An Ensemble Deep Transfer Learning Approach for Classifying COVID-19, Pneumonia, and Healthy Lungs in Chest X-ray Images with GRAD-CAM Visualization

COVID-19 and pneumonia both affect the human respiratory system and can cause symptoms ranging from mild respiratory issues to severe conditions. Radiological imaging techniques such as X-ray is really effective to diagnose these diseases. Though some patterns can overlap on a chest X-ray (CXR), pneumonia usually appears as a dense area in one part of the lung, while COVID-19 often shows up as hazy, cloud- like areas and scattered spots in both lungs. In this paper, we worked with a customized dataset which contains 21,000 labeled CXR images and for classification part several pre-trained networks such as MobileNet, ResNet (50 and 152 layers) and ViT were implemented. We used Gradient-weighted Class Activation Mapping (Grad-CAM) before fully connected layer to generate heat maps that show the most focused region of the image the model considers for its prediction. The results of the ensemble framework of these transfer learning approaches outperformed most state-of-the-art models with an accuracy of 0.9956% which is reliable and robust for classifying thoracic diseases from CXR images.