Self-supervised learning with self-distillation on COVID-19 medical image classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• COVID-19 diagnosis relies on radiology, but deep learning models require extensive annotated data.
• Annotated COVID-19 imaging data is scarce due to privacy and labeling challenges.
• Developing effective deep learning models for COVID-19 diagnosis with limited data is crucial.

Purpose of the Study:

• To propose an effective deep learning model for COVID-19 diagnosis using limited annotated medical images.
• To assist specialist physicians in improving the efficiency and accuracy of COVID-19 detection.

Main Methods:

• Introduced Masked Autoencoder (MAE) for pre-training and fine-tuning on small datasets.
• Proposed Self-Supervised Learning with Self-Distillation on COVID-19 medical image classification (SSSD-COVID).
• Implemented reconstruction loss and self-distillation loss on latent representations to transfer knowledge from decoder to encoder.

Main Results:

• Achieved 97.78% accuracy on the SARS-COV-CT dataset (2481 images) and 81.76% on the COVID-CT dataset (746 images).
• Incorporating external knowledge improved accuracies to 99.6% and 95.27% on the respective datasets.
• The SSSD-COVID model demonstrated superior performance compared to other models.

Conclusions:

• SSSD-COVID effectively diagnoses COVID-19 from limited data, outperforming existing methods.
• The model can be further enhanced by introducing external information, significantly boosting performance.
• This approach aids physicians in decision-making, improving COVID-19 detection efficiency.