Context-Aware Detection and Grading of Intracranial Aneurysms in DSA Images

Tiny aneurysms in digital subtraction angiography (DSA) are challenging to detect and classify due to their low contrast, complex vascular backgrounds, and small size. We propose a three-stage, context-aware framework for automated aneurysm analysis. First, aneurysm candidates are localized using an RT-DETR detector, which provides accurate anchor-free detection and robust sensitivity to small objects. Second, the detected regions, including the aneurysm patch, are expanded to include the surrounding vessel context and are enhanced using a pretrained RealESRGAN model to restore vascular detail and improve geometric fidelity without additional training. Third, each super-resolved patch is passed to a modified ResNet50 classifier for categorical severity grading (mild, moderate, severe, critical). The classifier integrates local and contextual cues to better capture risk-related morphology beyond lesion size alone. The framework is evaluated on a 2D DSA dataset annotated with bounding boxes and severity labels, reporting both detection and multi-class classification performance. Results show that combining RealESRGAN-based enhancement and context-aware classification yields substantial gains in accuracy and recall, enabling more reliable and clinically interpretable aneurysm assessment.