Abstract
BACKGROUND: Intracranial aneurysm (IA) is a focal cerebral artery dilatation affecting 2-5% of the population, with rupture leading to high mortality and disability. Early, accurate classification from computed tomography angiography (CTA) is crucial for management but is challenged by small datasets and limited interpretability. We evaluate a hybrid deep transfer learning framework with integrated Grad-CAM to improve both discrimination and explainability in CTA-based IA classification. METHODS: In this retrospective study, 83 eligible patients from two centers underwent CTA. We employed stratified 5-fold cross-validation to compare: a baseline deep learning model (DL), a transfer learning-enhanced model (DL + TL), and radiologist assessment. Both AI models used a hybrid ResNet-18 architecture with LASSO feature selection and logistic regression. Performance was assessed using AUC, accuracy, calibration, decision curve analysis, NRI, and IDI. Interpretability was quantified via Grad-CAM using Intersection-over-Union (IoU) and Dice similarity coefficient. RESULTS: The DL + TL model achieved superior performance with a mean AUC of 0.853 (95% CI: 0.789-0.912) and accuracy of 84.0%, outperforming both DL (AUC: 0.744, p = 0.012) and radiologists (AUC: 0.731, p = 0.008). Grad-CAM analysis showed DL + TL had significantly higher attention precision (IoU: 0.68 vs. 0.45 for DL, p < 0.001) and was rated more clinically relevant by blinded radiologists (4.2/5 vs. 2.8/5). CONCLUSION: Integrating transfer learning with quantitative interpretability assessment improves both accuracy and transparency of IA classification in limited-data settings. This framework offers a validated, interpretable approach for neurovascular imaging, pending further multi-center validation.