Unsupervised sub-segmentation for pigmented skin lesions


Zhao Liu
Machine Vision Lab (Dupont building)
University of the West of the England
BS16 1QY Bristol
Tel: +44 011 7328 3500
Fax: +44 011 7328 3636


Background: Early identification of malignant melanoma with the surgical removal of thin lesions is the most effective treatment for skin cancers. A computer-aided diagnostic system assists to improve the diagnostic accuracy, where segmenting lesion from normal skin is usually considered as the first step. One of the challenges in the automated segmentation of skin lesions arises from the fact that darker areas within the lesion should be considered separate from the more general suspicious lesion as a whole, because these pigmented areas can provide significant additional diagnostic information.

Methods: This paper presents, for the first time, an unsupervised segmentation scheme to allow the isolation of normal skin, pigmented skin lesions, and interesting darker areas inside the lesion simultaneously. An adaptive mean-shift is first applied with a 5D spatial colour–texture feature space to generate a group of homogenous regions. Then the sub-segmentation maps are calculated by integrating maximal similarity-based region merging and the kernel k-means algorithm, where the number of segments is defined by a cluster validity measurement.

Results: The proposed method has been validated extensively on both normal digital photographs and dermoscopy images, which demonstrates competitive performance in achieving automatic segmentation. The isolated dark areas have proved helpful in the discrimination of malignant melanomas from atypical benign nevi. Compared with the results obtained from the asymmetry measure of the entire lesion, the asymmetry distribution of the isolated dark areas helped increase the accuracy of the identification of malignant melanoma from 65.38% to 73.07%, and this classification accuracy reached 80.77% on integrating both asymmetry descriptors.

Conclusion: The proposed segmentation scheme gives the lesion boundary closed to the manual segmentation obtained by experienced dermatologists. The initial classification results indicate that the study of the distributions of darker areas inside the lesions is very promising in characterizing melanomas.