Cancerous Lesion Detection from Nevoscope Skin Surface Images via Parametric Color Clustering

This paper describes a new approach to analyze the spectral information of the samples of skin tissue that are localized in the spatial plane of microscopic image for discrimination of three different skin cancerous lesion prognoses. First, a cancerous lesion image is segmented from the skin surface based on Otsu’s optimal histogram thresholding technique. This allows us to localize the abnormal area in the skin tissue that is affected most as compared to the surrounding cells that appear brighter in color. Color clusters of the segmented darker lesions are used to obtain the three-dimensional (3D) spectral distribution function in the (R, G, B) color space. The Maximum Likelihood (ML) parameter estimation is utilized for calculation of the mean vector and co-variance matrix of the Gaussian (or normal) density approximation of skin samples and with the Mahalonobis distance as similarity measure in the learning and the testing phases of the pattern recognition system.