A Compact Unified Model for Moderate-to-Severe Carotid Artery Stenosis Using Image Segmentation: A Multicenter Study

This study aimed to evaluate the diagnostic performance of a new artificial intelligence model, the compact unified model (CU-Model), for moderate-to-severe carotid artery stenosis using digital subtraction angiography (DSA) images compared to clinical reference standards.

In this retrospective multicenter study, 156 patients with confirmed moderate-to-severe carotid artery stenosis were included. A pretrained CU-Model (trained on public non-clinical datasets for optical flow and stereo matching) was applied to raw DSA images without any fine-tuning on clinical data to avoid data leakage. The model served exclusively as a post-acquisition support tool: cardiologists applied it immediately after procedures for automated vessel segmentation and highlighting. Final North American Symptomatic Carotid Endarterectomy Trial (NASCET) measurements and interpreta-tions were performed manually under expert supervision by 2 blinded researchers.

The CU-Model identified left internal carotid artery (ICA) stenosis in 84 patients and right ICA stenosis in 72. It misclassified only 1 moderate left ICA stenosis case as normal, achieving 99.4% sensitivity (155/156) for any stenosis detection. For stenosis severity grading, strong agreement with DSA was observed (moderate: 57.68 ± 5.94 vs. 57.71 ± 6.19, r = 0.83, P < .001, 95% CI: 0.76-0.88; severe: 86.96 ± 5.97 vs. 87.02 ± 6.18, r = 0.91, P < .001, 95% CI: 0.85-0.95). Of 156 cases, 106 were moderate (50%-69%) and 50 severe (≥70%), using NASCET-adapted thresholds accounting for measurement variability. Hypertension (77%), smoking, and syncope were common, with significant association in symptomatic patients (P < .05).

The CU-Model demonstrates strong concordance with DSA as a practical, post-procedure support tool for vessel segmentation. Pending larger validation studies, it shows promise for enhancing clinical workflow efficiency.