Transfer Learning Approaches for Neonate Head Localization from Pressure Images

This paper explores the use of two types of transfer learning for the task of neonatal head localization from pressure images: 1) The pretrained CNN portion of the PressureNet model, a deep learning model that estimates adult pose given a pressure image, is used for transfer learning for a neonatal population. 2) Annotation of the training patient head locations was completed in the RGB image domain, then transferred to the pressure image domain of application. A multi-modal neonatal patient dataset suitable for this task was used. Data was simultaneously collected from a RGB-D video camera placed above the patient and a pressure sensitive mat (PSM) beneath the patient. Geometric transforms were used to achieve spatial registration between the video image plane and the PSM plane. Patient localization is important in the application of noncontact monitoring for vital sign estimation and movement detection. In testing on unseen patients, 54% of detections made by the object detection model achieved an IoU of 0.5 or greater. This is higher than the accuracy (33%) achieved using a pre-trained ResNet model trained with pressure images converted to RGB. This study demonstrates the potential for cross-domain transfer learning between RGB image and PSM domains.