Intracranial Aneurysm Rupture Risk Assessment Based on Multi-scale Deep Transfer Learning and Clinical Feature Fusion

Background: Rupture of intracranial aneurysms significantly increases the risk of non-traumatic subarachnoid hemorrhage (SAH), a condition often associated with high fatality. Accurately assessing the rupture risk of intracranial aneurysms is crucial for effective clinical interventions. Objective: This study presents an innovative, semi-automated framework that combines multi-scale deep transfer learning features (MDTLFs), enhanced by super-resolution reconstruction, along with clinical data, to improve the accuracy of aneurysm rupture risk assessment. By applying super-resolution to CTA images, we enhance feature extraction, thereby increasing the model’s sensitivity to anatomical structures that are prone to rupture. Materials and Methods: We analyzed 366 cases (226 unruptured, 140 ruptured), collecting clinical data alongside CTA images. For the first time, we applied super-resolution techniques with the Med3D model, extracting deep features from both original and super-resolved images and fusing them to form high-detail, multi-scale deep transfer learning features (MDTLFs). This approach captures structural details that are otherwise difficult to detect. The MDTLFs, along with clinical features, were optimized using LASSO regression to identify rupture-critical patterns, and the final feature set was evaluated across five machine learning classifiers (LR, RF, SVM, XGBoost, and AdaBoost) for predictive accuracy. Results: The multi-modal feature fusion framework achieved an accuracy of $89.5 \%$ and an AUC of 0.904, significantly outperforming singlefeature models (clinical AUC: 0.774, MDTLFs AUC: 0.855). Notably, super-resolution reconstruction proved essential in enhancing MDTLFs, leading to increased model accuracy by highlighting subtle image details that correlate with rupture risk. Conclusion: This study is a novel framework that combines MDTLFs, radiomics, and clinical data to enhance aneurysm rupture risk prediction. The innovative use of super-resolution reconstruction with multi-scale deep transfer learning marks a significant step forward, enabling more accurate and clinically relevant assessments.