Deep Learning of Child Abuse Imaging: Improving Outcomes of Children Evaluated for Physical Abuse

About This Grant

PROJECT SUMMARY Fractures are a common manifestation of physical abuse, with children <2 years at highest risk. The identification of healing fractures is crucial in the evaluation of physical abuse in a young child as these can suggest ongoing violence within the home and have serious implications for child protection. However, estimating time-since-injury of healing fractures based on imaging is often difficult and imprecise. Although deep learning (DL) models could vastly improve accurate dating of healing fractures in children presenting with suspicious injuries, a critical gap remains for accessible large digital pediatric imaging datasets and needed artificial intelligence (AI) infrastructure. Notably, this gap has recently been designated a critical pediatric health priority by the American College of Radiology. This project closes this gap by establishing the framework for deidentified image sharing and storage between three PEDSnet sites (Nationwide Children’s Hospital, Cincinnati Children’s Hospital Medical Center, Riley Hospital for Children) via a Secure File Transfer Protocol and providing the AI infrastructure needed for better image interpretation and diagnosis. We will train and validate DL models with state-of-the-art transformers such as DINOv3 and benchmark to the well-established convolutional neural network architecture ResNet-50 using skeletal imaging of accidental fractures of long bones in children <4 years to directly and accurately age healing fractures. In parallel, we will use meta- learning with a combination of labeled accidental fractures and unlabeled abuse fractures, followed by few-shot learning to regress the age of abuse fractures. Deliverables include establishing the framework for image sharing within pediatric health systems and the development of DL algorithms for aging of healing fractures that could be implemented widely as a virtual consultant for radiologists faced with the task of interpreting imaging completed in children presenting with high-risk injuries. This is the first study to propose the development of DL algorithms for aging healing fractures by 1) training on multicenter imaging data and 2) using real-world data of patients evaluated for abuse. This proposal is a key first step towards development of a national resource to stimulate and support high-quality, collaborative imaging research within pediatrics, dramatically improving patient outcomes within both pediatric and community settings. By providing a mechanism for cross-site image sharing, this project enables future scalable multi-institutional model development and validation for improved interpretation of imaging completed in child abuse evaluations.