Application of Artificial Intelligence in Vulnerable Carotid Atherosclerotic Plaque Assessment—A Scoping Review

Background and Objectives: Accurate evaluation of vulnerable carotid atherosclerotic plaques remains essential for preventing ischemic stroke. Conventional imaging modalities such as ultrasound and computed tomography angiography (CTA) have limited capacity to identify histopathological features of plaque instability, including fibrous cap rupture, lipid-rich necrotic core, and intraplaque hemorrhage. Artificial intelligence (AI) techniques—particularly deep learning (DL) and radiomics—have recently emerged as valuable adjuncts to standard imaging, achieving AUC values of 0.83–0.99 across modalities in identifying vulnerable plaques. This scoping review summarizes the available evidence on the application of AI in the detection and assessment of vulnerable carotid plaques. Methods: A systematic search of the English-language literature was conducted in MEDLINE, SCOPUS, and CENTRAL from 2000 to 30 June 2025, following the PRISMA-ScR framework. Eligible studies applied AI-based approaches (machine learning, deep learning, or radiomics) to evaluate carotid plaque vulnerability using ultrasound, CTA, or MRI. Extracted outcomes included diagnostic performance, correlation with histopathology or neurological events, and predictive modeling for stroke risk. Results: Of 201 records screened, 12 studies met inclusion criteria (ultrasound = 6; CTA = 4; high-resolution MRI = 2; publication years 2021–2025). All reported receiver operating characteristic area-under-the-curve (ROC-AUC) values for endpoints related to plaque vulnerability (symptomatic versus asymptomatic status, presence of intraplaque hemorrhage or lipid-rich necrotic core, fibrous-cap surrogates, and, less frequently, short-term cerebrovascular events). For ultrasound, contrast-enhanced videomics achieved an AUC of 0.87 (10 centers; n = 205), B-mode texture/radiomics reached 0.87 (n = 150), and segmentation-assisted models 0.827 (n = 202); other ultrasound models reported AUCs of 0.88–0.91. For CTA, a symptomatic-plaque machine-learning model yielded AUC 0.89 (n = 106); a perivascular-adipose-tissue (PVAT) radiomics nomogram achieved AUC 0.836 on external validation; a histology-referenced pilot attained AUC 0.987; and one mild-stenosis TIA model reported ROC performance. For high-resolution MRI (HR-MRI), radiomics-based models showed AUC 0.835–0.864 in single-modality cohorts and up to 0.984 with multi-contrast inputs. Across modalities, AUC ranges were: ultrasound 0.827–0.91, CTA 0.836–0.987 (external 0.836), and HR-MRI 0.835–0.984. Only two out of twelve studies performed external validation; calibration and decision-curve analyses were rarely provided, and most cohorts were single-center, limiting generalizability. Conclusions: AI demonstrates strong potential as a complementary tool for evaluating carotid plaque vulnerability, with high diagnostic performance across imaging modalities. Reported AUCs ranged from 0.83 to 0.99 based primarily on internal or hold-out validation, representing the upper bound of theoretical rather than real-world performance. Nonetheless, large prospective multicenter studies with standardized protocols, histopathological correlation, and external validation are required before clinical integration into stroke prevention pathways.