GroupLearning: Label Noise Mitigation Through Multi-Expert Collaboration for Medical Image Analysis

Deep learning in medical imaging requires accurately labeled data. However, incorrect labels are unpreventable in real word deep learning data collection procedure. To address this issue, we introduce the GroupLearning framework, a novel approach that utilizes the collective wisdom of multiple experts to identify and mitigate label noise. Extensive experiments across diverse datasets validate GroupLearning's efficacy, notably reducing label noise impact, especially when the distinction between images is based on inherent features. This positions GroupLearning as a valuable asset for medical image analysis, enhancing diagnostic accuracy in real-world applications.