CTA-Based 3D Thrombus Radiomics and Machine Learning for Differentiating Stroke Etiology in Acute Middle Cerebral Artery Occlusion

Abstract Purpose: To develop and evaluate machine learning models integrating three-dimensional (3D) thrombus radiomic features from preoperative computed tomography angiography (CTA) with clinical variables to differentiate large-artery atherosclerosis (LAA) from cardioembolism (CE) in patients with acute middle cerebral artery occlusion (MCAO). Materials and Methods: This retrospective two-center study included 453 patients with acute MCAO who underwent endovascular treatment (EVT). The 3D thrombus region was segmented on preoperative CTA, and radiomic features reflecting intensity, texture, and morphology were extracted. Multiple machine learning models integrating radiomic and clinical variables were developed using the training cohort and evaluated in a held-out test cohort. Discrimination, calibration, SHAP-based interpretability, feature importance, and decision curve analysis were assessed. Results: Among the evaluated models, the random forest (RF) model showed competitive performance in the held-out test cohort, with an area under the receiver operating characteristic curve of 0.822, sensitivity of 0.867, and the lowest Brier score (0.171) among the evaluated models. Both clinical and 3D thrombus radiomic features contributed to model predictions. Intraprocedural tirofiban use, admission LDL-C level, sex, admission NIHSS score, and selected thrombus texture and morphological features were important predictors. Decision curve analysis indicated positive net benefit over treat-all and treat-none strategies across threshold probabilities of approximately 0.05–0.60. Conclusion: Three-dimensional thrombus radiomics from routine preoperative CTA combined with clinical variables may help differentiate LAA from CE in acute MCAO. The interpretable RF model may serve as a complementary tool for acute-phase stroke etiological assessment.