ResSE-UNet: A convolution network for automated skin lesion segmentation

Abstract:
Dermatosis is one of the most common human diseases. However, there are a number of challenges with automatic skin lesion segmentation, including blurred edges of lesions, low contrast between normal and lesion areas, and hair occlusion. In this paper, we propose a convolutional network combining residual connections and channel attention mechanism based on U-Net to automatically segment skin lesions. We introduce residual blocks in U-Net encoder and decoder to enhance feature transfer. In addition, the SE blocks are added to the skip connection, so that when the feature information of the shallow network is transmitted, the feature response can be adaptively adjusted from the channel dimension, and the feature extraction efficiency of the network can be improved. We also use Group Normalization (GN) for normalization to avoid the increase of errors caused by batch normalization (BN) when the batch is too small. We evaluated our network architecture on the public dataset ISIC 2018. The results demonstrate that the proposed framework significantly enhances segmentation performance for skin lesions.