A fully automated deep learning pipeline system for carotid plaque segmentation and vulnerability classification on multi-contrast MRI: a multicenter study

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Background Carotid vulnerable plaque rupture is a major cause of ischemic stroke. However, accurate identification of high-risk plaques using medical images heavily relies on radiologists’ expertise. This study aimed to develop an artificial intelligence (AI) system for automated segmentation and classification of carotid plaques using high-resolution vessel wall imaging (HR-VWI) to assist in diagnostic workflow. Methods A total of 1610 carotid arteries from 1315 patients who underwent multi-contrast HR-VWI between January 2019 and July 2025 were included from four centers. The dataset was partitioned into training, internal test, pooled external test, and prospective test sets. A carotid fully automated pipeline system (CFAPS) was developed to perform plaque segmentation and vulnerability classification. Segmentation performance was evaluated using the Dice similarity coefficient (DSC), while classification performance was assessed using the area under the curve (AUC), visualized with Grad-CAM heatmaps and histopathology. Diagnostic performance and efficiency were compared between CFAPS and radiologists. In the prospective cohort who underwent 18 F-FDG PET/MRI, maximum standardized uptake value (SUVmax) and tissue-to-background ratio (TBR) were compared between CFAPS-classified vulnerable and stable plaque groups. Additionally, Kaplan-Meier analysis for stroke-free survival was performed on patients from two follow-up centers. Results CFAPS achieved DSCs of 0.863 ± 0.064, 0.855 ± 0.122, and 0.857 ± 0.069 on internal, external, and prospective test sets, respectively. For vulnerability classification, it attained AUCs of 0.942 (95% confidence interval [CI]: 0.899–0.975), 0.884 (95% [CI]: 0.837–0.924), and 0.923 (95% [CI]: 0.872–0.963), respectively. Grad-CAM localized to high-risk features, corroborated by histopathology (accuracy 0.881, κ = 0.73). CFAPS outperformed junior radiologists (2 years of plaque imaging experience) and reduced analysis time from 71.63 s to 0.53 s per case, while improving inter-reader agreement (κ: 0.49–0.57 to 0.72–0.75). Vulnerable plaques identified by CFAPS exhibited significantly higher SUVmax and TBR than stable plaques (both p < 0.001). Furthermore, the CFAPS risk classification was significantly associated with patients’ stroke-free survival (Center 1: HR = 4.34, p < 0.001; Center 2: HR = 8.89, p < 0.001). Conclusions The CFAPS system enables fully automated, robust, and interpretable segmentation and risk stratification of carotid plaques using multi-contrast HR-VWI, demonstrating strong generalizability and clinical utility.