Skin Image Segmentation Using a Self-Organizing Map
and Genetic Algorithms

Abstract

In order to distinguish malignant from benign skin lesions dermatologists use a microscope that shows the pigmented structure of the skin. However, it can be difficult to classify a skin lesion as benign or malignant using a dermoscopic image alone. This motivates computer analysis of dermoscopic images by digital image processing. The first step for a computer analysis is the segmentation of the image into regions of the same color, i.e. regions of the same color should be assigned the same gray level and regions of different colors should be assigned different gray levels. The number of colors is not known in advance. This paper presents a color clustering method that determines the number of colors automatically. First the RGB image is transformed into the L^∗u^∗v^∗ color space and segmented by a self-organizing map (SOM). After completion of the training a genetic algorithm groups the SOM neurons into clusters searching for a grouping that optimizes the Davies-Bouldin index. Various genetic algorithms are presented and evaluated for this purpose.