Non-Contact Respiratory Rate Estimation in Newborns During Quiet Sleep Using Video Magnification Techniques and a 3D Convolutional Neural Network

In this paper, we present a new non-contact strategy to estimate the respiratory rate (RR) in a neonatal intensive care unit (NICU) based on the Eulerian motion video magnification technique and a 3D Convolutional Neural Network (3D CNN). The magnification procedure was carried out using the Hermite decomposition. The RR is estimated using a 3D CNN and a region of interest (ROI) detected manually. We have tested the method on 8 infants in NICU during quiet sleep. A contact respiratory signal is acquired synchronously to the videos to compute the RR as reference for training the CNN. To compare the performance of the method, we compute the Mean Absolute Error, the Root Mean Squared Error and metrics from the Bland and Altman analysis to investigate the agreement of the method with respect to the respiratory signal reference. The proposed solution shows an agreement with respect to the reference of 95% and root mean squared error of 2.88.