Boosting automatic COVID-19 detection performance with self-supervised learning and batch knowledge ensembling.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Chest X-ray (CXR) analysis is a rapid method for COVID-19 detection.
• Existing methods often rely on supervised transfer learning, failing to address unique COVID-19 features and similarities with other pneumonias.

Purpose of the Study:

• To develop a novel, high-accuracy COVID-19 detection method using CXR images.
• To account for unique COVID-19 features and differentiate from other pneumonia patterns.

Main Methods:

• A two-phase approach: self-supervised learning-based pretraining and batch knowledge ensembling-based fine-tuning.
• Self-supervised pretraining learns representations from unlabeled CXR data.
• Batch knowledge ensembling refines detection by leveraging visual similarities within data batches during fine-tuning.

Main Results:

• The method demonstrated promising performance on public COVID-19 CXR datasets (large and unbalanced).
• High detection accuracy was maintained even with a significant reduction in annotated training images (down to 10%).
• The approach showed insensitivity to hyperparameter variations.

Conclusions:

• The proposed method surpasses current state-of-the-art COVID-19 detection techniques.
• This AI-driven approach can potentially alleviate the workload for healthcare professionals and radiologists.