Integrated radiomics and image-based deep learning framework using carotid CT angiography ROI datasets for plaque stability assessment as a predictor of ischemic stroke risk

Carotid plaque vulnerability drives ischemic stroke, yet preoperative assessment remains imprecise. A retrospective multi-center study was conducted using carotid CT angiography (CTA) datasets with pathology-validated carotid endarterectomy cases for radiomics analysis and an independently annotated CTA dataset for image-based deep learning. Radiomic features were screened using ANOVA–Kruskal–Wallis analysis, correlation filtering, and LASSO followed by development of Random Forest, Support Vector Machine, k-Nearest Neighbor, Naïve Bayes and Logistic Regression models. Image-based models included MLP, CNN and ImageNet-pretrained VGG16, VGG19, and ResNet50. Model performance was evaluated using ROC/AUC, calibration analysis and external validation. The main cohort comprised 260 consecutive CEA patients (200 for model development/internal validation and 60 for external validation), while an independent annotated CTA dataset of 236 cases (7394 ROIs) was used for image-based deep learning. On the internal test set, radiomics achieved AUCs of 0.858 (RF), 0.857 (SVM), 0.855 (Naïve Bayes), 0.843 (KNN) and 0.821 (Logistic Regression) with small train–test gaps. External validation confirmed comparable discrimination: SVM 0.839, Logistic Regression 0.835, KNN 0.830, Naïve Bayes 0.817, Random Forest 0.817; calibration curves showed close predicted–observed agreement. Image models reached test AUCs of 0.767 (VGG16), 0.754 (ResNet50), 0.718 (VGG19), 0.669 (MLP) and 0.654 (CNN); external AUCs were 0.776, 0.751, 0.719, 0.684, and 0.687, respectively. Grad-CAM localized hyperattenuating regions consistent with pathology, most clearly for VGG16/ResNet50. Radiomics classifiers particularly SVM/Logistic Regression and transfer-learning backbones (VGG16/ResNet50) offer reproducible plaque stability prediction from CTA with good calibration and external generalizability, supporting their potential use in stroke risk stratification. This framework may aid in preoperative risk stratification and guide surgical decision-making.