According to a new study, convolutional neural networks trained to identify abnormalities on upper extremity radiographs are susceptible to a ubiquitous confounding image feature that could limit their clinical utility: radiograph labels.