DCSAU-Net: A Deeper and More Compact Split-Attention U-Net for Medical Image Segmentation

• DOI: 10.48550/arxiv.2202.00972
• Publication Date: 2022-01-01
• Authors: Qing Xu; Wenting Duan; He Na
• Keywords: Upsampling; Net (polyhedron); Feature (linguistics); Code (set theory)

Deep learning architecture with convolutional neural network (CNN) achieves outstanding success in the field of computer vision. Where U-Net, an encoder-decoder structured by CNN, makes a great breakthrough biomedical image segmentation and has been applied wide range practical scenarios. However, equal design every downsampling layer encoder part simply stacked convolutions do not allow U-Net to extract sufficient information features from different depths. The increasing complexity medical images brings new challenges existing methods. In this paper, we propose deeper more compact split-attention u-shape (DCSAU-Net), which efficiently utilises low-level high-level semantic based on two novel frameworks: primary feature conservation block. We evaluate proposed model CVC-ClinicDB, 2018 Data Science Bowl, ISIC-2018 SegPC-2021 datasets. As result, DCSAU-Net displays better performance than other state-of-the-art (SOTA) methods terms mean Intersection over Union (mIoU) F1-socre. More significantly, demonstrates excellent challenging images. code for our work technical details can be found at https://github.com/xq141839/DCSAU-Net.