WO2020151355A1 - 一种基于深度学习的磁共振波谱重建方法 - Google Patents

一种基于深度学习的磁共振波谱重建方法 Download PDF

Info

Publication number
WO2020151355A1
WO2020151355A1 PCT/CN2019/120101 CN2019120101W WO2020151355A1 WO 2020151355 A1 WO2020151355 A1 WO 2020151355A1 CN 2019120101 W CN2019120101 W CN 2019120101W WO 2020151355 A1 WO2020151355 A1 WO 2020151355A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
convolutional neural
signal
magnetic resonance
spectrum
Prior art date
Application number
PCT/CN2019/120101
Other languages
English (en)
French (fr)
Inventor
屈小波
Original Assignee
厦门大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 厦门大学 filed Critical 厦门大学
Publication of WO2020151355A1 publication Critical patent/WO2020151355A1/zh
Priority to US17/385,104 priority Critical patent/US11782111B2/en

Links

Images

Classifications

    • GPHYSICS
    • G01MEASURING; TESTING
    • G01RMEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R33/00Arrangements or instruments for measuring magnetic variables
    • G01R33/20Arrangements or instruments for measuring magnetic variables involving magnetic resonance
    • G01R33/44Arrangements or instruments for measuring magnetic variables involving magnetic resonance using nuclear magnetic resonance [NMR]
    • G01R33/48NMR imaging systems
    • G01R33/54Signal processing systems, e.g. using pulse sequences ; Generation or control of pulse sequences; Operator console
    • G01R33/56Image enhancement or correction, e.g. subtraction or averaging techniques, e.g. improvement of signal-to-noise ratio and resolution
    • G01R33/5608Data processing and visualization specially adapted for MR, e.g. for feature analysis and pattern recognition on the basis of measured MR data, segmentation of measured MR data, edge contour detection on the basis of measured MR data, for enhancing measured MR data in terms of signal-to-noise ratio by means of noise filtering or apodization, for enhancing measured MR data in terms of resolution by means for deblurring, windowing, zero filling, or generation of gray-scaled images, colour-coded images or images displaying vectors instead of pixels
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01RMEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R33/00Arrangements or instruments for measuring magnetic variables
    • G01R33/20Arrangements or instruments for measuring magnetic variables involving magnetic resonance
    • G01R33/44Arrangements or instruments for measuring magnetic variables involving magnetic resonance using nuclear magnetic resonance [NMR]
    • G01R33/46NMR spectroscopy
    • G01R33/4625Processing of acquired signals, e.g. elimination of phase errors, baseline fitting, chemometric analysis
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01RMEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R33/00Arrangements or instruments for measuring magnetic variables
    • G01R33/20Arrangements or instruments for measuring magnetic variables involving magnetic resonance
    • G01R33/44Arrangements or instruments for measuring magnetic variables involving magnetic resonance using nuclear magnetic resonance [NMR]
    • G01R33/46NMR spectroscopy
    • G01R33/4633Sequences for multi-dimensional NMR
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00Image enhancement or restoration
    • G06T5/50Image enhancement or restoration by the use of more than one image, e.g. averaging, subtraction
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02ATECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/30Assessment of water resources

Definitions

  • the invention relates to a magnetic resonance spectrum reconstruction method, in particular to a magnetic resonance spectrum reconstruction method based on deep learning.
  • Magnetic resonance spectroscopy (Magnetic Resonance Spectroscopy, MRS) is a technique for determining molecular structure, which has important applications in the fields of medicine, chemistry, and biology. In magnetic resonance spectroscopy, how to ensure the quality of the spectrum signal while reducing the sampling time is the key to magnetic resonance spectroscopy.
  • Deep learning is an emerging method of data processing and reconstruction. Since Lecun et al. (Lecun Y, Bottou L, Bengio Y, Haffner P. Gradient-based leaming applied to document recognition [J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.) proposed convolutional neural networks (Convolutional Neural Networks, CNN) has received a lot of attention and developed rapidly. Jo Schlemper etc. (Schlemper J, Caballero J, Hajnal J V, Price A, Rueckert DA deep cascade of convolutional neural networks for dynamic MR image reconstruction[J]. IEEE Transactions on Medical Imaging, 2018, 37(2): 491-503 .) A neural network structure for compressed sensing reconstruction using the data collected from the actual measurement of the equipment as the training set is proposed.
  • the purpose of the present invention is to provide a magnetic resonance spectrum reconstruction method based on deep learning.
  • the present invention includes the following steps:
  • the specific method for generating the time-domain signal of the magnetic resonance spectrum by using the exponential function may be: generating the time-domain full sampling signal of the magnetic resonance spectrum according to the exponential function among them Represents a set of complex numbers, N and M represent the number of rows and columns of the time signal.
  • the specific expression of the fully sampled signal T is:
  • T n, m represents the data in the nth row and mth column of the signal T
  • R represents the number of spectral peaks
  • a r represents the amplitude
  • ⁇ t 1 and ⁇ t 2 represent the time increment
  • f 1, r and f 2 R represents the normalized frequency
  • ⁇ 1, r and ⁇ 2 r represents the attenuation factor
  • step 2) the specific method for establishing the training set of the under-sampled time-domain signal and the full-sampled spectrum can be:
  • U is used to indicate the under-sampling operation in the time domain, and the white in the template indicates that the corresponding data point is sampled. The data points in black are not sampled.
  • represents the index subset of U.
  • the unsampled signal in T is filled with 0 to obtain the complementary time-domain signal Tu , and the Fourier transform of Tu is performed to obtain the aliased spectrum signal Su ;
  • the Fourier transform obtains the fully sampled spectrum S, and saves the real and imaginary parts of S separately, namely among them, Represents a real number, the training set is composed of Tu and S
  • the specific method for the design data to verify the convolutional neural network in the convolutional neural network structure may be: the convolutional neural network module will include L convolutional layers, and each convolutional layer has 1 filter Convolutional layers are densely connected. The input of each layer in the module is the union of the outputs of all previous layers. In all convolutional layers, the size of the convolution kernel is k.
  • the convolutional neural network module complete the output signal S cnn,l from the input signal S l of the lth layer (1 ⁇ l ⁇ L) after passing through the convolutional neural network, which is defined as:
  • is the training parameter of the convolutional neural network
  • ⁇ ) represents the non-linear mapping from S l to S cnn,l of the training.
  • the specific method for the design data to verify the bottleneck layer in the convolutional neural network structure can be: the bottleneck layer mainly completes the function of changing the number of feature maps in the network structure, and the bottleneck layer is located in the convolutional neural network module Before and after entering the convolutional neural network module, the signal will pass through a bottleneck layer of Ki filters to increase the number of feature maps, and the output signal of the convolutional neural network module will also pass through a bottleneck layer of k o filters to reduce features Number of figures.
  • the specific method of designing the data verification layer in the data verification convolutional neural network structure can be: the data verification layer mainly completes the data verification function in the network structure, and will come from the i th volume
  • the output signal S cnn,l of the product neural network is used as input, and the input signal S cnn,l is transformed back into the time domain using the inverse Fourier transform F H to obtain the signal T l , the formula is as follows:
  • the specific method for establishing a data verification convolutional neural network structure as a spectral reconstruction model may be: multiple convolutional neural networks and data verification layers are cascaded in the data verification convolutional neural network structure
  • the combined module completes the input of the under-sampled magnetic resonance time-domain signal Tu , and outputs the reconstructed magnetic resonance spectrum signal
  • the function of the whole constitutes an end-to-end deep neural network structure.
  • the loss function of the data verification convolutional neural network structure will be defined as:
  • step 8 the specific method of training network optimization parameters can be: using Adam algorithm, which is common in deep learning, and training the model parameters in step 5) to obtain the optimal value of the model with
  • step 9 the under-sampled magnetic resonance time-domain signal of the target is
  • the specific method of reconstruction can be: input the under-sampled time-domain signal to the model After the forward propagation of the model, the complete spectrum signal is reconstructed Expressed by formula (6) as:
  • the invention provides a new method for reconstructing a complete spectrum from under-sampled magnetic resonance spectrum data by using a deep learning network.
  • the exponential function is used to generate the time-domain signal of the magnetic resonance spectrum, and the completed time-domain signal is obtained after the under-sampling operation is completed in the time domain, and the completed time-domain signal and the complete spectrum corresponding to the full sampling are combined to form a training data set;
  • establish a data verification convolutional neural network model for magnetic resonance spectrum reconstruction use the training data set to train the neural network parameters to form a trained neural network; finally, input the under-sampled magnetic resonance time-domain signal to be reconstructed In the trained data verification convolutional neural network, a complete magnetic resonance spectrum is reconstructed.
  • the present invention has a faster reconstruction speed, does not require the measured data set collected by the device as the training set, but uses the time signal generated by the exponential function as the training set of the magnetic resonance spectrum and designs the corresponding neural network structure .
  • This method of reconstructing the magnetic resonance spectrum through the data verification convolutional neural network has the characteristics of fast reconstruction speed and high quality of the reconstructed spectrum.
  • Figure 1 is the data verification convolutional neural network structure of MRI spectrum reconstruction.
  • Figure 2 is the under-sampling template.
  • Fig. 3 is a spectrum obtained after reconstruction of the full-sampling spectrum and the under-sampling magnetic resonance time-domain data of Example 1.
  • (a) is the full-sampling spectrogram
  • (b) is the reconstruction spectrum of the present invention in Example 1.
  • Figure 4 shows the correlation between the peak intensity of the full sampled spectrum and the peak intensity of the reconstructed spectrum.
  • an exponential function is used to generate a magnetic resonance signal training network, and then a two-dimensional magnetic resonance spectrum is reconstructed from the under-sampled magnetic resonance time-domain signal.
  • the specific implementation process is as follows:
  • This embodiment generates 5200 free induction attenuation signals. Generate the time-domain fully sampled signal of the magnetic resonance spectrum according to the exponential function Its expression is:
  • N and M represent the number of rows and columns of the time signal
  • R represents the number of spectral peaks
  • a r represents the magnitude of the amplitude
  • ⁇ t 1 and ⁇ t 2 represent time increments
  • f 1, r and f 2 where r represents the normalized frequency
  • ⁇ 1, r and ⁇ 2 where r represents the attenuation factor.
  • the number of peaks R is 2 to 52.
  • the range of amplitude a r is 0.05 ⁇ a r ⁇ 1, and the frequency f 1, r and f 2, r
  • the value range is 0.05 ⁇ f 1,r , f 2,r ⁇ 1.
  • the attenuation factor ⁇ 1,r and ⁇ 2, the value range of r is 19.2 ⁇ 1,r , ⁇ 2,r ⁇ 179.2.
  • M represents the under-sampling operation in the time domain.
  • Figure 2 is a schematic diagram of the under-sampling template. In the template, the white indicates that the corresponding data point is sampled, and the black indicates that the data point is not sampled. A total of 30% of the data points are collected. . ⁇ represents the index subset of U.
  • the unsampled signal in T is filled with 0 to obtain the complementary time-domain signal Tu , and the Fourier transform of Tu is performed to obtain the aliased spectrum signal Su .
  • Perform Fourier transform on the fully sampled signal T to obtain the fully sampled spectrum S, and save the real and imaginary parts of S separately, namely among them Represents a real number.
  • the training set is composed of both Tu and S
  • the convolutional neural network module will contain 8 convolutional layers, each with 12 filters.
  • the convolutional layers are densely connected, and the input of each layer in the module is the union of the outputs of all the previous layers.
  • the size of the convolution kernel is 3 ⁇ 3.
  • is the training parameter of the convolutional neural network
  • ⁇ ) represents the non-linear mapping from S l to S cnn,l of the training.
  • the bottleneck layer mainly completes the function of changing the number of feature maps in the network structure.
  • the bottleneck layer is located before and after the convolutional neural network module. Before entering the convolutional neural network module, the signal will pass through a bottleneck layer of 16 filters to increase the number of feature maps, and the output signal of the convolutional neural network module will also pass through a bottleneck layer of 2 filters to reduce the number of feature maps.
  • the data verification layer mainly completes the data verification function in the network structure. Taking the output signal S cnn,l from the l-th convolutional neural network as input, using the inverse Fourier transform F H to transform the input signal S cnn,l back into the time domain to obtain the signal T l , the formula is as follows:
  • the feedback function in the network structure makes the output of each module combined with the convolutional neural network and the data verification layer closer to the true spectrum signal.
  • the optimal value of the model can be obtained by training the model parameters in step 5) with
  • Fig. 3(b) The spectrogram obtained after reconstruction of the under-sampled magnetic resonance time-domain data in the embodiment is shown in Fig. 3(b), and compared with the full-sampling spectrogram in Fig. 3(a). It can be concluded that by using the designed data to verify the convolutional neural network and the network trained by the synthetic simulated magnetic resonance spectrum, high-quality magnetic resonance spectrum can be obtained.
  • Figure 4 shows the correlation between the peak intensity of the full-sampled spectrum and the peak intensity of the reconstructed spectrum.
  • the invention provides a new method for reconstructing a complete spectrum from under-sampled magnetic resonance spectrum data by using a deep learning network.
  • a finite exponential function is used to generate a time signal.
  • an aliased spectrum in the frequency domain is obtained, and then the aliased spectrum and the complete spectrum corresponding to the full sampling are combined to form a training data set.
  • a data verification convolutional neural network model for magnetic resonance spectrum reconstruction is established, and the training data set is used to train the neural network parameters to form a trained neural network.
  • This method of reconstructing the magnetic resonance spectrum through the data verification convolutional neural network has the characteristics of fast reconstruction speed and high quality of the reconstruction spectrum, and has good industrial practicability.

Abstract

提供一种利用深度学习网络从欠采样的磁共振波谱数据中重建出完整波谱的新方法。首先,利用有限指数函数来生成时间信号,在时域完成欠采样操作后得到频域的带混叠的波谱,然后将带混叠的波谱与全采样对应的完整波谱共同组成训练数据集。之后,建立用于磁共振波谱重建的数据校验卷积神经网络模型,用训练数据集来训练神经网络参数,形成训练好的神经网络。最后,将需重建的欠采样磁共振波谱信号输入到训练好的数据校验卷积神经网络中,重建出完整的磁共振波谱。这种通过数据校验卷积神经网络重建磁共振波谱的方法具有重建速度快和重建波谱质量高的特点。

Description

一种基于深度学习的磁共振波谱重建方法 技术领域
本发明涉及磁共振波谱重建方法,尤其是涉及一种基于深度学习的磁共振波谱重建方法。
背景技术
磁共振波谱(Magnetic Resonance Spectroscopy,MRS)是一种测定分子结构的一项技术,在医学、化学和生物学等领域有着重要应用。在磁共振波谱中,如何保证波谱信号质量的同时降低采样时间是磁共振波谱的关键。
传统的磁共振重建方法主要利用磁共振时间或者频率信号的数学特性来重建频谱。Qu Xiaobo等(Qu X,Mayzel M,Cai J,Chen Z,Orekhov V.Accelerated NMR spectroscopy with low-Rank reconstruction[J].Angewandte Chemie International Edition,2015,54(3):852-854.)提出了一种基于低秩汉克尔矩阵的磁共振波谱重建方法,在欠采样过程中重建出的高质量波谱信号,解决了对不同宽度的谱峰重建效果差的问题。该方法还扩展到了三维及更高维的波谱重建中(Ying J,Lu H,Wei Q,Cai J,Guo D,Wu J,Chen Z,Qu X.Hankel matrix nuclear norm regularized tensor completion for N-dimensional exponential signals[J],IEEE Transactions on Signal Processing,2017,65(14):3702-3717.),还通过对汉克尔矩阵的范德蒙分解(Ying J,Cai J,Guo Di,Tang G,Chen Z,Qu X,Vandermonde factorization of Hankel matrix for complex exponential signal recovery-application in fast NMR spectroscopy[J],IEEE Transactions on Signal Processing,2018,66(21):5520-5533.)和奇异值操作(Guo D,Qu X.Improved reconstruction of low intensity magnetic resonance spectroscopy with weighted low rank Hankel matrix completion[J].IEEE Access,2018,6:4933-4940)和(Qu X,Qiu T,Guo Di,Lu H,Ying J,Shen M,Hu B,Orekhov V,Chen Z.High-fidelity spectroscopy reconstruction in accelerated NMR[J],Chemical Communications,2018,54(78):10958-10961.)提高了波谱重建对密集谱峰和低强度谱峰的重建能力。但是,这类低秩汉克尔矩阵重建方法在迭代计算中的奇异值分解的时间消耗高,因此导致波谱重建时间较长。Guo Di等(Guo D,Lu H,Qu X.A fast low rank Hankel matrix factorization reconstruction method for non-uniformly sampled magnetic resonance spectroscopy[J].IEEE Access,2017,5:16033-16039.)成功地将低秩矩阵进行因式分解并引入并行计算,避免时间复杂度高的奇异值分解方法。Lu Hengfa等(Lu H,Zhang X,Qiu T,Yang J,Ying J,Guo D,Chen Z,Qu X.Low rank enhanced matrix recovery of hybrid time and frequency data in fast magnetic resonance spectroscopy[J].IEEE Transactions on Biomedical Engineering,2018,65(4):809-820.)则提出利用费罗贝尼乌斯范数项进行矩阵因子分解来避免奇异值分解,完成对欠采样多维磁共振波谱信号的快速和高质量波谱重建。
深度学习是一种新兴的数据处理与重建方法。自Lecun等(Lecun Y,Bottou L,Bengio Y,Haffner P.Gradient-based leaming applied to document recognition[J].Proceedings of the IEEE,1998,86(11):2278-2324.)提出卷积神经网络(Convolutional Neural Networks,CNN)受到许多关注并且发展迅速。Jo Schlemper等(Schlemper J,Caballero J,Hajnal J V,Price A,Rueckert D.A deep cascade of convolutional neural networks for dynamic MR image reconstruction[J].IEEE Transactions on Medical Imaging,2018,37(2):491-503.)提出了利用设备实测采集数据作为训练集的压缩感知重建的神经网络构。
发明内容
本发明的目的在于提供一种基于深度学习的磁共振波谱重建方法。
本发明包括以下步骤:
1)利用指数函数生成磁共振波谱的时域信号;
在步骤1)中,所述利用指数函数生成磁共振波谱的时域信号的具体方法可为:根据指数函数生成磁共振波谱时域的全采样信号
Figure PCTCN2019120101-appb-000001
其中
Figure PCTCN2019120101-appb-000002
表示复数的集合,N和M表示时间信号的行数和列数。全采样信号T的具体表达式为:
Figure PCTCN2019120101-appb-000003
其中,T n,m表示信号T的第n行,第m列的数据,R表示谱峰个数,a r表示幅度大小,Δt 1和Δt 2表示时间增量,f 1,r和f 2,r表示归一化频率,τ 1,r和τ 2,r表示衰减因子;表达式(1)同样适用于一维自由感应衰减全采样信号,此时有n=1,m>1或m=1,n>1。
2)建立欠采样时域信号与全采样波谱的训练集;
在步骤2)中,所述建立欠采样时域信号与全采样波谱的训练集的具体方法可为:采用U表示在时域中的欠采样操作,模板中白色表示对应的数据点被采样,黑色表示的数据点未被采样,Ω表示U的索引子集,若某一个信号点的索引(p,q)出现在集合Ω中,则(p,q)∈Ω;若某一个信号点的索引(p,q)没有出现在集合Ω中,则
Figure PCTCN2019120101-appb-000004
根据欠采样模板M对T中未被采样的信号通过填0得到补全的时域信号T u,对T u进行傅里叶变换获得带混叠的波谱信号S u;对全采样信号T进行傅里叶变换得到全采样波谱S,并将S的实部和虚部分开保存,即
Figure PCTCN2019120101-appb-000005
其中,
Figure PCTCN2019120101-appb-000006
表示实数,由T u和S两者共同组成训练集
Figure PCTCN2019120101-appb-000007
3)设计数据校验卷积神经网络结构中的卷积神经网络;
在步骤3)中,所述设计数据校验卷积神经网络结构中的卷积神经网络的具体方法可为:卷积神经网络模块将包含L个卷积层,每个卷积层I个滤波器;卷积层间采用密集连接的方式,模块中每一层的输入都是前面所有层输出的并集,在所有的卷积层中,卷积核大小为k。通过卷积神经网络模块,完成从第l层(1≤l≤L)的输入信号S l经过卷积神经网络后输出信号S cnn,l,它的定义为:
S cnn,l=f(S l|θ)    (2)
其中,θ是卷积神经网络的训练参数,f(S l|θ)表示训练的从S l到S cnn,l的非线性映射。
4)设计数据校验卷积神经网络结构中的瓶颈层;
在步骤4)中,所述设计数据校验卷积神经网络结构中的瓶颈层的具体方法可 为:瓶颈层在网络结构中主要完成改变特征图数量的功能,瓶颈层位于卷积神经网络模块的前后,进入卷积神经网络模块前信号会通过一个Ki个滤波器的瓶颈层以提高特征图数量,卷积神经网络模块的输出信号也会通过一个k o个滤波器的瓶颈层以减少特征图数量。
5)设计数据校验卷积神经网络结构中的数据校验层;
在步骤5)中,所述设计数据校验卷积神经网络结构中的数据校验层的具体方法可为:数据校验层在网络结构中主要完成数据校验功能,将来自第ι个卷积神经网络的输出信号S cnn,l作为输入,利用傅里叶逆变换F H将输入信号S cnn,l变换回时域中,获得信号T l,公式如下:
T l=F HS cnn,l            (3)
数据校验层的表达式如下:
Figure PCTCN2019120101-appb-000008
最后输出频域波谱
Figure PCTCN2019120101-appb-000009
其中最后一次,即第L层(L>1)的波谱
Figure PCTCN2019120101-appb-000010
即为整个深度学习网络的输出
Figure PCTCN2019120101-appb-000011
6)设计数据校验卷积神经网络结构中的反馈功能;
在步骤6)中,所述设计数据校验卷积神经网络结构中的反馈功能的具体方法可为:反馈功能在网络结构中使得每个卷积神经网络和数据校验层组合的模块输出更逼近真实谱信号,且使下一模块的输入更具可解释性;将每个数据校验层的输出与真实谱信号S=FT进行比较并反馈每个模块的参数更新,其中T是公式(1)中的全采样时域信号,F表示傅里叶变换。
7)建立数据校验卷积神经网络结构作为波谱重建模型;
在步骤7)中,所述建立数据校验卷积神经网络结构作为波谱重建模型的具体方法可为:数据校验卷积神经网络结构中级联了多个卷积神经网络和数据校验层组合的模块,完成输入欠采样的磁共振时域信号T u,输出重建后的磁共振波谱信号
Figure PCTCN2019120101-appb-000012
的功能,整体上构成一个端到端的深度神经网络结构。数据校验卷积神经网络结构的损失函数将定义为:
Figure PCTCN2019120101-appb-000013
其中,
Figure PCTCN2019120101-appb-000014
表示训练集,||·|| F表示矩阵的F-范数(Frobenius范数),
Figure PCTCN2019120101-appb-000015
θ是卷积神经网络的训练参数,λ是数据校验层的数据校验参数,两个参数θ和λ都需要训练。
8)训练网络最优化参数;
在步骤8)中,所述训练网络最优化参数的具体方法可为:采用深度学习中常见的Adam算法,对步骤5)的模型参数经过训练可得到模型的最优取值
Figure PCTCN2019120101-appb-000016
Figure PCTCN2019120101-appb-000017
9)对目标的欠采样磁共振时域信号
Figure PCTCN2019120101-appb-000018
进行重建;
在步骤9)中,所述对目标的欠采样磁共振时域信号
Figure PCTCN2019120101-appb-000019
进行重建的具体方法可为:给模型输入欠采样的时域信号
Figure PCTCN2019120101-appb-000020
经过模型的正向传播后,重建出完整的波谱信号
Figure PCTCN2019120101-appb-000021
用公式(6)表示为:
Figure PCTCN2019120101-appb-000022
10)在时频域进行欠采样操作的同时,利用卷积神经网络的强拟合能力和数据校验层数据校验的能力,完成对欠采样磁共振波谱信号的快速且高质量的重建。
本发明提供一种利用深度学习网络从欠采样的磁共振波谱数据中重建出完整波谱的新方法。首先,利用指数函数生成磁共振波谱的时域信号,在时域完成欠采样操作后得到补全的时域信号,将补全的时域信号与全采样对应的完整波谱共同组成训练数据集;然后,建立用于磁共振波谱重建的数据校验卷积神经网络模型,用训练数据集来训练神经网络参数,形成训练好的神经网络;最后,将需重建的欠采样磁共振时域信号输入到训练好的数据校验卷积神经网络中,重建出完整的磁共振波谱。
与传统方法相比,本发明具有更快的重建速度,不需要设备采集的实测数据集作为训练集,而是根据指数函数产生的时间信号作为磁共振波谱的训练集并设计了对应神经网络结构。这种通过数据校验卷积神经网络重建磁共振波谱的方法具有重建速度快和重建波谱质量高的特点。
附图说明
图1是磁共振波谱重建的数据校验卷积神经网络结构。
图2是欠采样模板。
图3是实施例1的全采样波谱和欠采样磁共振时域数据重建后获得的谱图。在图3中,(a)为全采样波谱图,(b)为实施例1的本发明重建谱。
图4是全采样波谱的谱峰强度和重建波谱的谱峰强度相关性。
具体实施方式
以下实施例将结合附图对本发明作进一步的说明。
本发明实施例将利用指数函数生成磁共振信号训练网络,然后在欠采样的磁共振时域信号中重建出二维磁共振波谱。具体实施过程如下:
1)利用指数函数生成磁共振波谱的时域信号
本实施例产生5200个自由感应衰减信号。根据指数函数生成磁共振波谱的时域的全采样信号
Figure PCTCN2019120101-appb-000023
其表达式为:
Figure PCTCN2019120101-appb-000024
其中,
Figure PCTCN2019120101-appb-000025
表示复数的集合,N和M表示时间信号的行数和列数,T n,m表示信号T的第n行,第m列的数据,R表示谱峰个数,a r表示幅度大小,Δt 1和Δt 2表示时间增量,f 1,r和f 2,r表示归一化频率,τ 1,r和τ 2,r表示衰减因子。实施例中N=256和M=116,谱峰的个数R为2~52个。对固定的谱峰,将尘成带有不同的幅度、频率和衰减因子的200个自由感应衰减信号,幅度a r取值范围0.05≤a r≤1,频率f 1,r和f 2,r取值范围0.05≤f 1,r,f 2,r≤1。衰减因子τ 1,r和τ 2,r取值范围19.2≤τ 1,r,τ 2,r≤179.2。
2)建立欠采样时域信号与全采样波谱的训练集
M表示在时域中的欠采样操作,图2是欠采样模板的示意图,模板中白色表示对应的数据点被采样,黑色表示的数据点未被采样,共计有30%的数据点被采集到。Ω表示U的索引子集,若某一个信号点的索引(p,q)出现在集合Ω中,则(p,q)∈Ω,若某一个信号点的索引(p,q)没有出现在集合Ω中,则
Figure PCTCN2019120101-appb-000026
根据欠采样模板U对T中未被采样的信号通过填0得到补全的时域信号T u,对T u进行傅里叶变换获得带混叠的波谱信号S u。对全采样信号T进行傅里叶变换得到全采样波谱S,并将S的实部和虚部分开保存,即
Figure PCTCN2019120101-appb-000027
其中
Figure PCTCN2019120101-appb-000028
表示实数。由T u和S两者共同组成训练集
Figure PCTCN2019120101-appb-000029
3)设计数据校验卷积神经网络结构中的卷积神经网络。
卷积神经网络模块将包含8个卷积层,每个卷积层12个滤波器。卷积层间采用密集连接的方式,模块中每一层的输入都是前面所有层输出的并集。在所有的卷积层中,卷积核大小为3×3。通过卷积神经网络模块,完成从第l层(1≤l≤L)的输入信号S l经过卷积神经网络后输出信号S cnn,l,它的定义为:
S cnn,l=f(S l|θ)       (2)
其中,θ是卷积神经网络的训练参数,f(S l|θ)表示训练的从S l到S cnn,l的非线性映射。
4)设计数据校验卷积神经网络结构中的瓶颈层
瓶颈层在网络结构中主要完成改变特征图数量的功能。瓶颈层位于卷积神经网络模块的前后。进入卷积神经网络模块前信号会通过一个16个滤波器的瓶颈层以提高特征图数量,卷积神 经网络模块的输出信号也会通过一个2个滤波器的瓶颈层以减少特征图数量。
5)设计数据校验卷积神经网络结构中的数据校验层
数据校验层在网络结构中主要完成数据校验功能。将来自第l个卷积神经网络的输出信号S cnn,l作为输入,利用傅里叶逆变换F H将输入信号S cnn,l变换回时域中,获得信号T l,公式如下:
T l=F HS cnn,l         (3)
数据校验层的表达式如下:
Figure PCTCN2019120101-appb-000030
最后输出频域波谱信号
Figure PCTCN2019120101-appb-000031
其中最后一次,即第8层的波谱
Figure PCTCN2019120101-appb-000032
即为整个深度学习网络的输出
Figure PCTCN2019120101-appb-000033
6)设计数据校验卷积神经网络结构中的反馈功能。
反馈功能在网络结构中使得每个卷积神经网络和数据校验层组合的模块输出更逼近真实谱信号。将每个数据校验层的输出与真实谱信号S=FT进行比较并反馈给每个模块的参数更新,其中T是公式(1)中的全采样时域信号,F表示傅里叶变换。
7)建立数据校验卷积神经网络结构作为波谱重建模型
数据校验卷积神经网络结构中级联了多个卷积神经网络和数据校验层组合的模块,完成输入欠采样的磁共振时域信号T u,输出重建后的磁共振波谱信号
Figure PCTCN2019120101-appb-000034
的功能,整体上构成一个端到端的深度神经网络结构。数据校验卷积神经网络结构的损失函数将定义为:
Figure PCTCN2019120101-appb-000035
其中,
Figure PCTCN2019120101-appb-000036
表示训练集,||·|| F表示矩阵的F-范数(Frobenius范数),
Figure PCTCN2019120101-appb-000037
θ是卷积神经网络的训练参数,λ是数据校验层的数据校验参数,两个参数θ和λ都需要训练。最终设计的数据校验卷积神经网络的结构图如图1所示。
8)训练网络最优化参数。
采用深度学习中常见的Adam算法,对步骤5)的模型参数经过训练可得到模型的最优取值
Figure PCTCN2019120101-appb-000038
Figure PCTCN2019120101-appb-000039
9)对目标的欠采样磁共振时域信号
Figure PCTCN2019120101-appb-000040
进行重建
给模型输入欠采样的时域信号
Figure PCTCN2019120101-appb-000041
经过模型的正向传播后,重建出完整的波谱信号
Figure PCTCN2019120101-appb-000042
用公式表示为:
Figure PCTCN2019120101-appb-000043
对实施例中的欠采样磁共振时域数据重建后获得的谱图分别如图3(b),对比图3(a)的全采样波谱图。可以得出,利用设计的数据校验卷积神经网络和合成仿真磁共振波谱训练出来的网络,可获得高质量的磁共振波谱图。图4是全采样波谱的谱峰强度与重建波谱的谱峰强度的相关性。
工业实用性
本发明提供一种利用深度学习网络从欠采样的磁共振波谱数据中重建出完整波谱的新方法。首先,利用有限指数函数来生成时间信号,在时域完成欠采样操作后得到频域的带混叠的波谱,然后将带混叠的波谱与全采样对应的完整波谱共同组成训练数据集。之后,建立用于磁共振波谱重建的数据校验卷积神经网络模型,用训练数据集来训练神经网络参数,形成训练好的神经网络。最后,将需重建的欠采样磁共振波谱信号输入到训练好的数据校验卷积神经网络中,重建出完整的磁共振波谱。这种通过数据校验卷积神经网络重建磁共振波谱的方法具有重建速度快和重建波谱质量高的特点,具有良好的工业实用性。

Claims (10)

  1. 一种基于深度学习的磁共振波谱重建方法,其特征在于包括以下步骤:
    1)利用指数函数生成磁共振波谱的时域信号;
    2)建立欠采样时域信号与全采样波谱的训练集;
    3)设计数据校验卷积神经网络结构中的卷积神经网络;
    4)设计数据校验卷积神经网络结构中的瓶颈层;
    5)设计数据校验卷积神经网络结构中的数据校验层;
    6)设计数据校验卷积神经网络结构中的反馈功能;
    7)建立数据校验卷积神经网络结构作为波谱重建模型;
    8)训练网络最优化参数;
    9)对目标的欠采样磁共振时域信号
    Figure PCTCN2019120101-appb-100001
    进行重建;
    10)在时频域进行欠采样操作的同时,利用卷积神经网络的强拟合能力和数据校验层数据校验的能力,完成对欠采样磁共振波谱信号的快速且高质量的重建。
  2. 如权利要求1所述一种基于深度学习的磁共振波谱重建方法,其特征在于在步骤1)中,所述利用指数函数生成磁共振波谱的时域信号的具体方法为:根据指数函数生成磁共振波谱时域的全采样信号
    Figure PCTCN2019120101-appb-100002
    的表达式为:
    Figure PCTCN2019120101-appb-100003
    其中,
    Figure PCTCN2019120101-appb-100004
    表示复数的集合,N和M表示时间信号的行数和列数,T n,m表示信号T的第n行,第m列的数据,R表示谱峰个数,a r表示幅度大小,Δt 1和Δt 2表示时间增量,f 1,r和f 2,r表示归一化频率,τ 1,r和τ 2,r表示衰减因子;表达式(1)同样适用于一维自由感应衰减全采样信号,此时有n=1,m>1或m=1,n>1。
  3. 如权利要求2所述一种基于深度学习的磁共振波谱重建方法,其特征在于在步骤2)中,所述建立欠采样时域信号与全采样波谱的训练集的具体方法为:采用U表示在时域中的欠采样操作,模板中白色表示对应的数据点被采样,黑色表示的数据点未被采样,Ω表示U的索引子集,若某一个信号点的索引(p,q)出现在集合Ω中,则(p,q)∈Ω;若某一个信号点的索引(p,q)没有出现在集合Ω中,则
    Figure PCTCN2019120101-appb-100005
    根据欠采样模板M对T中未被采样的信号通过填0得到补全的时域信号T u,对T u进行傅里叶变换获得带混叠的波谱信号S u;对全采样信号T进行傅里叶变换得到全采样 波谱S,并将S的实部和虚部分开保存,即
    Figure PCTCN2019120101-appb-100006
    其中,
    Figure PCTCN2019120101-appb-100007
    表示实数,由T u和S两者共同组成训练集
    Figure PCTCN2019120101-appb-100008
  4. 如权利要求1所述一种基于深度学习的磁共振波谱重建方法,其特征在于在步骤3)中,所述设计数据校验卷积神经网络结构中的卷积神经网络的具体方法为:卷积神经网络模块将包含L个卷积层,每个卷积层I个滤波器;卷积层间采用密集连接的方式,模块中每一层的输入都是前面所有层输出的并集,在所有的卷积层中,卷积核大小为k;通过卷积神经网络模块,完成从第l层(1≤l≤L)的输入信号S l经过卷积神经网络后输出信号S cnn,l,它的定义为:
    S cnn,l=f(S l|θ)  (2)
    其中,θ是卷积神经网络的训练参数,f(S l|θ)表示训练的从S l到S cnn,l的非线性映射。
  5. 如权利要求4所述一种基于深度学习的磁共振波谱重建方法,其特征在于在步骤4)中,所述设计数据校验卷积神经网络结构中的瓶颈层的具体方法为:瓶颈层在网络结构中主要完成改变特征图数量的功能,瓶颈层位于卷积神经网络模块的前后,进入卷积神经网络模块前信号会通过一个Ki个滤波器的瓶颈层以提高特征图数量,卷积神经网络模块的输出信号也会通过一个k o个滤波器的瓶颈层以减少特征图数量。
  6. 如权利要求5所述一种基于深度学习的磁共振波谱重建方法,其特征在于在步骤5)中,所述设计数据校验卷积神经网络结构中的数据校验层的具体方法为:数据校验层在网络结构中主要完成数据校验功能,将来自第ι个卷积神经网络的输出信号S cnn,l作为输入,利用傅里叶逆变换F H将输入信号S cnn,l变换回时域中,获得信号T l,公式如下:
    T l=F HS cnn,l  (3)数据校验层的表达式如下:
    Figure PCTCN2019120101-appb-100009
    最后输出频域波谱
    Figure PCTCN2019120101-appb-100010
    其中最后一次,即第L层(L>1)的波谱
    Figure PCTCN2019120101-appb-100011
    即为整个深度学习网络的输出
    Figure PCTCN2019120101-appb-100012
  7. 如权利要求1所述一种基于深度学习的磁共振波谱重建方法,其特征在于在 步骤6)中,所述设计数据校验卷积神经网络结构中的反馈功能的具体方法为:反馈功能在网络结构中使得每个卷积神经网络和数据校验层组合的模块输出更逼近真实谱信号,且使下一模块的输入更具可解释性;将每个数据校验层的输出与真实谱信号S=FT进行比较并反馈每个模块的参数更新,其中T是公式(1)中的全采样时域信号,F表示傅里叶变换。
  8. 如权利要求6或7所述一种基于深度学习的磁共振波谱重建方法,其特征在于在步骤7)中,所述建立数据校验卷积神经网络结构作为波谱重建模型的具体方法为:数据校验卷积神经网络结构中级联了多个卷积神经网络和数据校验层组合的模块,完成输入欠采样的磁共振时域信号T u,输出重建后的磁共振波谱信号
    Figure PCTCN2019120101-appb-100013
    的功能,整体上构成一个端到端的深度神经网络结构;数据校验卷积神经网络结构的损失函数将定义为:
    Figure PCTCN2019120101-appb-100014
    其中,
    Figure PCTCN2019120101-appb-100015
    表示训练集,||·|| F表示矩阵的F-范数(Frobenius范数),
    Figure PCTCN2019120101-appb-100016
    θ是卷积神经网络的训练参数,λ是数据校验层的数据校验参数,两个参数θ和λ都需要训练。
  9. 如权利要求8所述一种基于深度学习的磁共振波谱重建方法,其特征在于在步骤8)中,所述训练网络最优化参数的具体方法为:采用Adam算法,对步骤5)的模型参数经过训练可得到模型的最优取值
    Figure PCTCN2019120101-appb-100017
    Figure PCTCN2019120101-appb-100018
  10. 如权利要求9所述一种基于深度学习的磁共振波谱重建方法,其特征在于在步骤9)中,所述对目标的欠采样磁共振时域信号
    Figure PCTCN2019120101-appb-100019
    进行重建的具体方法为:给模型输入欠采样的时域信号
    Figure PCTCN2019120101-appb-100020
    经过模型的正向传播后,重建出完整的波谱信号
    Figure PCTCN2019120101-appb-100021
    用公式(6)表示为:
    Figure PCTCN2019120101-appb-100022
PCT/CN2019/120101 2019-01-25 2019-11-22 一种基于深度学习的磁共振波谱重建方法 WO2020151355A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/385,104 US11782111B2 (en) 2019-01-25 2021-07-26 Method for reconstructing magnetic resonance spectrum based on deep learning

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910075573.3 2019-01-25
CN201910075573.3A CN109903259B (zh) 2019-01-25 2019-01-25 一种基于深度学习的磁共振波谱重建方法

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/385,104 Continuation US11782111B2 (en) 2019-01-25 2021-07-26 Method for reconstructing magnetic resonance spectrum based on deep learning

Publications (1)

Publication Number Publication Date
WO2020151355A1 true WO2020151355A1 (zh) 2020-07-30

Family

ID=66944158

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/120101 WO2020151355A1 (zh) 2019-01-25 2019-11-22 一种基于深度学习的磁共振波谱重建方法

Country Status (3)

Country Link
US (1) US11782111B2 (zh)
CN (1) CN109903259B (zh)
WO (1) WO2020151355A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112630709A (zh) * 2020-11-03 2021-04-09 成都易检医疗科技有限公司 金属植入物磁共振成像方法、系统、终端及存储介质
CN113361819A (zh) * 2021-07-08 2021-09-07 武汉中科牛津波谱技术有限公司 一种线性预测方法及装置

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109903259B (zh) 2019-01-25 2020-05-29 厦门大学 一种基于深度学习的磁共振波谱重建方法
CN110490063A (zh) * 2019-07-11 2019-11-22 厦门大学 基于神经网络算法的不均匀场下脑模型定域谱校正方法
CN110598579B (zh) * 2019-08-23 2022-08-05 厦门大学 一种基于深度学习的超复数磁共振波谱重建方法
US20230194640A1 (en) * 2019-09-13 2023-06-22 Cedars-Sinai Medical Center Systems and methods of deep learning for large-scale dynamic magnetic resonance image reconstruction
CN110658484B (zh) * 2019-10-17 2021-07-20 东北大学 一种磁共振波谱重建方法及系统
CN111123183B (zh) * 2019-12-27 2022-04-15 杭州电子科技大学 基于复数R2U_Net网络的快速磁共振成像方法
CN111324861B (zh) * 2020-02-28 2022-05-03 厦门大学 一种基于矩阵分解的深度学习磁共振波谱重建方法
CN111915007B (zh) * 2020-07-29 2022-06-14 厦门大学 一种基于神经网络的磁共振谱降噪方法
CN113143243B (zh) * 2021-02-25 2023-08-22 厦门大学 一种基于谱分解的深度学习磁共振波谱重建方法
CN114417937A (zh) * 2022-01-26 2022-04-29 山东捷讯通信技术有限公司 一种基于深度学习的拉曼光谱去噪方法
CN115290761B (zh) * 2022-09-27 2023-01-17 中国科学院地质与地球物理研究所 基于深度学习的声发射数据重构方法及装置
CN115601611B (zh) * 2022-10-12 2023-05-09 武汉纺织大学 一种适应曝光变化的深度学习光谱重建方法及系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030078739A1 (en) * 2001-10-05 2003-04-24 Surromed, Inc. Feature list extraction from data sets such as spectra
WO2017113205A1 (zh) * 2015-12-30 2017-07-06 中国科学院深圳先进技术研究院 一种基于深度卷积神经网络的快速磁共振成像方法及装置
CN107749061A (zh) * 2017-09-11 2018-03-02 天津大学 基于改进的全卷积神经网络脑肿瘤图像分割方法及装置
CN108898568A (zh) * 2018-04-25 2018-11-27 西北大学 图像合成方法与装置
CN109903259A (zh) * 2019-01-25 2019-06-18 厦门大学 一种基于深度学习的磁共振波谱重建方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108535675B (zh) * 2018-04-08 2020-12-04 朱高杰 一种基于深度学习和数据自洽的磁共振多通道重建方法
CN108828481B (zh) * 2018-04-24 2021-01-22 朱高杰 一种基于深度学习和数据一致性的磁共振重建方法
CN108629784A (zh) * 2018-05-08 2018-10-09 上海嘉奥信息科技发展有限公司 一种基于深度学习的ct图像颅内血管分割方法及系统
US11346911B2 (en) * 2018-08-01 2022-05-31 Siemens Healthcare Gmbh Magnetic resonance fingerprinting image reconstruction and tissue parameter estimation

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030078739A1 (en) * 2001-10-05 2003-04-24 Surromed, Inc. Feature list extraction from data sets such as spectra
WO2017113205A1 (zh) * 2015-12-30 2017-07-06 中国科学院深圳先进技术研究院 一种基于深度卷积神经网络的快速磁共振成像方法及装置
CN107749061A (zh) * 2017-09-11 2018-03-02 天津大学 基于改进的全卷积神经网络脑肿瘤图像分割方法及装置
CN108898568A (zh) * 2018-04-25 2018-11-27 西北大学 图像合成方法与装置
CN109903259A (zh) * 2019-01-25 2019-06-18 厦门大学 一种基于深度学习的磁共振波谱重建方法

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112630709A (zh) * 2020-11-03 2021-04-09 成都易检医疗科技有限公司 金属植入物磁共振成像方法、系统、终端及存储介质
CN113361819A (zh) * 2021-07-08 2021-09-07 武汉中科牛津波谱技术有限公司 一种线性预测方法及装置
CN113361819B (zh) * 2021-07-08 2023-04-07 武汉中科牛津波谱技术有限公司 一种线性预测方法及装置

Also Published As

Publication number Publication date
CN109903259B (zh) 2020-05-29
US20210382127A1 (en) 2021-12-09
US11782111B2 (en) 2023-10-10
CN109903259A (zh) 2019-06-18

Similar Documents

Publication Publication Date Title
WO2020151355A1 (zh) 一种基于深度学习的磁共振波谱重建方法
Ying et al. Vandermonde factorization of Hankel matrix for complex exponential signal recovery—Application in fast NMR spectroscopy
CN110378980B (zh) 一种基于深度学习的多通道磁共振图像重建方法
Ying et al. Hankel matrix nuclear norm regularized tensor completion for $ n $-dimensional exponential signals
CN111324861B (zh) 一种基于矩阵分解的深度学习磁共振波谱重建方法
Yger et al. Wavelet kernel learning
Hou et al. Extracting a shape function for a signal with intra-wave frequency modulation
Huang et al. Exponential signal reconstruction with deep hankel matrix factorization
CN107423543B (zh) 一种超复数磁共振波谱的快速重建方法
CN110133554B (zh) 一种基于分数阶模型的磁共振指纹成像方法、装置及介质
CN112991483B (zh) 一种非局部低秩约束的自校准并行磁共振成像重构方法
Liu et al. DIIK-Net: A full-resolution cross-domain deep interaction convolutional neural network for MR image reconstruction
Cui et al. K-UNN: k-space interpolation with untrained neural network
CN110598579B (zh) 一种基于深度学习的超复数磁共振波谱重建方法
CN113109780B (zh) 基于复数密集连接神经网络的高分辨距离像目标识别方法
Fan et al. An interpretable MRI reconstruction network with two-grid-cycle correction and geometric prior distillation
Schwab et al. Global optimality in separable dictionary learning with applications to the analysis of diffusion MRI
CN113143243B (zh) 一种基于谱分解的深度学习磁共振波谱重建方法
Wang et al. DCT-net: Dual-domain cross-fusion transformer network for MRI reconstruction
Shi et al. Compressed sensing magnetic resonance imaging based on dictionary updating and block‐matching and three‐dimensional filtering regularisation
Wang et al. A nonlocal enhanced Low-Rank tensor approximation framework for 3D Magnetic Resonance image denoising
Lang et al. A fast Hankel matrix nonconvex factorization reconstruction method with inertia momentum for non-uniformly sampled NMR spectroscopy
Ma et al. Dynamic MRI exploiting partial separability and shift invariant discrete wavelet transform
Arun et al. Efficient directionality-driven dictionary learning for compressive sensing magnetic resonance imaging reconstruction
Ji et al. A new surrogate for tensor multirank and applications in image and video completion

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19911403

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19911403

Country of ref document: EP

Kind code of ref document: A1