Transformer-based fusion of radiomics-habitat and deep learning for assessing unruptured intracranial aneurysm instability.

Objectives: To develop and validate a prediction model that integrates radiomics-habitat and deep learning (DL) features derived from vessel wall MRI (VWI) for evaluating unruptured intracranial aneurysms (UIAs) instability. Methods: First, from January 2022 to January 2024, 519 consecutive patients with suspected UIAs were screened. After applying exclusion criteria, 293 patients with 312 UIAs were ultimately enrolled. 197 UIAs were stable (from 188 patients) and 115 UIAs were unstable (from 105 patients). Second, aneurysm regions were segmented, and K-means clustering was used to partition them into three habitat subregions. Third, a Transformer-based fusion model for assessing UIA instability was developed to integrate radiomics-habitat features, DL features, and clinical variables. Model performance was evaluated using AUC, calibration curves, and clinical gain metrics, including Net Reclassification Index (NRI) and Integrated Discrimination Improvement (IDI). Last, SHAP (SHapley Additive exPlanations) was applied to enhance model interpretability. Results: The Transformer-based fusion model assessing UIA instability exhibited superior performance (validation AUC = 0.844) compared with the optimal radiomics-habitat model (AUC = 0.721) and the top-performing DL model (DenseNet169, AUC = 0.816). The model demonstrated superior clinical utility, with an NRI of 0.282 and an IDI of 0.558 compared to the Radiomics-Habitat model. Decision curve analysis showed a high net clinical benefit across a range of threshold probabilities. Conclusion: The Transformer-based fusion model provides an exploratory risk-assessment model and has the potential to assist in clinical decision-making.