Evaluation of the performances of fuzzy cluster approaches in clustering burn wound images

The World Health Organization reported that the annual number of deaths caused by burn wounds was 265,000. Hence, it is important to develop utilities to help diagnosis of burn wounds. Successful burn treatment requires accurate estimation of some vital parameters like percentage of burn wound. In this thesis, a method for identifying optimum fuzzy clustering approach is proposed to provide initial step for system which aims to calculate percentage of burn wounds by separating burn and normal skin regions from burn wound images. 120 digital images (2D) were collected from The Burn Unit of the Karadeniz Technical University Faculty of Medicine Farabi Hospital. The data set is divided into train and test sets having 100 and 20 images, respectively. The system proposed in the thesis aims to determine the most successful fuzzy clustering approach on burn images between seven different distance metrics, Euclidean, Mahalanobis, Manhattan, Minkowski, Chebishev, Jaccard and Cosine and three different color spaces as RGB, HSV and LAB. The system performs exhaustive series of experiments to find the optimal number of clusters for the images in the train set, using fuzzy cluster validity indices. In addition, the best smoothing filter tried to detect to eliminate negative effect of image noise. Experimental results show that proposed system could successfully clustered burn images.