A new approach to classification applications with limited dataset

The first twenty-eight days after birth is called the neonatal period. Congenital anomalies or diseases that develop later cause the mortality rate to increase during this period. Rapid detection of suddenly changing body temperatures of babies who are cared for and treated in incubators in neonatal intensive care units is important in monitoring their health status. Evaluation of body temperature and thermal symmetry in the newborn is important in monitoring health conditions and predicting potential risks. If thermography, which is a harmless and non-contact method, and appropriate artificial intelligence techniques are used together, it may be possible to detect neonatal diseases at an early stage. Incubators are living spaces where babies are treated and cared for, and parameters such as temperature, humidity, and oxygen are kept under control. For newborns, there are factors such as the immaturity of their immune systems against diseases, their dependence on respiratory support and devices where their physiological parameters are monitored, and the need for thermal insulation. In addition, while some disease categories are common, some diseases are rarely seen. Due to the variety of cases encountered in neonatal intensive care units and the necessity of applying newborn imaging carefully, databases containing sufficient number of image populations cannot be created in all disease groups. In this thesis, using techniques aiming to obtain information about neonatal diseases with little data, pre-diagnosis systems have been developed with efficient classification methods for problems with limited data. Thus, the classification of diseases with insufficient data was provided. All thermograms of newborn babies were recorded in Selçuk University, Faculty of Medicine, Neonatal Intensive Care Unit, and a thermal image database was created. Classification algorithms based on traditional artificial intelligence methods require training on thousands of images and very large data sets. One of the common situations encountered in classification problems is the insufficient number of data. To overcome this problem, it is necessary to collect more data. In cases where it is not possible to collect more data, it is necessary to increase the small amount of data available in various ways and to search for ways to learn with fewer data. In the thesis, efficient classification approaches have been proposed in response to the problem of low and unbalanced data. In this context, cardiovascular diseases, pulmonary anomalies, necrotizing enterocolitis, intestinal atresia, infectious diseases, and esophageal atresia were classified with deep learning and data augmentation approaches. Neonatal diseases were considered in this study as multi-class for the first time and diseases with limited data samples were distinguished with high performance. Applications were carried out under four headings. Experiments using different convolutional neural network models and Siamese neural networks approach have been carried out. In the first of these experiments, 'abdominal & renal diseases', 'cardiovascular diseases' and 'pulmonary anomalies' were classified with 83% accuracy with the help of data augmentation techniques with convolutional neural networks using artificial neural networks and support vector machines in the fully connected network layer. In the second method proposed, six different approaches are proposed to classify four diseases. The highest accuracy of 94.55% was obtained for this experiment. In the study conducted to determine pulmonary anomalies, one of the most common conditions in newborns, the convolutional neural network model was developed and the effects of data augmentation were examined. In this application, it was determined that data increase on images taken from 34 newborns increased the classification accuracy from 84% to 91%. In another application, a multi-class classification study was carried out to provide pre-diagnosis to experts in the detection of diseases (necrotizing enterocolitis, esophageal atresia, and intestinal atresia, etc.) using Siamese neural networks and one-shot learning approach. By using two different optimization techniques and data augmentation, critical diseases with only a few sample data were classified using the method tested in 2-class and 3-class evaluation approaches. According to the results of the one-shot learning approach where classes that are not seen by the neural network model can be tested, the layer-based stochastic gradient descent algorithm used with data enhancement gave an average accuracy of 92.50%. Results based on disease class show that 100% accuracy is achieved in infectious diseases and esophageal atresia, 99.17% in intestinal atresia, and 94.17% in necrotizing enterocolitis.