Multi-Conv attention network for skin lesion image segmentation

To address the trade-off between segmentation performance and model lightweighting in computer-aided skin lesion segmentation, this paper proposes a lightweight network architecture, Multi-Conv Attention Network (MCAN). The network consists of two key modules: ISDConv (Inception-Split Depth Convolution) and AEAM (Adaptive Enhanced Attention Module). ISDConv reduces computational complexity by decomposing large kernel depthwise convolutions into smaller kernel convolutions and unit mappings. The AEAM module leverages dimensional decoupling, lightweight multi-semantic guidance, and semantic discrepancy alleviation to facilitate the synergy between channel attention and spatial attention, further exploiting redundancy in the spatial and channel feature maps. With these improvements, the proposed method achieves a balance between segmentation performance and computational efficiency. Experimental results demonstrate that MCAN achieves state-of-the-art performance on mainstream skin lesion segmentation datasets, validating its effectiveness.