Reproduce “XAG-Net: A Cross-Slice Attention and Skip Gating Network for 2.5D Femur MRI Segmentation” on OpenResearch

About this paper

Byunghyun Ko, Anning Tian, Jeongkyu Lee

XAG-Net, a 2.5D U-Net based architecture, enhances femur MRI segmentation accuracy and boundary precision by integrating pixel-wise cross-slice attention (CSA) and skip attention gating (AG) mechanisms. The model achieved a Dice Similarity Coefficient (DSC) of 0.9535 and an IoU of 0.9160, outperforming 3D U-Net by 14.78 percentage points in DSC while using 56.3% fewer parameters.

Key insights

• The pixel-wise softmax normalization in the CSA module effectively models fine-grained inter-slice relationships, crucial for structures with rigid boundaries like the femur.
• Combining CSA modules and AG blocks within skip connections demonstrates a synergistic effect, where CSA enables AG to more effectively refine intra-slice features, leading to enhanced segmentation.
• A thoughtfully designed 2.5D architecture incorporating specialized attention mechanisms can achieve superior segmentation accuracy and boundary precision while maintaining computational efficiency over 3D models.

Summary by alphaXiv.