Self Supervised Cerebrovascular Segmentation Using Shared Weight DDPM UNet with Adversarial Representation Learning

Dense voxel-wise annotation for Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) vessel segmentation is costly and often infeasible. We present a selfsupervised Denoising Diffusion Probabilistic Model (DDPM) UNet that eliminates the need for manual labels. A single shared encoder feeds three decoder heads-background inpainting, vessel segmentation, and angiogram synthesis-trained jointly via adversarial cycle-consistency on pseudo-masks generated by intensity thresholding. Pretrained on the Lausanne TOF-MRA Aneurysm Cohort and evaluated on the ADAM subset of COSTA, our model achieves Dice Similarity Coefficient (DSC) and Intersection Over Unions (IoU) scores of 0.6755 and 0.5286, respectively, rivalling fully supervised baselines. When fine-tuned with only one labeled slice, it attains a Dice of 0.6379 versus 0.3701 for a randomly initialized U-Net. This approach addresses the scarcity of annotated TOF-MRA datasets and paves the way for volumetric 3D diffusion extensions and task-specific attention modules to further enhance segmentation fidelity.