Automated Quantification of Clinically Significant Colors in Dermoscopy Images and Its Application to Skin Lesion Classification

Dermoscopy is a noninvasive skin imaging technique, which permits visualization of features of pigmented melanocytic neoplasms that are not discernable by examination with the naked eye. Color information is indispensable for the clinical diagnosis malignant melanoma, the most deadly form of skin cancer. For this reason, most of the currently accepted dermoscopic scoring systems either directly or indirectly incorporate color as a diagnostic criterion. For example, both the asymmetry, border, colors, and dermoscopic (ABCD) rule of dermoscopy and the more recent color, architecture, symmetry, and homogeneity (CASH) algorithm include the number of clinically significant colors in their calculation of malignancy scores. In this paper, we present a machine learning approach to the automated quantification of clinically significant colors in dermoscopy images. Given a true-color dermoscopy image with N colors, we first reduce the number of colors in this image to a small numberK, i.e., K << N, using the K-means clustering algorithm incorporating a spatial term. The optimal K value for the image is estimated separately using five commonly used cluster validity criteria. We then train a symbolic regression algorithm using the estimates given by these criteria, which are calculated on a set of 617 images. Finally, the mathematical equation given by the regression algorithm is used for two-class (benign versus malignant) classification. The proposed approach yields a sensitivity of 62% and a specificity of 76% on an independent test set of 297 images.