Deep Learning Radiomics Signature from Multicontrast MRI for Automated Identification of Symptomatic Carotid Plaques: A Multicenter Study.
Authors
Affiliations (5)
Affiliations (5)
- From the Department of Radiology and Nuclear Medicine (Q.G., J.Z., Y.Z., B.C., R.Q., F.Y., M.F., S. Zhang, C.Z., J.L.), Xuanwu Hospital, Capital Medical University, Beijing, China.
- Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics (Q.G., J.Z., Y.Z., B.C., R.Q., F.Y., M.F., S. Zhang, C.Z., J.L.), Beijing, China.
- Department of Radiology (S. Zhao), The Second Hospital of Hebei Medical University, Shijiazhuang, Hebei, China.
- Department of Radiology (X.W.), Shandong Provincial Hospital, Jinan, Shandong, China.
- From the Department of Radiology and Nuclear Medicine (Q.G., J.Z., Y.Z., B.C., R.Q., F.Y., M.F., S. Zhang, C.Z., J.L.), Xuanwu Hospital, Capital Medical University, Beijing, China [email protected].
Abstract
Ischemic stroke poses a significant global health burden. Accurately identifying symptomatic carotid atherosclerotic plaques, beyond relying solely on stenosis degree, remains a critical challenge for precise stroke risk stratification. We aimed to develop and validate a deep learning radiomics (DLR) signature based on multicontrast MRI to identify symptomatic carotid plaques accurately. In this retrospective multicenter study, 409 carotid arteries from 355 patients with carotid atherosclerosis were enrolled (219 training, 95 internal validation, 95 external test). Deep learning (DL) and radiomics features were extracted and combined from automatically segmented plaque regions on multicontrast MRI. The optimized DLR signature derived from a 3-stage feature selection pipeline was leveraged to train diverse machine learning classifiers for robust identification of symptomatic carotid plaques. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC) and compared against clinical models, radiomics-only models, and DL-only models. Subgroup analysis across stenosis severities and comparison of MRI-based American Heart Association lesion types between DLR-defined risk groups were performed. The DLR model with logistic regression demonstrated excellent performance in identifying symptomatic plaques, achieving AUROCs of 0.975 (95% CI, 0.954-0.992), 0.933 (95% CI, 0.876-0.976), and 0.881 (95% CI, 0.807-0.939) in the training, internal validation, and external validation cohorts, respectively. It significantly outperformed the clinical model (AUROCs of 0.701, 0.749, 0.711; <i>P</i> < .05), radiomics-only model (AUROCs of 0.877, 0.839, 0.789; <i>P</i> < .05), and DL-only model (AUROCs of 0.948, 0.894, 0.845; <i>P</i> < .05 in training/external). Performance remained consistently high across stenosis severity subgroups (AUROCs of 0.895-0.982 for severe, 0.863-0.971 for mild-moderate stenosis). DLR-defined symptomatic groups showed significantly higher prevalence of complex type VI lesions (internal: 50.0% versus 14.8%, <i>P</i> < .001; external: 48.7% versus 20.7%, <i>P</i> = .004) and lower prevalence of predominantly calcified type VII lesions (external: 8.1% versus 43.1%, <i>P</i> < .001) compared with asymptomatic groups. The developed multicontrast MRI-based DLR signature provides a highly accurate and robust tool for the automated identification of symptomatic carotid plaques, underscoring its potential value as a noninvasive tool to guide personalized stroke prevention strategies.