Machine learning to identify distal tibial classic metaphyseal lesions of infant abuse: a pilot study

• Published in: Pediatric radiology 2022-05, Vol.52 (6), p.1095-1103
• Main Authors: Tsai, Andy, Kleinman, Paul K.
• Format: Article
• Language: English
• Subjects: Abuse; Algorithms; Babies; Child; Child Abuse - diagnosis; Humans; Imaging; Infant; Infants; Learning algorithms; Machine Learning; Medical records; Medicine; Medicine & Public Health; Neuroradiology; Nuclear Medicine; Oncology; Original Article; Pediatrics; Pilot Projects; Radiographs; Radiography; Radiology; Retrospective Studies; Risk; Risk groups; Statistical analysis; Tibia; Tibia - diagnostic imaging; Tibia - injuries; Tibial Fractures - diagnostic imaging; Ultrasound
• ISSN: 0301-0449; 1432-1998
• DOI: 10.1007/s00247-022-05287-w

Background The classic metaphyseal lesion (CML) is an injury that is highly specific for infant abuse, and the distal tibia is one of the most common sites of occurrence. A machine learning tool that identifies distal tibial CMLs on infant skeletal surveys could assist radiologists in the diagnosis of infant abuse. Objective To develop and evaluate a machine learning-based classification algorithm to identify distal tibial CMLs on skeletal surveys performed for suspected infant abuse. Materials and methods We reviewed medical records of infants (≤1 year old) who had skeletal surveys for suspected abuse at a large tertiary children’s hospital over the past 13 years to identify those at low and high risk for abuse. Normal distal tibial radiographs from the low-risk group formed the normal study cohort; radiographs with distal tibial CMLs from the high-risk group formed the abnormal study cohort. We used these two cohorts to train a machine learning algorithm to classify distal tibial radiographs as normal or abnormal. We systematically evaluated this algorithm using a fivefold cross-validation procedure and statistically analyzed the results. Results The normal study cohort consisted of 177 radiographs from 89 infants, and the abnormal study cohort consisted of 73 radiographs from 35 infants. Our machine learning algorithm showed an overall performance accuracy of 93% and Kappa of 0.84. The overall sensitivity and specificity of the model were 88% and 96%, respectively. Conclusion Our developed machine learning model shows encouraging results as an automated tool to identify CMLs of the distal tibia on skeletal surveys performed for suspected infant abuse.