QMix: Quality-Aware Learning With Mixed Noise for Robust Retinal Disease Diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical datasets often contain noise due to complex acquisition and annotation processes.
• Noise, including label noise (mislabeled images) and data noise (missing features), degrades deep neural network performance.
• Existing noise learning methods primarily focus on label noise, neglecting data quality issues.

Purpose of the Study:

• To propose QMix, a novel noise learning framework designed for robust disease diagnosis models under mixed noise conditions.
• To effectively handle both mislabeled images and images with missing discriminative features in medical datasets.
• To improve the generalization and robustness of deep learning models in medical image analysis.

Main Methods:

• QMix employs an alternating strategy of sample separation and quality-aware semi-supervised training.
• Sample separation utilizes a joint uncertainty-loss criterion to classify images into correctly labeled, mislabeled high-quality, and mislabeled low-quality categories.
• Semi-supervised training incorporates a sample-reweighing loss and a contrastive enhancement loss to mitigate the impact of low-quality images and improve feature distinction.

Main Results:

• QMix achieved state-of-the-art performance across six public retinal image datasets.
• The framework demonstrated significant improvements in model robustness against mixed noise scenarios.
• Effective separation of different noise types was achieved, leading to better model training.

Conclusions:

• QMix provides an effective solution for learning robust disease diagnosis models in the presence of mixed noise.
• The proposed method addresses limitations of previous noise learning techniques by considering both label and data quality.
• The framework shows strong potential for enhancing the reliability of deep learning in medical image analysis.