应用对抗神经网络的胶质瘤MR图像分割  

MR images segmentation of gliomas using generative adversarial network

作　　者：李丰卓 符玲玲 郭金鑫 卢玉红 唐奇伶[1] LI Fengzhuo;FU Lingling;GUO Jinxin;LU Yuhong;TANG Qiling(College of Biomedical Engineering,South-Central University for Nationalities,Wuhan 430074,China)

机构地区：[1]中南民族大学生物医学工程学院,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2022年第1期64-70,共7页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：湖北省自然科学基金资助项目(2019CFB629)。

摘　　要：针对胶质瘤在结构上的多样性给分割带来的不精确等问题,提出一种应用对抗网络的胶质瘤MR图像分割方法,使用改进的U-Net网络作为生成器的基础架构,获得逐像素的分割结果,判别器是一个卷积神经网络结构.利用对抗机制优化生成器与判别器,直到两者同时收敛为止.训练好的生成器即可完成胶质瘤MRI分割.实验结果表明:提出的方法相比于传统U-net方法,Dice系数提高了4.42%,提高了分割的准确度.automatic segmentation algorithm of low-level glioma image based on generative antagonism network is proposed.The U-net network is used as the basic structure of generator to obtain pixel by pixel segmentation results. The discriminator is a convolutional neural network structure. In the process of network training,the adversarial training mechanism is used to iteratively optimize the generator and the discriminator until converging at the same time. The trained generator can complete MRI segmentation of glioma. Compared with the traditional U-net method,Experimental results show that the Dice coefficient of the proposed method is increased by 4.42%,which improves the accuracy of segmentation.

关 键 词：脑肿瘤 胶质瘤 分割 磁共振成像 对抗神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]