A Hybrid 2D–3D Deep Learning Framework for Automated Intracranial Aneurysm Segmentation and Morphological Risk Stratification from TOF-MRA

Intracranial aneurysm assessment from time-of-flight MR angiography (TOF-MRA) is clinically important, yet automated aneurysm-only segmentation remains challenging because aneurysms are often small, have weak boundaries, and lie adjacent to complex vascular structures. This paper presents a hybrid 2D-3D deep learning framework that performs volumetric aneurysm segmentation and then derives an interpretable morphology-based rupture-risk indicator from the predicted mask. The pipeline standardizes each TOF-MRA volume by resampling to$1 \times 1 \times 1 \text{mm} 3$isotropic spacing and applying robust intensity normalization using percentile clipping followed by z -score scaling. The segmentation model is optimized with a hybrid objective that combines Dice overlap and voxel-wise binary cross-entropy, enabling stable learning under severe foreground sparsity. Experiments on the Lausanne TOF-MRA dataset (284 volumes; train =200, test =84) demonstrate consistent segmentation performance (mean Dice =0.9326 ± 0.0284, Hausdorff distance =0.7185 ± 0.2743), with size-stratified analysis confirming robustness across small, medium, and large aneurysms. Overall, the proposed segmentation-to-morphology pipeline provides a practical route to aneurysm delineation and interpretable morphology-based risk scoring from TOF-MRA.