Body Part Detection from Neonatal Thermal Images Using Deep Learning

Abstract

Controlling thermal environment in incubators is essential for premature infants because of the immaturity of neonatal thermoregulation. Currently, medical staff manually adjust the temperature in the incubator based on the neonatal skin temperature measured by a probe. However, the measurement by the probe is unreliable because the probe easily peels off owing to immature skin of the premature infant. To solve this problem, recent advances in infrared sensing enables us to measure the skin temperature without discomfort or stress to the premature infant by using a thermal camera. The key challenge is how to extract skin temperatures of different body parts such as left/right arms, body, head, etc. from the thermal images. In this paper, we propose a method to detect the body parts from the neonatal thermal image by using deep learning. We train YOLOv5 to detect six body parts from thermal images. Since YOLOv5 does not consider relative positions of the body parts, we leverage the decision tree to check consistency among the detected body parts. For evaluation, we collected 4820 thermal images from 26 premature infants. The result shows that our method achieves precision and recall of 94.8% and 77.5%, respectively. Also, we found that the correlation coefficient between the extracted neck temperature and the esophagus temperature is 0.82, which is promising for non-invasive and reliable temperature monitoring for premature infants.