基于深度神经网络的高泛化性MR快速成像技术

A fast MR imaging technique with decent generalization performance based on deep neural networks

目的:提出一种基于深度神经网络(DNN)重建欠采样MR图像的技术并验证其临床价值。方法:DNN模型的主体由残差卷积网络和保真网络两个模块构成,能够适应不同尺寸和不同分辨率的输入图像且有效学习图像中的噪声分布。收集符合MR扫描适应症的志愿者共150例,K空间满采图像和加速欠采样图像为一组随机扫描同一被试的头部、颈椎、腹部、盆腔和膝关节共5个部位的多种常规序列,共计2437组影像;其中,满采图像作为标签数据,无需额外标注。结果:将同部位不同序列及不同部位不同序列数据分别作为DNN模型的输入训练得到模型1(当前序列除外的图像作为DNN模型输入)、模型2(输入当前序列图像)、模型3(当前部位图像除外)和模型4(输入当前部位图像)的重建效果均很好(SSIM≥0.93,PSNR≥37.22)。DNN模型重建图像的采集时间平均减少16.2%,但CNR平均提升8.5%,SNR提升7.7%以上。此外,DNN重建图像具有同等甚至高于满采图像的质量。结论:DNN模型可重建高质量MR图像且具备高泛化性,帮助临床实现加速扫描。

Objective:To propose a technique for reconstructing undersampled MR images based on deep neural network(DNN) and validate its clinical value.Methods:The main body of the DNN model consisted of two modules:residual convolutional network and fidelity network,which could adapt to input images of different sizes and resolutions and effectively learn the noise distribution in the images.A total of 150 volunteers who met the indications for MR scanning were included in this study.K-space full sampling images and accelerated undersampling images were a set of randomly scanned multiple routine sequences of the same subject's head,cervical spine,abdomen,pelvic cavity,and knee joint,totaling 2 437 sets of images.Among them,the fully captured images were used as labels without the need for additional annotation.Results:To evaluate the generalization of the DNN-based algorithm,four models were built and trained by changing the input images.The inputs of Model 1 employed all sequences(brain only) other than the current sequence as the output image,while the input of Model2 was the opposite.The input of Model 3 employed all sequences of various parts(cervical spine,abdomen,pelvic cavity,and knee) other than the current part as the output image,while the inputs of Model 4 were the opposite.The reconstructed results of four models were all very good(SSIM ≥0.93,PSNR ≥37.22).The average acquisition time was reduced by 16.2%,while the average contrast to noise ratio(CNR) was improved by 8.5%,and the signal to noise ratio(SNR) was improved by more than 7.7%.In addition,the DNN reconstructed images have the same or even higher quality than fully-sampled images.Conclusion:The DNN model can reconstruct high-quality MR images with excellent generalization,which can facilitate fast MR scanning in clinical practice.