This document presents a method to automatically measure abdominal aortic aneurysm (AAA) wall thickness from medical images using a regression algorithm. Two reference models (a porcine aorta and synthetic AAA phantom) were used to train and validate the algorithm. A support vector machine regression model using normalized pixel intensity profiles achieved a mean error of 0.262 mm, providing more accurate wall thickness estimates than direct measurements from low-resolution CT images. Further optimization and testing on real patient data is needed to validate the algorithm's ability to provide new insights into AAA development and rupture risk.

1. Toward Automatic Wall Thickness Measurements for Abdominal Aortic Aneurysm Images
Bara Aldasouqi1, Amin Jourabloo2, Joseph Roth2, Xiaoming Liu2, Seungik Baek1
Departments of 1Mechanical Engineering and 2Computer Science and Engineering, Michigan State University, East Lansing, MI
Introduction
Abdominal Aortic Aneurysm (AAA) is a leading
cause of death in the elderly population. Medical
images such as Computed Tomography (CT) are
used to diagnose AAA. Biomechanical research
also utilizes CT scans of AAA patients to extract
patient-specific vessel geometry for predicting the
risk of its rupture. However, the low resolution of
CT images often makes it difficult to estimate
reliable thickness measurements. This work
seeks to develop a regression algorithm that
utilizes the intensity profile across the AAA wall
boundary to estimate the local wall thickness. A
porcine aorta model and a synthetic AAA
phantom serve as reference data sets to train the
algorithm and validate its performance.
Methods
Two models were created to train the algorithm.
First, porcine aorta specimens were embedded in
paraffin molds. A micro-CT scan was performed
on the specimens for precise wall thickness
measurements. Secondly, a AAA phantom was
manufactured, replicating the various features
and densities present in a real patient’s AAA
(e.g., intraluminal thrombus). It was designed
such that wall thickness could be determined at
any location. For both cases, the specimens were
scanned using a clinical machine. The data was
split into 130 training samples and 40 testing
samples. To predict the wall thickness, we
trained a regression method using the pixel
intensity along profile lines perpendicular to the
wall as input features. Initially, a simple linear
regression (with regularizer λ = 0.1) was trained
and tested. This was also compared to results
from a 3rd degree polynomial Support Vector
Machine (SVM) regression (trained with γ = 1 /
length of feature vectors). In order to optimize the
performance, several feature representations
were tested such as normalized data, histogram
of data, and Fast Fourier Transform (FFT).
Results
Table 1 lists the accuracy for each type of
regression on each data model. The SVM
regression was found to perform better than the
linear regression, so it was chosen for further
optimization of the algorithm. Several feature
representations were tested as previously
mentioned. The best results are listed in Table 2.
The porcine model and the phantom model have
different characteristics, so each is optimized
using different feature representations. When
both are combined, normalization of the input
features yields the best results.
Conclusions
The developed algorithm is the first to estimate
AAA wall thickness using a data-driven
regression model, providing increased
confidence in resulting estimations. The algorithm
was successfully trained and validated on
reference data, yielding a mean error magnitude
of 0.262 mm, which is significant on images with
pixel size ranging from 0.47 to 0.70 mm.
The algorithm will be further optimized by testing
different features (including 2nd and 3rd order
statistics) and additional regression models.
Furthermore, the algorithm will be tested on real
patient data. We anticipate finding patterns in
wall thickness values in the spatial domain and in
the time domain. The existence of such trends
would further validate this method and, most
importantly, provide new insight into AAA
development.
Tables
Table 1 Accuracy results of the two regression models
applied to the three data models. Best results are bolded.
Data
Model
Regression
Model
RMS Error
(mm)
Error Mean
(mm)
Porcine
Linear 0.628 0.383
SVM 0.527 0.351
Phantom
Linear 0.506 0.363
SVM 0.297 0.222
Both
Linear 1.554 1.027
SVM 0.644 0.403
Table 2 Accuracy results for feature representations used
within the SVM regression. Only the best results are listed.
Data
Model
Feature
Transform
RMS Error
(mm)
Error Mean
(mm)
Porcine Histogram 0.354 0.250
Phantom Normalized 0.256 0.219
Both Normalized 0.336 0.262