Self-supervised small vessel segmentation with shape-aware geometric models and attention

Small vessels play a critical role in the development of various brain diseases. Significant advances have been made in high-resolution vascular imaging, but it remains a challenge to precisely segment small vessels to quantitatively assess their integrity. Existing vessel segmentation approaches typically assume, implicitly or explicitly, a tubular representation suitable for large vessels but suboptimal for small vessels because they often exhibit shape irregularity, weak contrast, and discontinuity even in high-resolution imaging data. To address these challenges, we propose a generalized and explicit small vessel representation and present a novel self-supervised small vessel segmentation network termed S3U-Net. First, S3U-Net is trained based on a novel shape-aware flux measure to estimate the direction and shape profiles of small vasculature with non-circular and irregular appearances. Second, multiple modules for local contrast attention (LCA) are incorporated to enhance small vessel responses in regions with weak contrast. Third, we develop a propagation-based post-processing algorithm based on the parallel transport frame (PTF) to enhance small vessel connectivity. To evaluate the efficacy of S3U-Net, comprehensive experiments were conducted on seven datasets from different modalities. The results demonstrated that our approach yields a considerable improvement in small vessel segmentation and their connectivity compared with existing methods. Finally, we applied S3U-Net to 7-Tesla brain images from elderly adults and demonstrate the successful detection of significant associations between small vessel density and cognitive functions.