| 引用本文： |
-
许广,管军霖,甘才军,汪华登.一种新的基于U-Net和ResNet的病理图像细胞核分割方法[J].广西科学院学报,2021,37(4):372-379,400. [点击复制]
- XU Guang,GUAN Junlin,GAN Caijun,WANG Huadeng.A New Method of Pathological Image Nuclei Segmentation Based on U-Net and ResNet[J].Journal of Guangxi Academy of Sciences,2021,37(4):372-379,400. [点击复制]
|
|
| 摘要: |
| 医学图像分割是图像处理的重要环节，而细胞核分割结果是病理学家进行癌症分类和评级的重要依据，提高其分割的准确率一直是研究的热点。但由于同器官的不同细胞核存在形态可能不一样、细胞之间相互重叠、细胞边界不清楚等现象，导致细胞核图像难以准确分割。为提高相互接触和重叠细胞核分割的准确性和精确率，本研究提出一种新型的细胞核分割网络模型。该模型首先是对原始细胞图进行ZCA白化预处理，并基于经典的U-Net网络结构，通过U-Net和ResNet残差模块进行训练，使用Batch Normalization方法实现数据归一化处理，解决训练过程中梯度震荡问题。在MoNuSeg和ISBI2018Cell两个数据集上的实验结果表明，本研究所提出的模型的分割准确率较高，分割出的细胞没有出现细胞核大面积粘连的现象，细胞核轮廓更加清晰。本研究所提的分割网络基于经典的U-Net网络结构，通过构造ResNet残差模块实现对细胞核上下文特征的提取，同时在残差模块使用Batch Normalization使得梯度的传输更加便捷，减少了训练时间，而且在分割相互接触的细胞核时，具有精确定位和准确分割的能力，是一种有效的细胞核分割方法。 |
| 关键词: 细胞核分割 残差网络 U-Net 深度学习 卷积网络 |
| DOI：10.13657/j.cnki.gxkxyxb.20211216.003 |
| 投稿时间：2021-03-30 |
| 基金项目:国家自然科学基金项目（NF170471）和广西重点研发计划项目（桂科AB19110038）资助。 |
|
| A New Method of Pathological Image Nuclei Segmentation Based on U-Net and ResNet |
|
XU Guang1, GUAN Junlin1, GAN Caijun1,2, WANG Huadeng1,2
|
| (1.School of Computer Science and Information Security, Guilin University of Electronic Technology, Guilin, Guangxi, 541004, China;2.Guangxi Key Laboratory of Image and Graphic Intelligent Processing, Guilin, Guangxi, 541004, China) |
| Abstract: |
| Medical image segmentation is an important part of image processing.And the result of nuclei segmentation is an important basis for cancer classification and grading by pathologists.Improving the accuracy of segmentation has always been a hotspot in research.However,due to the fact that different nuclei in the same organ may have different morphology,overlapping between cells and unclear cell boundaries,it is difficult to accurately segment the nuclear image.In order to improve the accuracy and precision of nucleus segmentation of mutual contact and overlapping cell,a new model of nuclei segmentation network is proposed in this article.In this model,the original cell graph is preprocessed with ZCA bleaching,and based on the classical U-Net network structure,the U-Net and ResNet residual modules are trained.The Batch Normalization method is used to realize data normalization and solve the problem of gradient oscillation in the training process.The experimental results on MoNuSeg and ISBI2018 cell data sets show that the segmentation accuracy of the model proposed in this article is high.Meanwhile,the cells separated do not show large area of nuclear adhesion,and the nuclear contour is clearer.The segmentation network proposed in this article is based on the classical U-Net network structure,and the ResNet residual module is constructed to extract the context features of the nucleus.At the same time,Batch Normalization is used in the residual module to make the gradient transmission more convenient and reduce the training time.Moreover,it has the ability of precise positioning and accurate segmentation when segmenting the nuclei contacted to each other,which is an effective method of nuclei segmentation. |
| Key words: nuclei segmentation residual network U-Net deep learning convolution network |
