Ensemble EfficientNet-B7 for High-Accuracy Multi-Class Chest X-Ray Classification with Grad-CAM and Uncertainty-Aware Interpretability

Abstract

Getting chest X-ray images classified quickly and accurately is vital for spotting COVID-19, viral pneumonia, and lung opacity, especially in places where medical resources are tight. In this study, we put together a solid system using the opensource EfficientNet-B7 model, tweaked on a public dataset with 21,165 X-ray images covering COVID-19, lung opacity, normal cases, and viral pneumonia. We used 5-fold cross-validation, data augmentation via Albumentations, and test-time augmentation with rotations, hitting an average test accuracy of 96.21% and an ensemble accuracy of 98.22%, plus a macro F1-score of 98.48% for all classes. GradCAM heatmaps help explain what the model focuses on, like opacities, making it more useful for doctors, while Monte Carlo Dropout gives uncertainty estimates (standard deviation between 0.0055 and 0.0076) for trustworthy results. We also pruned and quantized the model to make it work on edge devices. Built in PyTorch, this open-source solution offers a scalable, interpretable tool for multi-class chest radiography, with future validation planned on external datasets