Deep Bayesian active learning-to-rank with relative annotation for estimation of ulcerative colitis severity.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Accurate image-based severity estimation is crucial for computer-aided diagnosis.
• Deep learning models require extensive data, and discrete severity labels are costly and difficult to obtain for ambiguous cases.
• Relative annotation offers an alternative but faces challenges with the sheer volume of potential image pairs.

Purpose of the Study:

• To develop an automated method for selecting optimal image pairs for relative severity annotation.
• To improve the efficiency and accuracy of deep learning-based severity estimation models.
• To address the challenges of high annotation costs and ambiguous severity in medical images.

Main Methods:

• Proposed a deep Bayesian active learning-to-rank framework.
• Implemented automated selection of unlabeled image pairs based on model uncertainty and learning efficiency.
• Validated the theoretical adaptation of Bayesian neural networks for pairwise learning-to-rank.

Main Results:

• Demonstrated high performance in severity estimation using endoscopic images of ulcerative colitis.
• Showcased the method's effectiveness on both private and public datasets.
• Achieved robust performance even with significant class imbalance by prioritizing minority class samples.

Conclusions:

• The proposed deep Bayesian active learning-to-rank method efficiently selects informative image pairs for relative annotation.
• This approach significantly enhances severity estimation accuracy, particularly in scenarios with imbalanced datasets.
• The method offers a cost-effective and accurate solution for computer-aided diagnosis requiring severity assessment.