Thermogram Classification Using Deep Siamese Network for Neonatal Disease Detection With Limited Data

Abstract

Monitoring the body temperatures and evaluating the thermal asymmetry of newborns give an idea about neonatal diseases. Infrared thermography is a non-invasive, non-harmful, and non-contact modality that allows the monitoring of the body temperature distribution. Early diagnosis using a limited data set is extremely vital due to the high mortality rate in newborns and some difficulties in neonatal imaging. Thermography stands out as a useful tool in detecting neonatal diseases compared to other techniques. However, creating a thermogram database consisting of thousands of images from each class required by traditional artificial intelligence methods, is impossible due to the sensitivity of newborns. One of the meta-learning models that has recently gained success in applying limited data learning, especially one-shot, in various fields is Siamese neural networks. In this work, we perform a multi-class classification to provide pre-diagnosis to experts in disease detection using Siamese neural networks. By using two different optimisation techniques and data augmentation, critical diseases with only a few sample data are classified using the method tested in two- and three-class evaluation approaches. The results based on the disease type achieve 99.4% accuracy in infection diseases and 96.4% oesophageal atresia, 97.4% in intestinal atresia, and 94.02% in necrotising enterocolitis.