Edge Detection And Segmentation Of Adherent Cell Nuclei Based On Caps-Unet

Accurate detection of nuclei edges,especially the adherent ones,on histopathological images is crucial for the accurate segmentation of nuclei regions,precise counting of diseased cells,and 3D reconstruction which has been a hot spot issue in medical image processing.It has great practical significance for the improvement of modern medical technology for computer medical image-assisted clinical diagnosis.The traditional medical image segmentation method is difficult to achieve precise segmentation due to its weak representation capability.The deep learning method could overcome many drawbacks of traditional image segmentation method.So we hire the deep learning as the main research method in this paper.Most deep learning algorithms as the training ground truth that has many difficulties in adherent nuclei segmentation.In the third chapter of this paper,we hire edges information of nuclei as deep learning model's training ground truth rather than region information.What's more,we proposed a deep learning model named Caps-Unet aiming at handle the adherent nuclei segmentation problem of.The proposed Caps-Unet model replaces the normal convolution layers in the U-Net model with Capsule layers(denoted as Caps-layers).Each Caps-layer consists of four different-scale convolution units and a concatenated unit incorporating the feature maps of the four units.Experiments on MICCAI2017 adherent nuclei data set shows that,the Caps-Unet model could detects the adherent nuclei edges that are hardly detected by the U-Net model.Compared with the U-Net model,the Caps-Unet model achieves a 5%lower loss value and a 2%higher dice coefficient performance.Since the beginning was only to verify the performance of the Caps-Unet model.The marked adherent nuclei edges label are not rigorous.Therefore,we remade the edge label for the data set with stricter standards(we mark the edges with two pixel widths on the outside of nuclei)and increase training iterations,expecting a better edge detection results.The experiment results shows that Caps-Unet model performance has been improved after the training labels are marked with a stricter standard.The loss function value of the Caps-Unet model is reduced by 1%one more time while the evaluation function value is increased by about 0.3%in the final stage.What's more and the nuclei edges detection results are more clear and accurate.Since the Caps-Unet model in Chapter 3 achieves the expected results in adherent nuclei edge detection,so we train the Caps-Unet model with nuclei region information label in chapter 4 of this paper,and organizes the contrast experiment with the U-Net model.Experiment results shows that,compared to U-Net model the loss function value of the Caps-Unet model is reduced by about 3%and the evaluation function value is increased by about 3%.However,accurate segmentation of the adherent nuclei region is still not achieved.It is validates the research status that it is so difficult for the method to make a accurate segmentation of adherent nuclei which employed the nuclei region information as the training ground truth.Aiming at making accurate adherent nuclei region segmentation and meet the requirements of detecting nuclei in general medical image processing.we proposed a two-stream Caps-Unet model based on Caps-Unet model in chapter 5 of this paper,which aims make a accurate region segmentation of the adherent nuclei.The idea is as follows:Firstly,we respectively train two sub-Caps-Unet models with the edge information label and region information label,and then we give a binarization and dot product processing for the obtained nuclei edges map and region map,fusion of the region information and edge information of the adherent nuclei,ultimately achieving the goal of adherent nuclei region segmentation.In addition,we employ evaluation index D1CLE2 to evaluate the model segmentation performance.The experiment results show that the double-stream Caps-Unet model has a better performance in term of evaluation index DICE2 on four test images than the TernausNet model and the U-Net model,achieved accurate segmentation for adherent nuclei on histopathological images.In the chapter 6 of this paper,we designed and implemented an image semantic segmentation system based on U-Net.In this system,users can use various kinds of data sets to train the underlying model,achieving the purpose of segmenting various scenes images efficiently.In addition,we designed and implemented an image format converter that allows users to convert their own data set image format to the image format specified by the system.