Purpose or Case Report: Foreign body ingestion is common in children. Timely diagnosis of the nature and location of the foreign body is important. A button battery which lodges in the esophagus can quickly cause severe esophageal and mediastinal injury. Machine learning that can detect anatomical regions of interest is an important step in computerized foreign body localization and may result in prioritization of radiographs with mediastinal foreign bodies. The purpose of this study is to develop a machine learning model to identify anatomical regions of interest on pediatric foreign body series radiographs.
Methods & Materials: The institutional review board approved this retrospective study. The training dataset was created from 357 images selected from foreign body series radiographs acquired between 2007-2017. The images consisted of a variety of frontal views of the neck, chest, abdomen, and pelvis. These images were labeled with bounding box annotations for the following anatomical regions: neck (262), mediastinum (317), abdomen (352), and pelvis (302). 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 frontal radiographs acquired between 2017-2018, a total of 1679 images. The model was tested on the testing dataset and model sensitivity and specificity for each anatomic region was calculated. Only completely imaged anatomic regions were considered in the analysis; partially imaged anatomic regions were not considered (e.g., the abdomen on a chest radiograph or the mediastinum on an abdominal radiograph).
Results: Sensitivities for identifying anatomical regions of the neck, mediastinum, abdomen, and pelvis were 93.9%, 97.8%, 88.1%, and 90.3%, respectively. Specificities for identifying anatomical regions of the neck, mediastinum, abdomen, and pelvis were 99.8%, 99.8%, 100.0%, and 99.7%, respectively. Overall model accuracy was 94.1%.
Conclusions: Machine learning can be used to identify anatomical regions of interest on pediatric foreign body series radiographs. This model could be combined with a foreign body detection model to prioritize those exams with mediastinal foreign bodies for radiologist interpretation.