Analysis of regularity in skin pigmentation and vascularity by an optimized feature space for early cancer classification

Melanin pigmentation of the skin and the underlying blood vasculature is strongly correlated as biomarkers for the progressive stages in skin cancer. With the surge in keen interest among private and public organizations for research and development of optical scanners for imaging the skin and early cancer detection, it is also crucial from the patient usage point of view to have a reliable diagnostic algorithm that supports the underlined medical instrumentation. Here, we describe an optimally discriminative feature space to create a linearly separable cluster for each class of skin sample data points that are inherently seen to be highly random in nature. The dermoscopic skin lesion images utilized to train the system belong to three classes; namely, benign, dysplastic nevi, and malignant. Features are extracted from the spatial and spectral appearance of a skin lesion and its respective Fourier domain representation. Feature attributes are clustered in a three dimensional space with a classification accuracy of 93.3%. Thereby this method has high potential for device independent skin anomaly detection.