COVID-19 radiograph prognosis using a deep CResNeXt network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• COVID-19, caused by SARS-CoV-2, has led to a global epidemic, often resulting in severe viral pneumonia.
• Accurate COVID-19 diagnosis is crucial for effective patient treatment and differentiating it from other respiratory infections.

Purpose of the Study:

• To propose a novel CResNeXt model for COVID-19 prediction using chest radiographs.
• To evaluate the model's efficiency in terms of hyper-parameters and training time compared to existing architectures.

Main Methods:

• Development of a CResNeXt model based on residual network architecture for chest radiograph analysis.
• Comparative analysis of the proposed model against VGG19, ResNet-50, and ResNeXt regarding hyper-parameters and training duration.

Main Results:

• The CResNeXt model demonstrates high accuracy in binary classification (COVID-19 vs. No-Finding) at 98.63% (original data) and 99.99% (augmented data).
• Multi-class classification (COVID-19, Pneumonia, No-Finding) achieved accuracies of 97.42% (original) and 99.27% (augmented).
• The model requires fewer hyper-parameters and exhibits significantly reduced training time per epoch.

Conclusions:

• The CResNeXt model presents a computationally efficient and accurate tool for diagnosing COVID-19 from chest radiographs.
• This AI-driven approach shows promise for rapid and reliable identification of viral pneumonia, aiding clinical decision-making.