in this paper we propose a mathematical transform, based on several one-class support vector machines (SVM) models, to modify images at pixel level on the preprocessing stage in order to emphasize the difference of pixels between dissimilar regions. We show experimentally that the proposed transform does improve segmentation results of automatic thresholding algorithms such as Otsu, Mixture of Gaussians and k-means on biomedical images; specially in the presence of noise, clumped objects and difficult ROI identification.