RA-Net: Region-Aware Attention Network for Skin Lesion Segmentation

Abstract

The precise segmentation of skin lesion in dermoscopic images is essential for the early detection of skin cancer. However, the irregular shapes of the lesions, the absence of sharp edges, the existence of artifacts like hair follicles, and marker color make this task difficult. Currently, fully connected networks (FCNs) and U-Nets are the most commonly used techniques for melanoma segmentation. However, as the depth of these neural network models increases, they become prone to various challenges. The most pertinent of these challenges are the vanishing gradient problem and the parameter redundancy problem. These can result in a decline in Jaccard index of the segmentation model. This study introduces a novel end-to-end trainable network designed for skin lesion segmentation. The proposed methodology consists of an encoder-decoder, a region-aware attention approach, and guided loss function. The trainable parameters are reduced using depth-wise separable convolution, and the attention features are refined using a guided loss, resulting in a high Jaccard index. We assessed the effectiveness of our proposed RA-Net on four frequently utilized benchmark datasets for skin lesion segmentation: ISIC 2016, ISIC 2017, ISIC 2018, and PH2. The empirical results validate that our method achieves state-of-the-art performance, as indicated by a notably high Jaccard index.