Automatic Segmentation And Classification Of Microscopic Cell Images Based On Feature Analysis

Cancer has become the first killer that threatens human health,and the number of patients is increasing year by year.Early detection and early diagnosis of cancer is the key to improve treatment efficiency.Pathological examination of microscopic image of histological slide remains the main method for cancer diagnosis.With the rapid development of computer-aided diagnosis technology,quantitative analysis techniques of microscopic images were proposed to assist the pathological diagnosis of cancer.Due to the problems of poor contrast,uneven dyeing,and overlapped cells,reliable cell segmentation of the images is still a challenging task.In the stage of classification and identification,due to the diversity and complexity of cell microscopic images,the improvement of classification accuracy is limited by feature extraction and selection methods.To solve these problems,an automatic segmentation algorithm based on dual-criteria joint localization and improved mathematical morphology method,and a feature selection algorithm based on maximum correlation-minimum multicollinearity was proposed by this thesis.In order to study the feasibility of early cancer detection,Apc^MININ mouse model was used for quantitative analysis.The research work of this thesis mainly includes:（1）Aiming at the difficulty of segmentation in microscopic cell images,an automatic algorithm of cell nuclear segmentation based on dual criteria joint location and improved morphology was proposed.The top hat-bottom cap transform was used to enhance the contrast and improve the image quality.Then the wavelet transform and mean-shift clustering were collaborated to locate the nucleus.Then through the voting mechanism,the areas of interest were selected.By screening for circularity and area,the regions of interest were divided into individual nucleus and adhesion areas.Finally,the improved mathematical morphology method was applied to separate the adhesion areas with different degrees of adhesion.（2）Aiming at the problem of feature extraction of cells,various types of features were extracted in order to quantitatively describe different states of the cell nucleus.Not only the morphological features and color features which are often used in traditional application,but also the Gabor and GMRF features were extracted from the nuclear region.The Gabor feature analyzes the nuclear area from both time and frequency domains.The GMRF feature describes the nuclear area based on the structural relationship of the pixel points in the image.A comprehensive description of the structure and characteristics of the nucleus can be achieved by extracting various types of features.（3）Aiming at the problem of feature selection,the feature selection algorithm based on maximum correlation-minimum multicollinearity is used to filter the extracted initial feature parameters.By this method,features are selected,features with high class relevance are selected,and features with high redundancy are removed.The final best feature subset is input into the classifier SVM for classification.The results were compared with other feature selection algorithms.The effectiveness of the feature extraction algorithm was verified and the optimal feature subset was selected.By quantitatively analyzing the microscopic images of intestinal tissue sections of normal mice and Apc^MIN mice.We demonstrated the promise of the technique for improving the cancer diagnosis by characterizing the feature of cell nuclei.The experimental results show that the automatic segmentation algorithm based on the dual-criteria joint localization and improved morphology can segment the nuclei automatically and accurately.The four different types of features can describe the characteristics of the nuclear region from various aspects.The maximum relevance-minimum multicollinearity method is better than the other comparison methods in terms of classification accuracy and running time.The proposed automatic segmentation and recognition algorithms provide a new solution for the automatic segmentation and recognition of cell nuclei in microscopic cell images.Besides,the algorithms provide new theories and methods for the computer-aided diagnosis of early cancer,so it has certain theoretical significance and application value.