Abstract
Skull stripping is a critical preprocessing step for reliable neuroimaging analysis. Although recent deep learning methods have made remarkable progress in accurate brain extraction, their generalization capability remains limited across heterogeneous imaging protocols and diverse pathological conditions. This limitation arises from their predominant reliance on voxel-intensity information without incorporating explicit anatomical constraints. To address this issue, we introduce DiffShape, a conditional diffusion framework that explicitly decouples brain shape generation from direct voxel-wise intensity-to-label prediction and uses the generated shape as a prior to strengthen existing skull-stripping models. Specifically, DiffShape represents brain geometry using a one-dimensional polar-radius vector, providing a compact and tractable shape representation that enables efficient diffusion modeling over a learned low-dimensional shape manifold. Conditioning the diffusion process on the input image, DiffShape learns a noise-to-shape denoising trajectory that progressively transforms random noise into anatomically plausible brain geometries. The input image serves as a conditioning signal that steers the denoising trajectory toward patient-specific geometry, while the learned diffusion dynamics drive noisy shape samples toward high-probability regions of the conditional brain shape distribution, thereby preserving anatomical plausibility under heterogeneous contrasts or pathological variations. We evaluate DiffShape across datasets containing both healthy and pathological cases, and it consistently improves segmentation accuracy over state-of-the-art skull-stripping methods, demonstrating a promising direction for enhancing segmentation generalization in medical imaging.