ENHANCING CLINICAL APPLICABILITY OF CNN MODELS FOR PNEUMONIA DETECTION IN CHEST X-RAYS

Abstract

In order to lower the mortality rates linked to pneumonia which continues to be a major danger to global health quick and precise diagnostic tools are required. Deep learning combined with chest X-ray imaging has become a revolutionary method for automatic pneumonia identification. By carefully assessing and refining four well-known architectures AlexNet, ResNet18, DenseNet201 and SqueezeNet this study aims to improve the clinical usability of convolutional neural network models. In order to replicate real-world clinical limitations, we used transfer learning and trained each model only on CPU-based hardware using a three-class chest X-ray dataset of 5,863 pictures classified into normal, bacterial pneumonia and viral pneumonia. Classification accuracy computing efficiency and deployment feasibility in resource-constrained environments were used to evaluate performance. Our results show that while all models achieve impressive accuracy ResNet18 and DenseNet201 are especially well-suited for real-world clinical implementation without requiring expensive GPU infrastructure because they strike the ideal balance between diagnostic precision and computational feasibility. In order to enable efficient and trustworthy pneumonia detection in a variety of healthcare settings this study offers a useful and comparative approach for CNN model selection and optimization.