Optimization of pneumonia CT classification model using RepVGG and spatial attention features.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Pneumonia, including COVID-19, poses significant health risks, necessitating early detection.
• Image complexities in pneumonia CT scans hinder accurate classification by standard models.
• Distinguishing COVID-19 from other pneumonias is challenging due to shared characteristics.

Purpose of the Study:

• To develop an optimized CT classification model for COVID-19 detection.
• To enhance classification accuracy despite image intricacies and inter-pneumonia similarities.
• To improve upon existing deep learning models for COVID-19 screening.

Main Methods:

• A novel method based on RepVGG architecture was proposed.
• The model integrates a feature extraction backbone and a spatial attention block.
• This approach extracts spatial attention features while leveraging RepVGG's strengths.

Main Results:

• The optimized model demonstrated superior learning ability and reduced inference time compared to RepVGG.
• It outperformed advanced models like VGG-16, ResNet-50, and ViT.
• Achieved high performance metrics: 0.951 accuracy, 0.952 F1 score, and 0.902 Youden index.

Conclusions:

• The proposed method shows significant advantages in classifying and screening COVID-19 CT scans.
• It outperforms many basic models and networks with residual structures.
• This approach holds substantial reference value for clinical applications in COVID-19 detection.