Purpose or Case Report: The purpose of this study is to develop a deep learning algorithm for detecting COVID-19 in chest x-rays of pediatric patients.
Methods & Materials: A total of 212 pediatric chest x-rays are collected from the Emergency Department of Rhode Island Hospital. Among this cohort, 136 pediatric chest x-rays (64.2%) are confirmed positive for COVID-19 via RT-PCR. 47 pediatric chest x-rays (22.2%), of which 36 (76.6%) are positive for COVID-19, are randomly selected to be held out as an independent test set. The remaining 165 pediatric chest x-rays are augmented by 535 adult chest x-rays from the COVID-19 Image Data Collection, of which 342 (63.9%) are COVID-19 positive, to yield a total of 700 chest x-rays for model cross-validation. Data augmentation techniques, including horizontal flipping and rotation, are applied to the cross-validation cohort to increase sample size to 2,100 and improve model invariance to noise. All chest x-rays are downloaded at their original resolution before resizing and padding to ensure uniform size. The radiograph images are converted into 3-channel data and normalized using channel-wise ImageNet means and standard deviations to ensure standardization across multiple image sources. Deep learning models are trained via 5-fold cross-validation without patient overlap using EfficientNet-B0 initialized with ImageNet pre-trained weights. Within each fold, the model with the lowest validation loss is selected as part of a 5-fold ensemble to prevent overfitting. This 5-fold ensemble will serve as the overall diagnostic model and will be used to compute accuracy and AUC on the cross-validation and test cohorts.
Results: The diagnostic model achieved a cross-validated accuracy of 92.3% with an AUC of 0.916. On the withheld test set, the diagnostic model achieved an accuracy of 78.7% with an AUC of 0.864.
Conclusions: Existing literature has detailed the development of deep learning algorithms for diagnosing COVID-19 using cohorts with little to no pediatric representation. Pediatric patients are susceptible to serious COVID-19 complications including respiratory failure and Multisystem Inflammation Syndrome in Children. AI-assisted COVID-19 diagnosis in pediatric patients using chest radiographs would allow physicians to intervene earlier to prevent viral spread and complications. As the pandemic continues to inundate hospital emergency departments with patients, this diagnostic model can serve as a clinical tool to help physicians triage pediatric patients in an efficient manner.