Identification of Cancerous Lesions in Unconstrained Images

The incidence of melanoma rises rapidly in Caucasians after the age of 20, and US statistics show about 1 million new cases every year. Specialists in the field are highly accurate in determining whether a skin lesion is cancerous or not based solely on a visual inspection. No systems exist for accurately classifying skin spots. The first stage in the development of such a system is to identify the region of interest. This paper reviews approaches to using three edge detection algorithms for edge detection - and therefore extraction of the lesion from the surrounding skin. The three edge detection algorithms used are Sobel, Marr-Hildreth and Canny. Their performance is compared for 136 images of both cancerous and non-cancerous lesions. Depending on the images, the best results are obtained either by Canny or by the Marr-Hildreth algorithm, however the edges produced by the latter are indistinct and the processing time is four times that of the other algorithms.