A Hybrid Multi-Stage Framework For Intracranial Aneurysm Detection From Cta Images: A Comparative Study With End-To-End Deep Learning

Intracranial aneurysm detection from brain CTA images remains challenging due to the small size of vascular abnormalities and the limited availability of expert annotations. This paper proposes a hybrid multi-stage framework and presents a direct comparison with a baseline end-to-end deep learning classifier. The framework integrates image preprocessing, synthetic mask generation, attention-based segmentation, hybrid feature extraction, and independent classification. Synthetic vessel masks generated using vesselness filters and morphological operations were employed as surrogate labels to support training. An Attention U-Net enhanced with CBAM and optimized using Tversky loss achieved reliable segmentation performance, with a Dice score of 0.8798 and an IoU of 0.8056, enabling accurate localization of aneurysm regions even in low-contrast cases. For classification, deep features extracted from ResNet50 were combined with handcrafted structural and textural descriptors and classified using an XGBoost model. This hybrid strategy achieved an accuracy of 98.46% and eliminated false-negative aneurysm cases, outperforming the baseline end-to-end ResNet50 classifier. The results demonstrate that separating segmentation, feature extraction, and classification while integrating deep and structural information leads to improved diagnostic reliability and clinically safer predictions compared to conventional CNN-based approaches.