Publications

Skin cancer is the most common type of cancers in humans [38]. There are two types of skin cancer; melanoma and non-melanoma. Melanoma is more dangerous and can be fatal if not treated. A non-melanoma tumor is a benign tumor and it is unable to spread. To rule out skin cancer, five main ABCDE properties (Asymmetry, Border, Color, Diameter, and Evolving) in the lesion, are considered [39]. Several segmentation techniques have been proposed to measure and analyze these properties but none of them is able to give the optimal result for every image. For that purpose, an efficient and robust system that detects skin cancer details is needed. Image segmentation is an important preparatory process, in image processing. It consists of partitioning an image into several segments to make it more significant and easier to analyze [16]. In each segment, pixels should be homogeneous in terms of level of intensities, to obtain an optimum segmentation [15]. One of the easiest image segmentation approaches is Thresholding [16]. In their simplest case, bi-level thresholding methods consist in finding one optimum threshold value T for the pixel's intensity. The pixels are divided into two classes based on their intensity values; bigger than T belongs to the first class, and the rest belongs to the second class. Usually, one class is for the background region, and another for the object region of the image [15]. However, sometimes, one threshold is not sufficient to extract the objects of interest from multi-classes images, and two thresholds and more would be required. Most of the skin cancer images are composed of three main classes: cancerous lesion, new growth or precancerous lesion and the normal skin. Then, applying bi-level segmentation on these images, will result in a loss of information. MCET approach is the most popular thresholding technique, because of its accuracy, efficiency and ease of implementation [17]. It consists of minimizing the distance between the original image and the thresholded in order to estimate the optimum threshold [15]. In addition, the accuracy of MCET results will be certainly affected by the number of image classes and the prediction of distribution type applied on each class. In other words, the histogram of an image is a combination of different statistical distributions, and the prediction of this combination plays an essential role in detecting the best thresholds. One drawback of the MCET technique is that its complexity growths when the number of threshold points increases, which makes this method a time-consuming one [41]. In this paper, a novel multilevel segmentation model based on heterogeneous MCET is designed. The designed model uses a combination of statistical distributions: Gamma, Gaussian and lognormal. Additionally, a parallel processing method is implemented to boost the performance of the proposed model and to minimize its computational cost in term of time while detecting the optimum thresholds of the image. This research study deals with: (1) Constructing a multilevel segmentation model using a heterogeneous MCET based on Gamma, Gaussian and lognormal distributions. (2) Optimizing the objective function of cross entropy to obtain the optimum threshold points. (3) Developing a parallel processing algorithm to enhance the performance of the suggested model in term of time consuming. (4) Significant simulation using two benchmark skin cancer datasets: ISIC and PH2, to test the efficiency of the suggested model.