COVID-19 recognition from chest X-ray images by combining deep learning with transfer learning

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Deep Learning

Background:

• Accurate COVID-19 detection from chest X-rays (CXRs) is crucial for timely diagnosis and treatment.
• Existing deep learning models face challenges in computational complexity and generalization across varied datasets.

Purpose of the Study:

• To propose Covid-DenseNet, a novel deep learning model for efficient and accurate COVID-19 detection from CXR images.
• To enhance model generalization, reduce computational load, and improve performance on diverse datasets.

Main Methods:

• Utilized transfer learning for multi-scale feature extraction.
• Incorporated an attention mechanism to prioritize salient image features.
• Employed a multi-scale fusion architecture for richer semantic information and improved efficiency.

Main Results:

• Achieved high recognition accuracy (96.89% - 98.02%) on benchmark and public datasets.
• Outperformed advanced models, including DenseNet121, on external test sets (80% and 77.5% accuracy).
• Demonstrated improved robustness and performance with increased training data and sample feature enrichment.

Conclusions:

• Covid-DenseNet exhibits superior performance and generalization capabilities compared to existing models.
• The model's robustness increases with data enrichment, indicating strong potential for clinical application.
• The proposed architecture offers a promising tool for reliable COVID-19 detection in medical imaging.