International Journal of Science, Engineering and TechnologyAuthors: Ayushi Rathour
Abstract: Pneumonia remains a leading cause of mortality in pediatric populations globally, with an estimated 740,000 deaths annually in children under 5 years. Early accurate diagnosis is critical for timely intervention, yet diagnosis remains challenging in re-source-limited settings where radiologist expertise is scarce. While chest radiography is the primary diagnostic tool, interpretive variability and limited radiologist availability constrain diagnostic accessibility in low- and middle-income countries. This study developed and validated a lightweight deep learning model for automated pediatric pneumonia detection from chest X-rays, incor-porating rigorous cross-operator validation to assess real-world generalizability. Using MobileNetV2 transfer learning, the model was trained on 1,750 balanced chest radiographs and evaluated on internal validation (n=259) and cross-operator validation (n=485) datasets from the Guangzhou Women and Children's Medical Center. The model achieved 94.8% accuracy with 89.6% sensitivity on internal validation. Critically, on cross-operator validation with different radiologists and imaging equipment, the model maintained 96.4% sensitivity (242/251 pneumonia cases detected correctly) with 86.0% overall accuracy, representing an acceptable 8.8% degradation and demonstrating robust real-world performance. The lightweight 14MB architecture enables sub-second inference on mobile devices, and the maintained high sensitivity demonstrates the model learned generalizable disease patterns rather than dataset artifacts. The combination of high sensitivity (96.4%), strong ROC-AUC (0.964), and deployment fea-sibility through a prototype clinical framework demonstrates this approach can augment pneumonia screening in resource-limited pediatric clinics. These results bridge academic validation with practical clinical deployment, suggesting that rigorously validated AI-assisted diagnosis can improve childhood pneumonia detection in global health contexts where radiologist availability remains constrained.
DOI: https://doi.org/10.5281/zenodo.17531598
