COVID-19 prognosis using limited chest X-ray images

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• The COVID-19 pandemic necessitates rapid and accurate diagnostic tools.
• Reverse Transcription Polymerase Chain Reaction (RT-PCR) has limitations including high turnaround time and complex kits.
• Automated diagnostic methods using Chest X-rays (CXRs) are being explored due to the lack of large-scale annotated COVID-19 CXR datasets.

Purpose of the Study:

• To develop an automated COVID-19 diagnostic methodology using CXRs.
• To overcome the challenge of limited annotated COVID-19 CXR data.
• To enhance the explainability of AI models in medical diagnostics.

Main Methods:

• Utilized a large-scale pre-COVID era CXR dataset for self-supervised feature extraction using deep neural networks.
• Employed attention maps between global and local features of a convolutional network during fine-tuning with limited COVID-19 CXR data.
• Conducted thorough ablation studies to validate the contribution of each component.

Main Results:

• Empirically demonstrated the effectiveness of the proposed self-supervised learning approach for COVID-19 detection from CXRs.
• Provided visualizations of saliency maps highlighting critical image regions for model predictions.
• Showcased the model's ability to aid radiologists in localizing infected areas.

Conclusions:

• The proposed method effectively detects COVID-19 from CXRs by leveraging pre-existing data and self-supervised learning.
• Attention mechanisms and saliency maps contribute to explainable AI and practical clinical application.
• This approach offers a viable alternative or supplement to traditional COVID-19 diagnostic methods.