A CNN-transformer fusion network for COVID-19 CXR image classification

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Science

Background:

• The COVID-19 pandemic poses significant global health and economic challenges.
• Rapid and accurate detection of COVID-19 is crucial for controlling its spread.
• Current diagnostic methods face limitations in speed and scalability for large populations.

Purpose of the Study:

• To develop and evaluate a novel CNN-transformer fusion framework for automated pneumonia classification on chest X-rays.
• To enhance the accuracy and efficiency of detecting COVID-19 and differentiating it from bacterial pneumonia.

Main Methods:

• A two-part framework involving data processing and image classification was designed.
• The image classification stage utilizes a multi-branch network with custom convolution and transformer modules for feature extraction, focus, and classification.
• The network processes chest X-ray images to identify pneumonia and specific types like COVID-19.

Main Results:

• The proposed framework achieved high performance metrics: 97.09% accuracy, 97.16% precision, 96.93% recall, and 97.04% F1 score on benchmark datasets.
• Comparative analysis demonstrated superior performance over existing methods in accuracy, precision, and F1 score.
• The network effectively differentiates between COVID-19 and bacterial pneumonia.

Conclusions:

• The CNN-transformer fusion framework shows significant promise for accurate and rapid COVID-19 detection using chest X-rays.
• The developed AI tool can assist clinicians in diagnosing pneumonia and specifically COVID-19.
• Further advancements could establish this network as a valuable tool in clinical settings for infectious disease diagnosis.