A novel markerless multi-camera tool to early predict neurodevelopmental diseases like cerebral palsy in infants-at-risk

Abstract

General movement assessment (GMA) is the most predictive tool for physical impairments like Cerebral palsy (CP). However, GMA has its limitations which might be diminished with the utilisation of modern deep-learning tools. The software packages DeepLabCut (DLC) and Anipose offer a markerless multi-camera approach to build an extensive and dependable open-source database. Fifteen congenital heart patients and prematures, infants-at-risk for CP, were performing a markerless GMA surrounded by three cameras. With a MATLAB 2D analysis, we show the achievement of DLC on markerless labelling of limbs on par with human labelling. Retraining the neural networks offer refinement of more challenging markers like smaller joints. The 3D reconstruction, obtained with Anipose, showed good tracking as indicated by a constant length of rigid bodies. Furthermore, our preliminary MATLAB results show the possibility to analyse positional data in 3D and kinematics of limbs and joints. We compared those aspects of one infant, with the neurological and MOS-score outcomes. Those findings established that a 3D reconstruction can reveal different precise kinematic parameters, but the database and the parameters should both be expanded.