Automatic Segmentation of Skin Cancer Images using Adaptive Color Clustering

[attachment=17585]

Introduction
Skin cancer is one of the most common types of cancer and it can affect people at any age. It is a malignant tumor that develops changes in the skin texture and color, but it can be cured in more than 90% of cases, if the skin tumor is detected and treated in the early stages. There are two types of skin cancer, namely malignant melanoma and non-melanoma (basal cell and squamous cell carcinoma) [1]. Melanoma is more dangerous and can be fatal if untreated and a number of commercially available systems are designed for the analysis of pigmented skin lesions. Two representative systems are the SolarScan [2] and the microDERM dermoscopy unit [3]. It is useful to note that these systems are designed primarily to accurately capture skin images and not for automated detection of skin cancer images which is the aim of the image segmentation technique detailed in this paper.


Image Segmentation Algorithm
The main components of the developed image segmentation algorithm are illustrated in Fig. 1. The key component of the algorithm is the Adaptive Spatial K-Means clustering algorithm that is included in the development of a split and merge color-texture segmentation framework.


Experiments and Results
To evaluate the performance of the proposed algorithm, we use six representative skin lesion images depicted in Figs. 2 and 3. It can be noticed that the boundaries of some lesions are not well defined since parts of melanoma have characteristics of healthy skin tissue. To be able to determine the accuracy of the developed algorithm we constructed the ground truth by tracing manually the outline of the melanoma (see Fig. 4).


Conclusions
The aim of this paper is to present a novel algorithm for segmentation of skin cancer images by evaluating adaptively the color and texture information. The main novelty of this approach is the development of an adaptive spatially coherent color-clustering scheme (A-SKM) that is included in the implementation of a color texture segmentation algorithm. The resulting color-texture algorithm proved to produce accurate segmentation of low-resolution skin cancer images that are defined by large color and texture non-uniformities. This research is on-going and we plan to investigate ways of improving accuracy and to evaluate the performance of our algorithm when applied to large collections of skin cancer images.