Deep convolutional neural networks for COVID-19 automatic diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Disease Diagnostics

Background:

• The global increase in COVID-19 cases necessitates efficient diagnostic tools.
• Limited availability of traditional COVID-19 test kits highlights the need for automated solutions.
• X-ray imaging offers a potential alternative for rapid COVID-19 detection.

Purpose of the Study:

• To evaluate the effectiveness of convolutional neural network (CNN) models for automated COVID-19 diagnosis using X-ray images.
• To compare the performance of different CNN architectures and transfer learning strategies.
• To assess the impact of varying training and testing ratios on diagnostic accuracy.

Main Methods:

• Utilized CNN-based transfer learning with pre-trained models (ResNet18, ResNet50, ResNet101) for COVID-19 detection from X-ray images.
• Compared model performance across different training/testing ratios (80/20, 70/30, 60/40, 50/50).
• Trained a proposed CNN model from scratch as a secondary approach.

Main Results:

• ResNet models (ResNet18, ResNet50, ResNet101) demonstrated superior classification accuracy on two datasets.
• With a 70/30 ratio, accuracies reached 97.67%, 98.81%, and 100% for ResNet18, ResNet50, and ResNet101 on the first dataset.
• The second dataset yielded accuracies of 99%, 99.12%, and 99.29% for ResNet18, ResNet50, and ResNet101, respectively.
• Training CNNs from scratch also proved effective for identifying COVID-19 signs.

Conclusions:

• Transfer learning using ResNet architectures is highly effective for automated COVID-19 diagnosis from X-ray images.
• The proposed CNN models show significant potential for clinical application in rapid disease screening.
• Automated AI-driven diagnostic systems can supplement traditional methods, aiding in managing infectious disease outbreaks.