Automated Diagnosis of COVID-19 and Pneumonia Using Deep Learning Techniques on Radiological Images

Abstract

Respiratory disease, such as COVID-19 and pneumonia, are among the leading global causes of morbidity and mortality. Inexpensive yet universally applied chest X-ray (CXR) imaging is still difficult to interpret due to overlapping radiographic findings between diseases. In this paper, we propose an improved deep learning framework based on the EfficientNet-B3 architecture, aided by transfer learning, Grad-CAM visualizations, and data augmentation for autonomous diagnosis of respiratory diseases from CXR images. Two publicly available datasets were merged, cleaned, and balanced to create a heterogeneous training corpus of four classes of diagnostic conditions: COVID-19, bacterial pneumonia, viral pneumonia, and normal conditions. The proposed model was achieved accurate in test set 98.69% and high macro-averaged precision, recall, and F1-scores. The use of Grad-CAM visualizations enhanced the concentration of the model on clinically relevant lung regions, making it more explainable. These findings suggest the model's viability as a reliable clinical decision support system, especially in resource-limited settings, and are an advancement towards explainable AI in medical diagnosis.