A Lightweight Multi-Branch Skin Lesion Segmentation Network

In the task of skin lesion segmentation, there is often a trade-off between the demand for high precision and the desire for low computational cost. To address this issue, we propose a lightweight multi-branch network for skin lesion segmentation, named LMBNet. LMBNet primarily consists of three blocks: Multi-scale Axial Refinement Convolution (MARC), Efficient Axial Refinement Convolution (EARC), and Multi-scale Edge Feature Fusion (MEFF). The proposed MARC is designed to effectively blend and extract spatial and positional information from features. During the decoding stage, we developed the EARC to efficiently aggregate rich local features and global contextual information. In the lower-level stages, the MEFF is utilized to capture contextual semantic cues and edge information provided by the lower layers. Experiments on the ISIC2017 skin lesion segmentation dataset demonstrate that, compared to the baseline model BaseLineUNet, LMBNet reduces the number of parameters by 71%, decreases the computational cost by 62%, and improves the mIoU, DSC and Accuracy by 2.34%, 1.47% and 0.57% respectively. Moreover, on the ISIC2017 and ISIC2018 skin lesion segmentation datasets, LMBNet outperforms nine commonly used skin lesion segmentation networks—UNet, UNet++, Attention-UNet, TransFuse, UTNetV2, SANet, UNeXt-S, MALUNet, and EGE-UNet—in terms of Params, GFLOPs, Accuracy, and shows superior mIoU and DSC. Thus, our model achieves optimal overall performance by effectively balancing parameter count, computational demands, and segmentation precision.