LCM-Net: LLM-Driven Cross-Modality MoE Feature Fusion Network for Cancer Survival Analysis.

Area of Science:

• Oncology
• Bioinformatics
• Medical Imaging

Background:

• Cancer survival analysis is crucial for treatment evaluation and prognosis.
• Current cross-modal learning methods struggle with data redundancy and high-dimensional genetic data representation.
• Pathology image analysis is computationally intensive, and data heterogeneity hinders multimodal fusion.

Purpose of the Study:

• To propose a novel Large Language Model (LLM)-driven Cross-Modality MoE-feature Fusion Network (LCM-Net) for enhanced cancer survival prediction.
• To address challenges in integrating genomic and pathology data, including data redundancy and computational intensity.
• To improve the accuracy and efficiency of predicting cancer survival outcomes.

Main Methods:

• Developed a Genomic Language Alignment (GLA) module using LLMs to encode genomic features into concise representations.
• Introduced a Pathological Feature Refinement (PFR) module to filter irrelevant regions in pathology images.
• Proposed a Multimodal Expert Integration (MEI) module to fuse processed genomic and pathological features.

Main Results:

• LCM-Net demonstrated superior performance compared to state-of-the-art methods across five public datasets.
• Ablation studies confirmed the significant contribution of each proposed module (GLA, PFR, MEI).
• The approach effectively handles data redundancy and computational challenges in multimodal cancer data analysis.

Conclusions:

• The proposed LCM-Net offers a powerful and effective approach for cancer survival prediction by integrating diverse data modalities.
• The innovative modules significantly enhance feature representation and fusion for improved prognostic accuracy.
• This LLM-driven framework provides a promising direction for future research in computational oncology.