Classification of thermal images of neonates with deep learning methods

Early detection of diseases and monitoring of treatment process of neonates are critically important in terms of maintaining their health status. Thermal imaging, which is a non – ionized, no harmful radiation and non – contact method, has been used in medical applications for decades. When conventional studies based on thermal symmetry balance of the body are examined, it is seen that classical image processing techniques are applied. These techniques include image preprocessing (resizing, gray level transform), image enhancement (noise reduction, histogram equalization), manual region of interest (ROI) selection and monitoring of changes in ROI. A problem-based algorithm can achieve successful results using these techniques, but any situation that may occur during imaging should be defined in the algorithm. For example, because the infants move inside the incubator, the ROI must change shape, re-match and continue to monitor. When all the possibilities are taken into account, it will be seen that the cost of process will increase and therefore the real time applications will be removed. With the development of deep learning methods such as multi-layer perceptions, convolutional neural nets and generative adversarial nets, processes such as image classification and image generation are carried out on the basis of philosophy of learning from image. Thus, instead of creating a code index corresponding to each situation, it is sufficient to have images representing each situation. Within the scope of this thesis, thermal images were obtained belonging to 40 neonates who were treated in Selcuk University, Faculty of Medicine, Neonatal Intensive Care Unit and thermal images both were classified by using machine learning and deep learning methods and the thermal images were regenerated. Thousands of images are needed to train deep learning methods, since such a set of images would take a very long time in the medical field, data augmentation methods have been used to augment images. The results were obtained according to the 10 – fold cross validation technique and evaluated with various evaluation criteria (confusion matrix, specificity, accuracy, sensitivity, receiver operating characteristic, area under curve, structural similarity index and peak signal noise ratio). The best classification results were obtained by using convolutional neural nets and data augmentation with 99.85 % sensitivity, 99.82 % specificity and 99.84 % accuracy. These results show that deep learning methods are quite successful in the classification of thermal images.