摘要
针对传统磁共振成像MRI重建模型存在重建图像细节特征表征不足,导致重建效果不佳的问题,提出在卷积神经网络CNN的基础上,搭建一个基于Swin Transformer的双域生成对抗网络的脑部MRI重建模型,SwinGAN。首先,在SwinGAN生成模型中引入Swin Transformer模块,取代原有的CNN网络,以增强网络全局特征信息;然后增加一个上下文图像相对位置编码器(ciRPE),通过其提高网络捕获上下文信息的能力;最后设计频域和图像域生成器分别对数据中的信息进行处理。结果表明,本模型的PSNR指标和SSIM指标分别取值为33.06和0.954 7,相较于传统的Wasserstein-GAN模型、U-net模型和Complex-CNNDC明显更高。由此分析可知,本模型能够提升脑部MRI成像的细节特征表征能力,图像重建质量显著提升,重建图像与真实图像的相似度更高,满足MRI成像实际应用需求。
In view of the problem that the traditional reconstruction model of MRI reconstruction model is insufficient in the reconstruction image,it is proposed to build a brain MRI reconstruction model based on Swin Transformer based on convolutional neural network CNN,SwinGAN.First,the Swin Transformer module is introduced into the SwinGAN generation model to replace the original CNN network to enhance the network global feature information;then,a context image relative position encoder(ciRPE)is added to improve the ability of the network to capture the context information;finally,the frequency domain and image domain generator are designed to process the information in the data respectively.The results show that the PSNR index and SSIM indexes of this model are 33.06 and 0.9547,respectively,which are significantly higher compared to the traditional Wasserstein-CNN model,U-net model and Complex-CNNDC.According to this analysis,this model can improve the detailed feature characterization ability of brain MRI imaging,significantly improve the quality of image reconstruction,and improve the similarity between reconstructed images and real images,to meet the practical application requirements of MRI imaging.
作者
徐飞
邓亚萍
罗钦
陈兴
XU Fei;DENG Yaping;LUO Qin;CHEN Xing(Shenzhen Nanshan District Medical Group Headquarters,Shenzhen,Guangdong 518052,China)
出处
《自动化与仪器仪表》
2024年第12期56-61,共6页
Automation & Instrumentation
基金
区级《HIS系统的端到端监控》(NS2024010)。
