An accurate estimation of preterm infants’ limb pose from depth images using deep neural networks with densely connected atrous spatial convolutions

Continuing monitoring preterm infants' spontaneous motility during and after the hospitalization is of primary importance to early recognizing preterm-birth neurological complications. This work proposes a convolutional neural network (CNN)-based pipeline to estimate preterm infants' limb pose from depth images acquired in the neonatal intensive care unit (NICU) during the actual clinical practice. The limb-pose estimation relies on two consecutive CNNs with dense-structured pathways of atrous spatial convolutions. The first CNN (DeA detection) has the key role of roughly detecting joint and joint-connection position while the second (DeA regression) refines the previously found predictions to trace the limb-skeleton. The architectural design of the CNNs allows smooth information flow while gathering multi scale knowledge thus enriching CNNs' ability when predicting challenging frames (e.g., frames with self- or external- occlusions). The proposed pipeline was validated on a custom-dataset of 27000 depth frames acquired in the NICU. When applied to joint-detection, it achieved a median recall of 0.894 while the limb-pose estimation achieved a median root mean square distance error equal to 10.789 pixel. A study of efficiency on the DeA detection CNN was also conducted to derive an effective and efficient architecture for addressing the problem. We derived a CNN that achieved the same recall of the DeA detection with 2M fewer parameters. Continuing to improve the research in the field of preterm infants' contactless monitoring will pave the way for the development of increasingly cutting-edge decision support systems embedded with deep learning for the assessment of infants' spontaneous motility.