Counterfactual Bidirectional Co-Attention Transformer for Integrative Histology-Genomic Cancer Risk Stratification

Area of Science:

• Computational pathology
• Bioinformatics
• Machine learning in oncology

Background:

• Predicting patient survival from histological whole-slide images (WSIs) and genomic data is complex due to tumor heterogeneity.
• Current deep learning methods face challenges with learning biases from incomplete or irrelevant genomic guidance, leading to suboptimal prognostic predictions.

Purpose of the Study:

• To develop an advanced deep learning framework that effectively integrates histological and genomic data for improved patient prognostic survival prediction.
• To address and mitigate learning biases inherent in existing multimodal cancer data analysis methods.

Main Methods:

• Proposed the CounterFactual Bidirectional Co-Attention Transformer (CFBCT) framework.
• Integrated a bidirectional co-attention layer for enhanced feature interaction between genomic and histology data.
• Employed counterfactual reasoning and causal modeling for cancer risk stratification and bias reduction.

Main Results:

• The CFBCT framework demonstrated superior performance across eight diverse cancer benchmark datasets from The Cancer Genome Atlas (TCGA).
• Achieved an average 2.5% improvement in c-index performance compared to 18 state-of-the-art models in patient prognosis prediction.
• Showcased effective feature interaction and consistent identification of prognostic features from WSIs.

Conclusions:

• The proposed CFBCT framework offers a significant advancement over existing histology-genomic learning methods for patient survival prediction.
• The causal and counterfactual approach enhances understanding of feature influence on survival outcomes and reduces model bias.
• The framework's robust validation across multiple cancer types highlights its potential for clinical application.