Uk Kang, S-J. Bae, N. Oboukhova, G. Papayan

The automatic method of basic images characteristics estimation for TV system of melanoma diagnosis is offered. The method allows: To answer the following questions: has the image one color or multi color, and is white-blue veil present in the image. To estimate the shape and texture asymmetry degree of area with pathology. Before color image analysis the area with pathology is segmented. The segmentation based on the brightness feature with adaptive Otsu threshold. For color image analysis systems YCrCb and HSV (the coordinates of the H and Cr) were chosen. The main steps of method: The analyzing area is divided into blocks. For each block the average value of coordinate H and coordinate Cr is determined. Based on these estimations obtained for each block of analyzing area the H coordinate distribution histogram is got; the value of histogram density and Cr coordinate variance are calculated. Classification is realized according to Mahalonobis distance. Features array: H histogram density and Cr variance. Also H histogram allows to answer the question about the blue-white veil presence. In images with veil H histogram has significant estimates in the right side. The shape asymmetry estimation based on the area of non-overlapping regions respect to symmetry axes. Algorithm finds the position of the horizontal and vertical symmetry axis for the selected image area corresponding to the pathology. The angle between the symmetry axes and the coordinate axes is defined according to the assumption that all blocks of original image are random variables with normal distribution. In this case, the axis angle is determined by covariance matrix. Then the area of non-overlapping domains with respect to the major and minor axes are defined. Simultaneously with the shape asymmetry estimation the asymmetry of pathology area texture was got. At this stage of the analysis the most important information is not a specific type of texture, but the symmetry of the inhomogeneous areas. To characterize the inhomogeneous of the image block the Rosenfeld-Troy measure was used. Relative to the major and minor axes found previously the area of regions with different levels of inhomogeneous were defined. Accepted that the unit is a non-overlapping, if value of its inhomogeneous measure has a different form inhomogeneous measure of block in symmetry position. Based on the estimates of shape and texture asymmetry shape and texture common index of asymmetry is calculated. To test the algorithm has been used a database consisting of 118 samples. Conducted computer simulations showed the probability of correct classification: one color / multi-color images - 0.82; the image with blue-white veil / image without the blue-white veil 0.8. The task of determining the axes asymmetry number was correctly solved for 85% of the samples. According to the literature [1], the probability of correct classification melanoma from other new-formations via the visual assessment is 0.65-0.7. The results obtained for offered algorithms (0.82?0.85), so we can do the conclusion: the algorithms are effective and further research in this direction are actual and important.