Development of 3D Ultrasound Measurement Tools for Longitudinal Assessment of Carotid Atherosclerosis


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date18 May 2023


Stroke is a leading cause of death and disability throughout the world, and carotid atherosclerosis is a major source of emboli leading to ischemic stroke. The development of 3D ultrasound carotid imaging has allowed machine learning techniques-based biomarkers for the longitudinal assessment of carotid atherosclerosis. Improved identification of high-risk patients and sensitive techniques for monitoring carotid plaque response to therapy will enhance the management of these patients and reduce the risk of stroke. Therefore, there is a critical requirement for measurement tools for the cost-effective serial monitoring of carotid atherosclerosis for stroke risk stratification and treatment of atherosclerosis.

The compositions of lipid, calcification, and fibrous tissue determine a plaque’s vulnerability. Textural features have been utilized as proxies for plaque composition assessment. In this thesis, we developed a biomarker based on the changes of 376 plaque textural features measured from 3D ultrasound carotid images. To generate a scalar biomarker for each subject, principal component analysis (PCA) was first applied to reduce the dimensionality of feature vectors, and elements in the reduced feature vectors produced by PCA were then weighted using locality preserving projections (LPP) to capture essential patterns exhibited locally in the feature space. In addition, a physical understanding of the biomarker was established by identifying important textural features that contribute to the sensitivity of the biomarker. The proposed biomarker was more able to discriminate plaque changes exhibited by the pomegranate and placebo groups than total plaque volume (TPV) according to the result of t-tests (TPV: 𝑝 = 0.34, Proposed biomarker: 𝑝 = 0.000015). The sample size required by the new biomarker to detect a significant effect was 20 times smaller than that required by TPV.

We developed a new method to measure the voxel-based vessel-wall-plus-plaque volume (VVol). In addition to quantifying local thickness change as in the previously introduced vessel wall-plus-plaque thickness (VWT) metric, VVol further considers the circumferential change associated with vascular remodeling. 3D ultrasound images were acquired at baseline and one year after. The vessel wall region was divided into small voxels with the percent VVol-Change (Δ𝑉𝑉𝑜𝑙%) computed by taking the percent volume difference between corresponding voxels in the baseline and follow-up images. A 3D carotid atlas was developed to allow visualization of the local thickness and circumferential change patterns in the pomegranate versus the placebo groups. A new subject-based biomarker was obtained by computing the mean Δ𝑉𝑉𝑜𝑙% over the entire 3D map for each patient (meanΔ𝑉𝑉𝑜𝑙%). meanΔ𝑉𝑉𝑜𝑙% detected a significant difference between patients randomized to pomegranate juice/extract and placebo groups (𝑝 = 0.0002). The number of patients required by meanΔ𝑉𝑉𝑜𝑙% to establish statistical significance was approximately a third of that required by the local VWT biomarker.

Traditional machine learning-based methods are limited by their handcrafted features, while current deep learning-based methods are like black boxes which cannot provide interpretability of the results. To address the problems, we proposed a novel interpretable deep automatic biomarker learning framework for automatically learning disease-related volume and texture-based features and providing interpretability for aiding medical decisions. First, we developed an interpretable Siamese network (INS-Net) to automatically compute volume and texture-based biomarkers (𝐴𝑢𝑡𝑜𝑉𝑇) that leverage the slice-wise information. Second, to address the problems of lack of ground truth values for biomarkers, we proposed a weakly-supervised contrastive loss with a plaque-focus constraint, which incorporates domain expertise into the model. Third, we proposed a new visual interpretation method, interpretable slice score activation maps (ISAM), to show important regions relevant to 𝐴𝑢𝑡𝑜𝑉𝑇. The experimental results demonstrated that the proposed 𝐴𝑢𝑡𝑜𝑉𝑇 was more sensitive to the difference exhibited between patients in pomegranate juice/extract and placebo groups, and 𝐴𝑢𝑡𝑜𝑉𝑇 required a lower number of patients to establish statistical significance, compared with traditional biomarkers. In addition, our interpretable results showed a superior visualization compared with existing methods.

The proposed measurement tools in this thesis will allow more sensitive detection of treatment effect and more accurate risk stratification. The increased discriminative power conferred by the proposed measurement tools could substantially reduce the sample size, duration, and therefore cost required to establish the efficacy of novel dietary/lifestyle interventions.