Reproduce “Multi-modal Evidential Fusion Network for Trustworthy PET/CT Tumor Segmentation” on OpenResearch

About this paper

Yuxuan Qi, Li Lin, Jiajun Wang, Bin Zhang, Jingya Zhang

Researchers at Soochow University and collaborating institutions developed a Multi-modal Evidential Fusion Network (MEFN) for trustworthy PET/CT tumor segmentation. The framework, which integrates uncertainty quantification via Dempster-Shafer Theory, consistently outperformed state-of-the-art methods, achieving Dice Similarity Coefficients of 82.45% on AutoPET and 83.35% on Hecktor datasets, while providing interpretable uncertainty estimates for clinical decision-making.

Key insights

• Dempster-Shafer Theory provides an interpretable mechanism for decision-level fusion, allowing dynamic weighting of modality contributions based on their individual uncertainty estimates.
• Alternating training of a GAN-based Cross-Modal Feature Learning (CFL) stage effectively learns robust, aligned feature representations across PET and CT, mitigating inherent domain gaps.
• Novel loss functions, like Uncertainty Perceptual Loss (L_UP), and an Uncertainty Calibrator (UC) within the DFC module enhance model robustness by explicitly focusing on and learning from high-uncertainty regions.

Summary by alphaXiv.