Explainable Vision Transformers and Radiomics for COVID-19 Detection in Chest X-rays

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Disease Diagnostics

Background:

• The COVID-19 pandemic overwhelmed healthcare systems globally.
• Rapid and accurate identification of COVID-19 cases is crucial for containment.
• Current RT-PCR testing has limitations due to long turnaround times.
• Deep learning on chest X-rays (CXRs) shows promise but existing CNN models struggle with global context.

Purpose of the Study:

• To investigate the efficacy of Vision Transformers (ViT) for COVID-19 detection using CXR images.
• To compare ViT performance against Convolutional Neural Network (CNN) architectures.
• To develop a more efficient deep learning approach for COVID-19 screening.

Main Methods:

• Fine-tuning several ViT models for multiclass classification (COVID-19, Pneumonia, Normal).
• Utilizing a large dataset comprising 7598 COVID-19 CXR images, 8552 normal CXR images, and 5674 Pneumonia CXR images.
• Evaluating model performance using metrics such as Area Under Curve (AUC) and sensitivity.

Main Results:

• Achieved an Area Under Curve (AUC) of 0.99 for multiclass classification (COVID-19 vs. Pneumonia vs. Normal).
• Obtained a sensitivity of 0.99 for the COVID-19 class.
• Demonstrated superior performance compared to state-of-the-art CNN models for COVID-19 detection on CXRs.
• Attention maps confirmed the model's ability to identify key COVID-19 indicators in CXR images.

Conclusions:

• Vision Transformers (ViT) are highly effective for detecting COVID-19 in CXR images.
• ViT models offer improved performance over traditional CNNs by capturing global context.
• This approach presents a promising, rapid, and accurate method for COVID-19 screening.