Differential Feature Space in Mean Shift Clustering for Automated Melanoma Assessment

Abstract

Malignant melanoma is one of the most dangerous forms of skin cancer. However, an important aspect of this type of cancer is that, if detected early, it can be successfully treated. This characteristic makes automated early melanoma detection systems clinically desirable. Therefore, a novel technique for melanoma segmentation based in Mean Shift clustering methods along with a complete melanoma detection system was implemented. Mean Shift main benefit over other clustering techniques is that it does not need a known number of clusters, instead, a distance function and bandwidth value are defined affecting the number of clusters that are calculated. Conducted studies which focused on diagnostic assessment consider two implementations of the Mean Shift algorithm; one that takes the color of the image as the feature space and another that uses the color and texture information to create a two-dimensional feature space. The segmentation algorithm was used with a database of 200 images including unaltered dermatological images and illumination corrected ones in order to study the effects of two widely used illumination correction methods in the segmentation stage. The first illumination correction technique was developed by making use of morphological operators, specifically the morphological closing. The second illumination correction technique was the algorithm known as Retinex. Performance measurements of the Mean Shift implementations were included with further comparisons of two existing segmentation techniques; image thresholding using Otsu’s method and Gradient Vector Flow (GVF) Snakes. The classification results of using two different classifiers are also demonstrated as part of the complete melanoma detection system. The first classifier was based on Support Vector Machines (SVMs). The second classifier used linear discriminant analysis with various discriminant functions: linear, quadratic and mahalanobis distances. This approach demonstrates enhanced classification capabilities of melanoma detection which can also be extended to other dermatologic applications.KeywordsMean Shift clusteringmelanoma detectionlesion segmentationimage processing