COVID-Nets: deep CNN architectures for detecting COVID-19 using chest CT scans

Area of Science:

• Medical Imaging Analysis
• Deep Learning in Healthcare
• Computational Pathology

Background:

• Accurate and rapid diagnosis of COVID-19 is critical for patient management and public health.
• Computed Tomography (CT) imaging shows promise for COVID-19 detection, but automated analysis is needed.
• Existing deep learning models may require extensive training data and lack efficient transfer learning capabilities.

Purpose of the Study:

• To develop and evaluate novel deep convolutional neural network architectures, CovidResNet and CovidDenseNet, for COVID-19 diagnosis using CT images.
• To assess the effectiveness of transfer learning by initializing models with pre-trained weights from established architectures.
• To compare the performance of proposed models against standard approaches in classifying COVID-19, non-COVID-19 viral pneumonia, and healthy samples.

Main Methods:

• Proposed two novel deep convolutional network architectures: CovidResNet and CovidDenseNet.
• Implemented transfer learning by initializing models with weights from ResNet50 and DenseNet121.
• Trained and evaluated models on the SARS-CoV-2 CT-scan dataset (4173 images, 210 subjects) and the COVID19-CT dataset.

Main Results:

• Achieved up to 93.87% accuracy and 99.13% precision in binary classification tasks.
• Demonstrated strong performance in three-class classification (COVID-19, viral pneumonia, healthy) with up to 83.89% accuracy.
• CovidDenseNet achieved 81.77% accuracy on the COVID19-CT dataset for differentiating COVID-19 from other viral infections.

Conclusions:

• The proposed CovidResNet and CovidDenseNet models are effective for automated COVID-19 detection from CT images.
• Transfer learning significantly enhances diagnostic performance, overcoming limitations of training novel architectures from scratch.
• These models offer a promising, efficient, and accurate solution for COVID-19 diagnosis, outperforming standard models.