Research On Class Discovery Of Pathological Image And Parallelization

Generally,the doctor observes different pathological markers of tissue sections under the microscope and grades them in the diagnosis.However,time cost of manual analysis is expensive and the result contains the subjectivity of the doctors,which will lead to inappropriate treatment.Therefore,this paper has researched class discovery for pathological image and proposed a computer-aided approach.This paper proposed two segmentation algorithms for nucleus under different dye results: one based on region growing,the other based on watershed.In region growing algorithm,seeds grow in multiscale way to divide nuclei,which were obtained by SLIC and Otsu2 D for completeness.Multiscale growth is used to reduce repetition.In watershed algorithm,fast radial symmetry transform will be improved to get rid of noise、impurity and escape.Besides marks of improved FRST on H channel,gradient image is modified by multi-structure opening-closing filter and multi-scale opening-closing reconstruction.According to marks and gradient image,watershed can extract nuclei.Finishing segmentation,this paper extracts three features to construct random forest classifier,which is space structure,shape and texture.In the experiment,features will be filtered by importance of classification to obtain optimization subset,which depends on correlation of classification.The result showed new classifier is perfect and classification is successful.This paper presents a breadth first parallel method of RF to make training consumption substantial cut.The results of experiment showed it can take a cut and has improved efficiency of class discovery.