Deep learning uncertainty and confidence calibration for the five-class polyp classification from colonoscopy.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning models in medical imaging face interpretability challenges.
• Confidence calibration and classification uncertainty are key but their interaction and impact are unclear.

Purpose of the Study:

• Investigate the roles of confidence calibration and classification uncertainty in deep learning models.
• Propose a new Bayesian deep learning method integrating calibration and uncertainty.

Main Methods:

• Studied confidence calibration using post-process temperature scaling.
• Computed classification uncertainty from classification entropy and Bayesian predicted variance.
• Developed a novel Bayesian deep learning method.

Main Results:

• Confidence calibration and classification uncertainty improve model interpretation and accuracy.
• The proposed Bayesian method achieved state-of-the-art results in confidence calibration and classification accuracy on a polyp dataset.

Conclusions:

• Integrating confidence calibration and classification uncertainty enhances deep learning model interpretability and performance in medical image analysis.
• The novel Bayesian approach offers significant improvements for tasks like polyp classification.