A feasibility study of multispectral image analysis of skin tumors

To develop a noninvasive, early-detection method for skin cancers, a feasibility study of multispectral image analysis was investigated. The three most frequently occurring skin cancer types, ten basal-cell carcinomas (BCCs), ten squamous-cell carcinomas (SCCs) and five malignant melanomas (MMs) were studied, along with ten normal moles. Images were acquired by a charge-coupled device camera using eight narrow-band filters ranging from 450 nm to 800 nm, at 50-nm intervals. To extract main features of these tumors, principal components analysis (PCA) was performed, because it projects the multidimensional (here, eight-dimensional) data in the direction of maximum data variance. Then, the primary PCA components for red, green, and blue subset images were analyzed in terms of hue-saturation-intensity (HSI). By hue distributions, the BCCs and SCCs were differentiated from the MMs and normal moles. Texture information was used to further classify tumor types after the HSI analysis. The texture analysis, performed using a spatial gray-level co-occurrence matrix (SGCM), could separate MMs from normal moles. The BCCs and SCCs were further studied by Fisher's linear discriminant analysis. Distribution was described as a Gaussian mixture model. By this classification procedure, seven BCCs, eight SCCs, five MMs, and ten NMs were correctly classified. Three BCCs and two SCCs were unseparable. Thus, multispectral skin cancer image analysis has potential to diagnose skin cancers.