VSGNet: visual saliency guided network for skin lesion segmentation

Abstract

The accuracy of skin lesion segmentation is of great significance for the subsequent clinical diagnosis. In order to improve the segmentation accuracy, some pioneering works tried to embed multiple complex modules, or used the huge Transformer framework, but due to the limitation of computing resources, these type of large models were not suitable for the actual clinical environment. To address the coexistence challenges of precision and lightweight, we propose a visual saliency guided network (VSGNet) for skin lesion segmentation, which generates saliency images of skin lesions through the efficient attention mechanism of biological vision, and guides the network to quickly locate the target area, so as to solve the localization difficulties in the skin lesion segmentation tasks. VSGNet includes three parts: Color Constancy module, Saliency Detection module and Ultra Lightweight Multi-level Interconnection Network (ULMI-Net). Specially, ULMI-Net uses a U-shaped structure network as the skeleton, including the Adaptive Split Channel Attention (ASCA) module that simulates the parallel mechanism of biological vision dual pathway, and the Channel-Spatial Parallel Attention (CSPA) module inspired by the multi-level interconnection structure of visual cortices. Through these modules, ULMI-Net can balance the efficient extraction and multi-scale fusion of global and local features, and try to achieve the excellent segmentation results at the lowest cost of parameters and computational complexity. To validate the effectiveness and robustness of the proposed VSGNet on three publicly available skin lesion segmentation datasets (ISIC2017, ISIC2018 and PH2 datasets). The experimental results show that compared to other state-of-the-art methods, VSGNet improves the Dice and mIoU metrics by 1.84 % and 3.34 %, respectively, and with a 196 × and 106 × reduction in the number of parameters and computational complexity. This paper constructs the VSGNet integrating the biological vision mechanism and the artificial intelligence algorithm, providing a new idea for the construction of deep learning models guided by the biological vision, promoting the development of biomimetic computational vision as well as.