Medical Image Segmentation Using Spectral Clustering Based on Hypercomplex Analysis

Abstract

Images are destined to be segmented, either with our visual system (as we do it every day) or algorithmically using computing machines. However, as critical as medical images, image segmentation has to be consistent which we do not usually have. Therefore, to bring consistency, physicians need some kind of consistent clues about medical images they are analyzing. Particularly, when the images considered are colors with multiple channels/bands, developing such a consistent segmentation tool could take more effort and that often requires a rigorous mathematical computation and algorithm free from human cognition. To create such algorithm for color medical image segmentation, primarily the images have to be considered as collection of pixels that have to be represented properly and holistically with no separation of color components. Secondly, the relationships of pixels have to be defined holistically. Finally, based on these defined relationships, the pixels have to be grouped and segmented. Hence, in order to perform holistic segmentation, quaternion based representation of color pixels and quaternion based spectral clustering technique has been proposed in this thesis. Test results have shown that the proposed scheme can use and also can be used in machine learning applications of image segmentation and pattern recognition. Keywords: Quaternion Rotation, Spectral Clustering, Image Processing, Color Distance Measurement