Unified Benchmarking and Fusion of CNN Architectures for Multi-Class Lung Disease Classification in Chest X-Ray Imaging

Respiratory diseases such as pneumonia, tuberculosis, and COVID-19 impose a major burden on healthcare systems across the world. Unfortunately, the greatest burdens are in regions with limited radiology resources. Chest X-ray (CXR) remains among the fastest and most widely available tools for diagnosing respiratory diseases. However, image quality is often poor and there is a shortage of trained radiologists who can interpret these images. Deep learning approaches, specifically CNNs, have shown great potential to help practitioners and clinicians make correct decisions. This work presents an extensive evaluation of 15 CNN architectures trained on a large dataset containing 72,298 CXR images with four classes comprising COVID-19, pneumonia, tuberculosis, and normal. A uniform preprocessing pipeline is considered while training the models with identical configurations in order to present a fair comparison. Then Classification is performed using 15 CNN architectures to comprehensively evaluate multi-class respiratory disease detection. Among the individual architectures, DenseNet-201 achieved the highest validation accuracy of 96.79%, followed closely by RegNet 96.63% and MobileNet V3 96.54%. Building upon the complementary strengths of these topperforming models, a fusion framework is developed to enhance robustness and reduce prediction uncertainty in ambiguous cases. The proposed fusion model attained the highest validation accuracy of 97.12%, consistently outperforming all individual models and demonstrating improved reliability for multi-class chest X-ray classification. Finally, the proposed framework is integrated into a cloud-based diagnostic system helps in reduced queue time. The deployment demonstrates the framework for real-time, scalable clinical applications, supporting both advanced healthcare infrastructures and resource-constrained medical environments.