Digital analysis of pediatric chest radiographs for infiltrate detection using MobileNet V2

Abstract Background Detecting pulmonary infiltrates on pediatric chest radiographs remains challenging because of age-related anatomical variability and subtle radiographic findings. Artificial intelligence (AI) may assist radiologists in improving diagnostic efficiency and image interpretation. Objective To evaluate the performance of a MobileNet V2-based deep learning model in detecting pulmonary infiltrates on pediatric chest radiographs and assess its agreement with radiologist consensus. Material and methods A total of 600 pediatric chest radiographs (300 normal and 300 infiltrate images) were included. The dataset was divided into training (65%), validation (15%), and testing (20%) subsets. The MobileNet V2 model was trained using transfer learning for 50 epochs. Performance was evaluated using accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Agreement with radiologist consensus was assessed using Cohen’s kappa coefficient and the McNemar test. Results The model achieved 97% accuracy on the validation dataset and 90% on the testing dataset, with 95% sensitivity, 85% specificity, 86% PPV, and 94% NPV. Agreement with radiologist consensus was moderate (Cohen’s kappa = 0.408; p < 0.001), with an overall agreement of 71%. Conclusion The MobileNet V2 model demonstrated promising performance for detecting pulmonary infiltrates on pediatric chest radiographs and may serve as a decision-support or second-reader tool, particularly in resource-limited settings. Further validation with larger, more diverse datasets is required before broader clinical implementation.