Thermal image analysis for neonatal intensive care units

Measuring temperature has crucial importance for neonatal care. Especially keeping the body temperature of neonates who are premature and have low birth-weight constant is vital. Therefore, measuring the temperature correctly and detecting temperature changes quickly are quite important. Fast developments in medical infrared thermal imaging provided us with measuring temperature contactless and correctly. Using the infrared thermal imaging which is contactless, non-ionized, non-invasive, and harmless in neonatal intensive care units has gained importance recently. The medical thermal imaging is an imaging technique which is used to measure temperature distribution emitted by organs and tissues by catching infrared energy of the body. Thermal imaging can detect temperature changes which occurs at a body region where there are physiological dysfunctions. The detection of anomalies by the aid of the asymmetries observed at body temperature distribution revealed the idea of pre-diagnosing system installation by using thermal images. This thesis covers studies conducted to classify thermal images of neonates as a healthy or unhealthy. The thermal images (body temperature map images) used in this thesis were captured from the neonates in the Neonatal Intensive Care Unit of Selcuk University Faculty of Medicine within the scope of the project numbered 215E019 supported by TUBITAK. Firstly, the images were segmented to remove the background from the neonates' body region to avoid the effect of background on classification result. Secondly, multiresolution analysis methods were applied to segmented images to obtain feature vectors. Multiresolution analysis methods used in the thesis are Discrete Wavelet Transform (DWT), Discrete Ridgelet Transform (DRiT), Curvelet Transform (CuT), and Contourlet Transform (CoT). Finally, the obtained feature vectors were classified as healthy or unhealthy by using ANN. To the best of our knowledge, performance evaluation of multiresolution analysis methods used to classify thermal images of neonates as a healthy or unhealthy was carried out in this thesis for the first time. The evaluation results were presented comparatively. In the first stage of the thesis, the temperature map of 190 thermal images belonging to 19 unhealthy and 19 healthy neonates was classified. In the first implementation, multiresolution analysis methods (DWT, DRiT, CuT and CoT) were applied to the segmented 190 thermal images to extract the feature vectors. Then, these feature vectors were classified as healthy or unhealthy by using ANN. In the considered this implementation, the following statistical results were observed: the highest classification success (100%) with DWT, the best sensitivity rate (100%) was with DWT, DRiT, and CuT and the best specificity rate (100%) with DWT. In the second implementation, some statistical features (mean, standard deviation, variance, skewness, kurtosis, and moment) from approximation coefficients obtained by applying multiresolution analysis methods were extracted in order to reduce the dimension of feature vectors. After that, these statistical features were classified by using ANN. In this implementation, classification success for DWT, DRiT, CuT, and CoT was 74.74%, 72.11%, 68.42%, and 71.58%, respectively. The highest sensitivity rate (81.05%) was achieved with DRiT, while the highest specificity rate (86.32%) was reached with DWT. In the second stage of the thesis study, RGB color space conversion of thermal images belonging to 29 healthy premature (PRM) infants and 14 neonates who were diagnosed as heart disease (Aortic Coarctation, pulmonary artery) in the neonatal care unit were used. The feature vectors were extracted by using DWT, DRiT, CoT, and CuT. Then, these feature vectors were classified as healthy or unhealthy by using ANN. In this implementation, accuracy of classification for DWT, DRiT, CuT, and CoT was 81.4%, 74.42%, 72.09%, and 74.42%, respectively. The highest sensitivity rate (57.14%) was achieved with CuT, while the highest specificity rate (100%) was reached with DWT. Based on these results, it was thought that the number of thermal images used in the study was insufficient. As a result, a second implementation was performed. In this implementation, firstly the data augmentation method was used to increase the number of thermal images artificially. The data augmentation method was performed on 43 images which were used in the previous implementation in order to create further 5 different images from each image. Thus, the number of thermal images was increased from 43 to 258. In the second implementation, accuracy of classification for DWT, DRiT, CuT, and CoT was 90.7%, 90.7%, 91.09%, and 91.09%, respectively. The highest sensitivity rate (86.91%) was achieved with DWT, whereas the highest specificity rate (94.25%) was reached with CuT. The results of the second implementation show how the higher accuracy, sensitivity and specificity rates were achieved by using the data augmentation method.