GLA-UNet: A Light UNet with Grouped Laplacian-Attention for skin lesion segmentation

In recent years, most approaches tend to incorporate more complex modules to enhance the accuracy of medical image segmentation, which often entails high computational costs. Due to limited computational resources in mobile devices, models with heavy computational loads are difficult to deploy and apply in real clinical environments. To address this issue, we propose a light-weight model named GLA-UNet. Specifically, we design three key modules: (1) the Light Residual Convolution (LRConv) Module, which combines pointwise convolution with standard convolution equipped with residual connections to extract local features; (2) the Local Global Grouped Laplacian Attention Aggregation (LGGLAA) Module, based on a grouped lightweight Laplacian-Attention mechanism, effectively models long range dependencies and enhances edge perception while maintaining low computational complexity; (3) the Grouped Residual Multi scale Aggregation (GRMA) Module, primarily composed of four dilated convolutions with different dilation rates, designed to achieve global and local fusion of multistage features. The proposed GLA-UNet outperforms existing methods across multiple metrics, as validated on three public dermoscopy datasets: ISIC2017, ISIC2018, and PH2. More importantly, with only 0.035 M parameters and 0.055 GFLOPs computational cost, significantly lower than current models.