LGCF-Net: Local-Global Competitive Fusion for Ultra-Lightweight Medical Image Segmentation

Abstract

Although deep neural networks possess powerful feature representation capabilities, they typically demand massive parameter scales and heavy computational burdens, while current lightweight alternatives that reduce model sizes through operations like pruning or depthwise separable convolutions often struggle to maintain the original representation capacity. To address these challenges, we propose a ultra-lightweight medical image segmentation network based on local-global competitive fusion, termed LGCF-Net. Firstly, to simultaneously enhance the perception of localized boundary textures and capture long-range contextual dependencies under limited computational budgets, we design a parallel dual-branch architecture at each resolution layer, integrating an Asymmetric Depthwise Separable Convolution (ADSC) module and a Mamba-based Visual State Space (VSS) module. Secondly, to mitigate feature redundancy and representation conflicts arising from this dual-branch integration, we introduce a Branch Competitive Gating (BCG) module to achieve adaptive feature selection via branch-wise competitive weight allocation. Finally, to suppress cross-layer noise propagation during feature transmission, we develop a Decoder-Guided Attention Gating (DGAG) mechanism within the skip connections to filter and refine shallow encoder features using higher-level decoder semantic priors. Extensive experiments on four public datasets demonstrate that LGCF-Net consistently outperforms state-of-the-art methods. Notably, our network compresses the 34.52M baseline parameters by nearly 96% to just 1.53M. On the ACDC dataset, it achieves a 91.56% Dice score with only 2.82G FLOPs, establishing a effective trade-off between segmentation accuracy and computational efficiency.