Edge-Boosted U-Net for 2D Medical Image Segmentation

Automatic medical image segmentation has always been a heated study in computer-assisted diagnosis (CAD). It is a quite challenging task due to the diversity and complexity of medical images. In this paper, we propose an Edge-boosted U-Net (EU-Net) to address the problem of medical image segmentation. The architecture is basically a U-shape network, combined with three main parts: Edge Aggregation Path, Feature Fusion Block, and Feature Attention Block. The Edge Aggregation Path is to extract multilevel edge-relevant information. The Feature Fusion Block is designed to fuse features from different paths. And the Feature Attention Block is embedded in the network to generate more informative feature maps. The collaboration of these three parts effectively boosts the performance of the whole network. We verified the importance of each part by conducting several experiments. Meanwhile, we compared the proposed method with other state-of-the-art methods on three different modalities of public medical image datasets. Our method achieves the superiority with IoU and dice coefficient respectively on all the datasets. Notably, it attains 2% accuracy improvement over other methods on the challenging datasets.