Predicting intracranial angioplasty failure using vessel wall MRI habitat radiomics and deep learning: a multicenter study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, The First Affiliated Hospital of Soochow University, Suzhou, Jiangsu, China.
- Department of Radiology, The First Affiliated Hospital of Shandong First Medical University & Shandong Provincial Qianfoshan Hospital, Shandong Lung Cancer Institute, Shandong Institute of Neuroimmunology, Jinan, Shandong, China.
- Department of Radiology, Qilu Hospital of Shandong University, Jinan, Shandong, China [email protected] [email protected].
- Department of Radiology, The First Affiliated Hospital of Soochow University, Suzhou, Jiangsu, China [email protected] [email protected].
- Institute of Medical Imaging, Soochow University, Suzhou, Jiangsu, China.
Abstract
Primary balloon angioplasty (PBA) is a promising strategy for symptomatic intracranial atherosclerotic stenosis but carries the risk of procedural failure due to dissection or elastic recoil. Accurate preoperative risk stratification remains challenging. This study aimed to develop a multiscale high-resolution vessel wall imaging model integrating habitat radiomics and deep learning to predict PBA failure. This retrospective multicenter study analyzed 274 plaques (from 270 patients undergoing PBA). Plaques were stratified into training (n=150), internal validation (n=64), and external test (n=60) cohorts. The primary endpoint was PBA failure, defined as flow-limiting dissection or residual stenosis >50% requiring bailout stenting. A multiscale analysis was performed: (1) Clinical signature incorporating baseline demographics and radiological plaque metrics; (2) Habitat signature derived from radiomics features extracted within distinct subregions identified by unsupervised clustering; and (3) Deep learning signature utilizing a transformer-based network to capture global vessel wall dependencies. A combined fusion model was also developed and validated. The multiscale combined model demonstrated robust discrimination in both internal validation (AUC 0.88, 95% CI 0.80 to 0.96) and external test cohorts (AUC 0.83, 95% CI 0.73 to 0.94). It significantly outperformed the clinical model (P=0.006) and showed superior calibration and net clinical benefit. Habitat analysis revealed that signal variability and textural complexity in specific subregions were key predictors. By integrating habitat radiomics and deep learning, this vessel wall MRI-based fusion model quantifies multiscale plaque heterogeneity and biological burden to accurately identify high-risk patients for PBA failure, potentially facilitating personalized revascularization strategies.