Decouple, Reorganize, and Fuse: A Multimodal Framework for Cancer Survival Prediction

Area of Science:

• Biomedical Informatics
• Computational Biology
• Machine Learning for Healthcare

Background:

• Cancer survival analysis integrates diverse medical data for prediction.
• Current fusion methods (concatenation, attention, MoE) have limitations in dynamic feature combination and inter-feature information interaction.

Purpose of the Study:

• To propose a novel Decoupling-Reorganization-Fusion (DeReF) framework to address limitations in existing cancer survival analysis fusion methods.
• To enhance the dynamic fusion of decoupled features and improve information interaction among modalities.

Main Methods:

• Developed a DeReF framework featuring a random feature reorganization strategy.
• Incorporated dynamic Mixture-of-Experts (MoE) fusion modules.
• Integrated a regional cross-attention network within the modality decoupling module.

Main Results:

• The DeReF framework demonstrated improved generalization ability by increasing feature combination diversity.
• Overcame information closure issues, enabling expert networks to better capture inter-feature information.
• Achieved effective results on Liver Cancer and TCGA datasets.

Conclusions:

• The proposed DeReF framework offers a significant advancement in multimodal cancer survival analysis.
• Dynamic feature reorganization and fusion enhance prediction accuracy and model robustness.
• The method shows promise for improving clinical decision-making in cancer treatment.