Selective multimodal deep learning for reliable breast cancer subtype classification from histopathology and genomic...

Area of Science:

• Computational biology and bioinformatics
• Digital pathology and machine learning

Background:

• Accurate breast cancer subtype classification is crucial for personalized treatment and prognosis.
• Integrating histopathology and RNA-seq data presents challenges due to data heterogeneity.
• Existing methods struggle with combining multimodal data effectively for improved classification.

Purpose of the Study:

• To develop a robust multimodal framework combining histology and transcriptomic data for breast cancer subtype classification.
• To enhance classification accuracy, interpretability, and computational efficiency.
• To introduce an uncertainty-aware smart routing mechanism for optimized inference.

Main Methods:

• Utilized CTransPath vision transformer for histology feature extraction from whole-slide images (WSIs).
• Integrated standardized RNA-seq features into a shared latent space.
• Implemented and evaluated multimodal fusion techniques (gated attention, cross-attention, concatenation) with an uncertainty-aware smart routing mechanism.

Main Results:

• The routing-based multimodal model achieved 94.93% accuracy on the TCGA-BRCA dataset, outperforming unimodal and fixed fusion methods.
• The routing mechanism reduced multimodal inference to 38.2% of samples, yielding a 3.12× computational speedup.
• Attention rollout visualized discriminative histological regions, enhancing interpretability and improving distinction between similar subtypes (e.g., Luminal A/B).

Conclusions:

• Integrating CTransPath-derived histology with transcriptomic profiles via confidence-aware routing offers a practical and explainable approach.
• The proposed framework significantly improves breast cancer subtype classification accuracy and computational efficiency.
• This method is suitable for clinical decision support systems, providing interpretable insights for pathological assessment.