Multicenter clinical validation of a deep learning model for automated diagnosis and classification of aortic dissection

Aortic dissection (AD) is a life-threatening cardiovascular emergency requiring rapid and accurate diagnosis. This study aimed to clinically validate a VB-Net-based deep learning model for automated segmentation, detection and Stanford classification of AD on computed tomography angiography (CTA). This multicenter retrospective study enrolled 378 patients who underwent aortic CTA between September 2022 and July 2023 across three independent institutions. The gold standard was established by two senior radiologists, with discrepancies adjudicated by a third expert radiologist (≥25 years of experience). Model performance for AD diagnosis and Stanford classification was evaluated using area under the curve (AUC), sensitivity, specificity, and Cohen’s Kappa statistics. The inter-reader agreement between the two senior radiologists was substantial for both AD presence and Stanford classification (kappa: 0.96 and 0.99, respectively). The model demonstrated excellent diagnostic agreement across centers. For AD detection, Cohen’s kappa values were 0.99, 0.98, and 0.96, with AUCs of 1.00 (95% CI: 0.98–1.00), 1.00 (0.96–1.00), and 0.99 (0.95–1.00) at Centers 1–3, respectively. The lower bounds of the 95% CIs for sensitivity and specificity across all centers were ≥ 89.31% and ≥88.46%, respectively. For Stanford classification, kappa values were 0.93–0.96, with sensitivity and specificity consistently above 82%. Notably, AI-driven reconstruction reduced processing time from 10 min 18 s (manual) to 33.27 s per case ( p < 0.001). The model enables precise and rapid AD detection and Stanford classification using aortic CTA images, with high consistency and efficiency across multiple centers, demonstrating strong clinical applicability.