You are on page 1of 13

Reconstruction of 3D lung models from 2D planning data sets

for Hodgkin’s lymphoma patients using combined deformable
image registration and navigator channels
Angela Ng
Radiation Medicine Program, Princess Margaret Hospital, University Health Network, Toronto,
Ontario M5G 2M9, Canada
Thao-Nguyen Nguyen
Department of Medical Biophysics, University of Toronto, Toronto, Ontario M5G 2M9, Canada
Joanne L. Moseley
Radiation Medicine Program, Princess Margaret Hospital, University Health Network, Toronto,
Ontario M5G 2M9, Canada
David C. Hodgson and Michael B. Sharpe
Radiation Medicine Program, Princess Margaret Hospital, University Health Network, Toronto,
Ontario M5G 2M9, Canada and Department of Radiation Oncology, University of Toronto, Toronto,
Ontario M5G 2M9, Canada
Kristy K. Brock

Radiation Medicine Program, Princess Margaret Hospital, University Health Network, Toronto,
Ontario M5G 2M9, Canada; Department of Medical Biophysics, University of Toronto, Toronto,
Ontario M5G 2M9, Canada; and Department of Radiation Oncology, University of Toronto, Toronto,
Ontario M5G 2M9, Canada
͑Received 10 July 2009; revised 13 November 2009; accepted for publication 8 December 2009;
published 8 February 2010͒
Purpose: Late complications ͑cardiac toxicities, secondary lung, and breast cancer͒ remain a sig-
nificant concern in the radiation treatment of Hodgkin’s lymphoma ͑HL͒. To address this issue,
predictive dose-risk models could potentially be used to estimate radiotherapy-related late toxici-
ties. This study investigates the use of deformable image registration ͑DIR͒ and navigator channels
͑NCs͒ to reconstruct 3D lung models from 2D radiographic planning images, in order to retrospec-
tively calculate the treatment dose exposure to HL patients treated with 2D planning, which are now
experiencing late effects.
Methods: Three-dimensional planning CT images of 52 current HL patients were acquired. 12
image sets were used to construct a male and a female population lung model. 23 “Reference”
images were used to generate lung deformation adaptation templates, constructed by deforming the
population model into each patient-specific lung geometry using a biomechanical-based DIR algo-
rithm, MORFEUS. 17 “Test” patients were used to test the accuracy of the reconstruction technique
by adapting existing templates using 2D digitally reconstructed radiographs. The adaptation process
included three steps. First, a Reference patient was matched to a Test patient by thorax measure-
ments. Second, four NCs ͑small regions of interest͒ were placed on the lung boundary to calculate
1D differences in lung edges. Third, the Reference lung model was adapted to the Test patient’s
lung using the 1D edge differences. The Reference-adapted Test model was then compared to the
3D lung contours of the actual Test patient by computing their percentage volume overlap ͑POL͒
and Dice coefficient.
Results: The average percentage overlapping volumes and Dice coefficient expressed as a percent-
age between the adapted and actual Test models were found to be 89.2Ϯ3.9% ͑Right lung
=88.8%; Left lung=89.6%͒ and 89.3Ϯ2.7% ͑Right =88.5%; Left =90.2%͒, respectively. Paired
T-tests demonstrated that the volumetric reconstruction method made a statistically significant
improvement to the population lung model shape ͑pϽ0.05͒. The error in the results were also
comparable to the volume overlap difference observed between inhale and exhale lung volumes
during free-breathing respiratory motion ͑POL: p=0.43; Dice: p=0.20͒, which implies that the
accuracies of the reconstruction method are within breathing constraints and would not be the
confining factor in estimating normal tissue dose exposure.
Conclusions: The result findings show that the DIR-NC technique can achieve a high degree of
reconstruction accuracy, and could be useful in approximating 3D dosimetric representations of
historical 2D treatment. In turn, this could provide a better understanding of the biophysical rela-
tionship between dose-volume exposure and late term radiotherapy effects. © 2010 American
Association of Physicists in Medicine. ͓DOI: 10.1118/1.3284368͔
1017 1017 Med. Phys. 37 „3…, March 2010 0094-2405/2010/37„3…/1017/12/$30.00 © 2010 Am. Assoc. Phys. Med.
Key words: 3D lung model reconstruction, Hodgkin lymphoma, deformable image registration,
navigator channels
I. INTRODUCTION
Continuous improvements in the treatment of Hodgkin’s
lymphoma ͑HL͒ have significantly improved patients’ 5 yr
survival rates.
1,2
Late complications such as cardiac morbid-
ity and mortality, as well as secondary malignancies to the
lungs and breasts, occur at higher levels in HL survivors than
in the general population, 10–20 yr after their radiotherapy
͑RT͒ treatment.
1,2
Predictive RT-related dose-risk models of critical organs
could potentially estimate the risk of late effects. However,
the major challenge for developing and validating these
models is obtaining data about dose-volume exposures of
normal tissues, and the corresponding late toxicities. Patients
treated over 10 yr ago, for whom late outcome data are avail-
able, were planned with two-dimensional ͑2D͒ imaging ͑e.g.,
fluoroscopy͒. Consequently, three-dimensional ͑3D͒ dose-
volume exposures are unavailable for these patients. Due to
this limitation, previous studies into the late effects of RT HL
treatment have mostly been focused on the overall field
size,
3,4
treatment dose, and fractionation regimes.
3,5
Con-
versely, 3D CT-based RT planning is a relatively new inno-
vation, such that there is currently insufficient sample of late
complication patients with 3D-treatment plans. Numerous re-
cent studies on conformal RT have compared dosimetric out-
comes of 2D and 3D treatment plans by reconstructing 2D
plans from existing 3D treatment plans for various head and
neck,
6
prostate,
7,8
and gynecological
9
malignancies. How-
ever, to date, no study has attempted to construct 3D organ
volumes directly from historical 2D planning images, espe-
cially not for the treatment of HL.
In order to extract accurate patient anatomical information
from medical images, several investigations that have been
focused on the potentials of a finite element modeling de-
formable image registration ͑DIR͒ system. For example, in
surgical simulation, finite-element modeling ͑FEM͒-based
deformation has been used to model the heart
10
and
brain.
11,12
In oncology and radiotherapy, research has fo-
cused on how DIR can be used to improve treatment plan-
ning, delivery, and follow-up processes.
13–17
Particularly,
FEM-based DIR has been used to construct a four-
dimensional liver model to represent the temporal changes in
organ shape during a patient’s breathing cycle, in order to
calculate delivered dose to liver tumors during treatment.
18,19
The accuracy of the FEM deformable registration system,
when using a biomechanical-based platform ͑MORFEUS͒,
has been shown to be within the order of an image voxel size
of 0.25 cm.
20
DIR is a powerful tool for understanding the complex
motion, deformation, and response of soft tissue; however,
its use, as previously reported, requires a 3D volumetric im-
age of the patient and time-points of interest. This data, al-
though desirable, is not always available for patients. For
example, intrafraction motion, such as breathing, is often im-
aged using 2D imaging sequences ͑e.g., fluoroscopy and cine
MR͒. In attempts to develop an advanced approach for image
analysis using limited data, a navigator channel ͑NC͒ adap-
tation technique has recently been developed to model
patient-specific liver motion.
21
NCs were applied on 2D
coronal CT images to detect patient-specific motion of the
liver, and to adapt the motion changes to a population liver
motion model.
21
The absolute average accuracy of the NC to
predict liver motion was less than 0.11 cm.
21
The main objective of this study is to develop and evalu-
ate a novel technique that combines the use of DIR and NC
for regenerating 3D lung volumes from 2D RT planning im-
ages, as a step toward reconstructing the 3D dose-volume
exposure of HL patients for whom late outcome data is avail-
able. This technique utilizes NCs to extract patient-specific
structural lung information from 2D images, and uses this
information to adapt a population lung model that was cre-
ated from a biomechanical-based DIR system, MORFEUS,
20
as described in Nguyen et al.
21,22
Geometric validation of the
technique was performed by comparing the regenerated lung
models with the corresponding 3D lung contours from CT
using a volume overlap ͑VOL͒ test. As a benchmark evalu-
ation of geometric accuracy, the results were compared to the
volumetric changes typically observed in a patient’s breath-
ing cycle. This comparison was made because HL patients
treated with conventional RT were imaged and treated with
free breathing. This novel technique could offer an alterna-
tive to retrospectively perform 3D dosimetry on 2D plans,
for correlation with late toxicities in patients’ long-term
follow-ups. As a result, a better understanding of the dose-
volume association of various organs, as well as the relation-
ship between radiotherapy and various late effects, can be
established.
II. MATERIALS AND METHODS
II.A. Process overview
The technique combines the use of a biomechanical
model-based DIR system with NCs to generate 3D lung vol-
umes from 2D radiotherapy planning images. To evaluate the
accuracy of this process, patients with 3D CT images will be
used. The 3D CT images will be used to generate 2D DRRs,
similar to the 2D fluoroscopy images available from previ-
ously treated patients. The information from the 3D images
will then be available to evaluate the accuracy of the recon-
struction of the lung volume, however, no 3D data will be
used in the model construction, as this will not be available
for the retrospective patients.
The overall process involves generating a population lung
model, and adapting the model to become patient-specific
using information in the patient’s 2D radiographs. The pro-
cess involves ͑1͒ matching a Reference patient to a Test pa-
tient using lung measurements on DRRs, ͑2͒ using NCs on
matching patient’s DRRs to automatically detect lung differ-
1018 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1018
Medical Physics, Vol. 37, No. 3, March 2010
ences between Reference and Test patients, and ͑3͒ adapting
the population-Reference deformation model using the NC
difference to acquire the population-Test deformation to gen-
erate the 3D lung volume of the Test patient’s lungs. This
process was tested on currently treated HL patients who were
planned with a 3D CT-based treatment planning system. The
general workflow is shown in Fig. 1, and described in detail
below.
II.B. Patient and image data
3D CT images from 52 mediastinal HL patients ͑27 male,
25 female͒ treated between 1999 and 2008 were obtained.
For planning, a thoracic CT scan ͑GE Lightspeed, GE Medi-
cal Systems, Milwaukee, WI͒ was acquired ͑kV
=120; mA s=400͒ for each patient in the supine position in
an immobilization bag, in an arms-akimbo position,
23
under
free breathing conditions. In the arms-akimbo position, the
patient’s hands were on the hips and the elbows were bowed
outward.
23
The voxel size of the CT images was 0.1ϫ0.1
ϫ0.3 cm
3
. Each patient was planned for radiotherapy with
the 3D CT-based treatment planning system ͑PINNACLE3
v7.6, Philips Radiation Oncology Systems, Madison, WI͒.
From each plan, 2D DRRs were reconstructed retrospec-
tively from the CT images at an optimized resolution of
512ϫ512 pixels, a step size of 0.05 cm, and a source-to-
axis distance of 100 cm. A graticule was included on the
DRRs for image scaling. These DRRs were used to simulate
2D planning films used in historical HL treatments for vali-
dation.
In summary, a total of 52 3D patient images were ob-
tained. 12 image sets were used to generate the male and
female population model ͑six female, six male͒. 23 images
͑Reference patients͒ were used to generate the lung deforma-
tion template, constructed by deforming the population
model into each patient-specific lung geometry. An addi-
tional 17 patients ͑Test patients͒ were used to independently
test the accuracy of the adaptation of the population lung
model using NCs and 2D images only.
II.C. Population deformation models
A male and a female population lung model were gener-
ated using the process described in Nguyen et al.,
21
with the
modifications discussed below. Out of the 52 total HL pa-
tients, 12 “Base” patients ͑six male, six female͒ were ran-
domly selected to formulate the basis of the population lung
models. Left and right lung contours generated on the Base
patient’s planning CT were exported as binary masks from
the planning system. The mask files from each patient were
summated using a logical OR function, and converted into a
3D tetrahedral finite element volume mesh using the INTER-
ACTIVE DATA LANGUAGE ͑IDL͒ software ͑IDL v6.3, ITT,
Boulder, CO͒. Using an in-house biomechanical-based DIR
platform, MORFEUS, the meshes were smoothed ͑using a
Laplacian smoothing function͒ to form a FEM-based popu-
lation lung model. Components of MORFEUS are described
in detail in Brock et al. The population models encompassed
all original mask files and served as an initial template to
account for the general lung shape and volume.
21
Using MORFEUS, the male or female population model
was deformed into each of the 35 patients’ ͑12 Base that
made up the population model, 23 additional Reference not
used in construction of the population model͒ lungs to de-
velop patient-specific 3D lung models and deformation
maps, which describe the relative changes in the lung vol-
ume and structure from the population to each patient spe-
cific lung. The two groups of patients ͑Base and Reference͒
were used in this investigation as a proof of concept. Since
for application in historical clinical cases, volumetric infor-
mation would not be available to construct FEM models, a
sample population ͑Base patients͒ was used to construct the
population model, and another sample to construct deforma-
tion maps ͑Reference patients͒. These deformation maps
were then used to build 3D models for retrospective analysis
of patients with radiographic images only. The deformations
were carried out using the biomechanical properties of
Young’s modulus ͑5.0 kPa͒ and Poisson’s ratio ͑0.45͒.
20
The
nodes on the surface of the FEM of the reference lung were
deformed, using guided surface projection, onto a surface
defined by the secondary lung. The internal structure of the
lung was deformed according to the biomechanical proper-
ties assigned to the lung model. The larger population model,
which encompassed all original patient’s lungs allowed for
FIG. 1. Overall process of generating 3D models from 2D images. ͑a͒ Popu-
lation deformations, ͑b͒ three-step adaptation, and ͑c͒ evaluation of models.
1019 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1019
Medical Physics, Vol. 37, No. 3, March 2010
nondistorted deformation of the population to patient specific
lungs. The animation of each deformation was visually veri-
fied using the HYPERMESH software ͑HYPERMESH v9.0, Altair
Engineering, Troy, MI͒ to ensure accurate deformation.
In brief, to determine the Test patient’s 3D lung models
from the population model, deformation was done from the
3D population FEM model to each Reference 3D FEM
model. Using MORFEUS, deformation maps were generated
for each population to Reference patient 3D lung models.
These population-Reference 3D deformation FEM models
and maps were then adapted using the adaptation steps de-
scribed in Secs. II D–II F to determine Test specific maps
and 3D lung models.
II.D. Adaptation step 1: Matching of Reference and
Test patients using thorax measurements
To adapt the population lung model to match the Test
patient-specific lung geometry, the first step was to match
each Test patient to a similar Reference patient. The match
was based on similar thorax measurements taken on the pa-
tient’s coronal DRR. Thorax measurements were utilized to
help identify patients with similar lung sizes because first, by
necessity, lungs must fit within the confines of the thoracic
cage. Second, gender studies have found that the smaller
female lung volume is largely attributable to their smaller
ribcage dimensions.
24,25
Prior to taking the measurements,
DRRs were first distance scaled ͑reconstructed at SAD 100
cm͒, and then measured using the ruler tool in the treatment
planning system. Figure 2 illustrates the three measurements
taken include: ͑1͒ Width of superior thorax at vertebral body
level T2/ T3 ͑W1͒, ͑2͒ width of inferior thorax at T8/ T9
͑W2͒, and ͑3͒ length of thorax from T2/ T3 to T8/ T9 ͑L͒. 2D
measurements were taken directly on anterior DRRs to simu-
late the 2D planning conditions of historical HL treatments,
which utilized anterior and posterior chest radiographs. Lung
depth measurements were not taken because lateral planning
radiographs were not available in historical treatment plans.
In order to identify matching Reference and Test pairs
͑RTPs͒, thorax measurements within the same gender were
compared. This was performed by first grouping Test and
Reference patients of the same gender together. Second,
within each gender group, the least square difference ͑LSD͒
between every Test patient and each Reference patient was
calculated based on the three measurements ͑W1, W2, and L͒
as follows:
LSD=
ͱ
͑W1
test
− W1
ref
͒
2
+ ͑W2
test
− W2
ref
͒
2
+ ͑L
test
− L
ref
͒
2
.
͑1͒
A male or a female thorax similarity matrix was then con-
structed from the LSD scores. Based on the similarity matri-
ces, the lowest LSD score ͑i.e., highest similarity͒ for each
Test patient depending on gender was identified and used for
selecting its matching Reference patient to make up an RTP.
Figure 3 illustrates how the Reference patient for the first
RTP was selected. Tables I and II summarize the LSD scores
of the 17 RTPs. The two similarity matrices were gender-
based. LSD scores were the only parameters used in this first
TABLE I. Summary of LSD scores for the nine male RTPs, consisting of their ͑i͒ thorax width ͑W1, W2͒, ͑ii͒
thorax length ͑L͒ as measured ͑cm͒ on DRRs and their associated ͑iii͒ LSD scores.
Reference subjects Test subjects
RTP W1 W2 L W1 W2 L LSD
M1 15.0 24.8 14.3 15.1 24.1 13.9 0.81
M2 15.0 24.8 14.3 14.7 25.3 14.4 0.59
M3 15.0 24.8 14.3 15.6 24.9 14.4 0.62
M4 15.0 24.8 14.3 15.2 25.2 14.6 0.54
M5 15.0 24.8 14.3 15.3 25.0 14.0 0.47
M6 15.0 24.8 14.3 14.6 24.9 13.8 0.65
M7 15.1 27.2 16.5 15.3 27 15.6 0.94
M8 15.1 27.2 16.5 14.1 26.7 16.2 1.16
M9 15.1 27.2 16.5 14.3 26.9 15.3 1.47
FIG. 2. Matching of Reference and Test patients using thorax measurements
taken on DRR. The three measurements taken were ͑1͒ width of thorax at
T2/ T3 ͑W1͒, ͑2͒ width of thorax at T8/ T9 ͑W2͒, and ͑3͒ length of thorax
from T2/ T3 to T8/ T9 ͑L͒.
FIG. 3. Identification of a matching Reference patient ͑Ref. 3͒ for Test pa-
tient 1 based on the lowest LSD score calculated from the three thorax
measurements ͑W1, W2, and L͒.
1020 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1020
Medical Physics, Vol. 37, No. 3, March 2010
level of adaptation. For clinical application, if available,
other parameters such as patient’s age,
26,27
and or patient’s
weight ͑e.g., calculation of lean body mass͒
28
taken prior to
conventional simulation planning could also be investigated
for their use to identify similar lung volumes. The selected
Reference patient was further refined through subsequent ad-
aptation steps as described in the following sections.
II.E. Adaptation step 2: Quantification of structural
differences/shifts using navigator channels
To adapt each population-Reference deformation model to
match the Test patient, structural differences between the
Reference and Test patient’s lungs were quantified using NCs
on patient DRRs. A NC is a rectangular region of interest on
an image that captures the intensity values within the region
and calculates the difference in intensity gradients between
corresponding pairs of NCs.
21,22
To quantify edge differ-
ences, comparable DRRs of each RTP were first scaled by
graticule, then spatially defined by its crosshair. Four NCs
were placed at the most superior, inferior, left, and right
edges of the left and right lungs as shown on Fig. 1.
The vertical NCs ͑superior and inferior edges͒ were
0.5ϫ4.0 cm
2
͑widthϫlength͒, and the horizontal NCs ͑left
and right edges͒ were 2.0ϫ0.5 cm
2
in dimension. Each NC
was manually placed on the lung boundary of the Reference
image. A corresponding NC was then automatically placed
on the Test image at the same geometric position. The image
intensity within the NC was converted into a 1D intensity
plot function by averaging over the narrow dimension. Using
a second derivative calculation, the difference or shifts in
intensity plots of each pair of NCs were calculated based on
their points of inflection ͑POIs͒. Quantification of edge dif-
ferences was calculated as shifts in their intensity edge gra-
dients. Mathematically, the POI occurs where the rate of
change in the intensity values is the greatest, thus indicating
the edge of the organ. This technique reduces the effects of
noise and allows for comparison between gradient sizes due
to differences in contrast. The POI allows for finding the
location of the edge of the lung, which is of primary interest.
Figure 4 illustrates the process involved in this adaptation
step for one of the NCs in the superior edge position.
II.F. Adaptation step 3: Adaptation of population lung
model
Subsequently, Test patient lung volumes were recon-
structed through a series of adaptation calculations that in-
corporated the NC-detected shifts to adapt the Reference
lung volume. The deformation at each node of the model ץ
͑x
0
, y
0
͒ was adjusted using the following linearly weighted
equations to account for lateral ͓Eq. ͑2͔͒ and longitudinal
͓Eq. ͑3͔͒ adjustments.
ץ͑x
o
͒ = ⌬͑x
o
͒
ͫͩ

NC1
+ ץ
NC1

NC1
ͪ͑k
x
o
͒ + ͩ

NC2
+ ץ
NC2

NC2
ͪ͑1−k
x
o
͒
ͬ
,
͑2͒
ץ͑y
o
͒ = ⌬͑y
o
͒
ͫͩ

NC3
+ ץ
NC3

NC3
ͪ͑k
y
o
͒+ͩ

NC4
+ ץ
NC4

NC4
ͪ͑1−k
y
o
͒
ͬ
.
͑3͒
In Eq. ͑2͒, the deformation from population to Test patient
displacement ץ ͑x
0
͒ is predicted to be equal to the deforma-
tion from population to Reference patient at the same node,
⌬͑x
0
͒. This is adjusted by a linear weighting of the difference
in the lung edge between the Reference and Test patient
DRRs determined at each NCs, ץNC
1
͑right edge͒, and ץNC
2
,
͑left edge͒, against the predicted displacement of the popu-
lation to Reference patient deformation map at the NC node
positions ⌬NC
1
and ⌬NC
2
, respectively.
21,22
This is further
weighted by the relative distance of the adapting node to the
two NC positions, k and ͑1-k͒, where k varies between 0 and
1 ͑where 0 is the most right and superior edge of lung and 1
is the most left and inferior edge͒.
21,22
Equation ͑3͒ follows
the same logic, where NC
3
and NC
4
represent the superior
and inferior NC positions, respectively. The two equations
were applied iteratively to all the nodes of the Reference
deformation model to regenerate lung volumes for each of
the Test patients.
II.G. Evaluation of adapted models using volume
overlap tests
The NC-adapted models describe the 3D representations
of the 17 Test patient’s lungs and were evaluated by compar-
TABLE II. Summary of LSD scores for the eight female RTPs, consisting of their ͑i͒ thorax width ͑W1, W2͒, ͑ii͒
thorax length ͑L͒ as measured ͑cm͒ on DRRs and their associated ͑iii͒ LSD scores.
Reference subjects Test subjects
RTP W1 W2 L W1 W2 L LSD
F1 14.6 24.9 13.8 15.1 24.1 13.9 0.95
F2 14.6 24.9 13.8 14.7 25.3 14.4 0.73
F3 14.6 24.9 13.8 14.9 24.0 13.5 0.99
F4 14 26.1 16 14.1 26.7 16.2 0.64
F5 14 26.1 16 14.7 26 16.2 0.73
F6 14 26.1 16 14.3 26.9 15.3 1.10
F7 15.1 27.2 16.5 15.8 26.8 16.9 0.90
F8 15.1 27.2 16.5 15 27.2 15.7 0.81
1021 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1021
Medical Physics, Vol. 37, No. 3, March 2010
ing each adapted volume with the corresponding 3D lung
contours generated from planning CT. Comparing volumes
were aligned by their center of mass ͑COM͒, and their VOL
was computed. COM was applied because it helps rigidly
align the models in order to compare them volumetrically. In
reality, when 3D models are reconstructed retrospectively
from historical 2D planning films, COM would not be nec-
essary, as the reconstructed treatment beam would be set
relative to the anatomy on the Reference 3D images. The
VOL test consisted of two calculations. The first calculation
determined their percentage volume overlap ͑POL͒ with re-
spect to the single volume of interest ͑i.e., actual Test vol-
ume͒ as shown in Eq. ͑3͒. The second calculation ͑Dice
coefficient͒,
29,30
more commonly used for evaluating the
quality of a given segmentation, compared the overlapping
volume with respect to both comparing volumes as shown in
Eq. ͑4͒. In this paper, the Dice values were expressed as
percentages.
POL =
V
1
പV
2
V
2
͑3͒ Dice =
2͑V
1
പV
2
͒
V
1
+ V
2
ϫ100. ͑4͒
VOL tests were performed between the NC-adapted Test
͑V
1
͒ and actual Test ͑V
2
͒ volumes. For comparison, the VOL
tests were also performed between the original Reference
͑No NC adaptation͒ and Test volumes. The Test volumes
were based on contours generated from their CT images.
These results were compared using a Student’s paired T-test.
II.H. Evaluation of volume overlap results
For validation of the volume reconstruction method accu-
racy, the results were compared to the volumetric changes
observed during patient free-breathing cycle. This was per-
formed by evaluating the 4D-CT images of 12 former lung
cancer patients. At the time of planning, respiratory-
synchronized planning CT images ͑4D-CT; kV=120,
mA s=800͒ were acquired in cine mode, using a four-slice
fan beam CT scanner ͑Discovery LS, GE, Waukesha, WI͒.
CT images corresponding to the patient’s end inhale and ex-
hale phases were transferred to the treatment planning sys-
tem. Sets of inhale and exhale CT images were used to assess
the maximum variation in lung volume and position during a
typical patient’s breathing cycle. As described in Sec. II C,
lung contours generated on patients’ planning inhale and ex-
hale CT images were exported as binary masks from their
treatment plans, and used for VOL calculations, where V
1
and V
2
represent inhale and exhale lung volumes, respec-
tively. Since HL patients were under normal breathing during
RT, their lung volumes can vary substantially during treat-
ment, due to normal breathing motions. These results cap-
tured the typical variation in overlapping and nonoverlapping
lung volumes seen in patients’ breathing cycle, providing a
frame of reference to validate the accuracy of the technique.
The VOL results were evaluated using a Student’s T-test
͑two-sample unequal variances͒, performed between the NC-
adapted Test and free-breathing inhale-exhale lung VOL re-
sults.
FIG. 4. Quantification of lung edge differences ͑shift͒ ͑cm͒ between the Reference and Test lungs at the superior edge position. The NCs captured the image
intensity within the region and converted it into a 1D intensity plot. Intensity plots were aligned by their points of inflection ͑green͒ to calculate the shifts
between them.
1022 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1022
Medical Physics, Vol. 37, No. 3, March 2010
III. RESULTS
III.A. Population deformations
The deformation between the population and the Base or
Reference lung models were verified by visual inspection of
the models before and after the deformation. The population
lung model ͑mesh͒ and one of the Base patients’ lung model
͑solid on 5͑a͒ or 5͑b͒͒ are shown before and after the defor-
mation in Figs. 5͑a͒ and 5͑b͒, respectively. Similarly, Figs.
5͑c͒ and 5͑d͒ illustrate the deformation between the same
population model ͑mesh͒ and one of the Reference models
͑solid on 5͑c͒ or 5͑d͒͒. Qualitative assessment of the surface
deformations showed good volumetric and spatial agree-
ments between population and each of the respective Base or
Reference models.
III.B. Similarity matrix
A male or a female similarity matrix built from thorax
measurements ͑W1, W2, and L͒ taken from patients’ DRRs
was used to calculate LSD scores between patients to find
similar pairs of lungs within the same gender. Using the LSD
scores, a total of 17 RTPs were identified. The 17 RTPs, the
thorax measurements, and associated LSD scores are shown
on Table I ͑nine male sets͒ and Table II ͑eight female sets͒.
These five Reference subjects were selected, because they
often produce the lowest LSD scores with the Test subjects.
The LSD scores of the selected pairs ranged from 0.4 to 1.5
cm.
III.C. Evaluation of adapted models using VOL tests
A series of VOL tests were used to compare the NC-
adapted Test ͑adapted͒ or original Reference ͑not adapted͒
with the actual Test lung volumes from CT. The VOL results
for both lungs are summarized in Table III. Both measures of
POL and Dice show similar volume overlap results. Com-
parisons of the distributions of POL with and without adap-
tation for the 17 RTPs are shown in Fig. 6. As predicted, it
was found that generally, measures of overlap were higher
with than without NC adaptation ͑15/17 cases—88%͒. Dif-
ferences in model volumes can be demonstrated in graphical
representations. Figure 6 illustrates the overlapping volumes
between the three models ͑original Reference: Blue mesh,
NC-adapted Test: Black mesh, actual Test: Red opaque͒ for a
specific RTP ͑RTP 16͒, as captured by the NCs on the Ref-
erence or Test DRR. Noticeable NC adaptation could be seen
laterally ͑x͒ and longitudinally ͑y͒ between the blue and
black meshes in Figs. 7͑a͒ and 7͑b͒. Slight adaptation could
be seen in the vertical direction ͑z͒ in Figs. 6͑c͒ and 6͑d͒,
FIG. 5. Visual inspection of the spatial/volumetric agreement between popu-
lation ͑mesh͒ and one of the base ͑solid on aϩb͒ or reference ͑solid on cϩd͒
lung models before ͓͑a͒ and ͑c͔͒ and after ͓͑b and d͔͒ deformation.
TABLE III. Summarized POL and Dice results from the VOL tests between the reference or NC-adapted
reference and test volumes, as well as between free-breathing ͑FB͒ inhale and exhale lung volumes.
VOL Tests Ave. Min. Max. SD
POL Reference ͑not NC adapted͒ and test 84.3 63.6 94.8 9.5
NC-adapted reference and test 89.2 79.9 94.2 3.9
FB inhale and exhale lungs 90.2 85.7 96.5 3.0
Dice Reference and test 85.4 75.6 94.4 5.4
NC-adapted reference and test 89.3 83.8 94.0 2.7
FB inhale and exhale lungs 90.5 86.0 96.9 3.1
FIG. 6. Distribution of POL with and without NC adaptation. POL was
generally higher with than without adaptation. Exceptions were seen in
RTPs 7 and 8.
1023 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1023
Medical Physics, Vol. 37, No. 3, March 2010
particularly for the left lung. Exceptions to the above predic-
tion were seen in two RTP cases ͑RTPs 7 and 8͒. For these
cases, the difference in POL and Dice was approximately 1%
͓RTP 7: POL 90.9% ͑adapt͒ vs 91.9% ͑no adapt͒; Dice
90.8% ͑adapt͒ vs 91.8% ͑no adapt͒; RTP 8: POL 90.8%
͑adapt͒ vs 91.8% ͑no adapt͒; Dice 90.9% ͑adapt͒ vs 91.9%
͑no adapt͔͒. Figure 8 also illustrates lung models for one of
the exceptional cases ͑RTP 7͒. A visual inspection of their
nonoverlapping regions shows that the Reference and Test
volumes were very similar in lateral and longitudinal dimen-
sions. Volumetric mismatch was mainly located in the
anterior-posterior directions of the lungs. Table IV shows the
results of the paired T-tests between the adapted and non-
adapted models. The NC technique made significant changes
to the shape of the Reference model. Overall, there were
significant differences between the NC-adapted Test and
original Reference ͑without adaptation͒ lung volumes
͑pϽ0.05 for POL and Dice͒.
III.D. Evaluation of volume overlap results
HL patients undergoing radiotherapy treatment are nor-
mally treated under free-breathing. To provide insight into
how the accuracy of the volume reconstruction processes de-
scribed here compare to the inherent effect of breathing mo-
tion, VOL tests were performed between randomly selected
exhale and inhale lung volumes of 12 lung patients. This is a
comparative evaluation of the VOL results between NC-
adapted and actual Test lung volumes. Both measures of
POL and Dice show similar evaluation results. Figure 9 com-
FIG. 7. Visual inspection of the original reference ͑blue mesh͒, NC-adapted
test ͑black mesh͒ and actual test ͑red opaque͒ lung models for the RTP 16.
Figures 5͑a͒ and 5͑b͒ illustrate adaptation in the lateral ͑x͒ and longitudinal
͑y͒ direction. Figures 5͑c͒ and 5͑d͒ show slight adaptation in the vertical ͑z͒
direction. Ideally, the NC-adapted reference and test models should match.
FIG. 8. Visual inspection of one of the two exception cases ͑RTP 7͒ shows
that the original Reference ͑blue mesh͒ and actual Test lungs are very simi-
lar in lateral ͑x͒ and longitudinal ͑y͒ dimensions. Small adaptation noted on
superior portion of right lung. Volumetric nonoverlap mainly located in the
vertical direction ͑z͒. The NC-adapted lung model is represented in black
mesh.
TABLE IV. Statistical comparison of POL and nonoverlap ͑PNOL͒ between
the NC adapted and not adapted right and left lung models using paired
T-tests.
POL Dice
Adapted Not adapted Adapted Not adapted
Mean 89.2 84.3 89.3 85.4
Variance 15.2 90.3 7.4 28.9
SD 3.9 9.5 2.7 5.4
Observations 34 34 34 34
df 33 33
t Stat 4.3 5.1
P͑TϽ=t͒two-tail 0.0002 Ͻ0.0001
t Critical two-tail 2.0 2.0
FIG. 9. Comparison of the distribution of POL and Dice for the three con-
ditions: ͑i͒ Without NC adaptation, ͑ii͒ with NC adaptation, and ͑iii͒ free-
breathing inhale-exhale ͑FB͒ lung motion. Outliers ͑1.5ϫinterquartile
range͒ are represented as “+.”
1024 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1024
Medical Physics, Vol. 37, No. 3, March 2010
pares the POL and Dice results under the three conditions: ͑i͒
No NC adaptation, ͑ii͒ with NC adaptation, and ͑iii͒ free
breathing exhale-inhale cases. The range calculated for POL
values under the three conditions were 31.2%, 14.4%, and
10.9%, respectively. The range calculated for Dice values
under the three conditions were 31.2%, 18.8%, and 9.0%,
respectively. The VOL results with NC adaptation are com-
parable with those obtained for free-breathing lung volumes
͑POL: p=0.43; Dice: p=0.20͒.
IV. DISCUSSION
A novel technique that combines the use of DIR and NCs
to reconstruct 3D lung volumes from 2D planning data sets
has been developed and evaluated. The VOL results demon-
strate that our technique regenerated approximately 90% of
the actual Test volume, averaging a ten percent error in non-
overlapping volume. Ideally, the NC-adapted and Test vol-
umes would be identical ͑i.e., 100% POL and Dice͒. The
difference in volume between the NC-adapted and Test mod-
els is primarily due to the requirement to simulate and adhere
to the restrictive conditions of 2D treatment planning. Since
only coronal planning films were used, and no sagittal films
were obtained, depth of lung ͑z͒ information is unavailable
for these treatments. To simulate this limitation, only coronal
DRRs were used to obtain lateral ͑x͒ and longitudinal ͑y͒
lung difference for adapting existing 3D models. As a result,
NCs could not be applied in the vertical direction, and the
anterior-posterior portions of the lung could not be as accu-
rately modeled. While the model cannot overcome this defi-
ciency, it is able to utilize thorax measurements, which
helped to minimize the effects of depth differences in the
modeling process. A z-approximation would be favorable,
however, without a sagittal radiograph, this is difficult. Using
patient information such as patient thickness can be a mis-
leading approximation for their respective lung volumes due
to high vs low body mass index patients. The correlation
between length-width with thickness of the lung using this
population was investigated and showed very little linear
correlation. Future studies can investigate on defining a rela-
tionship between these measurements. In addition, the volu-
metric discrepancies observed between the NC-adapted and
Test models could be due to the reliability of the NC adap-
tation technique to detect “edges” based on sharp image in-
tensity gradients. The NC alignment was based on both bony
͑ribs, spine͒ and soft tissue ͑diaphragm͒ interfaces. These
edges provide sharp contrast between bone/liver and air in-
terface, allowing accurate matching of the feature edges. The
inferior NC at the diaphragm which is a soft tissue alignment
allowed for lung deformation. For clinical application, to in-
crease soft tissue and air contrast at the medial edge of the
lung, adjustable grayscale features are available to determine
optimal image contrast for better lung match on 2D planning
films. As discussed below, the VOL results were compared
with those in free breathing lung motion, to determine if the
accuracies were within acceptable limits.
Compared to results obtained without NC adaptation, the
NCs made a statistically significant improvement to the
population deformation models ͑pϽ0.05͒. The small adap-
tation seen in the anterior-posterior ͑AP͒ direction was a re-
sult of the incorporation of the population model into the
calculation. The population model provided a guide to the
general lung shape in the left-right ͑LR͒, superior-inferior
͑SI͒, and AP directions. Since the NC technique could only
adapt the population model to the Test specific model in the
SI and LR directions, the population model provided a rep-
resentative estimate of the lung shape in the AP direction. As
hypothesized, NC adaptation increased POL in 15 of the 17
cases ͑88%͒. For the two cases that fell outside the predic-
tion, the POL and Dice differences were only 1%. Although
the original Reference and Test lungs were very similar in
both LR and SI dimensions and the AP differences were
large, without sagittal images, these differences could not be
determined and incorporated into the adaptation. This could
be a major reason why the adaptation did not produce sig-
nificant changes. In addition, lung volume mismatch in the
medial regions due to variation in position of major blood
vessels ͑pulmonary arteries and superior vena cava͒ and bi-
furcation of the bronchus could influence the errors in the
lung adaptation process. As discussed below, using multiple
NCs on each side of the lung and adjusting image contrast
could better account for these local patient differences in
future investigations.
Comparison of the NC-adapted VOL results with the
volumetric changes observed during patient breathing dem-
onstrates that they are not significantly different. While re-
cent studies have investigated ways of immobilizing ventila-
tory motion in HL treatments,
31–33
patients undergoing
conventional HL treatments are typically treated freely
breathing, and as such internal respiration induced move-
ments will cause uncertainty in treatment delivery. This is
one of the underlying assumptions of this study. The posi-
tion, and volumetric variation observed in breathing lungs,
served as a baseline for evaluating and validating the results.
As predicted, the results show that lung volume and position
during free-breathing varied substantially ͑10% nonoverlap͒
between end inhale and exhale phases. This demonstrates
that the volumetric reconstruction accuracies are within ac-
ceptable limits, comparable to the free-breathing variation
͑Ͻ2% in POL, p=0.43 and Dice, p=0.20͒ in lung volumes
during respiration. This indicates that the accuracy of the
volumetric reconstruction method is not a limiting factor in
estimating normal tissue dose exposure.
In this study, the NC technique is favorable for use be-
cause, first, edge detection is automated; second, it associates
1D information to 3D organ information; and third, prelimi-
nary liver studies have found that images with low soft tissue
contrast, such as radiographs, are optimal for boundary
detection.
21,22
The NC technique has been previously studied
to accurately associate point edge information to a popula-
tion model to provide full organ information.
21,22
In addition,
NCs could potentially provide important structural details
that may be imperceptible to the human eye. Currently, the
NC technique is capable of detecting resolutions of 0.10
ϫ0.10ϫ0.25 cm
2
for CT images. Investigation is underway
1025 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1025
Medical Physics, Vol. 37, No. 3, March 2010
to increase its accuracy by interpolating image intensity val-
ues between voxel lengths.
21
For application of NCs on his-
torical radiographic films, it is anticipated that a high reso-
lution Vidar diagnostic film ͑VDF͒ scanner will be required
to digitize and preprocess the films prior to use with NCs.
The maximum optical resolution of the VDF scanner ͑VXR-
16, Vidar Systems Corporation, Herdon, VA͒ combined with
the Radiological Imaging Technology software ͑RIT113
v5.1, Colorado Springs, CO͒ used at our institution is 285
dots per inch ͑dpi͒, with a 16-bit grayscale. In comparison to
typical web graphics ͑72 dpi͒, higher resolution could poten-
tially allow for sharper edge detection ͑e.g., bones of the
thoracic cage and air cavities in lungs͒. In addition, the ad-
justable grayscale feature would allow us to better differen-
tiate tissues with similar densities ͑e.g., lung and heart inter-
faces͒. Future investigations on the effects of dpi resolution
and/or image contrast on radiographic image quality will
help determine optimal image parameters for NC application
on radiographic films.
The final step, adaptation of existing models using NC-
detected shifts, has a major advantage over simple linear
interpolation ͑without deformation͒ methods because it in-
corporates the population model and deformation maps into
the adaptation calculation. The population model provides an
estimate of the general lung shape and the population-
Reference patient deformation maps provide a guide of the
patient-to-patient deviations in lung shape. A previous study
on liver motion has demonstrated significant differences in
accuracy between a simple linear interpolation ͑errors
Ͼ0.50 cm͒ and the current technique ͑errorsϽ0.25 cm͒.
21
Additionally, this technique could be further improved with
the use of multiple NCs in each spatial direction. For this
study, only four NCs were utilized to capture differences on
each side ͑superior, inferior, left, and right͒ of the lung be-
cause this technique relies on a linearly weighted equation
that can only accommodate a single NC in each direction.
While this technique could be used for modeling lungs, as
they are somewhat rigidly shaped in the confines of the tho-
racic cage, other organs such as the heart and breasts are
more irregular in shape, with significant volumetric and po-
sitional differences between “similar” patients. Thus, using a
single NC on each organ side is insufficient to accurately
capture the differences in shape between patients. To over-
come this limitation, investigation is underway to develop a
more advanced adaptation technique using a bilinear interpo-
lation method that would accommodate for differences along
the same organ edge with multiple NCs in each direction.
Multiple NCs in various spatial directions would enable the
Reference model to better conform to the Test volume. There
are various lung segmentation techniques available, which
are similar to providing an infinite series of 1D NCs.
34–36
However, they are limited by the need for complete bound-
ary definition, which is challenged by regions of low soft
tissue contrast. The advantage of the NC technique is that the
complete boundary of the organ is not required to determine
the organ shape of the Test patient. This technique is applied
to sections of the edges that are visible with higher contrast
and the entire organ volume is regenerated using the adapta-
tion process.
Finally, the VOL tests are common methods for evaluat-
ing image segmentation and nonrigid registration techniques.
For this study, VOL was used to compare a pair of discretely
labeled structures ͑e.g., NC-adapted vs actual Test model͒ by
the voxels, and assumed that each voxel is either labeled or
not labeled ͑overlap or nonoverlap͒. Although these mea-
sures were straightforward to compute, overlap measures of
multiple partial volume labels to describe “fuzzy” overlaps
could be used in future studies, to provide a more compre-
hensive description of the degree of overlap between com-
paring volumes.
37,38
A technique that can help regenerate 3D plans from 2D
planning data sets could potentially provide a better under-
standing of the dose-volume association for many critical
organs, offering valuable insight on the dose-risk relationship
between RT and various late effects. Several studies have
investigated the late effects of radiation on pulmonary func-
tion. Smith et al.
39
prospectively evaluated 20 HL patients’
͑Stage I–III, mantle doses 3750 cGy͒ pulmonary test results
͑spirometry, lung volumes, and diffusing capacity͒ before re-
ceiving mantle irradiation and at yearly follow-up. Changes
in pulmonary function included gradual decrease in total
lung capacity ͑8% deficit at 6–10 yr͒, functional residual
capacity ͑5%–10% deficit at 6–10 yr͒, and diffusing capacity
͑10% deficit at 10 yr͒.
39
Similarly, Mefferd et al. examined
post-treatment pulmonary function in 34 children, ages 5–17,
treated with involved field treatment, and found that 55%
had abnormal values of diffusing capacity, and 10% showed
obstructive dysfunction at follow-up ͑mean follow-up 27.5
months͒.
40
Likewise, Gustavsson et al.
41
evaluated the long-
term effects on pulmonary function in 25 patients 10–20 yr
after mantle radiotherapy ͑35–43 Gy in 26 fractions͒. Al-
though there was little evidence of airflow obstruction, minor
restrictive ventilatory defects were found with decreased vi-
tal capacity, total lung capacity, and lung compliance in 18
patients ͑72%͒.
41
While the preceding studies provide evi-
dence of subclinical effects of radiation, no dose-effect rela-
tionship was deduced, as 3D dosimetric analysis of the 2D
treatment plans could not be performed at the time. Alterna-
tively, Boersma et al.
42
investigated the dose-effect relations
for local functional and structural changes in the lung after
irradiation for malignant lymphoma. 3D dose distributions
were calculated using CT and SPECT data. A change in ven-
tilation and perfusion was found in most patients in the dose
interval of 0–12 Gy, varying from an increase of 37% to a
decrease of 10%, followed by a decreasing trend at higher
doses.
42
Although this study offered a dose-effect relation-
ship, the examination period was relatively short, ranging
from only three to four months after irradiation. The excess
risk to pulmonary late effects, such as secondary lung cancer,
has been shown to begin 5 yr after radiation treatment and
persist for over 20 yr.
1,2,43
The risk of secondary lung cancer
in relation to treatment dosage have been recently quantified.
Travis et al. investigated the risk of treatment-associated
lung cancer in a study cohort of 19 046 HL patients, and
1026 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1026
Medical Physics, Vol. 37, No. 3, March 2010
found that treatment with alkylating agents without radio-
therapy was associated with increased lung cancer risk ͑rela-
tive risk ͓RR͔ =4.2͒, as was radiation dose of 5 Gy or more
without alkylating agents ͑RR=5.9͒.
43
Similarly, Gilbert et
al.
44
provided additional characterizations of the dose-
response relationship in radiation-induced lung cancer
through delineation of the secondary lung tumor from
follow-up CT images onto the initial 2D simulator films. The
estimated RR per Gy was reported 0.15 and the interaction of
radiation and chemotherapy that included alkylating agents
was additive.
44
However, this retrospective study is based on
2D dosimetry, where radiation absorbed dose was calculated
to a midpoint in the anterior-posterior direction from original
simulator films.
Based on the preliminary results, the proposed technique
offers a novel method to construct 3D lung models and treat-
ment plans directly from historical 2D planning images. This
is a preliminary study on the geometric validation of the
technique using the VOL test. Future studies will focus on
dosimetric validation of the technique, and application on
real 2D treatment plans. The first objective can be achieved
by regenerating 3D treatment plans using the NC-adapted
models and comparing them with their actual CT plans. A
normal tissue dose volume histogram can be used to define
the maximum, minimum, and mean dose delivered to the
lung, as well as mean total lung V5, V10, and V20. Further-
more, the conformity index can be used to assess the “quality
of coverage” of the organ by quantifying the volume of de-
lineation by an isodose line.
45
Aside from modeling lungs,
this technique also has potential in modeling other critical
organs, such as the heart and breasts for determining the
dose-volume late effects in the treatment of HL.
V. CONCLUSION
A novel technique to construct lung volumes from 2D
planning images using DIR and NC has been developed.
Validation of the technique demonstrates that its accuracy is
within the limits observed in respiratory motion during HL
treatment. This development could potentially allow for re-
generation of 3D lung models from 2D plans to perform
retrospective 3D dosimetric analysis for historic HL patients
for late outcome analysis. The VOL results are comparable to
the variation seen in respiratory motion, indicating that the
accuracy of our volumetric reconstruction method is accept-
able, and would not be the limiting factor in estimating nor-
mal tissue exposure. Pulmonary late effects are common in
HL patients 10–20 yr from their initial radiation treatment.
The described process could help provide a better under-
standing of the dose-volume association, and subsequently
provides valuable insight on the dose-risk relationship be-
tween RT and pulmonary late toxicities in HL patients.
ACKNOWLEDGMENTS
The authors would like to thank Mike Holwell and Lily
Chau for their Pinnacle support. Project funding provided by
the Canadian Institutes for Health Research ͑Grant No. MOP
77515͒. D.C. Hodgson and K.K. Brock are supported by a
Research Chair from Cancer Care Ontario.

Electronic mail: kristy.brock@rmp.uhn.on.ca
1
P. Lorigan, J. Radford, A. Howell, and N. Thatcher, “Lung cancer after
treatment for Hodgkin’s lymphoma: A systematic review,” Lancet Oncol.
6, 773–779 ͑2005͒.
2
A. Gustavsson, B. Osterman, and E. Cavallin-Stahl, “A systematic over-
view of radiation therapy effects in Hodgkin’s lymphoma,” Acta Oncol.
42, 589–604 ͑2003͒.
3
G. M. Chronowski, R. B. Wilder, S. L. Tucker, C. S. Ha, A. Younes, L.
Fayad, M. A. Rodriguez, F. B. Hagemeister, I. Barista, F. Cabanillas, and
J. D. Cox, “Analysis of in-field control and late toxicity for adults with
early-stage Hodgkin’s disease treated with chemotherapy followed by ra-
diotherapy,” Int. J. Radiat. Oncol., Biol., Phys. 55, 36–43 ͑2003͒.
4
M. T. Vlachaki, C. S. Ha, F. B. Hagemeister, L. M. Fuller, M. A. Rod-
riguez, P. C. Besa, M. A. Hess, B. Brown, F. Cabanillas, and J. D. Cox,
“Long-term outcome of treatment for Ann Arbor stage 1 Hodgkin’s dis-
ease: Patterns of failure, late toxicity and second malignancies,” Int. J.
Radiat. Oncol., Biol., Phys. 39, 609–616 ͑1997͒.
5
F. Villani, S. Viviani, V. Bonfante, P. De Maria, F. Soncini, and A. Laf-
franchi, “Late pulmonary effects in favorable stage I and IIA Hodgkin’s
disease treated with radiotherapy alone,” Am. J. Clin. Oncol. 23, 18–21
͑2000͒.
6
R. M. Chau, P. M. Teo, M. K. Kam, S. F. Leung, K. Y. Cheung, and A. T.
Chan, “Dosimetric comparison between 2-dimensional radiation therapy
and intensity modulated radiation therapy in treatment of advanced
T-stage nasopharyngeal carcinoma: To treat less or more in the planning
organ-at-risk volume of the brainstem and spinal cord,” Med. Dosim. 32,
263–270 ͑2007͒.
7
Y. M. Liu, C. Y. Shiau, M. L. Lee, P. I. Huang, C. M. Hsieh, P. H. Chen,
Y. H. Lin, L. W. Wang, and S. H. Yen, “The role and strategy of IMRT in
radiotherapy of pelvic tumors: Dose escalation and critical organ sparing
in prostate cancer,” Int. J. Radiat. Oncol., Biol., Phys. 67, 1113–1123
͑2007͒.
8
C. A. Sale, E. E. Yeoh, S. Scutter, and E. Bezak, “2D versus 3D radiation
therapy for prostate carcinoma: A direct comparison of dose volume pa-
rameters,” Acta Oncol. 44, 348–354 ͑2005͒.
9
E. Bednaruk-Mlynski, E. Senkus-Konefka, P. Szewczyk, A. Badzio, and
J. Jassem, “Parallel-opposed fields versus four fields, and two-versus
three-dimensional radiotherapy planning in thin patients with gynecologi-
cal malignancies,” Neoplasma 55, 151–157 ͑2008͒.
10
X. Papademetris, A. J. Sinusas, D. P. Dione, R. T. Constable, and J. S.
Duncan, “Estimation of 3-D left ventricular deformation from medical
images using biomechanical models,” IEEE Trans. Med. Imaging 21,
786–800 ͑2002͒.
11
M. I. Miga, K. D. Paulsen, P. J. Hoopes, F. E. Kennedy, A. Hartov, and D.
W. Roberts, “In vivo modeling of interstitial pressure in the brain under
surgical load using finite elements,” J. Biomech. Eng. 122, 354–363
͑2000͒.
12
K. D. Paulsen, M. I. Miga, F. E. Kennedy, P. J. Hoopes, A. Hartov, and D.
W. Roberts, “A computational model for tracking subsurface tissue defor-
mation during stereotactic neurosurgery,” IEEE Trans. Biomed. Eng. 46,
213–225 ͑1999͒.
13
A. Bharatha, M. Hirose, N. Hata, S. K. Warfield, M. Ferrant, K. H. Zou,
E. Suarez-Santana, J. Ruiz-Alzola, A. D’Amico, R. A. Cormack, R. Ki-
kinis, F. A. Jolesz, and C. M. Tempany, “Evaluation of three-dimensional
finite element-based deformable registration of pre- and intraoperative
prostate imaging,” Med. Phys. 28, 2551–2560 ͑2001͒.
14
T. Brückner, R. Lucht, and G. Brix, “Comparison of rigid and elastic
matching of dynamic magnetic resonance mammographic images by mu-
tual information,” Med. Phys. 27, 2456–2461 ͑2000͒.
15
K. K. Brock, M. Hawkins, C. Eccles, J. L. Moseley, D. J. Moseley, D. A.
Jaffray, and L. A. Dawson, “Improving image-guided target localization
through deformable registration,” Acta Oncol. 47, 1279–1285 ͑2008͒.
16
A. Kirilova, G. Lockwood, P. Choi, N. Bana, M. A. Haider, K. K. Brock,
C. Eccles, and L. A. Dawson, “Three-dimensional motion of liver tumors
using cine-magnetic resonance imaging,” Int. J. Radiat. Oncol., Biol.,
Phys. 71, 1189–1195 ͑2008͒.
17
M. R. Kaus, K. K. Brock, V. Pekar, L. A. Dawson, A. M. Nichol, and D.
A. Jaffray, “Assessment of a model-based deformable image registration
approach for radiation therapy planning,” Int. J. Radiat. Oncol., Biol.,
1027 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1027
Medical Physics, Vol. 37, No. 3, March 2010
Phys. 68, 572–580 ͑2007͒.
18
K. K. Brock, S. J. Hollister, L. A. Dawson, and J. M. Balter, “Technical
note: Creating a four-dimensional model of the liver using finite element
analysis,” Med. Phys. 29, 1403–1405 ͑2002͒.
19
K. K. Brock, D. L. McShan, R. K. Ten Haken, S. J. Hollister, L. A.
Dawson, and J. M. Balter, “Inclusion of organ deformation in dose cal-
culations,” Med. Phys. 30, 290–295 ͑2003͒.
20
K. K. Brock, M. B. Sharpe, L. A. Dawson, S. M. Kim, and D. A. Jaffray,
“Accuracy of finite element model-based multi-organ deformable image
registration,” Med. Phys. 32, 1647–1659 ͑2005͒.
21
T. N. Nguyen, J. L. Moseley, L. A. Dawson, D. A. Jaffray, and K. K.
Brock, “Adapting liver motion models using a navigator channel tech-
nique,” Med. Phys. 36, 1061–1073 ͑2009͒.
22
T.-N. Nguyen, J. L. Moseley, L. A. Dawson, D. A. Jaffray, and K. K.
Brock, “Adapting population liver motion models for individualized on-
line image-guided therapy,” Proceedings of Engineering in Medicine and
Biology Society, 2008 EMBS 30th Annual International Conference of
the IEEE, 3945–3948 ͑2008͒.
23
S. Pergolizzi, N. Settineri, M. Gaeta, E. Scribano, A. Santacaterina, G.
Ascenti, P. Frosina, and C. de Renzis, “What is the best position of the
arms in mantle field for Hodgkin’s disease?,” Int. J. Radiat. Oncol., Biol.,
Phys. 46, 119–122 ͑2000͒.
24
F. Bellemare, A. Jeanneret, and J. Couture, “Sex differences in thoracic
dimensions and configuration,” Am. J. Respir. Crit. Care Med. 168, 305–
312 ͑2003͒.
25
J. F. Bellemare, M. P. Cordeau, P. Leblanc, and F. Bellemare, “Thoracic
dimensions at maximum lung inflation in normal subjects and in patients
with obstructive and restrictive lung diseases,” Chest 119, 376–386
͑2001͒.
26
G. Stanley, D. Verotta, N. Craft, R. A. Siegel, and J. B. Schwartz, “Age
effects on interrelationships between lung volume and heart rate during
standing,” Am. J. Physiol. 273, H2128–H2134 ͑1997͒.
27
G. Stanley, D. Verotta, N. Craft, R. A. Siegel, and J. B. Schwartz, “Age
and autonomic effects on interrelationships between lung volume and
heart rate,” Am. J. Physiol. 270, H1833–H1840 ͑1996͒.
28
A. De Lorenzo, P. Petrone-De Luca, G. F. Sasso, M. G. Carbonelli, P.
Rossi, and A. Brancati, “Effects of weight loss on body composition and
pulmonary function,” Respiration 66, 407–412 ͑1999͒.
29
K. O. Babalola, B. Patenaude, P. Aljabar, J. Schnabel, D. Kennedy, W.
Crum, S. Smith, T. Cootes, M. Jenkinson, and D. Rueckert, “An evalua-
tion of four automatic methods of segmenting the subcortical structures in
the brain,” Neuroimage 47, 1435–1447 ͑2009͒.
30
L. R. Dice, “Measures of the amount of ecologic association between
species,” Ecology 26, 297–302 ͑1945͒.
31
G. Bentel, “Positioning and immobilization of patients undergoing radia-
tion therapy for Hodgkin’s disease,” Med. Dosim. 16, 111–117 ͑1991͒.
32
J. S. Stromberg, M. B. Sharpe, L. H. Kim, V. R. Kini, D. A. Jaffray, A. A.
Martinez, and J. W. Wong, “Active breathing control ͑ABC͒ for
Hodgkin’s disease: Reduction in normal tissue irradiation with deep in-
spiration and implications for treatment,” Int. J. Radiat. Oncol., Biol.,
Phys. 48, 797–806 ͑2000͒.
33
L. Claude, C. Malet, P. Pommier, P. Thiesse, S. Chabaud, and C. Carrie,
“Active breathing control for Hodgkin’s disease in childhood and adoles-
cence: Feasibility, advantages, and limits,” Int. J. Radiat. Oncol., Biol.,
Phys. 67, 1470–1475 ͑2007͒.
34
M. Kakar and D. R. Olsen, “Automatic segmentation and recognition of
lungs and lesion from CT scans of thorax,” Comput. Med. Imaging
Graph. 33, 72–82 ͑2009͒.
35
S. Ukil and J. M. Reinhardt, “Anatomy-guided lung lobe segmentation in
x-ray CT images,” IEEE Trans. Med. Imaging 28, 202–214 ͑2009͒.
36
J. Pu, J. Roos, C. A. Yi, S. Napel, G. D. Rubin, and D. S. Paik, “Adaptive
border marching algorithm: Automatic lung segmentation on chest CT
images,” Comput. Med. Imaging Graph. 32, 452–462 ͑2008͒.
37
W. R. Crum, O. Camara, and D. L. Hill, “Generalized overlap measures
for evaluation and validation in medical image analysis,” IEEE Trans.
Med. Imaging 25, 1451–1461 ͑2006͒.
38
W. R. Crum, O. Camara, D. Rueckert, K. K. Bhatia, M. Jenkinson, and D.
L. Hill, “Generalised overlap measures for assessment of pairwise and
groupwise image registration and segmentation,” Med. Image Comput
Comput Assist Interv Int Conf Med Image Comput Comput Assist Interv
8, 99–106 ͑2005͒.
39
L. M. Smith, N. P. Mendenhall, M. J. Cicale, E. R. Block, R. L. Carter,
and R. R. Million, “Results of a prospective study evaluating the effects
of mantle irradiation on pulmonary function,” Int. J. Radiat. Oncol., Biol.,
Phys. 16, 79–84 ͑1989͒.
40
J. M. Mefferd, S. S. Donaldson, and M. P. Link, “Pediatric Hodgkin’s
disease: Pulmonary, cardiac, and thyroid function following combined
modality therapy,” Int. J. Radiat. Oncol., Biol., Phys. 16, 679–685 ͑1989͒.
41
A. Gustavsson, J. Eskilsson, T. Landberg, H. Larusdottir, G. Svahn-
Tapper, T. White, and P. Wollmer, “Long-term effects on pulmonary func-
tion of mantle radiotherapy in patients with Hodgkin’s disease,” Ann.
Oncol. 3, 455–461 ͑1992͒.
42
L. J. Boersma, E. M. Damen, R. W. de Boer, S. H. Muller, C. M. Roos, R.
A. Valdes Olmos, N. van Zandwijk, and J. V. Lebesque, “Dose-effect
relations for local functional and structural changes of the lung after irra-
diation for malignant lymphoma,” Radiother. Oncol. 32, 201–209 ͑1994͒.
43
L. B. Travis, M. Gospodarowicz, R. E. Curtis, E. A. Clarke, M. Anders-
son, B. Glimelius, T. Joensuu, C. F. Lynch, F. E. van Leeuwen, E. Holo-
waty, H. Storm, I. Glimelius, E. Pukkala, M. Stovall, J. F. Fraumeni, Jr.,
J. D. Boice, Jr., and E. Gilbert, “Lung cancer following chemotherapy and
radiotherapy for Hodgkin’s disease,” J. Natl. Cancer Inst. 94, 182–192
͑2002͒.
44
E. S. Gilbert, M. Stovall, M. Gospodarowicz, F. E. Van Leeuwen, M.
Andersson, B. Glimelius, T. Joensuu, C. F. Lynch, R. E. Curtis, E. Holo-
waty, H. Storm, E. Pukkala, M. B. van’t Veer, J. F. Fraumeni, J. D. Boice,
Jr., E. A. Clarke, and L. B. Travis, “Lung cancer after treatment for
Hodgkin’s disease: Focus on radiation effects,” Radiat. Res. 159, 161–
173 ͑2003͒.
45
L. Feuvret, G. Noel, J. J. Mazeron, and P. Bey, “Conformity index: A
review,” Int. J. Radiat. Oncol., Biol., Phys. 64, 333–342 ͑2006͒.
1028 Ng et al.: 3D model reconstruction using deformable registration/navigator channels 1028
Medical Physics, Vol. 37, No. 3, March 2010
Copyright of Medical Physics is the property of American Association of Physicists in Medicine and its content
may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express
written permission. However, users may print, download, or email articles for individual use.