Deep Learning Based Nuclear Segmentation

The image of cell and histopathology is the basis of many biomedical research applications,so it is an important research field.Among them,nuclear division is the most basic and necessary step.Although many relevant methods have been proposed over the years,there are still some problems waiting for new methods to solve.For example,the segmentation effect of adjacent and overlapping nuclei is still unsatisfactory.Semantic segmentation is a key problem in the field of CV,and has important applications in many aspects.The medical image nucleus segmentation studied in this thesis is just such a problem.With the emergence of revolutionary deep learning,many problems in the field of CV have been far more effective than traditional technologies by using deep learning methods.The same is true of semantic segmentation.In this context,this thesis studies the algorithm of medical image nucleus segmentation based on depth neural network.Use image processing,deep learning,computer vision and other related knowledge to segment the nucleus of medical image,optimize the segmentation effect and improve the accuracy.The main contents of this study are summarized as follows:1.Research on the selection and enhancement of experimental datasets for preprocessing.This thesis uses two data sets.Data set A is derived from TNBC data set,which is composed of 55 hematoxylin and eosin(H&E)stained histopathology images provided by 11 patients,and was proposed by Peter Naylor and others.Dataset B is sourced from a U-Net++medical image nucleus segmentation practical project in the learning area of Station B.It is a publicly available dataset provided in the Kaggle 2018Data Science Bowl,with the address.Due to the average imaging quality of the data,this article enhances the relevant images.Due to the limited amount of medical image task data,in order to avoid overfitting,dataset A was randomly cropped to obtain 256*256 images to increase the data volume,and horizontal and vertical flipping were performed to enhance the robustness of the network.2.In this thesis,we first use FCN8s,DeepLabv3+,U-Net,U-Net++and Attention U-net and other classical semantic segmentation networks to perform experiments on the nuclear segmentation task of medical images.Observe the experimental results.Through thinking and learning about the experimental results that Attention U-net is not as good as U-net,this thesis uses U-Net as the basic network,and introduces the attention mechanism module combined with space and channel to improve the segmentation effect.However,the choice of attention module architecture is relatively difficult,which needs to be based on prior knowledge and some experiments,and different data sets are applicable to different modules and many other problems.This thesis introduces neural architecture search(NAS),which can automatically learn the optimal architecture of the network to solve these problems.3.Through observation and research on the results of classical semantic segmentation networks,it was found that the segmentation effect for adjacent and stacked nuclei is not good,and there is adhesion.In view of the strong learning ability of GAN neural network to image detail features,this thesis designs a U-net medical image nuclear segmentation network based on the confrontation training idea,and designs a reasonable confrontation training loss function to carry out confrontation training experiments,and obtains better experimental results than U-net.