Robust Deep Learning–Based Segmentation of Carotid Artery Structures in Ultrasound Images using various U-net architectures

Carotid artery ultrasound imaging is a cornerstone of non-invasive vascular assessment and is extensively used for the early detection, risk stratification, and monitoring of atherosclerotic cardiovascular disease. Stroke remains one of the leading causes of mortality and long-term disability worldwide, and a significant proportion of ischemic strokes are attributable to carotid artery atherosclerosis. Ultrasound imaging of the carotid arteries enables visualization of the arterial wall, lumen morphology, and atherosclerotic plaque characteristics in real time, without exposing patients to ionizing radiation. Quantitative analysis of carotid ultrasound images depends critically on the accurate segmentation of anatomical structures such as the lumen–intima boundary (LIB), the media–adventitia boundary (MAB), intima–media thickness (IMT), lumen diameter, and plaque regions. These measurements are clinically relevant biomarkers for assessing subclinical atherosclerosis, predicting cardiovascular events, and evaluating therapeutic response [1–4]. Various deep learning based U-net architectures have been evaluated for segmentation of carotid artery structures and validated against metrics like Dice Score, IOU, MAE etc.