CovXNet

CovXNet classifies chest X-ray images to detect COVID-19 and differentiate viral, bacterial, and normal cases using a multi-dilation convolutional neural network.

Key Features:

• Multi-dilation depthwise convolution: Architecture built around depthwise convolution with varying dilation rates to extract multi-scale features from chest X-ray images.
• Transfer learning and fine-tuning: Initial training on a large dataset of normal and traditional (viral/bacterial) pneumonia chest X-rays followed by fine-tuning on a smaller COVID-19-specific chest X-ray set.
• Multi-resolution network variants: Multiple forms of the network are designed and optimized for different image resolutions.
• Stacking ensemble: A stacking algorithm combines outputs from diverse architectures to enhance prediction accuracy.
• Gradient-based discriminative localization: Integration of a gradient-based discriminative localization technique to identify and distinguish abnormal regions associated with different pneumonia types.
• Performance: Reported accuracies of 97.4% for COVID-19 vs normal, 96.9% for COVID-19 vs viral pneumonia, 94.7% for COVID-19 vs bacterial pneumonia, and 90.2% for multiclass differentiation among COVID-19, normal, viral, and bacterial pneumonias.

Scientific Applications:

• Automated classification: Detection and differentiation of COVID-19, viral pneumonia, bacterial pneumonia, and normal cases from chest X-ray images.
• Abnormality localization: Identification and localization of abnormal regions in chest X-rays corresponding to different types of pneumonia.
• Model evaluation on chest X-rays: Assessment of deep learning classification performance using two distinct chest X-ray datasets.

Methodology:

Depthwise convolution with varying dilation rates; initial training on a large normal and traditional pneumonia chest X-ray dataset followed by transfer learning via fine-tuning on a smaller COVID-19 chest X-ray set; design of multiple network variants for different resolutions; stacking of model outputs; integration of gradient-based discriminative localization; evaluation on two distinct datasets.