HiCAF-Net: A Hierarchical Cross-Attention Fusion framework for cross-cancer subtype classification using...

Area of Science:

• Oncology
• Bioinformatics
• Computational Pathology

Background:

• Accurate cancer subtype classification is crucial for personalized medicine but challenged by tumor heterogeneity and limitations of unimodal analysis.
• Existing multimodal fusion methods struggle to integrate multi-scale morphological and genomic data, and often overlook pan-cancer commonalities.

Purpose of the Study:

• To develop a novel Hierarchical Cross-Attention Fusion and multitask learning framework (HiCAF-Net) for actionable knowledge discovery from fragmented oncology data.
• To effectively integrate multi-scale pathological features with high-dimensional genomic data for improved pan-cancer analysis.

Main Methods:

• HiCAF-Net employs a coarse-to-fine visual strategy using attention-based multiple instance learning and graph convolutional networks for multi-scale feature extraction.
• A dual-layer stepwise cross-attention mechanism bridges the semantic gap between pathological features and transcriptomic profiles.
• Adversarial domain adaptation synchronizes optimization across heterogeneous cancer types to uncover shared cross-cancer commonalities.

Main Results:

• HiCAF-Net significantly outperforms state-of-the-art single-modal and single-task baselines across eight distinct cancer types.
• The framework demonstrates robust generalization capabilities in joint pan-cancer analysis.
• Experiments were conducted on TCGA cohorts and a private PTMC dataset.

Conclusions:

• HiCAF-Net offers an efficient and deployable solution for integrating multi-scale pathological and genomic data in oncology.
• The proposed framework enhances actionable knowledge discovery for personalized medicine through improved pan-cancer analysis.
• The study highlights the potential of hierarchical cross-attention fusion and adversarial domain adaptation for complex biological data integration.