Methodology Research Of Nucleus Image Segmentation Based On Deep Learning

Identifying the cells' nuclei is the starting point for various computational pathology application,including nuclei morphology analysis,cell type classification,and cancer grading.Due to cancer cells' invasive nature,it is very complex to cure when cancer cells spreading.Hence,analyzing the early mutation in cells of the patient biopsy is an effective way to survive this disease completely.However,it is time-consuming and laborious for the pathologist to observe the specimen by microscopes,which will occupy much medical resource.Moreover,the location of the nucleus in tissue cells is the basic work of most medical research.Moreover,the location of the nucleus in tissue cells is the basic work of most medical research.Recently,medical image segmentation methods based on deep learning have become a hot issue in industry and academia.In the segmentation task of nuclear images,improving the accuracy of a single image is the key to improve the overall performance which is closely related to the quality of simples and labels.Those technologies recently based on CNNs almost are difficult to distinguish overlapping and close nuclei.Otherwise,most existing segmentation methods are only applicable to specific data sets and cannot be well migratedIn this work,we focus on image segmentation for cell nuclei,the main contributions of this thesis are:1.a data enhancement method that can simulate the real distribution of nuclear nuclei is proposed.It is well known that tissue slice images are difficult to obtain,labeling images time-consuming and labor-intensive,so the trainable samples are really seldom.The common images augmentation methods such as rotation,scaling,etc.,which cannot simulate the complex distribution of nuclear distribution.So this work proposed a new method for generating new samples by the labeled samples.To be specific,we first extract all the regions labeled as nuclei in the label images,and then randomly select a number of individual nuclei to form a new sample,which can simulate the actual situation between the nucleus.2.An iterative training strategy for solving samples' incomplete labeling problems is proposed.Labeling medical images require a rich background of pathology and will take a lot of time.In the labeling process,some problems such as no labeling or partial labeling of the nucleus,which will have a terrible impact on the supervised learning process of neural networks.Therefore,this paper proposes a learning method to alleviate the problem.First,we use the convolutional neural network to convolve the training samples,and then divide the front and back scene according to the probability values of the pixel points in the feature map.Finally,we manually select the excellent prediction results and fuse with the corresponding original label maps.Repeat the above steps to iteratively train to achieve the goal of precise segmentation.3.This paper proposes an image segmentation method named DcRes-U-Net,which based on U-Net and combines ResNet,dilated convolution,etc.The advantages of this method are as follows: First,we replace the VGG-like structure of the encoder with a more advanced convolutional neural network in the industry;Second,we add dilated convolutions to increase the receptive field,so that more semantic information can be extracted from each layer;Third,we modified the loss function in U-Net to better express the difference between predicted and labeled values.Fourth,we use the watershed algorithm in the traditional image processing algorithm to further improve the nuclear edge and enhance the IOU index of the single cell nucleus.In this paper,we segment the nuclear image based on U-Net and DcRes-U-Net models respectively.The experimental results show that the improved model has better segmentation effect than the original U-Net for the edge of the cell nucleus,overlapping nuclei,adjacent nuclei,and small cell nuclei.With the iterative training strategy,random data enhancement,K-means based sample partitioning,our DcRes-U-Net have been achieved high precision in the competition's test dataset.