Intelligent water signal removing method for magnetic resonance spectrum

Abstract

The invention discloses a magnetic resonance spectrum intelligent water signal removing method, and relates to a magnetic resonance spectrum water removing method. The method comprises the following steps of: 1) constructing simulation data by using a magnetic resonance spectrum signal model, wherein the simulation data comprises a magnetic resonance spectrum signal under a simulated non-ideal imaging condition, a corresponding metabolite component signal, a water signal and a background signal spectrum as a simulation training set; 2) designing a deep learning spectrum dehydration network and a loss function; 3) network training is carried out by using the generated training set, an optimizer with better performance in deep learning is adopted in training, and a network under the optimal network parameter set is obtained through a loss function; and 4) inputting the living body magnetic resonance spectrum signal into the dehydration network under the optimal parameter set to obtain a predicted no-water signal after water removal. In combination with the superior nonlinear learning ability of the neural network and the least square method based on the L2 norm, the water signal in the magnetic resonance spectrum signal can be quickly and accurately removed, the quantization error caused by the water signal is effectively reduced, and the accuracy of the subsequent spectrum quantization result is improved.

Description

Technical Field

[0001] The invention relates to a method for removing water from a magnetic resonance spectrum, in particular to a method for intelligently removing water signals from a magnetic resonance spectrum. Background Art

[0002] The proton magnetic resonance spectroscopy (MRS) signal of water is 3 to 5 orders of magnitude larger than the metabolite signal detectable in the brain (Dong, Zhengchao, Wolfgang Dreher, and DieterLeibfritz. "Toward quantitative short-echo-time in vivo proton MRspectroscopy without water suppression." Magnetic Resonance in Medicine , vol.55, pp. 672-679, 2006.) (Clayton, David B., Mark A. Elliott, and Robert E.Lenkinski. "In vivo proton spectroscopy without solvent suppression." Concepts in Magnetic Resonance , vol. 13, pp. 260-275, 2001.). Metabolites are present in such low concentrations of water in tissue that suppression of water signal is an important component of accurate quantitation by spectrometry. When water suppression fails or is inadequate, the accuracy of spectral content estimation will be greatly affected.

[0003] Suppressing water signals is a necessary step before quantification of metabolites based on spectroscopy. Problems with suppressing water signals include: (1) The measured MRS signal is close to or does not satisfy the ideal Lorentzian line shape, resulting in the SVD method based on water signal parameters (amplitude, frequency, phase and attenuation factor of the water signal) being inaccurate (Pijnappel, WWF, et al. "SVD-based quantification of magnetic resonance signals." Journal of Magnetic Resonance (1969) , vol.97, pp.122-134,1992.) (Vanhamme, Leentje, et al. "Fastremoval of residual water in proton spectra." Journal of Magnetic Resonance,vol. 132, pp. 197-203,1998.); (2) Filtering method of water signal (Marion, Dominique, MitsuhikoIkura, and Ad Bax. "Improved solvent suppression in one-and two-dimensionalNMR spectra by convolution of time-domain data." Journal of Magnetic Resonance (1969) , vol. 84, pp. 425-430, 1989.) (Sundin, Tomas, et al. "Accuratequantification of 1H spectra: from finite impulse response filter design for solvent suppression to parameter estimation." Journal of Magnetic Resonance , vol.139, pp.189-204, 1999.) (Coron, Alain, et al. "The filtering approach tosolvent peak suppression in MRS: a critical review." Journal of Magnetic Resonance ,vol.152, pp.26-40, 2001.) A large number of manual adjustment parameters are required (such as filter threshold and decomposition level); (3) When the spectral peak intensity and spectral peak line width of the water signal are too large, the dewatering method is likely to fail.

[0004] In recent years, some scholars have constructed a basis set of water signals and quickly solved the spectral signal after suppressing water signals through the second norm (Lin, Liangjie, et al. "Water removal in MR spectroscopic imagingwith L2 regularization." Magnetic resonance in medicine ,vol. 82, pp. 1278-1287, 2019.), but the accuracy of the constructed water signal basis set and the adjustment of the regularization parameters have a great influence on the water signal suppression effect. Summary of the invention

[0005] The purpose of the present invention is to provide a method for intelligently removing water signals from magnetic resonance spectroscopy, a method for fast solution of the least squares method based on the two-norm, combined with the nonlinear learning ability of the neural network to reduce the influence caused by the incompleteness of the constructed water signal basis set and the large sensitivity of the regularization parameter, effectively reduce the water signal estimation error, can quickly and accurately remove the water signal, and improve the subsequent spectral metabolite quantification accuracy.

[0006] In order to achieve the above-mentioned object of the invention, the specific technical scheme of the present invention is as follows:

[0007] A method for intelligently removing water signals by magnetic resonance spectroscopy comprises the following steps:

[0008] 1) Use the magnetic resonance spectroscopy signal model to construct simulation data, including the magnetic resonance spectroscopy signal under simulated non-ideal imaging conditions, the corresponding metabolite component signal, water signal and background signal spectrum as the simulation training set;

[0009] 2) Design a deep learning spectrum dehydration network and loss function;

[0010] 3) Use the training set generated in step 1) to train the network. Use the optimizer commonly used in deep learning (such as ADAM optimizer) to minimize the loss function in step 2). To obtain the optimal network parameter set The network below;

[0011] 4) In vivo magnetic resonance spectroscopy signals Enter the optimal network parameter set in step 3) The network under the prediction target water-free signal is obtained .

[0012] In step 1), the magnetic resonance spectroscopy signal model is used to construct simulation data, which includes the magnetic resonance spectroscopy signal under simulated non-ideal imaging conditions, the corresponding metabolite component signal, the water signal and the background signal spectrum as a simulation training set. The specific method is:

[0013] Magnetic resonance spectroscopy time domain signal It can be modeled as the linear sum of metabolite signal, water signal, background signal and noise, expressed as:

[0014] (1)

[0015] in, Indicates the length of the signal, Indicates the signal sampling points, represents the number of metabolites, Indicates metabolites, represents the sampling time interval of the signal, represents Gaussian white noise, Indicates that the parameter set The background signal modeled by the Gaussian model is generally composed of lipid macromolecules and other difficult-to-explain interference signals; water signal Through the parameter set The mixed model of the Lorentz line shape and the Gaussian line shape is composed of the specific formula:

[0016] (2)

[0017] in, represents an imaginary unit and satisfies , A set of parameters that describe the effect of the actual non-ideal imaging environment on the ideal MR spectroscopy water signal. It indicates the center frequency offset of the water signal caused by the inhomogeneous magnetic field. represents the water peak line broadening factor, represents the zero-order phase shift of water due to the non-ideal signal reception process, and Represent the signal strength of the Lorentzian and Gaussian line types of water signals, Indicates Metabolite signal, specifically expressed by the formula:

[0018] (3)

[0019] in, The parameter set that describes the effect of the actual non-ideal imaging environment on the metabolite signal in the ideal magnetic resonance spectrum is defined as ; The first parameter set Defined as , The inhomogeneity of the magnetic field causes The center frequency shift of each metabolite, Indicates the peak line broadening factor; the second parameter set Defined as , Indicates The concentration weighting factor of each metabolite; Indicates The ideal MR signal of each metabolite; It represents the multi-order phase shift introduced by non-ideal imaging conditions, which can be expressed by the formula:

[0020] (4)

[0021] in, represents a set of phase offsets of common J order, Indicates The phase shift of the order;

[0022] According to the time domain signal models (1) to (4), we take Get the corresponding signal sampling point , forming a vector , for the vector Perform Fourier transform to obtain the magnetic resonance spectrum signal under simulated non-ideal imaging conditions , the dimension is , as a column in the input data of the network training set; according to the time domain signal model (2), take Get the corresponding signal sampling point , forming a vector , for the vector Perform Fourier transform to obtain the water signal spectrum under simulated non-ideal imaging conditions , the dimension is , and used as a column in the label dataset 1 of the network training set;

[0023] Get the spectrum of the water-free signal , the dimension is , as a column in the label dataset 2 of the network training set.

[0024] In step 2), design the deep learning spectrum dehydration network and loss function:

[0025] The deep learning spectrum dehydration network can be divided into a water signal module and a water-free signal module:

[0026] a) Water signal module: used to predict water signals from input data; the function of the water signal module is expressed as:

[0027] (5)

[0028] in, Indicates the network parameters of the water signal module, Indicates the water signal module trained from arrive Nonlinear mapping; this module is composed of The first convolutional layer is the input layer, which consists of a The convolution layer consists of C1 convolution kernels, a batch normalization layer, and an activation layer; A convolution block consists of a convolution kernel with the number of C2, C 3,…, C Q of The convolution layer, a batch normalization layer, an activation layer and a one-fold mean downsampling layer; arrive The convolution block consists of a convolution kernel with the number of C Q+1 ,C Q+2,…, C H-1 of The convolutional layer, a batch normalization layer, an activation layer, and a one-fold zero-padding upsampling layer; The layer is the output layer, which consists of a The convolutional layer consists of , where the number of convolution kernels is 1; in particular, The convolution results of the layers are concatenated to The input of the layer; The output results of the first 1 / 2 of the total number of convolution kernels in the layer are taken as the first The input of the layer;

[0029] b) Anhydrous signal module: used to predict anhydrous signal from input data; the function of the anhydrous signal module is expressed as:

[0030] (6)

[0031] in, Indicates the network parameters of the waterless signal module, Represents the background signal module trained from arrive Nonlinear mapping; this module is composed of The stacked convolutional layers consist of: Layer 1 shared water signal module from 1 to layer; arrive The convolution block consists of a convolution kernel with the number of C Q+1 , C Q+2,…, C H-1 of The convolutional layer, a batch normalization layer, an activation layer, and a one-fold zero-padding upsampling layer; The layer is the output layer, which consists of a The number of convolutional layers is 1; in particular, The convolution results of the layers are concatenated to arrive The input of the layer; The output results of the last 1 / 2 of the total number of convolution kernels in the layer are taken as the first The input of the layer;

[0032] c) Overall network: The water signal module and the water-free signal module are integrated and described, and the entire network is expressed as:

[0033] (7)

[0034] in, represents the set of overall network parameters, Indicates the trained arrive The nonlinear mapping of represents the artificial water-based matrix, represents the regularization parameter;

[0035] The loss function of the network is defined as:

[0036] (8)

[0037] in, represents the set of overall network parameters, Represents the output of the water signal module, represents the spectrum of water signal under simulated non-ideal imaging conditions, represents the 2-norm of a vector, Indicates the output of the waterless signal module, represents the spectrum of the water-free signal, represents the identity matrix, represents the regularization parameter, represents the artificial water basis set matrix, represents the complex conjugate transpose operator, represents the matrix inversion, Regularization term designed for the two-norm based least squares method.

[0038] In step 3), the network is trained using the training set generated in step 1). The optimizer commonly used in deep learning (such as ADAM optimizer) is used in the training to minimize the loss function in step 2). To obtain the optimal network parameter set The network below.

[0039] In step 4), the in vivo magnetic resonance spectroscopy signal Enter the optimal parameter set in step 3) The quantized network under the above equation is used to obtain the predicted target water-free signal. .

[0040] The present invention combines the superior nonlinear learning ability of neural networks with the least squares method based on the L2 norm, which can quickly and accurately remove water signals from magnetic resonance spectroscopy signals, effectively reduce the quantization error caused by water signals, and improve the accuracy of subsequent spectral quantization results. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is the overall intelligent dehydration deep learning network structure in the embodiment of the present invention.

[0042] Figure 2 : This is a spectrum of the water-free signal measured in the embodiment of the present invention, wherein the horizontal axis is ppm, which is the unit of chemical shift; and the vertical axis is au, which is the signal intensity. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the following embodiments will further illustrate the present invention in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0044] According to the characteristics of magnetic resonance spectroscopy signals and the principle of PRESS sequence, the embodiment of the present invention generates non-ideal simulated magnetic resonance spectra based on quantum simulation, uses simulated metabolite spectra, background signals and water signal spectra as training sets, and obtains the optimal network model through iterative training. Finally, the actually collected in vivo magnetic resonance spectra are input into the network to predict the corresponding target water-free signal spectrum.

[0045] The embodiment of the present invention includes the following steps:

[0046] 1) Use the magnetic resonance spectroscopy signal model to construct simulation data, including the magnetic resonance spectroscopy signal under non-ideal imaging conditions, the corresponding metabolite component signal, water signal and background signal spectrum as the simulation training set. The specific method is as follows:

[0047] Magnetic resonance spectroscopy time domain signal It can be modeled as a synthesis of metabolite component signals, water signals, background signals, and noise, expressed as follows:

[0048] (1)

[0049] in, Indicates the length of the signal (in this embodiment ), Indicates the signal sampling time interval (in this embodiment ), represents Gaussian white noise (in this embodiment, the Gaussian noise signal-to-noise ratio is 30), It indicates that due to lipid macromolecules and other background signals that are difficult to explain, 10 random Gaussian mixture models are used for simulation in this embodiment. In this embodiment, one Lorentz and one Gaussian function are used for simulation, which is specifically expressed by the formula:

[0050] (2)

[0051] in, represents an imaginary unit and satisfies . A set of parameters to describe the effect of the actual non-ideal imaging environment on the ideal MR spectroscopy water signal , It indicates the center frequency offset of the water signal caused by the inhomogeneous magnetic field. represents the water peak line broadening factor, Represents the zero-order phase shift of water due to the non-ideal signal reception process. Respectively represent the signal strength of the Lorentz and Gaussian components of the water signal. The value range is [-10Hz, +10Hz], The value range is [20Hz,40Hz], The value range is [-45°, 120°], The value range of is [30, 180]. Indicates The metabolite component signal is specifically expressed by the formula:

[0052] (3)

[0053] in, The parameter set that describes the effect of the actual non-ideal imaging environment on the ideal magnetic resonance spectroscopy signal is defined as ; The first parameter set Defined as , subscript Indicates target metabolites, Indicates the total number of metabolites. , including alanine (Ala), aspartic acid (Asp), creatine (Cr), γ-aminobutyric acid (GABA), glucose (Glc), glutamine (Gln), glutamate (Glu), glycerolphosphatecholine (GPC), glutathione (GSH), lactic acid (Lac), myo-Inositol (mI), N-acetylaspartate (NAA), N-acetylaspartylglutamate (NAAG), phosphocholine (PCh), phosphocreatine (PCr) and taurine (Tau). This method is not only applicable to the above metabolites. The inhomogeneity of the magnetic field causes The center frequency shift of each metabolite, Indicates the peak line broadening factor; the second parameter set Defined as , Indicates The concentration weighting factor of each metabolite; Indicates The ideal magnetic resonance spectroscopy signal of each metabolite. The value range is [-10Hz, +10Hz], The value range is [5Hz,20Hz]. It represents the multi-order phase shift caused by the non-ideal signal reception process, which is specifically expressed by the formula:

[0054] (4)

[0055] in, express In this embodiment, Simulating zero-order phase , the value range is [-30°,+30°].

[0056] According to the time domain signal model, equations (1) to (4), take Get the corresponding signal sampling point , forming a vector , perform Fourier transform on the vector to obtain the magnetic resonance spectrum signal under simulated non-ideal imaging conditions , the dimension is , as the input data of the network training set.

[0057] According to the time domain signal model (Equation (2)), Get the corresponding signal sampling point , forming a vector , perform Fourier transform on the vector to obtain the water signal spectrum under simulated non-ideal imaging conditions , the dimension is , and used as the label data 1 of the network training set.

[0058] according to , forming a vector , perform Fourier transform on the vector and obtain the water-free signal spectrum , the dimension is , as the label data 2 of the network training set.

[0059] In this embodiment, a network training set {input data, {label data 1, label data 2}} is generated, with a total of 60,000 copies.

[0060] 2) Design deep learning spectral quantization network and loss function;

[0061] The deep learning spectral quantization network can be divided into two modules:

[0062] a) Water signal module: used to predict water signals from input data. The function of the water signal module is expressed as:

[0063] (5)

[0064] in, Indicates the network parameters of the water signal module, Indicates the water signal module trained from arrive This module is composed of 7 stacked convolutional layers: the first convolutional layer is the input layer, which consists of a The 2nd, 3rd and 4th convolution blocks consist of a convolution layer (the number of convolution kernels is 64), a BN layer and a ReLU (Rectified Linear Unit, ReLU) layer. The 5th and 6th convolution blocks consist of a convolution layer (the number of convolution kernels is 128, 256 and 512), a BN layer, a ReLu layer and a one-fold mean downsampling layer. The seventh layer is the output layer, which consists of a convolution layer (the number of convolution kernels is 256 and 128), a BN layer, a ReLu layer, and a one-time zero-padding upsampling layer. The convolutional layers (the number of convolution kernels is 1) are composed of . In particular, the convolution results of layers 2-4 are concatenated to the inputs of layers 5-7 respectively. The output results of the first 256 convolution kernels of layer 4 are used as the input of layer 5.

[0065] b) Anhydrous signal module: used to predict anhydrous signal from input data. The function of the anhydrous signal module is expressed as:

[0066] (6)

[0067] in, Indicates the network parameters of the waterless signal module, Represents the background signal module trained from arrive This module is composed of 7 stacked convolutional layers: the 1st to 4th layers share the 1st to 4th layers of the water signal module. The 5th and 6th convolutional blocks are composed of a The seventh layer is the output layer, which consists of a convolution layer (the number of convolution kernels is 256 and 128), a BN layer, a ReLu layer, and a one-time zero-padding upsampling layer. The convolutional layers (the number of convolution kernels is 1) are composed of . In particular, the convolution results of layers 2-4 are concatenated to the inputs of layers 5-7 respectively. The output results of the last 256 convolution kernels of layer 4 are used as the input of layer 5.

[0068] c) Overall network: The water signal module and the water-free signal module are integrated and described, and the entire network is expressed as:

[0069] (7)

[0070] in, represents the set of overall network parameters, Indicates the trained arrive The nonlinear mapping of represents an artificial water-based matrix. In this embodiment, It is a matrix composed of 1000 water signals randomly generated according to formula (2) in a column-by-column arrangement. represents the regularization parameter (100 in this embodiment), and the overall intelligent water removal network structure is as follows: Figure 1 shown.

[0071] The loss function of the network is defined as:

[0072] (8)

[0073] in, represents the set of overall network parameters, Represents the output of the water signal module, represents the spectrum of water signal under simulated non-ideal imaging conditions, represents the 2-norm of a vector, Indicates the output of the waterless signal module, represents the spectrum of the water-free signal, represents the identity matrix, represents the complex conjugate transpose operator, represents the matrix inversion, Regularization term designed for the least squares method based on the L2 norm.

[0074] 3) Use the training set generated in step 1) to train the network. Use common optimizers in deep learning (such as ADAM optimizer) to minimize the loss function in step 2). To obtain the optimal network parameter set The network below.

[0075] 4) In vivo magnetic resonance spectroscopy signals Enter the optimal parameter set in step 3) The quantized network under the above equation is used to obtain the predicted target water-free signal. .

[0076] In this embodiment, a simulated human brain magnetic resonance spectrum is used as the input of the network. , the dimension is 512×1, and the water-free signal result obtained by network prediction is as follows Figure 2 It can be seen that the error of the present invention is small and the result is highly accurate.

[0077] The above embodiments are only preferred embodiments of the present invention and cannot be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A method for intelligently removing water signals from magnetic resonance spectroscopy, characterized in that The following steps are involved: 1) Using the magnetic resonance spectroscopy signal model to construct a training set, the training set includes magnetic resonance spectroscopy signals, metabolite signals, water signals and background signals under simulated non-ideal imaging conditions as a simulation training set; 2) Design of deep learning network and loss function for spectral dehydration ,in is the set of parameters that the network can learn; 3) Use the training set generated in step 1) to train the network. Use the common optimizer in deep learning to minimize the loss function in step 2). Get the optimal network parameter set The network below; 4) In vivo magnetic resonance spectroscopy signals Enter the optimal network parameter set in step 3) The network under the prediction of no water signal .

2. A method for intelligently removing water signals from magnetic resonance spectroscopy as claimed in claim 1, characterized in that In step 1), the specific method of constructing a training set using the magnetic resonance spectroscopy signal model is: Magnetic resonance spectroscopy time domain signal It is modeled as the linear sum of metabolite signal, water signal, background signal and noise, expressed as: （1） in, Indicates the length of the signal, The signal sampling points, represents the number of metabolites, Indicates metabolites, represents the sampling time interval of the signal, represents Gaussian white noise, Indicates that the parameter set The background signal modeled by the Gaussian model is generally composed of lipid macromolecules and other difficult-to-explain interference signals; water signal Through the parameter set The mixed model of the Lorentz line shape and the Gaussian line shape is composed of the specific formula: （2） in, represents an imaginary unit and satisfies , A set of parameters that describe the effect of the actual non-ideal imaging environment on the ideal MR spectroscopy water signal. It indicates the center frequency offset of the water signal caused by the inhomogeneous magnetic field. represents the water peak line broadening factor, represents the zero-order phase shift of water due to the non-ideal signal reception process, and Represent the signal strength of the Lorentzian and Gaussian line types of water signals, Indicates Metabolite signal, specifically expressed by the formula: （3） in, The parameter set that describes the effect of the actual non-ideal imaging environment on the metabolite signal in the ideal magnetic resonance spectrum is defined as ; The first parameter set Defined as , Indicates the first The center frequency shift of each metabolite, Indicates the peak line broadening factor; the second parameter set Defined as , Indicates The concentration weighting factor of each metabolite; Indicates The ideal magnetic resonance signal of each metabolite; It represents the multi-order phase shift introduced by non-ideal imaging conditions, which can be expressed by the formula: （4） in, represents a set of phase offsets of common J order, Indicates The phase shift of the order; According to the time domain signal models (1) to (4), we take Get the corresponding signal sampling point , forming a vector , for the vector Perform Fourier transform to obtain the magnetic resonance spectrum signal under simulated non-ideal imaging conditions , the dimension is , as a column in the input data of the network training set; according to the time domain signal model (2), take Get the corresponding signal sampling point , forming a vector , for the vector Perform Fourier transform to obtain the water signal spectrum under simulated non-ideal imaging conditions , the dimension is , and used as a column in the label dataset 1 of the network training set; Get the spectrum of the water-free signal , the dimension is , as a column in the label dataset 2 of the network training set.

3. A method for intelligently removing water signals from magnetic resonance spectroscopy according to claim 1, characterized in that In step 2), the deep learning network and loss function for spectral dehydration are designed. ,Among them, the deep learning network for spectrum dehydration is divided into two modules: water signal module and water-free signal module; a) Water signal module: used to predict water signals from input data; the function of the water signal module is expressed as: （5） in, Indicates the network parameters of the water signal module, Indicates the water signal module trained from arrive Nonlinear mapping; this module is composed of The first convolutional layer is the input layer, which consists of a The convolution layer consists of C1 convolution kernels, a batch normalization layer, and an activation layer; A convolution block consists of a convolution kernel with the number of C2, C 3,…, C Q of The convolution layer, a batch normalization layer, an activation layer and a one-fold mean downsampling layer; arrive The convolution block consists of a convolution kernel with the number of C Q+1 ,C Q+2,…, C H-1 of The convolutional layer, a batch normalization layer, an activation layer, and a one-fold zero-padding upsampling layer; The layer is the output layer, which consists of a The convolutional layer consists of , where the number of convolution kernels is 1; in particular, The convolution results of the layers are concatenated to The input of the layer; The output results of the first 1 / 2 of the total number of convolution kernels in the layer are taken as the first The input of the layer; b) Anhydrous signal module: used to predict anhydrous signal from input data; the function of the anhydrous signal module is expressed as: （6） in, Indicates the network parameters of the waterless signal module, Represents the background signal module trained from arrive Nonlinear mapping; this module is composed of The stacked convolutional layers consist of: Layer 1 to 2 shared water signal modules layer; arrive The convolution block consists of a convolution kernel with the number of C Q+1 , C Q+2,…, C H-1 of The convolutional layer, a batch normalization layer, an activation layer, and a one-fold zero-padding upsampling layer; The layer is the output layer, which consists of a The number of convolutional layers is 1; in particular, The convolution results of the layers are concatenated to arrive The input of the layer; The output results of the last 1 / 2 of the total number of convolution kernels in the layer are taken as the first The input of the layer; c) Overall network: The water signal module and the water-free signal module are integrated and described, and the entire network is expressed as: （7） in, represents the set of overall network parameters, Indicates the trained arrive The nonlinear mapping of represents the artificial water-based matrix, represents the regularization parameter; The loss function of the network is defined as: （8） in, represents the set of overall network parameters, Represents the output of the water signal module, represents the spectrum of water signal under simulated non-ideal imaging conditions, represents the 2-norm of a vector, Indicates the output of the waterless signal module, represents the spectrum of the water-free signal, represents the identity matrix, represents the regularization parameter, represents the artificial water basis set matrix, represents the complex conjugate transpose operator, represents the matrix inversion, Regularization term designed for the two-norm based least squares method.

4. A method for intelligently removing water signals from magnetic resonance spectroscopy as claimed in claim 1, characterized in that In step 3), the common optimizer adopts ADAM optimizer.

Images