Assistant Professor Nationwide Children's Hospital 6741 Baronet Blvd Dublin OH 43017, Ohio, United States
Background: Physical abuse (PA) is a major source of morbidity for US children with >120,000 new victims yearly. Fractures are the most common manifestation of PA after soft tissue injuries. Estimating time-since-injury, vital for investigative agencies to confirm a clear timeline of traumatic events, is often difficult as current dogma for age estimates can be imprecise and vague. Additionally, timelines derived from healing of birth trauma are often extrapolated to fracture healing in older children although healing trajectories may differ. Deep Learning (DL) describes advanced statistical models that learn to recognize patterns in images and use them to draw high-level conclusions. Recently, DL models have been applied to a variety of radiologic tasks, though has yet to be widely investigated for the purpose of accurately estimating time-since-injury of fractures sustained in PA children. Objective: To assess the feasibility of developing a DL model that can distinguish between new and healing accidental fractures of the clavicle and diaphysis of long bones. Design/Methods: We used radiology images of accidental fractures in children 0-6 years identified between 2000-2016 at a large Midwestern tertiary children’s hospital. Chart review was conducted for each fracture to ensure accidental history was consistent with fracture and determine date of initial injury to establish the ground truth of the fracture age. Exclusion criteria included presence of any bone health comorbidity (e.g., metabolic bone disorder), internal fixation of fracture, and concern for PA. Additionally, all images were reviewed by 3 authors and any poor-quality images (e.g., obscured by casting material) were excluded. Fractures were manually identified, then tightly electronically cropped on all radiographs to focus on the fracture and immediate surrounding tissues. We fine-tuned a convolutional neural network (ResNet50) model to classify between new (0 days) and healing (10 days or older) fractures. We evaluated the model using a 5-fold cross-validation scheme at the patient-level. Results: We evaluated our model on 2059 images (444 individuals) including fractures of the clavicle, humerus, radius, ulna, femur, tibia, and fibula. 30% of images were new and 70% were healing. Our model was able to identify new fractures with a sensitivity of 83% and a specificity of 86%.
Conclusion(s): Our pilot study suggests that a DL model can distinguish between new and healing fractures across a variety of bones.
