Auxiliary Diagnosis for COVID-19 with Deep Transfer Learning

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Deep Learning for Medical Diagnosis

Background:

• Chest CT imaging is crucial for diagnosing COVID-19 and its complications.
• Accurate and rapid identification of COVID-19 on CT scans is essential for patient management.
• Deep learning offers potential for automating image analysis tasks in medical diagnostics.

Purpose of the Study:

• To develop and evaluate a deep transfer learning model for automatic COVID-19 recognition and classification in chest CT images.
• To assess the model's ability to differentiate COVID-19 from other pneumonias and normal cases.
• To classify the main manifestations of COVID-19 using CT imaging.

Main Methods:

• Utilized a dataset of 422 subjects including confirmed COVID-19, other pneumonia, and normal cases.
• Employed ResNet models pretrained on large image collections, fine-tuned for COVID-19 recognition.
• Applied a 70%-15%-15% data split for training, validation, and testing.
• Evaluated fine-grained classification of COVID-19 manifestations (ground-glass opacity, consolidation, fibrotic streaks).

Main Results:

• Achieved 94.87% sensitivity and 88.46% specificity for COVID-19 versus other groups using ResNet50.
• Obtained an overall accuracy of 89.01% for three-category classification (COVID-19, other pneumonia, normal).
• ResNet18 achieved 94.08% accuracy and 0.993 AUC for classifying COVID-19 manifestations.

Conclusions:

• Deep transfer learning models demonstrate high performance in recognizing and classifying COVID-19 on CT images.
• The models show promise for practical application, assisting physicians and reducing radiologist workload.
• Transfer learning is effective for COVID-19 CT image analysis, especially with limited training data.