Uncertainty quantification in skin cancer classification using three-way decision-based Bayesian deep learning.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Dermatology

Background:

• Automated medical image recognition is crucial but challenging.
• Deep learning excels in medical image analysis but often lacks uncertainty quantification (UQ).
• Overconfident predictions from deep learning models can have severe consequences in clinical settings.

Purpose of the Study:

• To address the limitations of current deep learning models in medical image analysis by incorporating uncertainty quantification.
• To develop and evaluate a novel hybrid dynamic Bayesian Deep Learning (BDL) model for skin cancer classification.
• To improve the reliability and reduce overconfident decision-making in automated skin cancer diagnosis.

Main Methods:

• Applied three UQ methods: Monte Carlo (MC) dropout, Ensemble MC (EMC) dropout, and Deep Ensemble (DE).
• Developed a novel hybrid dynamic BDL model based on Three-Way Decision (TWD) theory to further reduce uncertainty.
• Utilized DE and EMC methods in distinct classification phases for analyzing two skin cancer datasets.

Main Results:

• The proposed TWDBDL model demonstrated strong performance on two skin cancer datasets.
• Achieved an accuracy of 88.95% and F1-score of 89.00% on the first dataset.
• Achieved an accuracy of 90.96% and F1-score of 91.00% on the second dataset.

Conclusions:

• The proposed TWDBDL model effectively quantifies uncertainty in skin cancer image classification.
• The hybrid dynamic model enhances decision-making by preventing overconfident predictions.
• This approach shows significant potential for application across various stages of medical image analysis.