Research On Nuclear Segmentation Based On Pathological Images

Cell nucleus segmentation for tissue subtype classification and abnormal identification is not only the key starting point for feature extraction and classification inside cells,but also the basis for building cell distribution models.Therefore,accurate positioning of cell nuclei in pathological images is an important part of further analysis.However,due to the data characteristics of overlapping nuclei and unclear boundaries,existing methods cannot solve these problems well.Therefore,this thesis takes the nuclear pathological image as the main research object,and uses the deep learning technology to carry out the research on the segmentation of the nucleus.The main research contents of this thesis are as follows:In the preprocessing stage of the nuclear pathological image,in view of the problem of inconsistency in the color of the image,the method of maintaining the color standardization of the structure is adopted,so that the color of the image remains consistent and the structure does not change.In the cell nucleus segmentation task,five models of FCN,U-Net,Seg Net,Micro-Net Mask R-CNN and CPP-Net are used for experiments,and the performance evaluation and analysis of each model are completed by combining the three evaluation indicators of Dice coefficient,AJI and PQ.The experimental results show that compared with the other five models,the Mask R-CNN model shows relatively better performance on the task of cell nucleus segmentation.In order to improve the segmentation accuracy of the Mask R-CNN model,an improved Mask R-CNN network model is proposed for the shortcomings of the network model.First,in view of the two problems of inaccurate cell nucleus segmentation and slow running speed caused by large datasets,the semantic segmentation part of Mask R-CNN is replaced with a dual-channel attention segmentation network,and the backbone network ResNet101+FPN is replaced with Efficient Net+ Bi FPN can improve the propagation ability between features,so as to extract more feature information to the nucleus,and improve the computational efficiency of the model to a certain extent.Secondly,in view of the difficulty of accurate segmentation of the nucleus boundary,the Point Rend module for enhancing feature extraction is added to the Bi FPN structure to enhance the feature extraction capability of the nucleus boundary.The experimental results show that the Dice coefficient,AJI,and PQ of the improved Mask R-CNN model are 0.77,0.71,and 0.67,respectively,which are 0.07,0.03,and 0.03 higher than those of the Mask R-CNN model.In addition,the improved Mask R-CNN model Testing an image is 0.0016 s faster than Mask R-CNN.Therefore,the method in this thesis is not only suitable for the segmentation of cell nuclei,but also can improve the computational efficiency of the network.