Deep Learning Approach for Pediatic Pneumonia Classification in Chest Radiographs Using Convolutional Neural Networks and Transfer Learning Models

Abstract

Pneumonia, a severe respiratory infection characterized by high morbidity and mortality rates, poses a significant global health challenge, particularly among pediatric populations in resource-constrained settings, claiming the life of one child every 43 seconds globally. The United Nations' Sustainable Development Goals No. 3 prioritizes reducing child mortality, emphasizing the need for effective medical interventions. The COVID-19 pandemic has further increased the risks, making timely and accurate diagnosis of pneumonia even more critical. Chest radiographs are frequently used for pediatric assessments, but their interpretation can be challenging due to variability in reader expertise and patient age, often requiring a specialized radiographer. These challenges highlight the need for innovative solutions to improve the accuracy and speed of pneumonia diagnosis. In this study, we leveraged transfer learning techniques using pre-trained Convolutional Neural Networks (CNNs) to automate pneumonia detection from pediatric chest radiographs. Transfer learning offers the advantage of utilizing knowledge learned from large-scale image datasets, enabling efficient training and fine-tuning of deep learning models for specific tasks like pneumonia classification even with small datasets. We employed three popular pre-trained CNN models; VGG16, ResNet50, and DenseNet201 to classify over 5,000 pediatric chest radiographs, divided into pneumonia-positive and pneumonia-negative cases, with an 80/20 training-to-testing split. Each model was fine-tuned using various data augmentation and hyperparameter optimization techniques to enhance performance and generalization. The performance of the individual models was evaluated using standard metrics, including accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC-ROC). The models achieved accuracies of 78% (VGG16), 79% (ResNet50), and 86% (DenseNet201). These results highlight the effectiveness of transfer learning in efficiently adapting pre-trained models to the task of pneumonia classification. By leveraging these models, we were able to significantly reduce the training time and computational resources required, while still achieving high classification performance. This study demonstrates that transfer learning provides a powerful approach for accurate, timely diagnosis of pediatric pneumonia, supporting the goals of SDG#3 by improving patient outcomes and alleviating the burden on healthcare systems. With further optimization, such models have the

potential to be integrated into clinical workflows to assist radiologists in resource-constrained environments.