Automatic Segmentation And Cascade Classification Of Nucleus

In tumor diagnosis, pathological diagnosis of cells is one of the most commonly used methods, which can determine the benign and malignant tumor, histological type, degree of differentiation, proliferation and metastasis of malignant tumor. In the early stage of disease development, the changes of cell and tissue anatomy were relatively subtle. In order to identify these subtle changes, pathology doctors often need to carefully observe the pathological specimens through a microscope. This work is not only time-consuming but also very large. In addition, the discrimination of subtle changes is lack of objective criteria, and largely depends on the pathologist’s experience. So it puts forward a high demand for them. Computer aid analyze can improve the diagnosis efficiency, and reduce the pathology doctors’ workload and misdiagnosis.In this paper, in order to achieve the automatic segmentation and recognition classification of cancer cells, the segmentation of cell nuclei, feature calculation, and classification are deep researched, in particular:In the nucleus segmentation, local adaptive threshold segmentation is used for the segmentation. Based on local threshold segmentation, the segmentation algorithm of convex hull is used to obtain nucleus’s two dimensional convex hull.In characteristic parameter calculation and selection, a total of 107 characteristics are calculated, including morphological characteristics, optical density characteristics, texture features and other features about convex hull. Then the principal component analysis is used to extract new variables and the random forests sort the importance of characteristics. The classification results of the two methods are compared.In nucleus recognition and classification, a Adaboost-SVM cascade classification is constructed. At the same time, according to kinds of impurities in the pathological samples, the corresponding SVM is designed, and then they are cascaded. The experimental results show that the two methods can achieve better classification results.