Colour segmentation for lesion classification

Abstract

The incidence of skin cancer is increasing globally. Prognosis falls quickly as malignant skin cancers develop and so early detection is paramount in the fight against this disease. Colour is an important feature in the differential diagnosis of skin cancers in particular malignant melanoma. This paper describes a region based approach to segmentation of lesion images into areas of differing colour. The process initially divides the image into rectangular regions small enough to be considered as having only a single colour. This is then followed by conservative merging, where adjacent regions whose colours are very similar are coalesced. The segmentation is then completed by an iterative optimal merging process, the two most similar regions being merged at each step until the stopping condition is reached. Results show that the method produces a robust segmentation of regions of high colour contrast.