ReFixMatch-LS: reusing pseudo-labels for semi-supervised skin lesion classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Image Analysis

Background:

• Semi-supervised learning (SSL) often relies on domain-specific data augmentation.
• Pseudo-labeling methods can be limited by noisy training data.
• Existing SSL techniques require domain-specific data augmentation or suffer from noisy labels.

Purpose of the Study:

• To propose ReFixMatch-LS, a novel semi-supervised learning approach for medical image classification.
• To combine consistency regularization and pseudo-labeling using domain-independent augmentation.
• To improve the performance of medical image classification models by addressing noisy labels and increasing pseudo-label utilization.

Main Methods:

• Developed ReFixMatch-LS, integrating label smoothing and consistency regularization to mitigate noisy pseudo-labels.
• Employed domain-independent weak augmentation for generating pseudo-labels.
• Implemented a strategy to record and reuse high-confidence pseudo-labels across training epochs.

Main Results:

• ReFixMatch-LS effectively increases the number of high-confidence pseudo-labels.
• The method significantly improves model performance in medical image classification tasks.
• Achieved high Area Under the Curve (AUC) scores (91.54% to 95.47%) on ISIC 2018 skin lesion datasets.

Conclusions:

• ReFixMatch-LS offers a robust approach to semi-supervised learning for medical image analysis.
• The proposed method enhances pseudo-labeling by reducing noise and increasing label utilization.
• This technique shows strong potential for improving diagnostic accuracy in clinical applications like skin lesion diagnosis.