Combining edge guidance and feature pyramid for medical image segmentation

Automatic segmentation of medical images is very important for computer-aided diagnosis. The U-shaped and skip-connection based on convolution (UNet) has achieved the most advanced performance in the field of medical image segmentation. However, most existing UNet-based methods have the problem of coarse segmentation of tissue edges. We propose a novel edge guidance feature pyramid network (EGFPNet) for medical image segmentation with the following contributions. First, we synthesize local edge information and global location information to obtain tissue edge features. Second, in order to better utilize the edge information, we propose an edge guidance feature pyramid (EGFP). Edge features interact with area features of different scales to form complementary features of different scales. These complementary features of different scales also interact to represent the complete information of the tissue and improve the adaptability to different tissue scales. We compare with state-of-the-art medical image segmentation methods on the automated cardiac diagnosis challenge (ACDC) and the 2018 atrial segmentation challenge (2018 ASC). Our method achieved average dice score of 0.929 for right ventricle (RV), 0.886 for myocardium (Myo), 0.958 for left ventricle (LV), and 0.920 for left atrium (LA). Experimental results on two medical image segmentation datasets show that our method outperforms six state-of-the-art medical image segmentation methods. The code is available at https://github.com/jinancsl/EGFPNet.