Automated segmentation and classification of the atherosclerotic carotid plaque in ultrasound videos

Abstract

The automated and reliable delineation of atherosclerotic carotid plaques in ultrasound (CUS) videos is of significant clinical relevance for management of the disease and the prediction of future stroke events. To facilitate stroke risk assessment, in this study, we propose an integrated software system for the automated segmentation and classification of atherosclerotic carotid plaques in longitudinal CUS videos, which was evaluated using 10 CUS videos, from 10 patients (5 Asymptomatic, AS, and 5 Symptomatic, SY). The proposed methodology involves the following steps: a) CUS video frame (VF) resolution and intensity normalization, b) speckle reduction filtering, c) Motion-mode state-based cardiac cycle (CC) identification, d) deep learning (DL)-based plaque segmentation, e) extraction and selection of plaque region of interest (ROI)-specific textural features, and f) machine learning (ML)-based plaque classification. Initially, one CC (cardiac diastole-systole-diastole) was selected per CUS video, and the CC's consecutive VFs were identified and reduced in number to exclude redundant VFs. All standardized VFs per patient were extracted, cropped and resized to mainly accommodate the ROI and were fed into a priorly trained and evaluated 2-dimensional DL plaque segmentation model. For each VF, the DL-based segmented plaque ROI was projected onto its primary resolution-normalized VF counterpart, from which textural and amplitude modulation-frequency modulation (AM-FM) plaque ROI features were extracted. Statistical analysis on the total AS and SY VFs was used for feature selection. We identified 2 plaque-originating AM-FM features, which exhibited statistically significant differences between the AS and SY standardized VFs (p<0.05), followed by 3 textural features (p<0.05). To finalize our system, in a future study, the strong AM-FM AS/SY descriptors, identified here, will be evaluated alone or in combinations with other plaque-descriptive features, in machine learning (ML)-based plaque classification, using a larger CUS video sample.