Image processing and machine learning for telehealth craniosynostosis screening in newborns

OBJECTIVE The authors sought to evaluate the accuracy of a novel telehealth-compatible diagnostic software system for identifying craniosynostosis within a newborn (< 1 year old) population. Agreement with gold standard craniometric diagnostics was also assessed. METHODS Cranial shape classification software accuracy was compared to that of blinded craniofacial specialists us- ing a data set of open-source (n = 40) and retrospectively collected newborn orthogonal top-down cranial images, with or without additional facial views (n = 339), culled between April 1, 2008, and February 29, 2020. Based on image quality, midface visibility, and visibility of the cranial equator, 351 image sets were deemed acceptable. Accuracy, sensitivity, and specificity were calculated for the software versus specialist classification. Software agreement with optical cranio- metrics was assessed with intraclass correlation coefficients. RESULTS The cranial shape classification software had an accuracy of 93.3% (95% CI 86.8?98.8; p < 0.001), with a sensitivity of 92.0% and specificity of 94.3%. Intraclass correlation coefficients for measurements of the cephalic index and cranial vault asymmetry index compared to optical measurements were 0.95 (95% CI 0.84?0.98; p < 0.001) and 0.67 (95% CI 0.24?0.88; p = 0.003), respectively. CONCLUSIONS These results support the use of image processing?based neonatal cranial deformity classification software for remote screening of nonsyndromic craniosynostosis in a newborn population and as a substitute for optical scanner? or CT-based craniometrics. This work has implications that suggest the potential for the development of software for a mobile platform that would allow for screening by telemedicine or in a primary care setting.