Lightweight Skin Lesion Segmentation Network With Multi-Scale Feature Fusion Interaction

Abstract

The existing segmentation algorithms have many problems, such as a large number of parameters, a complicated calculation process, and difficulty in accurately segmenting skin lesion areas with hair interference, blurred edges, and unclear lesion features. We propose a lightweight skin lesions segmentation network (LSLS-Net) to address the above problems. In the part of encoded feature extraction, we extract multi-scale features through different sizes of convolution kernels to capture rich detailed features of the skin lesion area; then we use a feature fusion enhancement module to enhance the extracted features. That is, we design a lightweight feature extraction module that extracts global features, an edge feature enhancement module that enhances edge features, and a feature fusion attention module that fuses and enhances global features and edge features. At the same time, the obtained different feature information is interfused with the unenhanced features to obtain more abundant features. Experimental results on two public datasets, ISIC-2018 and PH2, show that compared with current mainstream medical image segmentation algorithms UNet, AttentionUNet, UNet++, DoubleU-Net, CACDU-Net, EIU-Net, and HmsU-Net, the proposed algorithm not only obtains excellent performance in the number of parameters and computational complexity but also has a clear outline and continuous edge for the segmentation of skin lesions, which has a better segmentation effect. Additionally, experiments on the PH2 dataset further show that LSLS-Net possesses strong generalization capabilities.