Improving Medical Image Classification in Noisy Labels Using only Self-supervised Pretraining.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Noisy labels degrade deep learning model performance by causing overfitting and corrupted feature extractors.
• Contrastive self-supervised pretraining improves natural image classification with noisy labels.
• The impact of self-supervised pretraining on medical image classification with noisy labels remains unexplored.

Purpose of the Study:

• To investigate the effectiveness of contrastive and pretext task-based self-supervised pretraining for medical image classification with noisy labels.
• To evaluate if self-supervised pretraining methods successful in natural image datasets are applicable to medical imaging.

Main Methods:

• Utilized contrastive and pretext task-based self-supervised learning to pretrain models.
• Applied pretraining to initialize deep learning classification models for medical datasets (NCT-CRC-HE-100K and COVID-QU-Ex) with induced noisy labels.

Main Results:

• Self-supervised pretraining effectively improved feature learning in medical image classification tasks.
• Initialized models demonstrated enhanced robustness against noisy labels compared to standard training.

Conclusions:

• Self-supervised pretraining is a viable strategy to mitigate the negative impact of noisy labels in medical image analysis.
• This approach holds promise for improving the reliability of deep learning models in clinical settings with imperfect data.