CSWinUNETR: Segmentation of Thin Anatomical Structures in Medical Images
CSWinUNETR is a general-purpose backbone for segmenting thin, tortuous structures like blood vessels or wrinkles in medical images. It uses cross-shaped stripe self-attention to capture long-range context, a detail-enhanced multi-scale attention module to preserve fine details, and sparse dynamic snake convolution to follow complex curves. This approach consistently outperforms existing methods across various medical imaging tasks.