Development of a segmentation method for dermoscopic images based on color clustering

Kobe University (Japan)
Faculty of Engineering, Department of Computer & Systems Engineering, Kitamura Lab

Development of a Segmentation Method for Dermoscopic Images Based on Color Clustering

Dissertation

by

Harald Galda

01.09.2003

Contents

Abstract ... iii
Acknowledgements ... v

Chapter 1 Introduction ... 1
1.1 Research Goal ... 1
1.2 Basics of Medical Image Classification ... 1
1.3 The Dermoscope and Dermoscopic Image Characteristics ... 2
1.4 Computer Analysis of Dermoscopic Images ... 4
1.5 Directions and Contributions of this Research ... 7

Chapter 2 Color Clustering Methodology ... 10
2.1 Color Spaces ... 10
2.2 Cluster Validity ... 12
2.4 Self-Organizing Maps ... 15
2.5 Genetic Algorithms ... 16

Chapter 3 Color Clustering by Self-Organizing Maps ... 18
3.1 Introduction ... 18
3.2 Self-Organizing Maps Used in this Research ... 20
3.3 Computational Experiments ... 21
3.4 Results ... 22
3.5 Conclusions ... 24

Chapter 4 Genetic Clustering Algorithms ... 25
4.1 Introduction ... 25
4.2 Genetic Algorithms in this Research ... 26
4.3 Computational Experiment ... 29
4.4 Results ... 30
4.5 Conclusions ... 32

Chapter 5 Objective Functions for Color Clustering ... 33
5.1 Introduction ... 33
5.2 Fuzzy Cluster Validity Indices ... 33
5.3 Computational Experiment ... 36
5.4 Results ... 38
5.5 Conclusions ... 46

Chapter 6 Classification of Dermoscopic Images ... 47
6.1 Introduction ... 47
6.2 Image Features ... 51
6.3 Feature Selection and Classification ... 55
6.4 Computational Experiments ... 57
6.5 Results ... 58
6.6 Conclusions ... 62

Chapter 7 Conclusions ... 64
List of Publications by the Author of this Dissertation ... 67

References ... 68

Abstract
Malignant melanoma is a very dangerous kind of skin cancer. In order to treat malignant melanoma it must be detected as early as possible. However, when looking at a malignant melanoma by the naked eye, it can be mistaken as a nevus (benign skin lesion). Therefore, dermatologists use a microscope that shows the pigmented structure of the skin. This microscope is called “dermoscope”. An irregular overall structure, an irregular border and several colors indicate that a skin lesion is malignant. A homogeneous structure, a regular border and few colors indicate that a lesion is benign. However, even when using a dermoscope a melanoma can be mistaken as a nevus. Therefore it is desirable to analyze dermoscopic images by a computer in order to classify them as malignant or benign.

Before a dermoscopic image is classified, usually the skin lesion border is extracted. For this purpose, previously developed methods segment the image into regions of the same color (color segmentation) or into regions that fulfill a homogeneity criterion (region based segmentation). In this research we focus on color segmentation by clustering. Color segmentation methods applied to dermoscopic images either assume a pre-specified number of colors or are based on histograms in which maxima are selected as initial centers for fuzzy c-means clustering. However, the number of colors visible in dermoscopic images is not known in advance. Moreover, when applying fuzzy c-means clustering a proper number of clusters must be specified. It is possible to run fuzzy c-means for several numbers of clusters and to compare the results. However, the partition resulting from fuzzy c-means depends on the initial cluster centers, which are usually randomly selected. Therefore, fuzzy c-means should be run many times for the same number of clusters. However, this would require a very long computation time since an image contains a few ten thousands to a few millions of pixels.

The goal of this research is developing a method that automatically determines the number of clusters in color space. The clustering accuracy is evaluated by cluster validity index. Cluster validity indices describe how well a partition (cluster center set) represents the “natural” clusters of a data set. The method proposed in this research optimizes cluster validity indices in the following way: The image is transformed into a perceptually uniform color space. The pixels of the transformed image are used to train a self-organizing map (SOM). The SOM neurons are grouped in this research a GA is applied to group the neurons into clusters optimizing a cluster validity index.

In this research the following topics are addressed:

• SOM dimensionality and weight vector initialization. We find that the SOM should have two dimensions and the initial weights should be the nodes of a regular grid adjusted to the pixel density in color space.
• Choice of the GA type (steady state, simple), initialization of the GA population, selection strategies and fitness scaling. We figure out that a simple GA with stochastic random sampling without replacement and linear fitness scaling and an initial population that contains approximately the same number of individuals per possible number of clusters is suitable for finding the number of colors.
• Choice of the cluster validity index. We investigate several cluster validity indices for the purpose of evaluating clustering accuracy in color space. The GA performs best when applying the Xie-Beni index.

An image can be segmented into regions of the same color using the color clustering method proposed in this research. This makes it possible to distinguish the pigmented structures visible in a dermoscopic image. Moreover, the skin lesions visible in dermoscopic images segmented by the method proposed in this research can be classified as malignant or benign based on color statistics and texture. Dermoscopic images are segmented by the method proposed in this research as well as by a previously developed skin image segmentation method that assumes a fixed number of 4 colors. Various feature set yielding a high accuracy (ratio of number of correctly classified images to total number of images) are determined. The classification performance when applying the segmentation method proposed in this research is compared to the classification performance when applying the previously developed method. When using the segmentation method proposed in this research, the accuracy is as high or even higher than when applying the previously developed segmentation method. This indicates that the segmentation method proposed in this research keeps color and texture information relevant and discards information irrelevant for classification. Therefore, the segmentation method proposed in this research is effective for dermoscopic image analysis.

Chapter 1
Introduction
1.1 Research Goal
Malignant melanoma is one of the most dangerous kinds of skin cancer. In order to cure malignant melanoma successfully it is necessary to recognize it as early as possible. However, a melanoma can easily be mistaken as a nevus even by a dermatologist. The diagnosis of melanoma can be improved by dermoscopy (a dermoscope is a microscope that makes pigmented structures of the skin visible) [1, 2, 3]. Dermoscopy is also called “dermatoscopy”, “epiluminescence microscopy” or “surface microscopy”. A medical doctor can classify skin lesions visible in a dermoscopic image by checking the presence or absence of certain skin characteristics like irregular pigment network or de-pigmented areas. However, the performance of computer based image analysis of digital dermoscopic images is better than the performance of a dermatologist using dermoscopy alone [1, 2]. In order to analyze and classify a skin image by a computer, it must be segmented first. The goal of this research is to find a new color image segmentation method that can be applied to dermoscopic images.

1.2 Basics of Medical Image Classification
In order to classify an object visible in an image parameters that describe its texture or the shape of its boundary are calculated. These parameters are called image features. Based on a vector containing these image features as components, the object can be classified. Usually some kind of preprocessing (e.g. noise removal, object segmentation, edge extraction) is to be done before the image feature vector can be calculated.

[...]