PC2064: Intelligent Radiograph Analysis and Digital Solutions 4 Kids (iRADS 4 Kids): Development and Validation of a Convolutional Neural Network for Pediatric Pneumonia Classification

Background: Accurate chest x-ray (CXR) interpretation can be a diagnostic challenge for pediatric clinicians. Numerous algorithms have been developed and validated for the automated interpretation of CXRs in adult populations using convolutional neural networks (CNNs). However, no robust algorithms have been validated for clinical use in pediatrics. Predictive analytics in pediatric medicine is more difficult than in adult medicine. Image acquisition is more difficult in children leading to variable image quality. Models for children must be trained to recognize the evolution in the normal anatomy of children as they age.

Methods: Using an open-source database of 5,232 pediatric CXRs annotated by experts in pediatric radiology to identify the presence or absence of pneumonia, a series of CNNs were developed. Data was split prior to analysis into training (n=4192), test (n=624), and validation (n=1040) datasets. Using pre-specified hyperparameters from ResNetImage’s ResNet50, IRV2, DenseNet121, and InceptionV3 models transfer learning was performed to the classification of pediatric CXRs. Models were run using an unfreezetop10 structure with a sigmoid activation, Adam optimizer and binary cross entropy loss. We balanced class weights inversely based on frequencies. The accuracy, precision, recall, and F1 of each CNN model was assessed. A real-world classification success metric was generated by having two practicing pediatricians reannotate a subset of images. CNN performance was then compared to real-world classification success. All analysis was done using Python in Google Colab and all relevant code is freely available on GitHub.

Results: Based on expert annotations, pneumonia was present in 3,883 (74.2%) of the CXRs in the database (Figure 1). Expert re-annotation and performance of the IRV2, ResNet50, DenseNet121, and InceptionV3 CNN models are shown in Figure 2. The ResNet model had the highest accuracy in classifying pneumonia correctly, identifying 91% of cases. The four models all had a higher accuracy but lower recall and precision compared to the practicing pediatricians.

Conclusion: This CXR model can accurately identify the presence of pneumonia in children using only raw images. This is important clinical work to bridge the gap in machine learning models for pediatric and adult CXR interpretation. Additional opportunities exist to improve the model recall and precision.