Abstract:
Abnormal neonatal General Movements (GMs) during 6-20 weeks of age are strong predictors of whether an infant is at-risk of developing cerebral palsy (CP). Current protocols for manual GM scoring are time-consuming and human resource intensive, require specialist training, and do not scale to wider application. In this work, we developed a robust markerless pose-estimation scheme, based on advanced deep-learning technology, to automatically track neonatal GMs in standard iPad video recordings. Video recordings from 6 infants (2-5 months) were used to assess generalization of learning. Twelve anatomical locations (3 per limb) were manually labelled in 2000 frames from 5 infants to shape the training set (total of 24,000 points). A Resnet152 deep-neural-network was trained using the annotated data. The network’s performance was then tested on the entire video from the 6th infant (train:test ratio: 19:1 frames). Results demonstrated generalization feasibility with exceptional accuracy of 98.84% in tracking body-parts in the novel data, calculated from the sensitivity and selectivity measures of >99.86% and 97.93%, respectively, associated with <10 false-negatives and 153 false-positives. Our preliminary results indicate the possibility of establishing a fully automated platform for accurate analysis of neonatal GMs, for early diagnosis of neurodevelopmental disorders (including CP) in early infancy.