COVID-Transformer: Interpretable COVID-19 Detection Using Vision Transformer for Healthcare.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• Accurate and rapid COVID-19 testing is crucial for disease control.
• Manual chest X-ray analysis is time-consuming and prone to errors.
• Limited availability of large datasets hinders AI model development for COVID-19 detection.

Purpose of the Study:

• To develop a Vision Transformer-based deep learning pipeline for efficient COVID-19 detection from chest X-rays.
• To create the largest publicly available dataset of chest X-ray images for COVID-19 research.
• To evaluate the performance of the proposed model against established deep learning architectures.

Main Methods:

• Aggregated 30,000 chest X-ray images from three open-source datasets.
• Developed and trained a Vision Transformer deep learning model for image classification.
• Compared the model's performance against baseline models like EfficientNetB0, InceptionV3, Resnet50, MobileNetV3, Xception, and DenseNet-121.
• Utilized Grad-CAM for model interpretability and visualization.

Main Results:

• The Vision Transformer model achieved 98% accuracy and 99% AUC for binary classification (COVID-19 vs. normal).
• The model demonstrated 92% accuracy and 98% AUC for multi-class classification (COVID-19, normal, pneumonia).
• The proposed transformer model outperformed all baseline models across all evaluated metrics.

Conclusions:

• The Vision Transformer-based pipeline offers a highly accurate and efficient method for COVID-19 detection using chest X-rays.
• The developed large-scale dataset and interpretable AI model can significantly assist radiologists in diagnosis and disease monitoring.
• This AI approach addresses the limitations of manual X-ray analysis and testing scarcity during pandemics.