Lightweight Unet with depthwise separable convolution for skin lesion segmentation

Abstract

Abstract Accurate segmentation of skin lesions is crucial for the early diagnosis of skin diseases, as clear lesion boundaries facilitate the comprehensive extraction of lesion features. However, in practical applications, the contours or boundaries of lesion areas in medical images often appear blurred and lose detail due to factors such as lighting conditions, imaging equipment, skin color differences, and physician experience. Meanwhile, existing deep learning models are often large in structure and computationally expensive, making them unsuitable for clinical environments that require efficiency and ease of deployment.To address these issues, this paper proposes a lightweight skin lesion segmentation model, LMSAUnet. Based on the traditional UNet architecture, this model removes standard convolution modules and introduces the ECDF module, which combines deep separable convolution and attention mechanisms. This significantly reduces the number of parameters and computational complexity while improving segmentation performance.Experimental results demonstrate that LMSAUnet outperforms multiple state-of-the-art segmentation models on several skin lesion datasets, significantly reducing computational resource consumption while exhibiting stronger generalization capabilities. The findings validate the model’s potential for efficiently and accurately performing skin lesion segmentation tasks under resource-constrained conditions, offering promising clinical application prospects.