A Machine Learning Approach For CTA-based Plaque Characterization and Stroke Risk Prediction in Carotid Artery Atherosclerosis

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT Carotid artery atherosclerosis is a major vascular risk factor and accounts for approximately 15% of all strokes. A major risk marker in patients with carotid atherosclerosis has been the degree of narrowing, or stenosis, of the carotid artery lumen. While stenosis is often quantified via angiography, imaging can also provide detailed assessments of plaque. Our project is motivated by converging data that correlate vulnerable plaque elements, which can be captured with imaging, with increased stroke risk. Identifying high-risk or vulnerable carotid plaques before a stroke occurs is important because stroke prevention treatments, like carotid endarterectomy or stenting, carry risks and ideally should only be performed only those patients at highest risk of stroke. CTA (computed tomographic angiography) is an attractive tool for plaque imaging since it is not operator dependent, can be quickly performed, and is more widely available than MRI. Although CTA offers significant potential to evaluate these plaque features, small studies have not reached a consensus regarding their reliability and clinical relevance. In this project, we plan to explore the utility of CTA for the detailed carotid vessel wall imaging by employing a unique, large-scale clinical dataset and advanced algorithms. Our overarching objective in this R21 project is to conduct developmental and interdisciplinary research that will lay the foundation for the implementation and validation of novel CTA-based technologies that can be adopted in the risk stratification of patients with carotid atherosclerosis. Our central hypothesis is that there are CTA-based carotid plaque features that can be reliably extracted and used for stroke risk stratification, which will be more sensitive and specific than standard stenosis grading. To pursue our objective, we will pursue two specific aims: In Specific Aim 1 we plan to optimize the use of human reader defined plaque features in predicting culprit carotid plaques. We will perform a blinded, multi-reader study of CTA-derived carotid plaque features in a large scale clinical dataset to test the association between CTA-derived human-defined features and stroke, and compute accuracy metrics. In Specific Aim 2, we plan to develop algorithms to automatically characterize and discriminate culprit carotid plaque in CTA. We will implement and test image processing algorithms that automatically compute from a CTA scan stroke-associated carotid artery plaque features (from Aim 1), and then train a machine learning algorithm to distinguish culprit from asymptomatic carotid artery plaques. We believe that this R21 study is significant because it will establish a novel, machine learning-aided imaging strategy which can aid in identifying high-risk carotid artery plaques before they cause stroke and when they can be properly treated to prevent stroke from occurring in the future.