• 首页
  • 期刊简介
  • 编委会
  • 投稿指南
  • 收录情况
  • 杂志订阅
  • 联系我们
引用本文:李 琳,浦贵阳,李 杨,王树超,蒋明峰.基于SA-CycleGAN 的3T磁共振图像生成方法[J].软件工程,2023,26(9):52-58.【点击复制】
【打印本页】   【下载PDF全文】   【查看/发表评论】  【下载PDF阅读器】  
←前一篇|后一篇→ 过刊浏览
分享到: 微信 更多
基于SA-CycleGAN 的3T磁共振图像生成方法
李 琳1, 浦贵阳2, 李 杨3, 王树超4, 蒋明峰3
[1.浙江理工大学信息科学与工程学院, 浙江 杭州 310018;
2.中移(杭州)信息技术有限公司, 浙江 杭州 310000;
3.浙江理工大学计算机科学与技术学院, 浙江 杭州 310018;
4.中国人民解放军联勤保障部队第九○三医院神经外科, 浙江 杭州 310004]
202030504092@mails.zstu.edu.cn; 18867101205@139.com; yangli@zstu.edu.cn; wsccgy@sina.com; m.jiang@zstu.edu.cn
摘 要: 磁共振成像(Magnetic Resonance Imaging, MRI)广泛应用于临床诊断,相较于1.5T MRI,3T MRI具有高对比度和高信噪比等优势。文章提出了一种基于生成对抗网络融合自注意力机制(SA-CycleGAN)的超场强磁共振图像生成方法,利用生成对抗网络从1.5T MRI生成3T MRI,并将自注意力机制嵌入生成对抗网络框架,引入谱归一化处理,在减少函数振荡的同时加速模型收敛;为提高生成图像的真实性,将先验信息引入网络,提出组合损失函数。使用50对3D磁共振图像训练网络,并用10对图像进行测试。实验结果表明:所提SA-CycleGAN方法生成的磁共振图像的峰值信噪比(PSNR)和结构相似性(SSIM )高于SRGAN、CycleGAN等对比方法。
关键词: 磁共振成像;生成对抗网络;自注意力机制;谱归一化;组合损失函数
中图分类号: TP391    文献标识码: A
A Method for Generating 3T MRI Based on SA-CycleGAN
LI Lin1, PU Guiyang2, LI Yang3, WANG Shuchao4, JIANG Mingfeng3
[1.School of Information Science and Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China;
2.China Mobile (Hangzhou) Information Technology Co., Ltd., Hangzhou 310000, China;
3.School of Computer Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China;
4.Neurosurgery of the 903th Hospital of PLA, Hangzhou 310004, China]

202030504092@mails.zstu.edu.cn; 18867101205@139.com; yangli@zstu.edu.cn; wsccgy@sina.com; m.jiang@zstu.edu.cn
Abstract: Magnetic Resonance Imaging (MRI) is widely used in clinical diagnosis, and compared to 1.5T MRI, 3T MRI has advantages such as high contrast and high signal-to-noise ratio. This paper proposes a super-field magnetic resonance image generation method based on the integration of Generative Adversarial Network and Self-attention mechanism (SA-CycleGAN). GAN is used to generate 3T MRI from 1.5T MRI, and the SA mechanism is embedded in the GAN framework. At the same time, spectral normalization processing is introduced, which reduces the function oscillation and accelerates the model convergence. In order to improve the authenticity of the generated image, prior information is introduced into the network and a combined loss function is proposed. The network is trained with 50 pairs of 3D MRI images and tested with 10 pairs of images. The experimental results show that the proposed SACycleGAN can generate higher Peak Signal-to-noise Ratio (PSNR) and Structural Similarity (SSIM) values of magnetic resonance images than contrast methods such as SRGAN (Super Resolution GAN) and CycleGAN.
Keywords: magnetic resonance imaging; generative adversarial network; self-attention mechanism; spectral normalization; combined loss function


版权所有:软件工程杂志社
地址:辽宁省沈阳市浑南区新秀街2号 邮政编码:110179
电话:0411-84767887 传真:0411-84835089 Email:semagazine@neusoft.edu.cn
备案号:辽ICP备17007376号-1
技术支持:北京勤云科技发展有限公司

用微信扫一扫

用微信扫一扫