Health status detection of neonates using infrared thermography and deep convolutional neural networks

Abstract

Protection of body temperature is critically important for health. Diseases and infections cause local temperature imbalances in the body. Infrared Thermography (IRT), which is a non-invasive and non-contact method, has been used in medical applications for decades. Pre-diagnosis and follow-up treatment systems can be realized by monitoring the temperature distribution in the body. In this study, IRT and deep Convolutional Neural Networks (CNNs) models were used together for the first time to detect the health status of neonates. Neonatal thermal images have been taken in the Neonatal Intensive Care Unit (NICU) of Selcuk University, Faculty of Medicine (Konya, Turkey), over a one-year period. Neonatal thermal images were obtained from selected 19 healthy and 19 unhealthy neonates. Data augmentation methods, such as brightness enhancement, color transformation, resolution and contrast changes, and the addition of different noises, were applied to the thermal images for the training of a CNN model. A number of 3800 thermal images taken from neonates in NICU were augmented to 15,200 and 30,400 thermal images. Then, using CNNs, 380, 3800, 15,200, and 30,400 neonatal thermal images were classified as healthy and unhealthy. The optimal result obtained was with 99.58% accuracy, 99.73% specificity, 99.43% sensitivity, and 0.996 AUC for the 30,400 thermal images employed, Using the proposed system, 15,159 of 15,200 thermograms belonging to healthy premature babies were classified as healthy, whereas 15,114 of 15,200 thermograms of premature babies, diagnosed with at least one disease, were determined as unhealthy.