Nucleus Image Segmentation Strategy For Automatic Screening System

In recent years,cervical cancer has become a social problem that threatens women's lives.In the diagnosis of cervical cancer,cervical smear pathology is an effective method.This method requires an experienced pathologist to diagnose diseased cells by microscope.It not only brings a heavy burden to doctors,but also causes degraded performance for the subjective factor and visual fatigue of doctors.Automatic screening developed in recent years which is aimed to solve this problem.By using image processing,this technique first recognizes cells and then measures the DNA contents accurately to provide auxiliary for diagnosis.Image segmentation is the key of automatic screening system,and it directly determines the performance of screening.However,the imaging environments under the microscope are complex.On the one hand,uneven illumination,background shading and uneven dyed nucleus exist.On the other hand,there are inevitably blood cells,lymphocytes,garbage,impurities and conglobation cells in cell images.These problems reduce the performance of image segmentation.In order to solve these problems,we put forward a method to segment cervical nuclei in complex background.It effectively relieves the influence of complex background conditions on cell segmentation.Then we propose a segmentation strategy of overlapping based on target recognition.It is used to segment overlapping cells in a cell image.After overlapping cell segmentation,we find that there is an abnormity of gray value and texture in the overlapping region.Therefore,we propose a method for pixel reconstruction in overlapping regions based on GMM-UBM(Gaussian Mixture Models-Universal Background Model).The work of this paper includes the following aspects:Firstly,a method to segment cervical nuclei in complex background is put forward.This method first employs the local threshold method to segment images.In this procedure we propose a parameter adapting method which adjusts its parameters automatically according to the function of local threshold window size and the binarized outline number.The local threshold method transforms an image into a binary image which is then passed to image corrosion operator to generate a marking image.With the binary image,the watershed algorithm was finally performed to segment the image.Secondly,a segmentation strategy based on object recognition is proposed,and this method consists of three stages: rough segmentation,recognition and fine segmentation.A cervical nuclei segmentation method in complex background is used to crude segment images.Then,the result of the rough segmentation are classified and processed according to their labels.Finally,the fine segmentation is further employed to process overlapping nuclei.A method for determining the number of concave points of overlapping nuclei based on prior knowledge of classification is proposed.This method is used to guide the segmentation of overlapping cells.Finally,a method to reconstruct the pixels in overlapping regions based on GMM-UBM is proposed.This method firstly uses a large number of data to train a GMM model(named UBM model),and then derives GMM of each cell by maximum a posteriori adaptation with the UBM and the normal grayscale value of this nuclei.The grey values are randomly generated by the GMM and filled into the overlapping area,with the constraint of cell grey values.Lastly,the image restoration algorithm is used to repair the connecting region.Experiments show that this method can overcome complex cell image conditions and accurately segment overlapping nuclei.This method also can effectively recover the cell features such as texture,gray,light density value,and improve the accuracy of cell measurement and classification.