Ne
where p(t) and p̂(t) are the sample means of the actual ob- Figure 3 shows the first three eigenvectors or main princi-
servation and the estimated tissue element positions, respec- ple components of the spatial variation of a right parotid gland
tively. The bracket  ·  indicates the average over all samples obtained using the PCA method in Sec. II.A.1 on all daily
for t > τ . CBCT images of a patient. The left plot indicates a mainly
The average deviation of the sample standard deviations, inferior-to-superior (IS) position variation, and the middle
Estd,τ ,k , for the patient k is defined as one indicates a mainly post-to-anterior (PA) position variation
viewed from patient right. The plot on the right describes the
1 
Ne
patient right-to-left (RL) shift mainly due to the organ shrink-
2
Estd,τ,k = (ŝp,i − sp,i )2 , (12)
Ne i=1 age. Vector on each tissue element has a different length due
to organ deformation. As shown on the RL eigenvector, the in-
where sp,i and ŝp,i are the sample standard deviations of the ferior portion of the parotid had a quicker shrinkage than the
actual observations and the estimated tissue element posi- superior portion during the treatment course. It implies that
tions, respectively. using PCA to model organ spatial variations may not only
Finally, the total errors for all patients, Emean,τ , Estd,τ , are simplify the description of an organ deformation, but also
calculated as the RMS of all individuals’ deviations. provide a means of studying the dose induced morphologi-
In order to perform an unbiased evaluation, the individual cal changes of the organ during the treatment course. Figure 4
OSG was constructed using the two model parameters, the
order of the polynomial L and the forgetting factor λ, deter-
mined by minimizing the average error in Eq. (7) over the
other ten patients. Additionally, the OSG was further vali-
dated by evaluating the discrepancy of the estimated vs the
actually delivered cumulative dose for each organ of interest.
The cumulative dose of each tissue element in organs of inter-
est was calculated using the actual or estimated position sam-
ples, respectively. In order to calculate daily dose distribution
in an observed organ sample, the actual daily CBCT image
after remapping the voxel intensity from the planning helical
CT image based on the deformable registration was used for
the dose calculation. In order to calculate the dose distribu-
tion in an estimated organ sample, a weighted average image
constructed using the observed images was employed for the
calculation. The weights were equal to those used in the OSG F IG . 4. Eigenvalues of different eigenvectors for all 11 patients. The sum of
to generate the organ sample. The equivalent uniform dose all eigenvalues of one patient is equal to 100%.

Medical Physics, Vol. 39, No. 12, December 2012

7334 Nie, Liang, and Yan: Organ sample generator
F IG . 5. The actual and the estimated coefficients of the first three eigenvec-
tors for one patient at the end of second week.
shows the eigenvalues of all eigenvectors obtained using all
daily images of the 11 patients. The PCA was performed us-
ing all tissue elements simultaneously for all 18 organs of in-
terest. As shown in Eq. (2), the eigenvalue is proportional to
the square of the eigenvector coefficient. This indicates that
the geometric variations of organs in h&n cancer radiotherapy
is dominated by a few (the first three or four) eigenvectors.
Figure 5 shows an example of the actual (red) and the es-
timated (green) coefficients of the first three eigenvectors for
one patient at the end of the second week of the treatment. The
third coefficient, which mainly indicates the organ LR posi-
tion displacement, shows a clear time trend due to primarily
the patient tissue shrinkage. In Fig. 6, the actual and the esti-
mated frequency distributions of the first three coefficients for
Medical Physics, Vol. 39, No. 12, December 2012
7334
F IG . 6. The actual PDF of the first three coefficients for one patient, and the
corresponding estimations after the first and the third weeks of the treatment.
the same patient are shown. The estimated coefficient distri-
bution at the end of the third week is closer to the actual one
than the one estimated at the end of the first week. Again, the
distribution of the third coefficient is asymmetric due to tissue
shrinkage during the treatment course.
The average deviations for all patients calculated for the
OSG estimation accuracy on the mean and the standard devi-
ation of the individual PDF are presented in Figs. 7 and 8. The
average deviation in the mean for all 18 organs was ≤ 2 mm
at the end of the first and the second week of the treatment for
most or organs, but slightly > 2 mm for CTV2 and parotids.
It was then further decreased with extra image feedback dur-
ing the later treatment. The estimation error of the standard
deviation for all organswas < 1 mm after the second week of
the treatment.
The mean and the standard deviation of the dose dis-
crepancy between the estimated EUD, D95 or D1 and those
actually delivered are shown in Table I. The planned dose
had the largest error in estimating the treatment delivered
dose. It is comparable to the published results.14 The dose in
parotids and cochlea was underestimated, and had the largest
7335 Nie, Liang, and Yan: Organ sample generator
F IG . 7. The patient averaged error (mm) of the mean of the estimated PDF.
deviations when using the planned dose to evaluate the treat-
ment dose. The improvements of dose estimation obtained
using different number of daily images in the OSG are also
shown in the table.
IV. DISCUSSIONS
The main purpose of this work is to construct a patient spe-
cific OSG that generates random organ samples from a distri-
bution obeying the individual organ variation PDF. The OSG
can be used to construct the expected treatment dose and per-
form adaptive inverse planning optimization during the adap-
tive treatment course. The study demonstrates that utilizing
conventional methods in mathematics and science, the OSG
can be constructed and updated recursively using treatment
daily volumetric images obtained during the treatment course,
and applied in the clinical practice for adaptive radiotherapy.
In addition, the OSG can also be useful for computer simula-
tion studies on patient/organ spatial variation, as well as for
the evaluation of treatment planning and delivery techniques.
PCA (Ref. 8) and time-varying LSR (Ref. 9) are the pri-
mary methods used in the construction of the OSG. The
PCA method can translate a complex problem of a pa-
tient’s anatomical deformation into a relatively simple struc-
ture which includes only few eigenvectors. In addition,
eliminating the nonsignificant eigenvectors which have a
small contribution to the organ spatial variation may poten-
tially reduce the uncertainty caused by the deformable image
F IG . 8. The patient averaged errors (mm) of the standard deviation of the estimated PDF.
Medical Physics, Vol. 39, No. 12, December 2012
7335
registration method. However, this is only a conjecture and
needs to be proven. The time-varying LSR has been the
foundation of most conventional dynamic filter theories. This
method is used to estimate the coefficients of the eigenvectors
in the OSG. The organ variation process in radiation treatment
is most likely dynamic and nonstationary due to the potential
dose response of patient and organs of interest. The complex-
ity of the dynamic variation pattern of each organ of inter-
est may not be described using a fixed function. Therefore,
we have selected the generic polynomial form in the LSR fit-
ting. In addition, the forgetting factor has also been used in
the LSR fitting to handle the time trend of the organ variation.
In principle, both the order of the polynomial and the forget-
ting factor could be determined for each patient during his/her
own treatment course. However, due to the limited number of
daily image samples, the estimation could be unstable. There-
fore, we chose to use the precollected patient daily image data
to determine the generic model parameters. Patient specific
model parameters could provide additional improvement in
the estimation, and will be included in our future studies.
For all subgroups (10 patients in each subgroup) of the
11 h&n patients, the optimized model parameter L, the or-
der of the polynomial, was always 0, and the forgetting factor
λ was very close to a constant of 0.72 with the mean and the
standard deviation λ̄ ± σλ = 0.71 ± 0.04, 0.71 ± 0.05, 0.69
± 0.02, 0.75 ± 0.01, 0.74 ± 0.02 for τ = 1, 2, 3, 4, 5 week, re-
spectively. This implies that the model parameters determined
using the precollected group of data could be time invariant
7336 Nie, Liang, and Yan: Organ sample generator 7336

TABLE I. The average error, μ [%], and deviation, σ [%], of the planned and the estimated treatment dose with respect to the actual delivered dose in each
organ of interest.

Plan Week 1 Week 2 Week 3 Week 4 Week 5

(%) μ σ μ σ μ σ μ σ μ σ μ σ

GTV − 0.3 0.6 − 0.6 1.1 − 0.3 0.6 − 0.1 0.3 0.0 0.2 0.0 0.1
CTV 0.0 0.6 − 0.4 1.4 − 0.1 0.9 − 0.0 0.5 0.1 0.2 0.1 0.1
CTV2 − 0.1 0.4 − 0.4 1.1 − 0.1 0.8 0.0 0.4 0.1 0.2 0.1 0.1
Left parotid − 4.5 5.3 − 2.4 4.9 − 2.3 3.5 − 0.9 2.1 0.0 1.0 0.3 0.4
Right parotid − 5.2 4.4 − 1.8 6.1 − 1.7 3.4 0.5 2.4 0.7 1.5 0.3 0.7
Mandible 0.0 2.1 0.3 1.2 0.1 1.4 0.1 0.6 0.2 0.4 0.1 0.2
Cord 0.3 2.3 0.4 1.4 1.0 1.6 0.3 0.5 − 0.0 0.3 − 0.1 0.3
Brain stem 1.5 3.4 − 0.8 1.7 − 0.1 1.7 0.2 1.4 − 0.4 0.9 − 0.2 0.3
Left cochlea − 12.8 13.5 1.3 5.9 0.1 4.2 − 1.5 1.9 − 0.1 2.2 0.3 1.1
Right cochlea − 15.9 16.8 1.6 3.7 0.4 5.4 0.1 2.1 0.4 2.0 0.5 1.3
Inferior constrictor − 1.1 1.0 − 0.6 0.9 − 0.4 0.7 − 0.3 0.5 − 0.1 0.4 − 0.1 0.1
Middle constrictor − 0.4 1.2 − 0.4 0.7 − 0.4 0.6 − 0.2 0.6 − 0.1 0.3 0.0 0.1
Superior constrictor 0.0 0.6 0.0 0.7 0.1 0.6 0.1 0.5 0.1 0.4 0.0 0.2
Supraglottic larynx − 0.0 1.3 − 0.2 0.9 − 0.0 0.7 − 0.2 0.8 − 0.1 0.3 − 0.0 0.1
Glottic larynx − 0.9 1.0 − 0.4 1.0 − 0.3 0.7 − 0.2 0.4 0.0 0.2 − 0.0 0.1
Oral cavity − 0.0 1.0 − 0.3 0.8 − 0.2 0.6 − 0.2 0.6 0.1 0.2 0.0 0.2
Left SMG − 1.0 1.4 − 0.3 0.7 − 0.3 1.2 − 0.1 0.8 0.2 0.3 0.1 0.1
Right SMG − 0.1 0.6 − 0.1 0.5 0.0 0.7 − 0.1 0.4 0.1 0.3 0.1 0.1

during the entire treatment course. In this case, the mean coef-
ficients were estimated most likely using the running average
with different weights from the daily observations. Of course,
these two specific model parameters were obtained from the
group of patients, instead of the patient specific. The patient
specific model parameters should be ideal, which needs to be
explored in future studies.
The evaluation results were quite promising. The organ
samples generated by the OSG showed a variation distri-
bution, and time trend similar to the actual organ varia-
tion. The results also demonstrated that the estimation was
continuously improved during the patient treatment course
with the additional treatment feedback images. The aver-
age deviation in the estimation of the individual PDF mean
was largely improved (<2 mm) for all organs after the
third week of the treatment. Meanwhile, the estimation er-
ror in the individual PDF standard deviation of all organswas
<1 mm. This implies that the OSG can produce the random
organ samples with a similar organ variation PDF as the ac-
tual measured one for each individual during the treatment
course.
It has been studied1 that the mean and the standard devi-
ation of an organ variation PDF are the dominating factors
in the quantification of the delivered dose in organs of inter-
est. Using the mean and the standard deviation of tissue el-
ement position alone without full knowledge of the PDF for
the treatment dose construction in an organ of interest will
give a reasonably good approximation on the treatment dose
distribution. Compared to the pretreatment planning dose, the
cally for the parotids and the cochlea. After the first 3 weeks
of the treatment, the dose estimation accuracy was within 1%
on the mean discrepancy for most of the organs except for the
left cochleae (1.5%). The corresponding standard deviation
was within 2% for parotids, the brain stem and the cochleae,
and within 1% for all other organs. This convergence is, of
course, greatly helpful for the adaptive treatment process with
the multiple treatment modification strategy.
In summary, the OSG with the patient specific organ
variation PDF is introduced for expected treatment dose
construction and adaptive inverse planning optimization. The
accuracy of the OSG can be improved continuously and re-
cursively during the treatment course using daily volumetric
image feedback. The OSG was evaluated using daily CBCT
images obtained during h&n cancer treatment. Both the
geometric and dosimetric results demonstrate the feasibility
of applying the OSG in adaptive radiotherapy process.
ACKNOWLEDGMENTS
This study is partially supported by the research grant from
Elekta Oncology Systems Co. None of the authors has a direct
or indirect financial incentive associated with the study and
publication of this paper.
1 D. Yan and D. Lockman, “Organ/patient geometric variation in external
beam radiotherapy and its effect,” Med. Phys. 28, 593–602 (2001).
2 D. Yan, “Image-guided/adaptive radiotherapy,” in Medical Radiology-

expected treatment dose estimated using the OSG is a clear Radiation Oncology, Volume: New Technologies in Radiation Oncology,
edited by W. Schlegel, T. Bortfeld, and A. L. Grosu (Springer-Verlag,
improvement. When the daily images obtained from the first Berlin, 2005), Chap. 4.5.
week of the h&n treatment were applied for the OSG, the dis- 3 D. A. Low et al., “Novel breathing motion model for radiotherapy,” Int. J.

crepancy in the dose estimation was largely reduced, specifi- Radiat. Oncol. Biol. Phys. 63, 921–929 (2005).

Medical Physics, Vol. 39, No. 12, December 2012

7337 Nie, Liang, and Yan: Organ sample generator
4 Q. Zhang et al., “A patient-specific respiratory model of anatomical motion
for radiation treatment planning,” Med. Phys. 34, 4772–4781 (2007).
5 D. Yan et al., “An off-line strategy for constructing a patient specific plan-
ning target volume in adaptive treatment process for prostate cancer,” Int.
J. Radiat. Oncol. Biol. Phys. 48, 289–302 (2000).
6 J. Barker et al., “Quantification of volumetric and geometric changes oc-
curring during fractionated radiotherapy for head-and-neck cancer using an
integrated CT/linear accelerator system,” Int. J. Radiat. Oncol. Biol. Phys.
59, 960–970 (2004).
7 R. M. Seibert, C. R. Ramsey, J. W. Hines, P. A. Kupelian, K. M. Langen,
S. L. Meeks, and D. D. Scaperoth, “A model for predicting lung cancer
response to therapy,” Int. J. Radiat. Oncol. Biol. Phys. 67, 601–609 (2007).
8 I. T. Jolliffe, Principal Component Analysis (Springer, New York, 2002).
9 K. J. Åström and B. Wittenmark, Adaptive Control, 2nd ed. (Dover, New
York, 2008).
Medical Physics, Vol. 39, No. 12, December 2012
7337
10 M. Sohn, M. Birkner, D. Yan, and M. Alber, “Modeling individual
geometric variation based on dominant eigenmodes of organ deforma-
tion: Implementation and evaluation,” Phys. Med. Biol. 50, 5893–908
(2005).
11 W. Lu et al., “Fast free-form deformable registration via calculas of varia-
tions,” Phys. Med. Biol. 49, 3067–3087 (2004).
12 T. Zhang, Y. Chi, E. Meldolesi, and D. Yan, “Automatic delineation of
online head and neck CT images: Towards online adaptive radiotherapy,”
Int. J. Radiat. Oncol. Biol. Phys. 68, 522–530 (2007).
13 A. Niemierko, “A generalized concept of equivalent uniform dose (EUD),”
Med. Phys. 26(6), 1100 (1999).
14 Q. Wu, Y. C. Chi, P. Y. Chen, D. J. Krauss, D. Yan, and A.
Martinez, “Adaptive replanning strategies accounting for shrinkage in
head and neck IMRT,” Int. J. Radiat. Oncol. Biol. Phys. 75, 924–932
(2009).