Self-Supervised Learning Application on COVID-19 Chest X-ray Image Classification Using Masked AutoEncoder.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computer-Aided Diagnosis

Background:

• The COVID-19 pandemic highlighted the need for rapid AI-driven diagnosis using medical imaging.
• Challenges include high accuracy demands and limited medical data for training AI models.

Purpose of the Study:

• To implement Masked AutoEncoder (MAE), a self-supervised learning method, for classifying 2D chest X-ray images.
• To evaluate MAE's performance against traditional training methods for AI-based medical image analysis.

Main Methods:

• Utilized a Vision Transformer (ViT) as a feature encoder for image reconstruction within the MAE framework.
• Fine-tuned the pretrained ViT encoder using labeled medical datasets.
• Compared MAE-based training with training from scratch and transfer learning using COVID-19 chest X-ray images.

Main Results:

• MAE-based training achieved superior performance with 0.985 accuracy and 0.9957 AUC.
• An optimal mask ratio of 0.4 was identified for MAE.
• MAE demonstrated efficiency, achieving comparable results with only 30% of the labeled training data.

Conclusions:

• MAE offers significant performance enhancements for AI-based disease diagnosis, particularly with limited medical imaging datasets.
• This approach has substantial implications for future diagnostic tools in data-scarce scenarios.