Ret-UNet: Enhancing medical image segmentation with self-retention

Medical image segmentation has advanced significantly due to the integration of deep learning techniques, particularly convolutional neural networks (CNNs) like U-Net. However, CNNs often struggle to capture global spatial relationships, which are crucial for accurately segmenting complex anatomical structures. To address this limitation, we propose Ret-UNet, a novel architecture that enhances the traditional U-Net framework by incorporating the Self-Retention mechanism. Self-Retention introduces an explicit shape prior related to the Euclidean distance, which effectively encode global spatial relationships within the image. The Ret-UNet leverages both local feature extraction and global context awareness by incorporating Ret Blocks into the U-Net like architecture, leading to improved segmentation performance. Evaluations on ACDC, CAMUS and Synapse datasets demonstrate that Ret-UNet achieves superior segmentation accuracy and robustness, outperforming state-of-the-art models. The code is available at https://github.com/weirdgit/RetUNet .