Author: Jingyun Chen, Yading Yuan 👨🔬
Affiliation: Columbia University Irving Medical Center, Department of Radiation Oncology 🌍
Purpose: To develop and evaluate the Scale-attention network (SANet) for automated pancreas segmentation on MR images.
Methods: To develop SANet, we extended the classic U-Net design with a dynamic scale attention mechanism that effectively combines low-level details with high-level semantics from feature maps at multiple scales, enhancing SANet’s ability to integrate information across different resolutions. To evaluate the performance of SANet, we utilized 384 T1-weighted MRI scans from the recently published PanSeg dataset, which includes contributions from five different sites. Following the experimental setup described in the PanSeg paper, we combined data from Sites 1 and 2 for model training and 5-fold cross-validation, while using data from 72 patients from the remaining three sites as external test cases. We evaluated the performance of SANet in comparison to the state-of-the-art model PanSegNet, the model proposed in the PanSeg paper. The average Dice Similarity Coefficient (DSC), 95% Hausdorff Distance (HD95) and Average Symmetric Surface Distance (ASSD) were used as evaluation metrics.
Results: With just 10.3M model parameters trained over 500 epochs, SANet outperformed PanSegNet in both internal validation (DSC: 85.35% vs 85.02%, HD95: 6.13 vs 6.37, ASSD: 1.38 vs 1.32) and external testing (DSC: 81.61% vs 79.27, HD95: 6.24 vs 9.23, ASSD: 1.75 vs 2.19), demonstrating its superior generalization capability. Compared to PanSegNet, which required 31.1M parameters and 1000 epochs, SANet achieved comparable or better performance with only one-third of the parameters and half the training epochs. This efficiency is particularly advantageous for clinical deployment, as it enables faster training convergence while reducing computational requirements and model complexity, all while maintaining high segmentation performance.
Conclusion: Our preliminary results demonstrate that SANet is a highly effective and efficient model for accurate pancreas segmentation in MRI images, which plays an importance role in the detection and classification of pancreas tumor.