MRI-based habitat radiomics combined with vision transformer for identifying vulnerable intracranial atherosclerotic plaques and predicting stroke events: a multicenter, retrospective study

Accurate identification of high-risk vulnerable plaques and assessment of stroke risk are crucial for clinical decision-making, yet reliable non-invasive predictive tools are currently lacking. This study aimed to develop an artificial intelligence model based on high-resolution vessel wall imaging (HR-VWI) to assist in the identification of vulnerable plaques and prediction of stroke recurrence risk in patients with symptomatic intracranial atherosclerotic stenosis (sICAS). Between June 2018 and June 2024, a retrospective collection of HR-VWI images from 1806 plaques in 726 sICAS patients across four medical institutions was conducted. K-means clustering was applied to the T1-weighted imaging (T1WI) and T1-weighted imaging with contrast enhancement (T1CE) sequences. Following feature extraction and selection, radiomic models and habitat models were constructed. Additionally, the Vision Transformer (ViT) architecture was utilized for HR-VWI image analysis to build a deep learning model. A stacking fusion strategy was employed to integrate the habitat model and ViT model, enabling effective identification of high-risk vulnerable plaques in the intracranial region and prediction of stroke recurrence risk. Model performance was evaluated using receiver operating characteristic (ROC) curves, and model comparisons were conducted using the DeLong test. Furthermore, decision curve analysis and calibration curves were utilized to assess the practicality and clinical value of the model. The fused Habitat + ViT model exhibited excellent performance in both the validation and test sets. In the validation set, the model achieved an area under the curve (AUC) of 0.949 (95% CI: 0.927-0.969), with a sensitivity of 0.879 (95% CI: 0.840-0.945), a specificity of 0.905 (95% CI: 0.842-0.949), and an accuracy of 0.897 (95% CI: 0.870-0.926). In the test set, the AUC increased to 0.960 (95% CI: 0.941-0.973), with specificity rising to 0.963 and an accuracy of 0.885 (95% CI: 0.857-0.913). The DeLong test revealed statistically significant differences in AUC between the fused model and the single-modal models (test set, vs. ViT p = 0.000; vs. Habitat p = 0.000) Cox regression analysis showed that the Habitat + ViT index, based on the prediction probability of the Habitat + ViT model, was an independent predictor of stroke recurrence (HR: 2.07; 95% CI: 1.12-3.81), with significant predictive power for stroke events at multiple time points. Specifically, measured by AUC values, the model's predictive performance at 1, 2, 3, and 4 years was 0.751 (95% CI: 0.679-0.823), 0.820 (95% CI: 0.760-0.876), 0.815 (95% CI: 0.753-0.877), and 0.780 (95% CI: 0.680-0.873), respectively. The integrated Habitat + ViT model based on HR-VWI demonstrated superior performance in identifying high-risk vulnerable plaques in sICAS patients and predicting stroke recurrence risk, providing valuable support for clinical decision-making. This study was supported by the National Natural Science Foundation of China (grant 82204933). Henan Key Laboratory of Neurorestoratology (HNSJXF-2021-004), 2019 Joint Construction Project of Henan Provincial Health Committee and Ministry of Health (SB201901061), and the Xin Xiang City Acute Ischemic Stroke Precision Prevention and Treatment Key Laboratory.