Purpose or Case Report: Timely diagnosis of foreign body ingestion in children is important, particularly in the case of an ingested button battery. A button battery which lodges in the esophagus can quickly cause severe esophageal and mediastinal injury. It is also important to distinguish an ingested button battery from a coin; a button battery may be misdiagnosed a coin because of its similar radiographic appearance. The purpose of this study is to develop a machine learning model to identify button batteries and coins on pediatric foreign body series radiographs.
Methods & Materials: The institutional review board approved this retrospective study. The training dataset was created from 228 images selected from foreign body series radiographs acquired between 2007-2017. It included 114 images with ingested button batteries, 57 images with ingested coins, and 57 normal images. For simplicity, only frontal radiographs were used. The type of foreign body was either endoscopically proven or confirmed by the consensus of three pediatric radiologists. The button batteries and coins were labeled with bounding box annotations. This training dataset was used to create the model using the object detector toolkit of the Turi Create framework and YOLOv2 model with a Darknet base network on an iMacPro. The testing dataset consisted of all foreign body series radiographs acquired between 2017-2018, a total of 1678 images (37 button batteries, 347 coins, 211 other foreign bodies, 1083 normal), however one image contained a superimposed button battery and coin and this was excluded. Coins and button batteries in the testing dataset were either endoscopically proven or confirmed by the interpreting radiologist in the setting of an appropriate history. The model was tested on the testing dataset and model sensitivity and specificity calculated.
Results: The sensitivity and specificity of the machine learning model for button battery detection was 81% and 92%, and for coin detection was 83% and 96%. Only 1% of normal images were false positive for a coin or button battery, but 58% of images with other foreign bodies (not a coin or button battery) were false positive for a coin or button battery. The overall accuracy of the model was 88%.
Conclusions: Machine learning can be used to identify and differentiate button batteries and coins on pediatric foreign body series radiographs. Further development with a larger training dataset is needed to improve model accuracy.