AI-Driven Pneumonia Detection: Evaluating CNN Models on Pediatric Chest X-Rays

Introduction: Pneumonia is a serious respiratory infection and an important contributor to global all-cause mortality in both children and adults. Existing methods to diagnose pneumonia using chest radiographs require expert radiology review, which poses a barrier to accessibility and speed of the clinical decision-making process. In recent years, improvements in artificial intelligence (AI) and machine learning - in particular Convolutional Neural Networks (CNNs) - have brought new opportunities for automated pneumonia diagnosis. Methods: In this study, we examined the ability of five different CNN architectures (MobileNet, DenseNet121, ResNet50, VGG16, and a custom iVGG13 model) to classify pediatric chest radiograph images as pneumonia or normal. The trained models were developed using a dataset of 5,856 labeled JPEG images. Results: MobileNet achieved the best predictive per- formance of all models tested, with an accuracy of 91%, a sensitivity of 96% and specificity of 84%. MobileNet outperformed the other models in each of these key metrics. Discussion: Our results demonstrate the potential of CNN models, especially MobileNet, to provide a rapid and accurate diagnosis of pneumonia. AI-based models represent a promising option to integrate into clinical practice, improving diagnostic efficiency, enabling this type of early intervention, and potentially addressing areas with poor access to expert radiologists.