Bayesian deep learning for reliable oral cancer image classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning achieves high performance in medical imaging but lacks reliability due to inability to quantify model uncertainty.
• Integrating deep learning into clinical workflows is hindered by the lack of reliable uncertainty estimation.

Purpose of the Study:

• To develop and evaluate a Bayesian deep network for quantifying uncertainty in oral cancer image classification.
• To improve the reliability and accuracy of deep learning models in medical diagnostics.

Main Methods:

• Utilized a Bayesian deep network architecture capable of estimating predictive uncertainty.
• Trained and evaluated the model on a large dataset of intraoral cheek mucosa images from a high-risk population.
• Implemented an uncertainty-informed referral strategy to identify difficult cases.

Main Results:

• The Bayesian deep network successfully produced meaningful uncertainty information for oral cancer image classification.
• Uncertainty-informed referral significantly improved classification accuracy, reaching approximately 90% for specific referral criteria.
• The model demonstrated capability in identifying challenging cases requiring expert review.

Conclusions:

• Bayesian deep networks offer a reliable approach to uncertainty estimation in medical imaging, addressing a key limitation of conventional deep learning.
• Uncertainty quantification enhances the practical utility of AI in clinical settings, particularly for early disease detection like oral cancer.
• The proposed method facilitates improved diagnostic workflows by prioritizing cases that require further clinical attention.