MARNet: Medical Image Segmentation Network Based on Multi-axis Attention Mechanism and Aggregated Supervision Enhancement

Recently, some pioneering works tend to apply more complex modules to enhance the segmentation performance. However, this is not conducive to real-world clinical settings due to limited computational resources. To address this challenge, we propose a more optimized high-performance lightweight U-shaped image segmentation model: Multi-axis Attention Mechanism and Aggregate Supervision Enhanced Image Segmentation Network (MARNet). It integrates the Enhanced Multi-axis Attention (EMIA) module and the Multi-featiture Aggregate Supervision Fusion Module (ARFM). EMIA utilizes memory units, residual connections, and element-wise operations to obtain enhanced image multi-view edge information while reducing the parameters of the model’s self-attention mechanism. The ARFM module effectively fuses multi-scale features by adapting to different scales and utilizing auxiliary masks. Our test results on the ISIC2017 and ISIC2018 datasets show that MARnet outperforms existing methods while maintaining minimal computational requirements, demonstrating its great potential in the field of medical image segmentation.