DSNET: A Lightweight Segmentation Model for Segmentation of Skin Cancer Lesion Regions

Currently, most skin disease segmentation tasks tend to use large models to achieve better segmentation performance. However, in real medical application scenarios, due to the limited computing and storage resources of hardware devices, it is difficult to deploy large models in business systems for applications. To address this problem, we propose a lightweight segmentation architecture that concentrates more on details. This model achieves optimal segmentation performance while maintaining low model parameters and computational complexity. To reduce the model size and guarantee model segmentation performance, we proposed a detail-enhanced separable difference convolution as a base module in the model. Second, we designed a dynamic gate attention module as a bridge between the encoding and decoding layers to further improve the expressiveness of flat texture information. Furthermore, we propose a boundary-interior correction loss function that focuses on both the boundary and the interior of the region, which effectively solves the dermatosis segmentation task for problems such as boundary blurring and interior missing. We conducted extensive experiments on two publicly available skin lesion segmentation datasets (ISIC2017 and ISIC2018). Our method achieved state-of-the-art results compared to previous studies. DSNET has only 0.291M parameters and 0.287 GFLops, and it achieves 81.30% mIoU and 89.81% DSC metrics, respectively. DSNET has two main features: lightweight and high efficiency, which not only highlights dermatological detail feature extraction but also supports deployment on medical mobile devices.