Lymphoma Tissue Pathological Images Classification Based On Manifold Learning

Accurate diagnosis on lymphoma tissue pathological has become one of the mostdifficult areas in clinical and pathological diagnosis for its complication on pathologicalimagination variability due to little difference between different types of lymphomapathological visualization. With digital image processing and pattern recognition appliedon analysis and processing lymphoma tissue pathological images can provide moreobjective and effective information for pathologists.The traditional systems based on neural network is not efficient due to thecomplication of network structure and long time to establish models, since featureextraction is too small, or dimensionality of the feature extracted is to high. In this paper,with the consideration of characteristics of lymphoma pathology images, firstly we extractthe overall pathological features, say, color features and texture features. Then manifoldlearning method for nonlinear dimensionality reduction is adopted to perform the datadimensionality reduction of high-dimensional features from pathological image. At last, BPneural network is used class various types of lymphoma pathology image. The mainresearch work and content is as follows:First, according to the pathological features of lymphoma, feature extraction methodsare respectively studied on color and texture of lymphoma pathology image.During the process of color feature extraction, we select the most suitable HSV color spacemodel for human to percept, and propose54handle quantitative methods aiming atLymphoma pathology images, and then extract the pathological image’s features, shown ascolorful histogram, by using quantized HSV based on uniform blocks; While during theprocess of texture feature extraction, we finally prefer uniformed LBP operator to extract59dimensional texture features from lymphoma tissue pathological images, as it not onlyintegrates the texture features of both structure and statistic, but also have rotationaldeformation and better robustness.Secondly, although the laplacian eigenmap (LE), one of the manifold learningmethods, is suitable for clustering and classification, LE does not take a good use of type information of samples and have troubles in generalizing new samples. Therefore, weproposed a Laplacian Eigenmap method with Supervised learning based on typeinformation and Kernel trick on the foundation of LE (SKLE) for feature reduction, whichhas been shown an effective performance by simulations.Finally, an effective performance of classification on various types of lymphomatissue pathology image of our proposed methodology called SKLE, which produces thefeature reduction on the fusion of color texture and feature, has shown by experimentscompared with use color features, texture features, integrated with color and texturefeatures only respectively as the input of BP neural network.