Classification of Neonatal Diseases With Limited Thermal Image Data

Abstract

Evaluation of body temperature and thermal symmetry in neonates is important in monitoring health conditions and predicting potential risks. With thermography, which is a harmless and noncontact method, diseases in neonates can be detected at an early stage using appropriate artificial intelligence techniques. Medical imaging is limited due to neonates' sensitivity to the thermal environment. This study proposes a classification model for classification problems with limited data (specifically, neonatal diseases) using data augmentation and artificial intelligence methodology. In the study, a multi-class classification was performed by combining images produced by data augmentation and employing the ability of convolutional neural networks to learn important features from the images, with 4 classes ranging from 8 to 16 newborns in each class. That is, there are four classes: 34 neonatal with abdominal, cardiovascular, and pulmonary abnormalities and 10 neonatal undiagnosed (premature). The dataset was created by taking 20 images from each of the 44 neonates. To test the performance of the proposed method, six different data separation experiments were conducted. Although the best classification accuracy is 94%, the 89% value obtained in the experiment when the model was tested with image samples of babies that had not been used in training the model is more significant for the model.