CovidMulti-Net

CovidMulti-Net classifies chest X-ray (CXR) images to identify COVID-19 and distinguish it from bacterial pneumonia, viral pneumonia, and normal cases.

Key Features:

• Parallel-Dilated Multi-Scale Feature Fusion Architecture: Employs parallel-dilated convolutions with multi-scale feature fusion to extract detailed features from CXR images.
• Transfer Learning: Uses pre-trained DenseNet-169, ResNet-50, and VGG-19 models for feature extraction from medical imaging data.
• Feature Concatenation: Concatenates features extracted from DenseNet-169, ResNet-50, and VGG-19 to form a comprehensive feature set for classification.
• Dataset Composition: Trained and validated on three publicly available datasets comprising 1341, 1341, and 446 CXR images of healthy individuals alongside 902, 1564, and 1193 images of patients with Viral Pneumonia, Bacterial Pneumonia, and COVID-19.
• Classification Performance: Achieves reported accuracies of 99.4% for 2-Class (COVID-19 vs normal), 95.2% for 3-Class (COVID-19, Viral Pneumonia, normal), and 94.8% for 4-Class (COVID-19, Bacterial Pneumonia, Viral Pneumonia, normal) datasets.

Scientific Applications:

• Clinical diagnostics: Rapid classification of COVID-19, bacterial pneumonia, viral pneumonia, and normal cases from CXR images to support clinical diagnosis and triage.
• Patient management and resource allocation: Supports decision-making for patient management and allocation of medical resources during respiratory disease outbreaks.

Methodology:

Parallel-dilated multi-scale feature fusion architecture; transfer learning using DenseNet-169, ResNet-50, and VGG-19; concatenation of extracted features; training and validation on three public CXR datasets for 2-, 3-, and 4-class classification.