Uncertainty-Aware Causally Adaptive Foundation Model Enhances Colorectal Cancer Diagnosis

Overview

The UAD-FM model integrates uncertainty quantification, causal adaptation, and confidence calibration to improve colorectal cancer pathology diagnosis. It demonstrates superior accuracy, robustness to domain shifts, and provides interpretable uncertainty maps across multiple public CRC datasets.

Background

Colorectal cancer (CRC) is a leading global malignancy diagnosed primarily through histopathological examination of H&E-stained whole-slide images (WSIs). Computational pathology models face challenges including domain shifts from multi-center variability, unreliable uncertainty estimation, and spurious correlations that limit clinical reliability. Existing models often lack robust confidence measures and adaptability, hindering their integration into clinical workflows. Advances in foundation models, uncertainty decomposition, and causal representation learning offer promising avenues to address these limitations.

Data Highlights

Key Findings

• UAD-FM decomposes uncertainty into epistemic and aleatoric components tailored for CRC histopathology.
• Incorporates causal test-time adaptation using do-interventions to remove spurious correlations and improve robustness.
• Achieves superior accuracy and calibration compared to existing foundation models across five diverse CRC datasets.
• Generates interpretable uncertainty maps facilitating human-AI collaboration and clinical trust.
• Employs entropy-based adaptation strategies at inference time to address domain shifts without requiring source data.
• Integrates post-hoc confidence calibration to defer uncertain cases to pathologists, enhancing safety in clinical deployment.

Clinical Implications

The UAD-FM framework offers a reliable and generalizable tool for CRC pathology diagnosis that adapts to multi-institutional variability and quantifies prediction confidence. Its interpretable uncertainty outputs can guide pathologists in decision-making, potentially reducing diagnostic errors and workload. By enabling safe deferral of uncertain cases, it supports seamless integration into clinical workflows, improving diagnostic accuracy and patient outcomes.

Conclusion

UAD-FM represents a significant advancement in computational pathology by unifying uncertainty modeling, causal adaptation, and confidence calibration within a foundation model. This approach enhances the reliability and generalizability of colorectal cancer diagnosis across heterogeneous clinical settings.

References

1. Global Cancer Statistics 2020 -- Colorectal Cancer Incidence and Mortality
2. TCGA-COAD/READ Dataset -- The Cancer Genome Atlas Colorectal Adenocarcinoma
3. DigestPath 2019 Challenge -- Colorectal Cancer Histopathology
4. NCT-CRC-HE-100K Dataset -- Colorectal Histology Images
5. LC25000 Dataset -- Histopathological Image Dataset for Cancer Classification