Biological Cell Image Analysis Methods Based On Hierarchical Information

In this dissertation,we mainly study on the processing methods of phase contrast microscopy cell images and myocardial immunohistochemical images.Cancer is one of the leading causes of death.However,death is not only caused by cancer itself,but also the failure of body function caused by the spread of cancer.If its proliferation and spread can be controlled,lethal ratio from cancer can be reduced.Studying the transformation of cancer cells with different drugs can help to evaluate and develop cancer drugs and also can promote automatic pathological study.Heart is one of the most important organs,which is also the drive of body system.Studying myocardial immunohistochemical images can help to evaluate myocardial function.Early identification,timely treatment.Since artificial analysis is error-prone and time-consuming,it is necessary to bring in automatic or semi-automatic methods for processing medical images.Many methods have been proposed,but they still cannot meet the application requirements.We propose the target hierarchical information concept.It states that a target is composed by many information layers of different importance,for special targets,they can just be defined by the most important information layer.Based on this concept,many different algorithms and frameworks have been developed.The main work and achievements are as follows:1.After analyzing the classical segmentation and detection methods,we propose the hierarchical mergence approach.First,the maximum ROI is constructed by grayscale or gradient information.This ROI only contains cells.Then,local grayscale homogeneous regions are formed by Gaussian blurring operations and morphological operations.Thus,different information layers can be generated by applying OTSU algorithm to different local regions.Finally,the segmentation and detection results are produced by mergence strategies.Experiment shows that it can provide better results,when compared to the classical methods.2.We propose the target hierarchical information concept and realize the interframe correlation algorithm with the foremost information.First,the foremost information is defined.Second,interframe correlation is achieved by local overlapping detection and optical flow detection.During the procedure,foremost information adhesion problem between targets willcome up.A newly developed region partition algorithm,which is called PairSplit,is applied to deal with these situations.Several restoration strategies are also provided to prevent degradation of foremost information.Finally,the approximated rectangle and convex regions are counted as the detection and segmentation results.Experiment shows that it can effectively correlate interframe information and segment targets.3.Since the classical detection and identification methods cannot provide more information other than location and category of targets,we define the intermediate information and subordinate information with the target hierarchical information concept and also develop a series of detection methods.First,the method of adding intermediate information based on the detected foremost information.Second,the method of detecting individual and adhesion targets based on the subordinate information.Third,the method of determining the adhesion mode between targets.Furthermore,the method of Hough circle detection and screening,the method of detecting missed targets and new targets,the method of splitting and grouping subordinate information and so on.Experiment shows that they can provide more information than classical methods.4.For myocardial immunohistochemical images,we propose an interactive algorithm based on local prior structural information,which is inspired by graph theory based interactive algorithms.First,hierarchical information is produced for different dyeing information.Then,individual nucleus and adhesion nucleus are determined by the relationship among information layers.PairSplit algorithm is also applied to deal with adhesion problems.Finally,the category of nucleus is determined by overlapping detection of different dyeing information.Experiment shows that the semi-automatic algorithm can accurately count the number and identify the category of nucleus.