Correlation of Imaging Characteristics of Carotid Plaque with Clinical and Histopathological Features and Application of Artificial Intelligence Techniques in Identifying Carotid Disease

Abstract

Listen

Translate article

Introduction: Stroke is the third-leading cause of death. Carotid artery stenosis (CAS) accounts for 8%–37% of strokes. Landmark carotid trials have shown carotid endarterectomy (CEA) superior to best medical therapy (BMT) for symptomatic moderate stenosis and asymptomatic high-grade stenosis. This study aims to correlate imaging characteristics of vulnerable plaque (VP) with clinical and histopathological features and to assess the ability of artificial intelligence (AI) techniques to identify VP and disease burden using duplex ultrasound (DU) data. Materials and Methods: This prospective study from August 2022 to April 2024 used DU and computerized tomographic angiography (CTA) or magnetic resonance angiography (MRA) to assess carotid plaques in patients planned for CEA and subjected plaques to histology. We also imaged carotid plaques using DU in asymptomatic patients on BMT. YOLOv8 deep learning and U-Net AI models were used to analyze ultrasound (US) images. Results: One hundred and fiftyOne hundred and fifty patients were included in the study. Sixty (40%) were symptomatic. The age range was 33–84 years (mean 58.9). One hundred and fifteen (76.7%) were males. Hypertension, diabetes, dyslipidemia, and smoking were seen in 59%, 56%, 49%, and 40%, respectively and were more prevalent in patients with VP (90%, 60%, 70%, and 50%). Of the 70 patients who underwent CEA, 60 were symptomatic and 10 asymptomatic. 70% of symptomatic patients had a RANKIN (mRS) score of 2–3. All surgical patients underwent DU, 49 CTA, and 21 MRA. Based on the US, 53/60 (88%) symptomatic patients had VP. 56/60 (93%) were confirmed VP on histology. Positive predictive value (PPV) for VP was highest for DU (95%), followed by MRA (89%) and CTA (73%). PPV of DU screening in operated asymptomatic patients was 100%. AI showed 96% PPV for identifying >50% stenosis. Model training in identifying VP accuracy increased from 0.73 to 0.94 metrics. Conclusion: DU is an excellent modality to identify the nature of carotid plaque and can be a good tool to identify VP in asymptomatic CAS. More data will improve AI accuracy.