Suppressing label noise in medical image classification using mixup attention and self-supervised learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) excel in medical image classification but require large, accurate datasets.
• Label noise is common in medical annotations due to annotator variability, degrading DNN performance.
• Overfitting to noisy labels is a significant challenge for DNNs in medical applications.

Purpose of the Study:

• To develop an innovative, noise-robust training approach for medical image classification.
• To mitigate the adverse effects of label noise on DNN performance.
• To enhance the reliability and accuracy of DNNs in clinical settings.

Main Methods:

• Incorporated contrastive learning to improve feature representation in DNNs.
• Developed an intra-group mixup attention module for noise suppression and sample interpolation.
• Integrated these strategies into a vanilla supervised learning framework.

Main Results:

• The proposed method effectively handles label noise in medical image classification.
• Comparative experiments demonstrated superiority over state-of-the-art methods on synthetic and real-world noisy datasets.
• An ablation study confirmed the significant contribution of both contrastive learning and mixup attention components.

Conclusions:

• The noise-robust training approach shows strong capability in curbing label noise in medical image classification.
• The method offers potential for real-world clinical applications by improving DNN reliability.
• This work advances robust deep learning techniques for noisy medical data.