Carotid Vessel Wall Segmentation and Characterization in Three-Dimensional Ultrasound Images by Domain Adaptation

Abstract

Identification of patients with high stroke risk with carotid atherosclerosis is crucial, as timely interventions greatly reduce their stroke risk. The use of three-dimensional ultrasound (3DUS) measurements are reported to be sensitive to risk stratification and treatment monitoring, but the requirement of manual segmentation is time-consuming and prone to observer variability. Although supervised deep-learning segmentation models have been proposed, training of these models requires a sizeable manually segmented training set, making larger clinical studies prohibitive.

In this thesis, first, an adversarial framework, called unsupervised shape-and-texture-based generative adversarial network (USTGAN), is proposed to optimize the pre-trained convolutional neural network (CNN) without requiring the manual segmentation to supervise the fine-tuning process. USTGAN integrates a novel texture-based discriminator with a shape-based discriminator, which together provide feedback for the CNN to segment the target images in a manner similar to the source images. The texture-based discriminator increases the accuracy of the segmentation network in locating the artery, thereby reducing the number of failed segmentations. The occurrence of failed segmentation is further reduced by a self-checking mechanism to detect longitudinal discontinuities of the artery and by self-correction strategies involving surface interpolation followed by a case-specific tuning of the CNN. USTGAN achieved a Dice similarity coefficient (DSC) of 85.7%±13.0% in lumen-intima boundary (LIB) and 86.2%±10.6% in media-adventitia boundary (MAB), showing improvement over the baseline performance of 74.6%±30.7% in LIB and 75.7%±28.9% in MAB.

Besides evaluating segmentation accuracy, the ability to predict events is also evaluated. The extracted carotid vessel wall region is compared with a method based on the volume and texture of manually segmented carotid plaque regions, as it was shown to predict cardiovascular events in a previous clinical study. The segmentation network, fine-tuned by USTGAN, was applied to extract the carotid artery region. Then a set of 376 measures was applied to extract features of the region of interest (ROI). Multi-view latent space projection (MVLSP) and linear discriminant analysis (LDA) were used to generate biomarkers in independent training sets using leave-one-out cross-validation. The proposed texture-change-based risk score (Δ𝑇𝑒𝑥𝑡𝑢𝑟𝑒 𝑠𝑐𝑜𝑟𝑒) and the baseline texture risk score (𝐵𝐿 𝑇𝑒𝑥𝑡𝑢𝑟𝑒 𝑠𝑐𝑜𝑟𝑒) predicted cardiovascular events (Kaplan-Meier logrank, 𝑃 < 0.001 for both), whereas the change of vessel wall volume (Δ𝑉𝑊𝑉 ) did not. The areas under the receiver operating characteristic curves of Δ𝑇𝑒𝑥𝑡𝑢𝑟𝑒 𝑠𝑐𝑜𝑟𝑒 and 𝐵𝐿 𝑇𝑒𝑥𝑡𝑢𝑟𝑒 𝑠𝑐𝑜𝑟𝑒 were 0.88 and 0.75, compared to 0.53 for Δ𝑉𝑊𝑉 and 0.74-0.79 for plaque texture and volume parameters previously reported.

The challenge of cross-domain application highlights a need for more generalizable models. Folding Fan ResNet (FFRN), a CNN designed to enhance cross-domain generalizability, is proposed. FFRN incorporates a novel skip connection pathway called Folding Fan Connection (FFC), which improves encoder-decoder feature fusion through multiple additions and convolutional operations. This design reduces parameter redundancy while maintaining efficient feature propagation, leading to increased generalizability compared to existing CNNs. Additionally, the quadratic mapping loss (QML) was proposed to improve model fine-tuning. QML addresses arterial mislocalization by deferring the optimization of incorrectly localized boundaries while prioritizing adjustments to correctly segmented regions. This strategy prevents the reinforcement of erroneous features and allows for better modeling of carotid structures. FFRN demonstrated higher generalizability than four state-of-the-art segmentation models, achieving DSCs of 84.0%±18.2% in LIB and 84.6%±17.1% in MAB before fine-tuning. With QML fine-tuning, the DSCs increased to 88.9%±8.2% in LIB and 90.6%±8.1% in MAB, outperforming all competing approaches, even after their performances were improved by QML.

The framework proposed in this thesis can improve the clinical utility of 3DUS carotid measurements in multi-center trials and in clinical practice. The proposed segmentation network provides accurate and consistent prediction across medical centers, and fine-tuning algorithms allow calibration for differences in 3DUS scanning parameters and patient characteristics. The proposed event prediction framework provides more sensitive biomarkers for identifying patients with a high risk of cardiovascular events.

Date of Award	10 Sept 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Ho Man CHAN (Supervisor) & Chi Yuen Bernard CHIU (External Co-Supervisor)