Uncertainty-informed deep learning models enable high-confidence predictions for digital histopathology.

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning in oncology

Background:

• Clinical deployment of computational biomarkers requires user confidence.
• Expressing predictive uncertainty is crucial for reliable AI in healthcare.

Purpose of the Study:

• To develop a clinically-oriented uncertainty quantification approach for whole-slide images in cancer digital histopathology.
• To improve the reliability and confidence of AI-driven diagnostic predictions.

Main Methods:

• Utilized dropout-based uncertainty estimation for whole-slide images.
• Established confidence cutoffs using thresholds calculated on training data.
• Trained models to differentiate lung adenocarcinoma from squamous cell carcinoma.

Main Results:

• High-confidence predictions demonstrated superior performance compared to predictions without uncertainty.
• Performance was validated through cross-validation and testing on two large, multi-institutional external datasets.
• Uncertainty thresholding proved reliable under domain shift, maintaining accurate predictions for out-of-distribution cohorts.

Conclusions:

• The developed uncertainty quantification method enhances the reliability of AI models in digital pathology.
• This approach improves diagnostic accuracy for lung cancer subtypes and maintains performance across different datasets and domains.
• Clinically-oriented uncertainty estimation is vital for trustworthy AI biomarker deployment.