SACU-Net: Shape-Aware U-Net for Biomedical Image Segmentation With Attention Mechanism and Context Extraction

With the advent of convolutional neural network in image processing, deep learning has been used for medical image segmentation, such as liver, retina vessel, Nuclei and COVID-19 lesion segmentation, etc. Even though, the accuracy and the interpretablility of these methods still need to be futher improved. In this paper, we propose a novel Shape-aware U-Net architecture with Attention Mechanism and Context Extraction named SACU-Net to address the aforementioned issues regarding segment performance and shape identification. Our model mainly includes three modifications on the standard U-Net: 1) design a new context extraction module block (MRC) to capture high-level context feature, 2) re-design the skip pathways with spatial and channel attentions which reduce the semantic gap between the feature maps of encoder and decoder sub-networks, 3) a novel loss function was proposed to emphasize the segmentation accuracy on segmentation tasks. We have evaluated the SACU-Net in comparison with U-Net Variants on four different medical image segmentation tasks: liver segmentation in abdominal CT scans, retina vessel, Nuclei and COVID-19 lesion, which obtains a relative improvement in performance of 7.30%, 5.5% and 8.40% compared with the state-of-the-art U-Net Variant.