DE-Net: A deep edge network with boundary information for automatic skin lesion segmentation

Abstract

Automatic skin lesion segmentation is one of the most important tasks for computer-aided diagnosis of skin cancer. Although many deep learning-based methods have been proposed, most of them do not pay attention to the boundary information that is critical for achieving more accurate segmentation. In this work, an end-to-end deep edge convolutional neural network (DE-Net) based on the encoder-decoder structure is proposed to pay more attention to the skin lesion boundaries. In the decoder process of the proposed DE-Net, an edge information guided module (EIGM) is designed to introduce the boundary information of the original image and fuse it with different levels of contextual information. Furthermore, to generate more accurate skin lesion boundaries without losing the entire performance and guide the module to highlight more boundary information, we propose an entirety-center-edge (ECE) loss function that can further optimize the boundary details on the basis of the necessary segmentation results. The proposed loss function is backbone-independent and has better performance than other commonly used loss functions in segmentation tasks. In the experiment, the ISIC-2017 dataset is employed to evaluate the effectiveness of the proposed method. It achieves the performance of 0.8792 and 0.8053 on the metrics Dice coefficients and Jaccard index, respectively. Furthermore, we also evaluate the proposed method on ISIC-2016 and PH2 datasets to demonstrate its generalizability. Experimental results demonstrate that the proposed method can outperform the state-of-the-art methods on all these three datasets and has strong generalizability.