Benchmarking common uncertainty estimation methods with histopathological images under domain shift and label noise

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Deep learning for medical imaging

Background:

• Deep learning (DL) shows promise in histopathology but requires robust uncertainty estimation for high-risk applications.
• Models must identify and reject uncertain classifications to prevent misdiagnosis in critical scenarios.

Purpose of the Study:

• To rigorously evaluate uncertainty and robustness methods for Whole Slide Image (WSI) classification.
• To assess the effectiveness of selective classification, where models reject uncertain predictions.
• To investigate performance under domain shift and label noise conditions.

Main Methods:

• Comparison of Deep Ensembles, Monte-Carlo Dropout, Stochastic Variational Inference, and Test-Time Data Augmentation.
• Evaluation on tile-level and slide-level data, considering domain shift and label noise.
• Analysis of ensemble approaches combining multiple uncertainty estimation techniques.

Main Results:

• Ensemble methods generally yield superior uncertainty estimates and robustness against domain shifts and label noise.
• Selective classification (rejecting uncertain samples) consistently improves accuracy on both in-distribution and out-of-distribution data.
• Unlike classical computer vision, no single method consistently outperformed others across all histopathology tasks.

Conclusions:

• Ensemble strategies are crucial for enhancing uncertainty quantification and robustness in DL-based histopathology.
• Selective classification is a vital strategy for improving diagnostic accuracy and reliability in medical AI.
• The study provides a code framework to advance research in uncertainty estimation for histopathological data.