L-SwinUNet：融合注意力增强与深度可分离卷积的轻量化分割模型

L-SwinUNet: A Lightweight Segmentation Model that Combines Attention Enhancement with Deep Separable Convolutions

摘要: 针对现有基于Transformer的分割网络普遍存在参数冗余、计算复杂度高、推理效率低等问题，本文提出一种轻量化医学图像分割网络L-SwinUNet。该网络从三个维度进行协同优化：首先，在编码器–解码器跳跃连接中嵌入无参数SimAM注意力模块，自适应增强浅层空间语义与边界敏感特征；其次，在解码器中采用深度可分离卷积替代标准卷积，以分离式特征提取策略大幅削减参数规模与浮点运算量；最后，在上采样阶段引入CARAFE内容感知重组算子，通过自适应核预测机制精细重建高分辨率边缘细节。实验在Synapse数据集上验证了方法的有效性。结果表明，相较原始Swin-UNet，L-SwinUNet在减少约48%参数与40%计算量的同时还提升了Dice与HD95指标，证明其在医学图像分割中的轻量化优势与精度潜力。

Abstract: To address the prevalent issues of parameter redundancy, high computational complexity, and low inference efficiency in existing Transformer-based segmentation networks, this paper proposes a lightweight medical image segmentation network, L-SwinUNet. The network undergoes synergistic optimisation across three dimensions: firstly, embedding parameter-free SimAM attention modules within the encoder-decoder skip connections to adaptively enhance shallow spatial semantic and boundary-sensitive features; secondly, it employs depthwise separable convolutions within the decoder instead of standard convolutions, significantly reducing parameter size and floating-point operations through a separable feature extraction strategy; finally, it introduces the CARAFE content-aware reorganisation operator during the upsampling stage, which employs an adaptive kernel prediction mechanism to finely reconstruct high-resolution edge details. Experiments on the Synapse dataset validate the method’s efficacy. Results demonstrate that compared to the original Swin-UNet, L-SwinUNet achieves approximately 48% fewer parameters and 40% reduced computational load while improving Dice and HD95 metrics, proving its lightweight advantages and accuracy potential in medical image segmentation.