Spectral-spatial classification for non-invasive health status detection of neonates using hyperspectral imaging and deep convolutional neural networks

Abstract

For premature babies, rapid and harmless early health status detection can both ensure survival and improve the quality of life. In this regard, the best method for detecting the health status is the least invasive process for the baby, using the principle of less contact, more observation. In the neonatal intensive care unit, the important factors in keeping neonates alive and reducing their sequelae are the preliminary diagnoses and follow-up systems that are created using development technology. Hyperspectral imaging is considered a powerful tool to determine the health status of neonates because it provides diagnostic information on the disease. In this study, deep convolutional neural network models using hyperspectral imaging in the visible and infrared regions (204 bands at 400-1000 nm) were used to detect the health status of neonates. Hyperspectral images were taken in a one-month period from different neonates at Selcuk University, Faculty of Medicine, neonatal intensive care unit (Konya, Turkey), and 6528 images (3264 unhealthy, 3264 healthy) in total. Data augmentation methods, such as rotation, flipping, shearing, zooming, shifting and brightness enhancement, were applied to the hyperspectral images for the training of the convolutional neural network model. By using all spectral bands, the number of 6528 images obtained from neonates in the neonatal intensive care unit was augmented to 65280 images. The final results were accuracy of 93.38%, kappa coefficient of 86.76%, specificity of 92.77% and sensitivity of 93.98% for the 65280 hyperspectral images employed. In this study, the objective was to determine the noninvasive health status of neonates with convolutional neural networks using visible and infrared spectral bands. The results show that hyperspectral imaging operating in the visible and infrared regions and the convolutional neural network are highly effective in detecting the health status of neonates.