Nuclear Segmentation Method Of Tissue Slice Images

With the advent of the slide imaging scanner,histopathology slides are progressively digitized and can be stored in the form of digital images.At the same time,some of the emerging technologies that detect the pathological cells that can be executed by relying on computer technology and assist the medical staff in the diagnosis of the disease also attract a large number of researchers.In clinical medicine,the segmentation and recognition of pathological slice images is of great value in the diagnosis of cancer and disease.The traditional segmentation method is manually sketched by the pathologist,but due to the complexity of the data and the intensity of the number,there is no doubt that the challenge of doctors' work is increased,and the situation of misdiagnosis is more likely to occur.This thesis deals with the tissue of glioma cell nuclei,and processes the clinical medical data and nuclear segmentation challenge dataset.The main research and innovations of this thesis are as follows:Combining the particularity of the slice in the production process,this thesis compares several denoising methods and adopts an adaptive median filter to denoise these noises.For the preliminary segmentation of the nucleus,a clustering algorithm is improved in the idea of fuzzy clustering for the case of the complex foreground and background of experimental data and the histological features of boundary ambiguity.Firstly,this thesis constructs a multi-class to two-category transformation model,and based on the model,the algorithm is improved to solve the segmentation problem at the fuzzy small boundary of complex images.At the same time,it reduces the negative impact of the randomness of the initial cluster center selection on the segmentation result;Then,the image is morphologically repaired to make the division of the nucleus more complete;Finally,the Dice coefficient is used to compare the implementation of the improved clustering algorithm with the unimproved algorithm.The results of clustering show that the boundary of the nucleus is sharply aliased.According to the nature of contour fitting by the level set algorithm,this thesis further divides the nucleus accurately by the improved level set method.When curve evolution is performed on multiple targets with similar distances,there will be mutual attraction between the curves.And the curve fusion is avoided by increasing the position constraint.At the same time,the thesis improves a fast implementation scheme to accelerates the rate of level set evolution,and further improves the segmentation efficiency while ensuring segmentation accuracy.Finally,the effects of the position constraint level set model are compared and the efficiency of the fast level set scheme is verified.For the adhesion nuclei,the distance transformation maps with both internal and external markers were used in the paper and successfully segmented by the watershed algorithm.The functions in the above methods are implemented in engineering,which facilitates the analysis and judgment of the pathologist.At the same time,the final result of this thesis is compared with the doctor's manual ground truth,and a relatively high accuracy and overlap is obtained.