Utilizing Deep Learning Models and Transfer Learning for COVID-19 Detection from X-Ray Images.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Infectious Disease Diagnostics

Background:

• The COVID-19 pandemic necessitated rapid and accurate diagnostic tools.
• Shortages in testing equipment and expert personnel highlighted the need for automated detection methods.
• Radiographic imaging, specifically X-rays, offers a potential avenue for COVID-19 diagnosis.

Purpose of the Study:

• To evaluate and compare various deep learning models for detecting COVID-19 from chest X-ray images.
• To develop and validate a modified ResNet50 model for both binary (COVID-19 vs. No COVID-19) and multi-class (COVID-19 vs. No COVID-19 vs. Pneumonia) classification.
• To interpret the model's decision-making process using explainability techniques like LIME and Grad-CAM.

Main Methods:

• Evaluation of classic convolutional neural network (CNN) architectures including VGG19, ResNet50, MobileNetV2, InceptionV3, Xception, and DenseNet121.
• Comparison with specialized COVID-19 detection models like DarkCOVIDNet and COVID-Net.
• Development of a modified ResNet50 model and assessment of its performance on binary and multi-class classification tasks.
• Application of LIME and Grad-CAM for model interpretability and visualization of classification features.

Main Results:

• ResNet50 emerged as the top-performing model among the evaluated classic architectures.
• The modified ResNet50 model achieved a binary classification accuracy of 99.20% for COVID-19 detection.
• The modified ResNet50 model achieved an 86.13% accuracy in multi-class classification, distinguishing COVID-19 from pneumonia.
• LIME provided better interpretability through contour maps compared to Grad-CAM's heatmaps, highlighting regions indicative of COVID-19.

Conclusions:

• The ResNet50 architecture, particularly the modified version, demonstrates significant potential for accurate COVID-19 detection from X-rays.
• The model's ability to differentiate between COVID-19 and pneumonia is crucial for clinical decision-making.
• Explainability methods confirm the model's focus on relevant lung regions, supporting its generalization capabilities.
• The proposed model is intended for free deployment to aid global diagnostic efforts.