Reliable Label-Efficient Learning for Biomedical Image Recognition.

Area of Science:

• Biomedical Image Analysis
• Machine Learning
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) for biomedical image analysis demand extensive labeled datasets.
• Manual annotation and proofreading by multiple experts are labor-intensive, hindering the creation of large, high-quality datasets.

Purpose of the Study:

• To develop a deep active learning framework for efficient and reliable biomedical image dataset annotation.
• To enable the active selection of informative data queries and trustworthy human experts.

Main Methods:

• A dropout-based strategy and similarity criterion were used to measure data uncertainty for selection and error elimination.
• An expertise estimator was employed to assess labeler reliability through offline testing and online consistency evaluation.
• The framework was applied to classification tasks on confocal endomicroscopy and gastrointestinal endoscopic images.

Main Results:

• The proposed deep active learning framework demonstrated efficiency in selecting informative data and reliable experts.
• Experimental results showed promising performance compared to baseline methods in medical image classification.
• The method effectively managed annotations from labelers with varying expertise levels.

Conclusions:

• The developed deep active learning framework addresses the bottleneck of creating large-scale annotated biomedical datasets.
• This approach enhances the efficiency and reliability of expert selection and data annotation for DNNs.
• The method shows significant potential for advancing biomedical image analysis through improved dataset generation.