ABSTRACT Automatic Rib Fracture Detection in Pediatric Radiography to Identify Non-Accidental Trauma PI: Adam M. Alessio Non-accidental trauma caused by physical abuse is a leading cause of death in children in the United States. Because rib fractures are highly predictive of child abuse and chest radiographs are commonly performed for multiple indications, pediatric chest radiographs can have a critical role in the identification of abuse. Detection of rib fractures on pediatric radiographs is challenging and a high percentage of fractures are missed, particularly in imaging centers with limited pediatric radiology experience. Currently, there are no viable computer assisted strategies for rib fracture detection on chest radiographs. The purpose of this proposal is to develop machine learning methodology to detect rib fractures on pediatric radiographs using images from a network of hospitals. These methods will rely on a two-stage approach including a thoracic cavity segmentation stage followed by a fracture detection stage. We will explore two fracture detection strategies using novel supervised learning approaches: a heterogeneous U-net and a multi-modal regional-convolutional neural network. These methods will be trained and tested with a large set of fracture-absent radiographs (N=1000) from Seattle Children's Hospital and a diverse set of labelled fracture-present radiographs (N=500) from collaborating sites. These methods will be developed with an intentionally diverse set of radiographs, representative of the variety of fracture presentations and image quality in clinical practice, in order to position this rib fracture detection method for rapid translation to clinical practice. The ultimate goal of this proposal is to provide a computer assisted rib fracture assessment tool that would be a rapid and widely-available add-on to all pediatric chest radiograph exams, improving detection of rib fractures and potentially leading to improved identification of child abuse.