COVID-19 CT image recognition algorithm based on transformer and CNN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• The COVID-19 pandemic significantly impacted global health and economies.
• Deep learning networks are increasingly used for medical image analysis, including chest CT scans for COVID-19 detection.
• Existing deep learning models struggle with the complex lesion distribution and local features in COVID-19 CT images.

Purpose of the Study:

• To develop an advanced deep learning model for improved COVID-19 diagnosis using chest CT images.
• To address the limitations of current models in capturing diverse lesion characteristics.
• To enhance the accuracy and efficiency of COVID-19 detection through medical imaging analysis.

Main Methods:

• Proposed a parallel bi-branch model, Trans-CNN Net, integrating Convolutional Neural Network (CNN) and Transformer modules.
• Utilized CNN for local feature extraction and Transformer for global feature extraction.
• Implemented a bi-directional feature fusion structure for cross-modal feature integration across different scales.

Main Results:

• The Trans-CNN Net achieved a classification accuracy of 96.7% on the COVIDx-CT dataset.
• This accuracy surpasses typical CNN (ResNet-152 at 95.2%) and Transformer (Deit-B at 75.8%) networks.
• Demonstrated significant improvement in diagnostic accuracy for COVID-19 detection.

Conclusions:

• The Trans-CNN Net offers a novel and effective deep learning approach for COVID-19 diagnosis from CT images.
• The model's ability to extract multi-scale features improves diagnostic performance.
• This fusion of deep learning and medical imaging advances real-time diagnosis of COVID-19 related lung diseases, potentially saving lives.