Purpose or Case Report: Machine learning that can identify and localize objects in an image using a labeled bounding box is called object detection. The purpose of this study is to demonstrate object detection in identifying rickets on pediatric wrist radiographs.
Methods & Materials: The institutional review board approved this retrospective study. The radiology information system was searched for radiographic examinations of the wrist performed for the evaluation of rickets from 2007-2018 in children less than 7 years old. Inclusion criteria were an exam type of “Rickets Survey” or “Joint Survey 1 View” with reports containing the words “rickets” or “rachitic.” Exclusion criteria were reports containing the words “renal,” “kidney,” or “transplant.” Two pediatric radiologists reviewed the images and classified them as either rickets or normal. Twenty-six images were excluded because of: healing rickets (10), excessive artifact (8), discrepant radiologist interpretation (6), abnormal positioning (1), and duplicate image (1). The remaining images were annotated according to their classification by drawing a labeled bounding box around the distal radial and ulnar metaphases. The training dataset was created from those images acquired between 2007-2017. This included 264 normal wrists on 142 images and 104 wrists with rickets on 61 images (most images had bilateral wrists). This training dataset was used to create the object detection model using the Turi Create framework and YOLOv2 model with a Darknet base network on an iMacPro. The testing dataset consisted of those images acquired between 2017-2018. This included 37 normal wrists on 20 images and 20 wrists with rickets on 10 images. The model was tested on the testing dataset and model sensitivity and specificity calculated.
Results: Of the 20 wrists with rickets in the testing set, 16 were correctly identified as rickets, 2 incorrectly identified as normal, and 2 were not labeled. Of the 37 normal wrists in the testing set, 33 were correctly identified as normal, 2 incorrectly identified as rickets, and 2 were not labeled. This yielded a sensitivity and specificity of 80% and 95% for wrists with rickets and 89% and 90% for normal wrists. Overall model accuracy was 86%.
Conclusions: Object detection can be used to identify rickets on wrist radiographs. Further development with a larger training dataset is needed to improve model sensitivity and specificity, and robust testing is needed for model validation.