Attention Mechanism Enhanced Multi-layer Edge Perception Network for Deep Semantic Medical Segmentation

Existing deep learning–based medical image segmentation methods have achieved gratifying progress, but they still suffer from the coarse boundaries with similar pixels of target. Because the boundary of medical images becomes blurred and the gradient is inconsistent and not apparent, high-resolution images are needed for more accurate segmentation. To tackle these problems, we propose an efficient multi-layer edge perception U-shaped structure for medical image segmentation. In this paper, we present a multi-layer edge perception network for describing more precise edges of medical targets. The U-structure architecture of our network embeds a multi-layer edge perception module, which has the following advantages: (1) connecting different scales and channels to help the network better learn the feature of the medical image via the combination of a pyramid structure and several edge perception modules; (2) a new downsampling block is designed to improve the network’s sensibility to the target boundary. We demonstrate the effectiveness of the proposed model on the DRIVE datasets, and achieve a Dice gain of 0.841 over other models. In this paper, we propose an efficient multi-layer edge perception U-shaped structure for medical image segmentation. A large number of experiments show that the performance of our proposed multi-layer edge perception U-shaped network is significantly better than the traditional segmented network structure.