Classification of health status of neonates with deep learning methods using hyperspectral imaging

Rapid and harmless methods for early detection of the health status of premature babies can both ensure survival and improve these babies' quality of life. In this regard, the best method for health status detection of premature babies is the least invasive process (the principle of less touch/much more observation). In the neonatal intensive care unit (NICU), one of the important factors in keeping neonates alive and reducing their sequelae is the preliminary diagnosis and follow-up systems that will be created by using technologies that are still in the developmental stages. Hyperspectral imaging (HSI) is seen as a powerful tool for determination of neonatal health status because it provides diagnostic information about the disease. The hyperspectral images used in the thesis study were obtained from 19 different neonates in Selcuk University Medical Faculty Neonatal Intensive Care Unit. There are 32 hypercubes in total and 6528 hyperpectral images obtained from these hypercubes. 2 dimensional Convolutional Neural Networks (2D-CNN) and 3 dimensional Convolutional Neural Networks (3D-CNN) models were used to detect the health status of neonates using HSI. Mini cubes were created using the neighbourhood extraction method, and classification was done with 3D-CNN. In order to evaluate the classification performance, general accuracy, Cohen's kappa coefficient, sensitivity and specificity values were calculated. Using the neighbourhood extraction method, 100% overall accuracy, 100% Cohen's kappa coefficient, 100% sensitivity and 100% specificity were reached, and all data were classified correctly. In addition, high accuracy rates were obtained by using less training data with the neighboring method. These results show that deep learning methods are very successful in classifying hyperspectral images of neonates.