Research On The Method Of Pathology Images Nuclei Segmentation Based On Deep Learning

Medical image segmentation is a critical step of medical image processing and analysis,the purpose of medical image segmentation is to segment the special meaning part of medical images and then extract their relevant features information,which can provide reliable basis for clinical diagnosis and pathology research.Accurate nuclei segmentation is the basis for cell detection,cell classification and tumor classification,which has drawn significant attention from researchers in recent years.Due to the variation in cell morphology,staining processes and scanning equipment,each pathology image has distinct characteristics,so correctly carrying out the nuclei segmentation is quite a challenging task.This thesis mainly focuses research on the method of pathology images nuclei segmentation which based on the deep learning,the major work of this thesis are as follows:(1)A pathology images nuclei segmentation model based on the ResNet residual block is designed.In order to capture the deep feature information of pathology images,the model uses ResNet residual blocks to replace its basic convolutional blocks in the encoder and decoder stages of U-Net.Comparative analysis of experiments on the Cancer Genomic Atlas(TCGA)dataset are carried out,the average Dice coefficient of the segmentation model is 0.7916,which is better than 0.6928 of CNN2,0.7623 of CNN3 and 0.7808 of U-Net.(2)A pathology images nuclei segmentation model based on the nuclei and contour information aggregation module is designed.In order to improve the segmentation accuracy of pathology images with cell overlapping,boundary adhesion and chromatin loosening.The model designed in this thesis adopts the DenseNet to replace the encoder part of U-Net,which can not only reduce the parameters of the network model,but also obtain more detailed pathology images feature information during the encoder stage.Output branch of this model is composed by two decoders,between the two decoders,many multi-level information aggregation modules exist which can fuse the feature information of nuclei and contour.On the TCGA data,the experimental results show that the average Dice coefficient of the model is improved by 3.88% than the basic U-Net network,it shows that in the process of aggregating information,cell contour can assist the model to identify the complex cell boundaries,which plays an important role in the recovery of compressed image feature information in the encoder.