Deep Learning-Based Segmentation of the Atherosclerotic Carotid Plaque in Ultrasonic Images

Abstract

Early stroke risk stratification in individuals with carotid atherosclerosis is of great importance, especially in high-risk asymptomatic (AS) cases. In this study, we present a new computer-aided diagnostic (CAD) system for the automated segmentation of the atherosclerotic plaque in carotid ultrasound (US) images and the extraction of a refined set of ultrasonic features to robustly characterize plaques in carotid US images and videos (AS vs symptomatic (SY)). So far, we trained a UNet model (16 to 256 neurons in the contracting path; the reverse, for the expanding path), starting from a dataset of 201 (AS = 109 and SY = 92) carotid US videos of atherosclerotic plaques, from which their first frames were extracted to prepare three subsets, a training, an internal validation, and final evaluation set, with 150, 30 and 15 images, respectively. The automated segmentations were evaluated based on manual segmentations, performed by a vascular surgeon. To assess our model’s capacity to segment plaques in previously unseen images, we calculated 4 evaluation metrics (mean ± std). The evaluation of the proposed model yielded a 0.736 ± 0.10 Dice similarity score (DSC), a 0.583 ± 0.12 intersection of union (IoU), a 0.728 ± 0.10 Cohen’s Kappa coefficient (KI) and a 0.65 ± 0.19 Hausdorff distance. The proposed segmentation workflow will be further optimized and evaluated, using a larger dataset and more neurons in each UNet layer, as in the original model architecture. Our results are close to others published in relevant studies.