GenoPath-MCA: Multimodal masked cross-attention between genomics and pathology for survival prediction

Area of Science:

• Computational pathology
• Bioinformatics
• Cancer research

Background:

• Survival prediction is crucial for cancer risk assessment and personalized treatment.
• Integrating whole slide images (WSIs) with genomic data presents challenges in modality alignment and fusion.
• Existing methods struggle with data heterogeneity and lack personalized fusion strategies.

Purpose of the Study:

• To develop a novel multimodal framework, GenoPath-MCA, for enhanced cancer survival prediction.
• To model dense cross-modal interactions between histopathology and gene expression data.
• To address patient-level heterogeneity and improve fusion strategies.

Main Methods:

• Proposed GenoPath-MCA, a multimodal framework utilizing masked co-attention for feature alignment.
• Introduced Multimodal Masked Cross-Attention Module (M2CAM) for high-order image-gene and gene-gene relationship capture.
• Developed Dynamic Modality Weight Adjustment Strategy (DMWAS) for adaptive fusion weights and importance-guided patch selection.

Main Results:

• GenoPath-MCA significantly outperformed existing methods on public multimodal cancer survival datasets.
• Achieved superior performance in concordance index and robustness.
• Visualizations confirmed biological interpretability and clinical potential.

Conclusions:

• GenoPath-MCA offers a robust and interpretable approach for multimodal cancer survival prediction.
• The framework effectively integrates histopathology and genomic data, addressing key challenges in the field.
• Demonstrates significant potential for improving automated risk assessment and personalized cancer treatment planning.