C-UQ: Conflict-based uncertainty quantification-A case study in lung cancer classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Uncertainty quantification is vital for reliable deep learning in medical diagnostics.
• Current methods may have limitations in accurately assessing model confidence.

Purpose of the Study:

• To introduce a novel conflict-based uncertainty quantification approach using Dempster-Shafer Theory and Deep Ensembles.
• To apply this method to lung cancer classification and evaluate its effectiveness.

Main Methods:

• Leveraging Dempster-Shafer Theory with Deep Ensemble methods.
• Converting softmax outputs to Basic Belief Assignments and applying the rule of combination.
• Using conflict as a measure of uncertainty in aggregated predictions.

Main Results:

• Achieved high accuracy (0.957) and U_Recall (0.819) in lung cancer classification on the LIDC-IDRI dataset.
• Demonstrated superior Out-of-Distribution detection with AUC scores up to 0.864.
• Showcased improved performance with increased ensemble size compared to entropy-based methods.

Conclusions:

• The conflict-based uncertainty quantification method effectively measures prediction confidence in medical deep learning.
• This approach enhances reliability for clinical decisions and improves Out-of-Distribution detection.
• Future work will focus on optimizing efficiency and exploring advanced Dempster-Shafer Theory applications.