MedDiffusionNet: A Geometric Deep Learning Network for Intracranial Aneurysm Segmentation Across Imaging Modalities

As intracranial aneurysms (IAs) are potentially fatal vascular abnormalities that can lead to severe strokes, their accurate segmentation from acquired imaging data is essential for diagnosis and treatment planning. Current automatic IA segmentation methods predominantly focus on directly processing medical images, whereas research exploring approaches beyond image-based modalities is limited. To address this gap, we propose medical diffusion network (MedDiffusionNet), a novel segmentation network incorporating geometric deep learning, designed to directly work on three-dimensional mesh models of aneurysms. The core of MedDiffusionNet comprises multiple diffusion blocks grounded in the heat diffusion equation, which enable efficient information propagation on graphs. By incorporating learnable diffusion time and gradient features, MedDiffusionNet performs frequency-domain convolution with an adaptive receptive field. Notably, this architecture aligns with the geometric properties of vascular aneurysm models, allowing MedDiffusionNet to learn a powerful nonlinear mapping from the feature space onto the label space on the associated mesh. We validate MedDiffusionNet on an extremely imbalanced dataset reconstructed from data acquired through multiple imaging modalities. Experimental results show that MedDiffusionNet outperforms eight state-of-the-art networks in IA segmentation, accurately delineating IA boundaries across diverse vascular shapes and sizes. Moreover, our network achieves higher values of the associated intersection over union and Dice similarity coefficient while maintaining geometric consistency with the ground truths.