3D ultrasound quantification of carotid vessel wall thickness for stroke risk stratification

Abstract

A biomarker is defined as a "a naturally occurring molecule, gene, or characteristic by which a particular pathological or physiological process, disease, etc. can be identified". It quantifies how much or how little of a certain process has occurred. This project aims to quantify a biomarker that describes how much plaque build has occurred in the carotid artery vessel wall. This information is highly crucial to determining the risk of stroke in a person.

Carotid atherosclerosis more commonly known as a stroke is a leading cause of death and disability around the world with more than 17.3 million people suffering a stroke each year. In the United States alone someone has a stroke every 40 seconds. A carotid stroke occurs when a blood clot or plug stops the brain from receiving any more blood. The lack of oxygen resulting from the negligent blood flow causes cells in the brain to die leading to either death or mental disability. Though the consequences of this disease are extremely serious, simple changes to diet and lifestyle can reverse the progression of the disease.

This project employs machine learning techniques like logistic regression, support vector classification and an artificial neural network to accurately determine a biomarker that can let one know how far the disease has progressed or regressed. Ultrasound scans of over 50 subjects that were equally separated into either treatment or placebo groups were taken on baseline dates and a follow up scan after at least a year was taken. The treatment that these subjects underwent was a regular consumption of pomegranate juice over the course of at least a year. These scans were converted into vessel wall plus plaque thickness (VWT) maps and were used as a dataset to perform machine learning.

The data was then used to train model to solve a classification problem and then the model was repurposed to output a value that would determine how well certain treatments have worked, pomegranate juice in this case. The models trained were then compared to traditional methods like VWTavg and VWTweighted and were shown to have a significantly lower p-value.