Preterm-infants’ movement monitoring in neonatal intensive care units (NICUs) has a strong diagnostic and prognostic role. Despite its importance, movement monitoring still relies on clinicians’ visual observation at the crib side. The goal of this work is to present an automatic, deep-learning based approach to movement detection from depth-image streams. The proposed approach relies on a pre-trained 3D convolutional neural network (CNN), which we proposed in a previous work for joint-pose estimation (joint-CNN). The joint-CNN is here modified (mov-CNN) to fulfill the movement detection task by combining the encoder path of the joint-CNN with 3 fully-connected layers. When tested on a dataset acquired in the actual clinical practice, the proposed mov-CNN achieved a recall of 0.84. The achieved results prompt the possibility of translating the developed methodology into the actual clinical practice, as a valuable tool to support clinicians in NICUs.
A 3D CNN for preterm-infants’ movement detection in NICUs from depth streams
Moccia S.;
2020-01-01
Abstract
Preterm-infants’ movement monitoring in neonatal intensive care units (NICUs) has a strong diagnostic and prognostic role. Despite its importance, movement monitoring still relies on clinicians’ visual observation at the crib side. The goal of this work is to present an automatic, deep-learning based approach to movement detection from depth-image streams. The proposed approach relies on a pre-trained 3D convolutional neural network (CNN), which we proposed in a previous work for joint-pose estimation (joint-CNN). The joint-CNN is here modified (mov-CNN) to fulfill the movement detection task by combining the encoder path of the joint-CNN with 3 fully-connected layers. When tested on a dataset acquired in the actual clinical practice, the proposed mov-CNN achieved a recall of 0.84. The achieved results prompt the possibility of translating the developed methodology into the actual clinical practice, as a valuable tool to support clinicians in NICUs.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
