Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
From the Society for Vascular Surgery
Local aortic aneurysm wall expansion measured with automated
image analysis
Jordan B. Stoecker, MD,
a
Kevin C. Eddinger, MD,
a
Alison M. Pouch, PhD,
b,c
Amey Vrudhula, BS,
d
and
Benjamin M. Jackson, MD,
a
Philadelphia, Pa; and New York, NY
ABSTRACT
Background: Assessment of regional aortic wall deformation (RAWD) might better predict for abdominal aortic aneu-
rysm (AAA) rupture than the maximal aortic diameter or growth rate. Using sequential computed tomography angio-
grams (CTAs), we developed a streamlined, semiautomated method of computing RAWD using deformable image
registration (dirRAWD).
Methods: Paired sequential CTAs performed 1 to 2 years apart of 15 patients with AAAs of various shapes and sizes were
selected. Using each patient’s initial CTA, the luminal and aortic wall surfaces were segmented both manually and
semiautomatically. Next, the same patient’s follow-up CTA was aligned with the first using automated rigid image
registration. Deformable image registration was then used to calculate the local aneurysm wall expansion between the
sequential scans (dirRAWD). To measure technique accuracy, the deformable registration results were compared with
the local displacement of anatomic landmarks (fiducial markers), such as the origin of the inferior mesenteric artery and/
or aortic wall calcifications. Additionally, for each patient, the maximal RAWD was manually measured for each aneurysm
and was compared with the dirRAWD at the same location.
Results: The technique was successful in all patients. The mean landmark displacement error was 0.59 60.93 mm with
no difference between true landmark displacement and deformable registration landmark displacement by Wilcoxon
rank sum test (
P
¼.39). The absolute difference between the manually measured maximal RAWD and dirRAWD was
0.27 60.23 mm, with a relative difference of 7.9% and no difference using the Wilcoxon rank sum test (
P
¼.69). No
differences were found in the maximal dirRAWD when derived using a purely manual AAA segmentation compared with
using semiautomated AAA segmentation (
P
¼.55).
Conclusions: We found accurate and automated RAWD measurements were feasible with clinically insignificant errors.
Using semiautomated AAA segmentations for deformable image registration methods did not alter maximal dirRAWD
accuracy compared with using manual AAA segmentations. Future work will compare dirRAWD with finite element
analysisederived regional wall stress and determine whether dirRAWD might serve as an independent predictor of
rupture risk. (JVSeVascular Science 2022;3:48-63.)
Clinical Relevance: Current abdominal aortic aneurysm (AAA) surveillance methods are limited to assessments of the
maximal diameter, which cannot accurately predict for AAA expansion and rupture risk. Automated assessment of AAA
expansion across the entire three-dimensional geometry of the aneurysm could better describe aneurysm growth and
could substantially inform management decisions, including the indications for repair. We have developed an accurate
and streamlined approach to assessing local three-dimensional AAA expansion with submillimeter accuracy using
computed tomography imaging obtained during routine aneurysm surveillance. This novel process does not require
significant user expertise nor computer processing power and can be performed using open-source software readily
accessible to both scientists and clinicians.
Keywords: Aneurysms; Biomechanics; Computational analysis and simulation
From the Division of Vascular Surgery and Endovascular Therapy, Department
of Surgery,
a
and Division of Radiology,
b
Hospital of the University of Pennsyl-
vania, Philadelphia; the Department of Bioengineering, University of Pennsyl-
vania, Philadelphia
c
; and the Icahn School of Medicine at Mount Sinai, New
York.
d
Author conflict of interest: none.
Presented as a poster abstract at the 2020 Biomedical Engineering Society
Annual Meeting, October 14-17, 2020 ONLINE and the 2020 Society for
Vascular Surgery Vascular Research Initiatives Conference, November 5-12,
2020 ONLINE.
Correspondence: Jordan B. Stoecker, MD, Division of Vascular Surgery and
Endovascular Therapy, Department of Surgery, Hospital of the University of
Pennsylvania, 3400 Spruce St, 4th FL, Silverstein Bldg, Philadelphia, PA
19146 (e-mail: jordan.stoecker@pennmedicine.upenn.edu).
The edito rs and revie wers of this article have no relev ant financial relationships to
disclose per the JVS-Vascular Science policy that requires reviewers to decline
review of any manuscript for which they may have a conflict of interest.
2666-3503
Copyright Ó2021 by the Society for Vascular Surgery. Published by Elsevier Inc.
This is an open access article under the CC BY-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).
https://doi.org/10.1016/j.jvssci.2021.11.004
48
The detected prevalence of asymptomatic abdominal
aortic aneurysms (AAAs) has increased as noninvasive
screening modalities have become more widely avail-
able. Aortic diseases, including aortic aneurysms, are
the 12th leading cause of death in the United States.
1
AAA rupture is a major source of morbidity and mortality,
with reported mortality rates of 40% to 50%.
2
Surgical
intervention for asymptomatic patients has generally
been reserved for those with an aortic diameter
>5.5 cm or with an aneurysm expansion rate >10 mm/
y. However, a previous autopsy study estimated that
10% to 24% of ruptured AAAs will be <5.5 cm in diam-
eter.
3
Another study reported that 7% had had a diam-
eter of <5.0 cm.
4
Thus, improved risk stratification of
AAAs could prevent rupture and save lives.
Maximum diameter measurements on computed to-
mography (CT) angiograms (CTAs) have been previously
shown to have an error in the range of 2 to 5 mm, which
can limit its clinical utility.
5,6
This has been especially
notable for noncircular aneurysms and when the center-
line of the aorta is not axially oriented, which will hamper
the accuracy of the measured aneurysm diameter in the
axial plane.
5,6
Given that the management recommen-
dations have been based on the thresholds of the size
and growth rate, a diameter-based measurement tech-
nique can lead to treatment recommendations that
are either overly aggressive or conservative owing to
measurement error. To better assess aneurysm growth
from surveillance imaging scans, prior literature has
examined the role of the aortic cross-sectional area or
volumetric changes. Although these appear more sensi-
tive to detecting overall aneurysm growth, they cannot
confer any information regarding localized aortic
changes.
7,8
It has been previously demonstrated that complex geo-
metric properties of AAAs, including the curvature,
shape, and degree of mural thrombus, can be superior
to the maximal diameter measurements alone in pre-
dicting for aneurysm growth and possible rupture.
Furthermore, changes in aneurysm tortuosity over time
have been suggested as a potential marker during aneu-
rysm surveillance.
9
However, many of these properties
have remained cumbersome to calculate, are nonstan-
dardized, and difficult to interpret for those without sig-
nificant bioengineering knowledge. Therefore, these
have not translated into clinical practice and appear un-
likely to become relevant in the near future.
10-14
There-
fore, although research into novel methods of
aneurysm surveillance has been significant, maximal
diameter measurements and growth in the maximal
diameter have remained the default, yet suboptimal,
clinical parameters.
Regional aortic wall deformation (RAWD) can be calcu-
lated using deformable image registration (dirRAWD).
13,15
This technique allows for quantification of aneurysm dis-
ease progression on surveillance imaging, with the ability
to calculate the degree of shape changes between two
rigidly registered CT images, a process previously called
vascular deformation mapping. The benefits of assessing
dirRAWD include the simple interpretation, high accu-
racy, and the ability to determine both radial and longi-
tudinal aortic changes. Furthermore, because aortic
rupture customarily occurs in a focal area of the aortic
surface, RAWD might better assess for the risk of the clin-
ically significant event one would like to predict.
Although deformable image registration is rapidly being
applied to neurologic and pulmonary disease processes,
it has infrequently been applied to the field of vascular
surgery, with no effects yet in clinical practice.
14,16-19
A continued need exists for a more sensitive and accu-
rate method of measuring the changes in aortic aneu-
rysm dimensions, considering that the accurate
detection of small magnitude changes has important
implications for improving the understanding of aortic
aneurysm progression and better informing treatment
decisions. The aim of the present study was to develop
a novel, semiautomated deformable image registration
pipeline that can be applied to routine clinical surveil-
lance CTA studies of patients with AAAs. We compared
the results of the RAWD analysis with routine clinical
CTA assessments to better understand the potential
benefits and limitations of this technique. Our hypothesis
was that the proposed deformation analysis technique
could be accurately implemented and require minimal
user input and that the results would favorably compare
with the results obtained from both manual deformation
and AAA diameter growth assessments.
METHODS
Population. Sequential surveillance CTAs of 15 patients
with infrarenal AAAs of various shapes, sizes, and the
extent of luminal thrombus from 2014 to 2018 were
selected by querying the University of Pennsylvania
ARTICLE HIGHLIGHTS
dType of Research: A prospective, experimental
research study
dKey Findings: We developed a streamlined, semiau-
tomated method of computing local aneurysm
expansion using deformable image registration in
patients with computed tomography scans per-
formed 1 to 2 years apart. Our algorithm demon-
strated high accuracy in a cohort of 15 patients,
with a submillimeter measurement error of 0.27 6
0.23 mm and a relative error of 7.9%.
dTake Home Message: Our work has demonstrated
that calculating local aneurysm expansion is feasible
using open-source software and minimal user
training and could serve as an independent predic-
tor of rupture risk in the future.
JVSeVascular Science
Stoecker et al
49
Volume 3, Number C
Health System PACS (picture archiving and communica-
tion system) database using Nuance mPower (Nuance
Communications Inc, Burlington, Mass). The initial CTA
scans were evaluated using radiology workstations and
Sectra IDS7 (Sectra AB, Linköping, Sweden) and TeraRe-
con (TeraRecon Inc, Foster City, Calif). The inclusion
criteria were age >18 years at the initial imaging study,
the presence of infrarenal fusiform AAAs without evi-
dence of rupture, and at least two sequential CTA studies
performed within the health care system separated by 12
to 24 months. The patients were excluded if the aneu-
rysm in question had ruptured or if the patient had un-
dergone endovascular or open repair of the aneurysm
in the interval between the imaging studies. The institu-
tional review board expedited the review and approved
the present study owing to the use of previously
collected and anonymized data.
RAWD algorithm. The complete image processing and
analysis pipeline, constituting the current approach to
measuring dirRAWD, has been diagrammed in Fig 1.
Raw data from the patients’CTA scans were anonymized
and exported into DICOM (digital imaging and com-
munications in medicine) format. ITK-SNAP (available at:
http://www.itksnap.org/pmwiki/pmwiki.php; Penn Image
Computing and Science Laboratory, Philadelphia, Pa)
was used for segmentation and rigid registration of each
patient’s baseline CTA scan
20
(Supplementary Fig). In
brief, ITK-SNAP is an open-source platform for medical
image analysis that includes functionalities for manual
image tracing and semiautomated segmentation using
three-dimensional (3D) active contours and machine
learning and a registration feature for the rigid alignment
of images. In the present study, the images did not
require application of filters or augmentation of
Hounsfield units. Multilabel segmentations of the aortic
lumen, aortic wall, and thrombus were created for all
baseline CTAs (Fig 2). The aortic segmentation was per-
formed by one physician, both manually and using a
semiautomated process (using either active contour
evolution based on Hounsfield unit thresholding or tissue
classification), with the two methods compared for ac-
curacy using the Dice coefficient. The semiautomated
segmentation was inspected and manually corrected if
required. The Dice coefficient is commonly used to
assess for agreement between segmentation images.
The Dice coefficient ranges from 0 (no overlap) to 1
(perfect overlap) and is defined as follows:
Dice ¼
2Number of Overlapping Voxels
Total Number of Voxels in Both Images Combined
After segmentation of the baseline scan, the baseline
and follow-up CTA images were aligned using auto-
mated rigid multiresolution image registration and the
mutual information similarity metric in ITK-SNAP. The
initial rigid registration was performed using a multireso-
lution scheme with 16to 8down-sampling to grossly
align the images. Rigid registration was then repeated
with computation of the similarity metric confined to
the dilated aortic segmentation, first with 4down-
sampling and then at full resolution. Manual inspection
was performed on all rigidly registered images to
confirm adequate alignment (Fig 2).
After rigid image registration, deformable registration
was performed using
greedy
(available at: https://sites.-
google.com/view/greedyreg/home; Penn Image
Computing and Science Laboratory) to automatically
generate a deformation field between the baseline and
follow-up CTA images.
21
This open-source application
has shown efficiency and accuracy for both CT and
Fig 1. Overview of algorithm used to calculate regional aortic wall deformation (RAWD) with deformable image
registration (dirRAWD) between sequential computed tomography angiograms (
CTAs
) and generation of aortic
deformation map.
50
Stoecker et al
JVSeVascular Science
--- 2022
magnetic resonance imaging. For deformable image
registration, we used a normalized cross-correlation sim-
ilarity metric with gradient descent optimization for 100
iterations at a course level (4), 50 iterations at an inter-
mediate level (2), and 20 iterations at full resolution.
For further algorithm optimization, the deformable regis-
tration was limited to a 10 10 10 neighborhood
around each voxel, and the baseline aortic segmentation
was dilated by 1.5 cm and used to restrict the deforma-
tion algorithm to the area of interest. A smoothing
gradient with a sigma of 1.732 voxel was applied to the
image match metric at each voxel, and the entire defor-
mation field was smoothed with a gradient sigma of
0.707 voxel after each iteration to limit error and
generate a smooth field. The deformable image registra-
tion equations and smoothing gradients used have been
previously described in great detail.
22,23
A mesh of the outer wall of the aorta was then auto-
matically generated from the segmentation of the base-
line CTA image using ITK-SNAP. The deformation field
obtained by registration of the baseline image and
follow-up scan was applied to this baseline outer aortic
wall mesh, producing an aortic deformation map with
local dirRAWD at every vertex on the mesh. This mesh
can be displayed as a “heatmap”of the outer aortic
wall, with the magnitude of deformation displayed at
every mesh vertex to describe the local change in that re-
gion of the aorta over the period between the baseline
and follow-up CTA scans. An example of the final results
from the image analysis pipeline are shown in Fig 3.
To assess the accuracy of this algorithm in determining
dirRAWD, the maximal RAWD was manually determined
(manRAWD) for each patient. This was performed using
the manual aortic segmentation and rigidly registered
images, with manual measurement of the magnitude
of displacement in the aortic outer wall between the
baseline and follow-up images and the region of most
expansion. This measured maximal manRAWD was
considered the “true maximal deformation”and was
compared to the algorithm-derived dirRAWD in the cor-
responding aortic wall region for assessment of accuracy.
To assess segmentation intraobserver variability, all
manual and semiautomated segmentations were per-
formed twice by the same individual. Comparisons
were then performed between the calculated dirRAWD
using the repeat segmentation compared with the orig-
inal to determine the relative intraobserver variability
introduced by the segmentation process.
Fig 2. Baseline computed tomography angiogram (CTA; A) and overlay of multilabel aortic aneurysm segmen-
tation (B). On the right, the rigidly registered follow-up CTA is displayed (C) with overlay of the baseline seg-
mentation (D). The
white arrow
denotes the local aortic expansion that occurred between performance of the
two CTAs.
JVSeVascular Science
Stoecker et al
51
Volume 3, Number C
Furthermore, to assess our algorithm’s accuracy, the
landmark displacement error was calculated using a fi-
duciary marker. First, the following landmark regions
were defined in the baseline CTA: the inferior mesenteric
artery and aortic wall calcifications >4 mm in size. Next,
the deformation field generated from the dirRAWD algo-
rithm was applied to displace the landmarks from the
baseline CTA. The deformed landmark locations on the
baseline CTA were compared against their true locations
on the follow-up CTA to determine the displacement er-
ror in our algorithm.
The maximal aortic diameter was determined for all
baseline and follow-up scans using axial measurements
of the aneurysm and confirmed by measurements of the
aortic segmentation (in mm). The aortic expansion rate
was defined as the difference in the maximal aortic diam-
eter between the baseline and follow-up CTAs during the
study period between images (mm/y). These manually
determined measurements were compared with the re-
sults from the radiology reports with no differences noted.
The image analysis and calculations were performed on a
system with 8 GB of RAM (random access memory) and
Intel Core i5-8350 central processing unit.
Statistical analysis and image generation. Statistical
analysis was performed using STATA (StataCorp, College
Station, Tex). Plots were generated using Tecplot (Tec-
plot, Inc, Bellevue, Wash) and MeshLab (MeshLab, Univer-
sity of Pisa, Pisa, Italy). Nonparametric statistical analysis
was used because of the small study population. Contin-
uous variables were assessed using the Wilcoxon signed
rank test when comparing two paired groups or Fried-
man’s analysis of variance test when comparing three
paired groups. Categorical variables were compared us-
ing the Fischer exact test. The relationships between
the initial maximal aortic diameter, aortic expansion
rate, and maximal deformation were calculated using
the Pearson correlation coefficient. All
P
values were
two-sided, and
P
<.05 was considered to indicate sta-
tistical significance. Logistic and linear regression
modeling were used for binary and continuous out-
comes, respectively. All results are presented as the
mean 6standard deviation, unless otherwise specified.
To eliminate the outliers, maximal deformation was
defined as the 97.5th percentile of each aorta. A Dice
score >0.7 has been previously suggested as an appro-
priate benchmark for image validation and was used in
the present study.
24
RESULTS
The technique was successful for all 15 patients, with
initial demographic data shown in Table I and the
Supplementary Table. The results for four example cases
are shown in Fig 4. The average patient age was 66.7 6
11.3 years, and the average interval between the sequen-
tial scans was 16.1 62.7 months. The average baseline
aortic diameter was 42.7 69.9 mm, the average follow-
up aortic diameter was 48.3 611.8 mm, the mean aortic
expansion rate was 4.5 62.6 mm/y, and the mean true
maximal deformation rate was 4.3 62.5 mm/y. Seven pa-
tients (47%) had had aortic expansion rates of >5 mm/y
(an indication of rapid growth). The manual aortic seg-
mentation and semiautomated aortic segmentation for
each aortic aneurysm showed agreement with a Dice
score of 0.92 60.02. On visual inspection, seven of the
semiautomated aortic segmentations (47%) required
no manual correction, six (40%) had
required <2 minutes of correction, two (13%) had
required <5 minutes of correction, and one aneurysm
with extensive intraluminal thrombus had required
10 minutes of manual correction. The total time for rigid
image registration was <2 minutes for all patients, and
the total time for deformable image registration
was <15 minutes in 14 of the 15 patients.
Fig 3. Local aortic deformation map shown from anterior (A) and posterior (B) perspectives. C, Slice of second
computed tomography angiogram (CTA) from the indicated level, with an overlay of the initial aorta size in
red
and a
white arrow
denoting the degree of local expansion between the sequential CTAs.
52
Stoecker et al
JVSeVascular Science
--- 2022
Maximal deformation accuracy assessment. The
maximal manRAWD, dirRAWD at the same location
derived using the manual AAA segmentation, and dir-
RAWD at the same location derived using the semiauto-
mated AAA segmentation are presented in Table II and
displayed graphically in Fig 5. The three groups were
tested for measurement concordance using Friedman’s
analysis of variance, with a resulting
P
value of .001 and a
Kendall’s coefficient of concordance of 0.992, indicating
excellent agreement between the three RAWD
methods. Compared with the maximal manRAWD, the
dirRAWD derived using manual aortic segmentation
yielded a mean error of 0.25 60.24 mm/y, a mean rela-
tive error of 6.5%, and no statistically significant differ-
ences between the two groups using the Wilcoxon
signed rank test (
P
¼.69). Comparing the manRAWD to
the dirRAWD derived using semiautomated aortic seg-
mentation yielded a mean error of 0.27 60.23 mm/y, a
mean relative error of 7.9%, and no statistically significant
difference between the two groups using the Wilcoxon
signed rank test (
P
¼.84). Additionally, no difference was
found between the maximal dirRAWD derived using
manual AAA segmentation compared with that using
semiautomated AAA segmentation by the Wilcoxon
Table I. Demographics of 15 included patients
Demographic Value
Age, years 66.7 611.3 (46-83)
Male sex 9 (60)
History of hypertension 10 (67)
History of statin use 7 (47)
Family history of vascular aneurysms 3 (20)
Follow-up duration, months 16.1 62.7 (13-22)
Initial aortic diameter, mm 42.7 69.9 (31-60)
Final aortic diameter, mm 48.3 611.8 (34-66)
Radiographic expansion rate, mm/y 4.5 62.6 (0.5-9.3)
True maximal deformation
(manRAWD), mm/y
4.3 62.5 (1.2-9.2)
manRAWD,
Maximal regional aortic wall deformation was manually determined.
Data presented as mean 6standard deviation (range) or number (%).
Fig 4. A-D, Four example results of regional aortic wall deformation (RAWD) with deformable image registration
(dirRAWD) aortic maps generated by our algorithm.
JVSeVascular Science
Stoecker et al
53
Volume 3, Number C
rank sum test (
P
¼.55). The intraobserver variability in
calculating dirRAWD was 4.4% for manual segmentation
and 5.2% for semiautomated segmentation.
Landmark displacement analysis: true vs semi-
automated RAWD. The true landmark deformation,
algorithm-derived landmark deformation using the
manual aortic segmentation, and algorithm-derived
landmark deformation using the semiautomated aortic
segmentation are shown in Table III for all patients. The
three groups were tested for measured landmark
displacement concordance using Friedman’s analysis of
variance with a resulting
P
value of .002 and Kendall’s
coefficient of concordance of 0.976, indicating excellent
agreement between the calculated landmark displace-
ment values between the methods. Compared with the
true landmark deformation, the algorithm-derived
landmark deformation using the manual aortic seg-
mentation resulted in a mean error of 0.67 61.44 mm, a
mean relative error of 13.1%, and no difference using the
Wilcoxon signed rank test (
P
¼.39). Compared with the
true landmark deformation, the algorithm-derived
landmark deformation determined using the semi-
automated aortic segmentation yielded a mean error of
0.59 60.93 mm, a mean relative error of 12.6%, and no
difference using the Wilcoxon signed rank test (
P
¼.58).
Aneurysm diameter change vs dirRAWD. The compar-
ison of the changes in the maximal aortic diameter
compared with the maximal dirRAWD is shown in
Fig 6. In the eight patients with an aneurysm expansion
rate <5 mm/y, excellent agreement was found between
the radiographic reported maximal aortic expansion and
maximal dirRAWD on linear regression (
P
¼.01), with less
agreement between the aortic expansion rate and
maximal dirRAWD in those with expansion rates >5mm/
y(
P
¼.54).
DISCUSSION
The described dirRAWD algorithm was successful in
all15patientsforwhomitwasattempted.These15pa-
tients included those with AAAs of various shapes, sizes,
and amount of luminal thrombus using open source
software and minimal required computational power.
This technique resulted in submillimeter errors in the
calculated maximal dirRAWD compared with the man-
RAWD (the true value), with no statistically significant
differences. Furthermore, the generation of aortic seg-
mentations was the only manual component of this
pipeline, and we demonstrated no statistically signifi-
cant differences in the deformable image registration
error using semiautomated AAA segmentations vs
manual AAA segmentations of the baseline CTA. Addi-
tionally, our methods were able to display the magni-
tude of local growth directly on the outer aortic wall
that was more readily interpretable than prior methods,
which displayed surrogated markers such as surface
stress.
14
Improvement in overall accuracy compared with
maximal diameter measurements. RAWD algorithms
use 3D image data along the entire length of the aorta,
rather than diameters placed at fixed locations along
the aortic length. The use of 3D image data results in
decreased measurement error compared with measure-
ments restricted to the axial plane, which have a re-
ported error range of 2 to 5 mm.
5-7
Modern
semiautomated rigid and deformable registration
Fig 5. Scatterplot with trendline comparing true maximal aortic deformation (manRAWD) against maximal
regional aortic wall deformation (RAWD) with deformable image registration (dirRAWD) for every aorta in the
present study. Maximal dirRAWD derived from a manual aortic segmentation is shown in
orange
, and the
maximal dirRAWD from a semiautomated aortic segmentation is displayed in
blue
.
54
Stoecker et al
JVSeVascular Science
--- 2022
techniques, which align each location along the 3D
aortic wall, have been shown to have a precision in the
range of 0.5 to 1.0 mm, with our technique having a
mean error of 0.24 mm (range, 0.1-0.9 mm).
15,25,26
Superiority with 3D measurement. This technique is
able to quantify a continuous range of aortic RAWD in
three dimensions with associated direction and magni-
tude. The calculated displacement field of dirRAWD are
vectors, allowing for the assessment of the direction of
deformation and strain measurements, in addition to
the overall magnitude. The maximal diameter tech-
niques assume an outward radial direction to all aortic
expansion in a fixed horizontal plane, which does not
capture the circumferential or longitudinal compo-
nents of aortic expansion. The unaccounted circumfer-
ential and longitudinal components of aortic
expansion are likely to be significant in aneurysms
that either have a high curvature or are a highly
tortuous. The 3D nature of the dirRAWD calculations
also allows for the creation of interpretable data visual-
ization models, which are approachable for surgeons,
can serve as a patient education tool, and could have
utility during surgical planning.
Growth assessment away from area of maximal
diameter. This algorithm models the deformation across
the entire wall of the abdominal aorta and, with its
increased sensitivity compared with maximal diameter
methods, is able to assess for areas of small magnitude
dirRAWD in regions independent of the area of the
maximal diameter. This was observed in several patients
in our cohort who had had subtle, but significant, aortic
expansion in the proximal neck of their aorta (Fig 7).
These smaller magnitude growth areas will often not be
appreciated when using maximal diameter
measurements but might require consideration during
planning for surgical treatment of the aneurysm to avoid
failure of stent-graft sealing.
Fig 6. Scatterplot with trendline comparing radiographic reported yearly aortic expansion rate against the
maximal regional aortic wall deformation (RAWD) with deformable image registration (dirRAWD) derived using
the semiautomated aneurysm segmentation.
Fig 7. Two examples of regional aortic wall deformation
(RAWD) with deformable image registration (dirRAWD)
aortic maps generated by our process, in which maximal
aortic deformation was noted to not be in the region of
the maximal aneurysm diameter.
JVSeVascular Science
Stoecker et al
55
Volume 3, Number C
Compared with prior RAWD techniques. Our proposed
dirRAWD algorithm has several advantages compared
with prior reported RAWD methods. Perhaps the most
significant is that our process does not use custom in-
house developed scripts or software, which have been
often used in prior studies.
14,25-27
The proposed method
uses open-source software applications with graphic
user interfaces, which are inexpensive, widely accessible,
and can be implemented even by those with limited
background in programming.
Previously described RAWD techniques also often
required both significant user expertise in the field and
dedicated high-performance computing capability,
which restricted the applications to smaller cen-
ters.
14,28,29
Our deformation method can generate a
RAWD heatmap on most standard powered computers
for most cases within <15 minutes, with no differences
in measured accuracy. Finally, prior techniques have
not assessed their deformable registration models for ac-
curacy, reporting only the accuracy seen with the rigid
registration process.
14,28,29
We developed a novel
approach for assessing dirRAWD accuracy in this model,
comparing the maximal dirRAWD to the true aortic
deformation on the rigidly registered images (man-
RAWD), and assessing landmark displacement error be-
tween the second time point (ground truth) and the
deformed model as a fiduciary marker.
Clinical correlation. Preliminary work by our group has
shown that the aortic wall stress calculated using a finite
element analysis (FEA) model might correlate and
regionally colocate with RAWD and aortic expansion.
27
However, FEA accuracy is limited by the variable aneu-
rysm wall thickness, degree of mural thrombus, and wall
calcifications, which all significantly increase the model
complexity and requires significant user expertise for
accurate results.
30-32
Assessment of RAWD using this
technique is faster than the calculation of FEA stress and
is easier to interpret and understand by clinical vascular
surgeons. Therefore, the technique might be a better
clinical measure for guiding clinical treatment. Given the
low computational power required for this novel defor-
mation technique, it would be much easier to incorpo-
rate it into medical image viewing software compared
with its FEA counterpart. As noted from our results, at
increased rates of radiographic aortic expansion
(>5 mm/y), we noted a significantly weaker correlation
between the change in the maximal aortic diameter and
aneurysm deformation. This could have resulted from
aneurysms thought to have rapid radiographic diameter
growth, displaying significant changes in their
morphology that might not reliably be assessed with
traditional maximal diameter techniques. Given that
deformation calculations can better assess the entire
aneurysm, this process might better guide management
for this subset of patients. Additionally, because the
technique uses rigid registration of the sequential CT
scans, it ensures that the same area of the aneurysm will
be assessed for interval changes and reduce any human
measurement error. Furthermore, work is ongoing to
assess the accuracy of this technique for the assessment
of aneurysm shrinkage or growth after endovascular
repair. This would be especially relevant for assistance in
managing type II and V endoleaks. The incorporation of
other previously noted geometric characteristics, in
addition to the degree of deformation, would also be
possible with our technique, given the aneurysm has
been segmented.
9
Finally, the effects of patient clinical
factors when determining the significance of aortic
deformation should be assessed (ie, it is unclear whether
one maximal deformation cutoff should be used for all
patients or the mean deformation over an entire aneu-
rysm wall would be more important clinically).
Technique and study limitations. The largest drawback
to determining RAWD is that it requires sequential CT
scans. In contrast, other methods such as FEA only
require a single time-point CT. In clinical practice, this
limitation will be mitigated owing to the routine AAA
imaging surveillance recommended for patients with
AAAs >3 cm in size and repair recommended for those
with AAAs >5.5 cm.
33
The described technique, therefore,
will likely be most applicable to patients requiring AAA
surveillance and undergoing repeated imaging studies.
Additionally, all the patients included in the present
study had undergone contrast-enhanced CT scans. Work
is ongoing to assess whether using nonecontrast-
enhanced CT scans with 2- to 3-mm slices will signifi-
cantly affect final dirRAWD accuracy. This technique still
requires user input for manual correction of the semi-
automated aortic segmentation. However, in the present
study, almost all the patients had required minimal user
input (<5 minutes) for accurate dirRAWD calculations.
Finally, although the present study only included 15 pa-
tients, limiting the generalizability to all patients with
AAAs, we have been able to characterize this developing
technique and demonstrate its accuracy.
CONCLUSIONS
The present study has demonstrated the feasibility of
accurate 3D RAWD calculations with application to
AAAs, which can be performed using open source soft-
ware and low computing power. This process accom-
plishes submillimeter accuracy in determining aortic
growth within a 1- to 2-year period, with minimal user
input. No significant differences were found in RAWD ac-
curacy when using manual aortic aneurysm segmenta-
tion compared with semiautomated aortic
segmentation. Thus, in the future, using either semiauto-
mated or fully automated aortic segmentation during
dirRAWD calculations should be preferred because of
the significant user input required for manual segmenta-
56
Stoecker et al
JVSeVascular Science
--- 2022
Table II. Comparison between manRAWD and algorithm-derived maximal RAWD calculated using our deformable image
registration pipeline
PT. NO.
Maximal man-
RAWD, mm/y
Maximal dirRAWD derived using manual
segmentation, mm/y
Maximal dirRAWD derived using semiautomated
segmentation, mm/y
1 1.5 1.6 1.7
2 2.5 3.1 2.9
3 1.2 1.3 1.1
4 1.4 1.5 1.7
5 5.1 5.0 5.4
6 4.1 3.7 4.9
7 2.3 2.3 2.2
8 1.9 1.8 2.0
9 5.9 5.6 6.0
10 6.4 6.4 6.3
11 6.7 6.4 6.3
12 5.4 5.8 5.2
13 3.8 4.0 3.8
14 9.2 8.3 8.5
15 7.0 6.8 6.7
MEDIAN
(IQR)
4.1 (2.1-6.2) 4.0 (2.1-6.1) 4.9 (2.1-6.2)
IQR,
Interquartile range;
dirRAWD,
regional aortic wall deformation using deformable image registration;
manRAWD,
manually determined maximal
regional aortic wall deformation;
Pt. No.,
patient number;
RAWD,
regional aortic wall deformation.
Table III. Comparison between true IMA deformation measured manually and deformable image registration-derived IMA
deformation
PT. NO.
True IMA deformation,
mm/y
IMA deformation using manual segmen-
tation, mm/y
IMA deformation using semiautomated segmen-
tation, mm/y
1 1.2 1.1 1.0
2 1.1 1.3 1.2
3 1.0 0.9 1.1
4 1.2 1.1 1.0
5 6.5 6.4 6.5
6 3.5 3.3 3.1
7 4.2 3.7 3.8
8 2.7 2.7 2.8
9 2.0 2.1 2.1
10 8.0 5.2 5.5
11 5.0 5.0 4.6
12 7.3 7.3 6.4
13 2.5 2.6 2.5
14 6.9 12.1 10.0
15 3.8 3.3 3.4
MEDIAN
(IQR)
3.5 (1.6-5.8) 3.3 (1.7 - 5.1) 3.1 (1.7 - 5.1)
IMA,
Inferior mesenteric artery;
IQR,
interquartile range;
Pt. No.,
patient number.
JVSeVascular Science
Stoecker et al
57
Volume 3, Number C
tion. The described technique for the assessment of dir-
RAWD might improve the accuracy of aortic imaging
surveillance, inform clinical decision-making, further
research of aortic research questions, and help provide
information on the natural history of aortic disease.
AUTHOR CONTRIBUTIONS
Conception and design: JS, KE, AP, AV, BJ
Analysis and interpretation: JS, KE, AP, BJ
Data collection: JS, KE
Writing the article: JS, KE, BJ
Critical revision of the article: JS, KE, AP, AV, BJ
Final approval of the article: JS, KE, AP, AV, BJ
Statistical analysis: JS
Obtained funding: Not applicable
Overall responsibility: JS
REFERENCES
1. Beckman JA. Aortic aneurysms: pathophysiology, epidemiology, and
prognosis. In: Creager MA, Dzau VJ, Loscalzo J, editors. Vascular
Medicine. Philadelphia, PA: Saunders Elsevier Inc; 2006.
2. Schermerhorn ML, Bensley RP, Giles KA, Hurks R, OʼMalley AJ,
Cotterill P, et al. Changes in abdominal aortic aneurysm rupture and
short-term mortality, 1995-2008: a retrospective observational study.
Ann Surg 2012;256:651-8.
3. Darling RC, Messina CR, Brewster DC, Ottinger LW. Autopsy study of
unoperated abdominal aortic aneurysms: the case for early resec-
tion. Circulation 1977;56:II161-4.
4. Nicholls SC, Gardner JB, Meissner MH, Johansen KH. Rupture in small
abdominal aortic aneurysms. J Vasc Surg 1998;28:884-8.
5. Hong H, Yang Y, Liu B, Cai W. Imaging of abdominal aortic aneu-
rysm: the present and the future. Curr Vasc Pharmacol 2010;8:
808-19.
6. Elefteriades JA, Farkas EA. Thoracic aortic aneurysm. J Am Coll Car-
diol 2010;55:841-57.
7. Davies RR, Gallo A, Coady MA, Tellides G, Botta DM, Burke B, et al.
Novel measurement of relative aortic size predicts rupture of
thoracic aortic aneurysms. Ann Thorac Surg 2006;81:169-77.
8. Parr A, Jayaratne C, Buttner P, Golledge J. Comparison of volume and
diameter measurement in assessing small abdominal aortic aneu-
rysm expansion examined using computed tomographic angiog-
raphy. Eur J Radiol 2011;79:42-7.
9. Pappu S, Dardik A, Tagare H, Gusberg RJ. Beyond fusiform and
saccular: a novel quantitative tortuosity index may help classify
aneurysm shape and predict aneurysm rupture potential. Ann Vasc
Surg 2008;22:88-97.
10. Pérez EA, Rojas-Solórzano LR, Finol E. Geometric predictors of
abdominal aortic aneurysm maximum wall stress. Chem Eng Trans
2016;49:73-8.
11. Georgakarakos E, Ioannou CV, Kamarianakis Y, Papaharilaou Y,
Kostas T, Manousaki E, et al. The role of geometric parameters in the
prediction of abdominal aortic aneurysm wall stress. Eur J Vasc
Endovasc Surg 2010;39:42-8.
12. Siika A, Lindquist Liljeqvist M, Hultgren R, Gasser TC, Roy J. Aortic
lumen area is increased in ruptured abdominal aortic aneurysms
and correlates to biomechanical rupture risk. J Endovasc Ther
2018;25:750-6.
13. Georgakarakos E, Ioannou CV, Papaharilaou Y, Kostas T,
Katsamouris AN. Computational evaluation of aortic aneurysm rupture
risk: what have we learned so far? J Endovasc Ther 2011;18:214-25.
14. Burris NS, Hoff BA, Kazerooni EA, Ross BD. Vascular deformation
mapping (VDM) of thoracic aortic enlargement in aneurysmal dis-
ease and dissection. Tomography 2017;3:163-73.
15. Klein S, Staring M, Pluim JPW. Evaluation of optimization methods
for nonrigid medical image registration using mutual information
and B-splines. IEEE Trans Image Proc 2007;16:2879-90.
16. Galbán CJ, Han MK, Boes JL, Chughtai KA, Meyer CR, Johnson TD,
et al. Computed tomography-based biomarker provides unique
signature for diagnosis of COPD phenotypes and disease progres-
sion. Nat Med 2012;18:1711-5.
17. Boes JL, Hoff BA, Hylton N, Pickles MD, Turnbull LW, Schott AF, et al.
Image registration for quantitative parametric response mapping of
cancer treatment response. Transl Oncol 2014;7:101-10.
18. Keith L, Ross BD, Galbán CJ, Luker GD, Galbán S, Zhao B, et al.
Semiautomated workflow for clinically streamlined glioma para-
metric response mapping. Tomography 2016;2:267-75.
19. Pouch AM, Tian S, Takebe M, Yuan J, Gorman R, Cheung AT, et al.
Medially constrained deformable modeling for segmentation of
branching medial structures: application to aortic valve segmenta-
tion and morphometry. Med Image Anal 2015;26:217-31.
20. Yushkevich PA, Piven J, Hazlett HC, Smith RG, Ho S, Gee JC, et al.
User-guided 3D active contour segmentation of anatomical struc-
tures: significantly improved efficiency and reliability. Neuroimage
2006;31:1116-28.
21. Yushkevich PA, Pluta J, Wang H, Wisse LEM, Das S, Wolk D. IC-P-174:
fast automatic segmentation of hippocampal subfields and medial
temporal lobe subregions in 3 Tesla and 7 Tesla T2-weighted MRI.
Alzheimers Demen 2016;12:P126-7.
22. Avants BB, Epstein CL, Grossman M, Gee JC. Symmetric diffeomor-
phic image registration with cross-correlation: evaluating automated
labeling of elderly and neurodegenerative brain. Med Image Analysis
2008;12:26-41.
23. Avants BB, Tustison NJ, Song G, Cook PA, Klein A, Gee JC.
A reproducible evaluation of ANTs similarity metric performance in
brain image registration. Neuroimage 2011;54:2033-44.
24. Zijdenbos AP, Dawant BM, Margolin RA, Palmer AC. Morphometric
analysis of white matter lesions in MR images: method and valida-
tion. IEEE Trans Med Imaging 1994;13:716-24.
25. Gao X, Boccalini S, Kitslaar PH, Budde RP, Tu S, Lelieveldt BP, et al. A
novel software tool for semi-automatic quantification of thoracic
aorta dilatation on baseline and follow-up computed tomography
angiography. Int J Cardiovasc Imaging 2019;35:711-23.
26. Kauffmann C, Tang A, Therasse E, Giroux M-F, Elkouri S, Melanson P,
et al. Measurements and detection of abdominal aortic aneurysm
growth: accuracy and reproducibility of a segmentation software.
Eur J Radiol 2012;81:1688-94.
27. Auer M, Gasser TC. Reconstruction and finite element mesh gener-
ation of abdominal aortic aneurysms from computerized tomogra-
phy angiography data with minimal user interactions. IEEE Trans
Med Imaging 2010;29:1022-8.
28. López-Linares K, García I, García A, Cortes C, Piella G, Macía I, et al.
Image-based 3D characterization of abdominal aortic aneurysm
deformation after endovascular aneurysm repair. Front Bioeng Bio-
technol 2019;7:267.
29. Subasic M, Loncaric S, Sorantin E. 3-D image analysis of abdominal
aortic aneurysm. Stud Health Technol Inform 2000;77:1195-200.
30. Shang EK, Nathan DP, Sprinkle SR, Vigmostad SC, Fairman RM,
Bavaria JE, et al. Peak wall stress predicts expansion rate in descend-
ing thoracic aortic aneurysms. Ann Thorac Surg 2013;95:593-8.
31. Gasser TC, Auer M, Labruto F, Swedenborg J, Roy J. Biomechanical
rupture risk assessment of abdominal aortic aneurysms: model
complexity versus predictability of finite element simulations. Eur J
Vasc Endovasc Surg 2010;40:176-85.
32. Polzer S, Gasser TC, Bursa J, Staffa R, Vlachovsky R, Man V, et al.
Importance of material model in wall stress prediction in abdominal
aortic aneurysms. Med Eng Phys 2013;35:1282-9.
33. Chaikof EL, Dalman RL, Eskandari MK, Jackson BM, Lee WA,
Mansour MA, et al. The Society for Vascular Surgery practice guide-
lines on the care of patients with an abdominal aortic aneurysm.
J Vasc Surg 2018;67:2-77.e2.
Submitted Mar 31, 2021; accepted Nov 17, 2021.
58
Stoecker et al
JVSeVascular Science
--- 2022
Supplementary Table. Detailed demographics of 15 included patients
Pt.
No.
Age,
years SexHypertension
Statin
use
eGFR,
mL/min
Family history of
aneurysm
True maximal defor-
mation, mm/y
Initial maximal aortic
diameter, mm
Final maximal aortic
diameter, mm
1 63 M Yes No 90 No 1.5 42 44
2 79 M No Yes 50 No 2.5 39 42
3 73 M Yes Yes 70 No 1.2 31 32
4 83 F Unknown Unknown 70 No 1.4 31 32
5 72 M Yes Yes 80 Yes 5.1 45 53
6 70 F No No 70 No 4.1 55 62
7 62 F Yes Yes 90 No 2.3 27 31
8 59 M Yes No 110 No 1.9 51 54
9 46 M Yes No 120 Yes 5.9 60 65
10 58 F Yes Yes 100 No 6.4 36 43
11 46 M Yes No 120 No 6.7 44 58
12 68 F No No 90 No 5.4 48 55
13 80 F No Yes 50 No 3.8 32 38
14 77 M Yes Yes 80 No 9.2 54 64
15 65 M Yes No 90 Yes 7.0 45 52
eGFR,
Estimated glomerular filtration rate (using the Chronic Kidney Disease Epidemiology Collaboration creatinine equation, 2021);
F,
female;
M,
male.
JVSeVascular Science
Stoecker et al
59
Volume 3, Number C
Supplementary Fig. A, Baseline computed tomography angiogram (CTA) as shown in ITK-SNAP. B, Baseline CTA
with single label segmentation of the aortic aneurysm. ITK-SNAP is able to automatically generate a 3D model of the
segmentation (lower left) and export this as a mesh of the outer wall of the aortic aneurysm. C, Baseline CTA and
aneurysm segmentation compared with follow-up CTA as shown in ITK-SNAP. Initially, a significant translational
difference was present between the two CTAs, as noted in the sagittal and coronal images, which had led to initial
poor alignment of the aneurysm between scans. D, Baseline CTA and dilated aortic aneurysm segmentation
compared with follow-up CTA, as shown in ITK-SNAP. Dilated segmentation was used during both rigid and
deformable registration to restrict the algorithms to the area of interest and improve both accuracy and efficiency.
E, Baseline CTA and dilated aneurysm segmentation compared with follow-up CTA, as shown in ITK-SNAP after rigid
registration. Initial rigid registration was performed using a multiresolution scheme with 16to 8down-sampling
to grossly align the images. Rigid registration was then repeated with computation of the similarity metric confined
to the dilated aortic segmentation, first with 4down-sampling and then at full resolution. F, Baseline CTA and
aneurysm segmentation compared with follow-up CTA, as shown in ITK-SNAP, after rigid registration. Excellent
alignment can now be visualized between the two time points in both the aneurysm and bony landmarks. The
difference between the size of the baseline aneurysm segmentation and the follow-up CTA aneurysm size indicated
interval growth of the aneurysm. The rigid registration matrix (detailing the amount of translation required to align
the images) was then exported to a text file.
Greedy
was then used with the baseline CTA, follow-up CTA, dilated
aneurysm segmentation, and rigid registration matrix to automatically generate a 3D deformation field. For
deformable image registration, we used a normalized cross-correlation similarity metric with a gradient descent
optimization for 100 iterations at a course level of 4, 50 iterations at an intermediate level of 2, and 20 iterations at
full resolution. For further algorithm optimization, the deformable registration was limited to a 10 10 10
neighborhood around each voxel, and the baseline aortic segmentation was dilated by 1.5 cm and used to restrict
the deformation algorithm to the area of interest. A smoothing gradient with a sigma of 1.732 voxels was applied to
the image match metric at each voxel, and the entire deformation field was smoothed with a gradient sigma of
0.707 voxel after each iteration to limit error and generate a smooth field. G, Baseline aneurysm segmentation on the
left compared with the 3D deformation field on the right (displayed as the magnitude of deformation, with
white
indicating increased deformation). F, As expected from the rigid registration results, maximal deformation was
noted at the anterior wall of the aneurysm. H, The patient’s baseline aneurysm mesh is displayed (Left) as shown in
MeshLab. (Center)
Greedy
was used to apply the 3D deformation field to this baseline aneurysm mesh. (Right)
Magnitude difference experienced by every vertex on the baseline aneurysm mesh with application of the defor-
mation field (ie, difference between the left and middle images) with smoothing to remove segmentation artifact.
60
Stoecker et al
JVSeVascular Science
--- 2022
