Analysis of burn images by multivariate statistical methods and deep learning approaches

Especially in rural emergency departments where burn specialists are rarely present, in order to decide on the first intervention to the patient, photographs of the burn wound are taken and sent to a full-fledged health center where the images can be examined by burn specialists. The aim of this thesis is to determine the healthy and burned skin regions in 105 digital (2D) images of burn patients. Two types of approaches have been proposed for this purpose. In the first approach consisting of two stages, segmentation and classification, the performance of 4 segmentation and 10 classification methods commonly used in the literature was evaluated in terms of changing color spaces (CIE L*a*b*, HSV, YCbCr) and the feature matrices (gray level histogram statistics, Haralick attributes, average pixel intenstiy values) used in the training phase of the classifiers. The best classification performance was achieved in the case when a* and b* color channels of the CIE L*a*b* color space were used as the input data in the FCM, and each segmented region was classified by the ANN classifier trained with 9 Haralick features that were selected by forward selection. The average F-score calculated by 5-fold cross-validation was 74.28%. In the second approach used in the thesis, U-Net and SegNet architectures were trained using two kinds of training sets and the SegNet model, which was trained with 64x64 pixel image blocks, achieved an average F-score of 80.5% in a 5-fold cross-validation environment.