MDLA-UNet: A Lightweight Multi-scale Directional Aggregation UNet for Skin Lesion Segmentation
摘要
Accurate segmentation of skin lesions is crucial for the early diagnosis and treatment planning of skin cancer. However, existing high-performance models often exhibit high computational complexity and large parameters, making them difficult to deploy on resource-constrained mobile medical devices. To address this challenge, this paper proposes an efficient CNN-based medical image segmentation model—the Lightweight Multi-Scale Directional Aggregation UNet (MDLA-UNet). Building upon the classic U-Net architecture, this network incorporates an Axial Dilated Multi-scale (ADM) module as its encoder-decoder backbone. It efficiently captures long-range contextual dependencies and local details with minimal parameters by utilizing parallel horizontal and vertical long-range convolutional branches and multi-directional dilated convolutions. Additionally, the Feature Embedding Bridge (FEB) module optimizes skip connections. Through grouped aggregation and multi-scale attention mechanisms, it effectively integrates high-level semantic information with low-level spatial details, mitigating common issues such as blurred lesion boundaries and irregular segmentation masks. Extensive experiments on the ISIC2017, ISIC2018, and PH2 datasets demonstrate that MDLA-UNet achieves competitive performance, with DSCs of