Method and device for improving signal-to-noise ratio of nuclear magnetic resonance spectrum based on deep iterative shrinkage threshold network

Abstract

The invention relates to the technical field of nuclear magnetic resonance, in particular to a method and device for improving the signal-to-noise ratio of a nuclear magnetic resonance spectrum based on a deep iterative shrinkage threshold network. The method comprises the following steps: step 1, acquiring a low signal-to-noise ratio nuclear magnetic resonance spectrum; and step 2, inputting the low-signal-to-noise-ratio nuclear magnetic resonance spectrum into the trained deep iterative shrinkage threshold network for processing to obtain a high-signal-to-noise-ratio nuclear magnetic resonance spectrum. According to the method, a deep iterative shrinkage threshold network is provided, an iterative shrinkage threshold algorithm is expanded into a deep neural network form, the structural advantages of a traditional optimization algorithm and the calculation efficiency of a deep learning network are combined, the signal-to-noise ratio of the nuclear magnetic resonance spectrum is improved, and the noise reduction performance of the nuclear magnetic resonance spectrum can be improved while the excellent spectrum denoising performance is kept. And the network training speed and the interpretability of the training process are improved. According to the method, the problems of high time and economic cost, complicated operation and unsatisfactory effect of the traditional method for improving the signal-to-noise ratio of the NMR spectrogram are solved.

Description

Technical Field

[0001] The present invention relates to the field of nuclear magnetic resonance technology, and more specifically, to: 1. A method for improving the signal-to-noise ratio of nuclear magnetic resonance spectrum based on a deep iterative shrinkage threshold network; 2. A device for improving the signal-to-noise ratio of nuclear magnetic resonance spectrum based on a deep iterative shrinkage threshold network. Background Art

[0002] Nuclear magnetic resonance (NMR) technology is an effective method to analyze the molecular structure and chemical environment of substances by detecting the resonance phenomenon of nuclear magnetic moments in an external magnetic field. NMR spectra provide rich chemical and physical information. The signal-to-noise ratio (SNR) is an important indicator for evaluating the quality of NMR spectra, which represents the ratio of signal intensity to noise intensity. A higher SNR means a clearer spectrum, which helps to improve the accuracy of molecular structure analysis.

[0003] In NMR spectra, signals usually appear as peaks in the frequency domain, which correspond to the resonance frequencies of different nuclei and their chemical environments, while noise appears as random fluctuations on the baseline. In order to obtain high-quality NMR spectra, it is necessary to maximize the ratio of signal to noise, that is, to improve the signal-to-noise ratio. High SNR not only provides clearer spectra, but also more accurately determines the position, intensity and shape of peaks, which is crucial for quantitative analysis and molecular structure analysis. Spectra with low SNR may cause signal peaks to be masked by noise, resulting in the loss or misinterpretation of important chemical information. In addition, high SNR helps to shorten experimental time. In NMR experiments, multiple scans are usually required to average random noise and thus improve SNR. Each scan time ranges from a few minutes to a few hours. When the SNR is high enough, the number of scans can be reduced, thereby shortening the experimental time, which is particularly important for high-throughput analysis or real-time monitoring experiments.

[0004] In recent years, domestic and foreign scholars have conducted a lot of research on improving the signal-to-noise ratio of NMR spectra and proposed a variety of methods and techniques. For example, using higher strength magnets to enhance the intensity of NMR signals, thereby increasing the signal-to-noise ratio; improving the signal-to-noise ratio by optimizing the parameters of the pulse sequence, such as relaxation time, duty cycle, pulse sequence time, etc. Although these traditional methods have made significant progress in improving the signal-to-noise ratio of NMR spectra, hardware upgrades are usually accompanied by high costs and complex operating requirements, while data acquisition optimization methods may require long experimental times and complex parameter settings. In addition, traditional methods are still limited in their effectiveness when dealing with complex and dynamic systems. Summary of the invention

[0005] Based on this, it is necessary to provide a method and device for improving the signal-to-noise ratio of nuclear magnetic resonance spectra based on a deep iterative shrinkage threshold network to address the problems of high time and economic costs, complex operations and unsatisfactory results of traditional methods for improving the signal-to-noise ratio of NMR spectra.

[0006] The present invention is implemented by the following technical solutions:

[0007] In a first aspect, the present invention discloses a method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network, comprising the following steps:

[0008] Step 1, obtaining a low signal-to-noise ratio nuclear magnetic resonance spectrum;

[0009] Step 2: Input the low signal-to-noise ratio NMR spectrum into the trained deep iterative shrinkage threshold network for processing to obtain a high signal-to-noise ratio NMR spectrum.

[0010] Among them, the deep iterative shrinkage threshold network includes: a data initialization layer, an iterative feature part, and a spectrum output layer.

[0011] The data initialization layer is used to process the low signal-to-noise ratio nuclear magnetic resonance spectrum to obtain the initial feature map X_input. The iterative feature part includes: N sequentially connected iterative feature layers, which are used to perform N feature iterations on X_input to obtain the output feature map X_output of the Nth iteration N The spectral output layer is used to N A convolution operation is performed to obtain a high signal-to-noise ratio NMR spectrum.

[0012] Among them, all convolutions in the deep iterative shrinkage threshold network are one-dimensional convolutions.

[0013] This method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network implements the method or process according to an embodiment of the present disclosure.

[0014] In the second aspect, the present invention discloses a device for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network, which uses the method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network disclosed in the first aspect.

[0015] The device for improving the signal-to-noise ratio of nuclear magnetic resonance spectrum based on deep iterative shrinkage threshold network includes: a spectrum acquisition module and a spectrum processing module.

[0016] The spectrum acquisition module is used to obtain a low signal-to-noise ratio nuclear magnetic resonance spectrum. The spectrum processing module is used to input the low signal-to-noise ratio nuclear magnetic resonance spectrum into the trained deep iterative shrinkage threshold network for processing to obtain a high signal-to-noise ratio nuclear magnetic resonance spectrum.

[0017] The device for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network implements the method or process according to an embodiment of the present disclosure.

[0018] In a third aspect, the present invention discloses a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network disclosed in the first aspect are implemented.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] (1) The present invention provides a deep iterative shrinkage threshold network, which expands the iterative shrinkage threshold algorithm into the form of a deep neural network, combining the structural advantages of traditional optimization algorithms and the computational efficiency of deep learning networks, and has the advantages of high speed and high precision. Compared with traditional optimization methods, the present invention uses a deep iterative shrinkage threshold network to process low signal-to-noise ratio nuclear magnetic resonance spectra, which can achieve higher reconstruction accuracy while retaining high computing speed.

[0021] (2) The present invention combines the advantages of the traditional iterative shrinkage threshold denoising algorithm, can flexibly adjust the number of iterations, and improves the network training speed and the interpretability of the iterative denoising process while retaining excellent spectral denoising performance.

[0022] (3) The present invention integrates deep learning, so that the key parameters of the network (such as shrinkage threshold, step size, etc.) are automatically optimized through end-to-end learning, so that the network can adaptively process different types of data, significantly improving the generalization ability and reconstruction effect of the network, and ensuring high performance in a variety of application scenarios.

[0023] (4) The present invention is based on multiple iterations and threshold shrinkage processing, exhibits excellent robustness, can effectively remove noise, and improve signal clarity and reliability; therefore, even in a noisy environment, the present invention can still obtain high-quality, high signal-to-noise ratio spectral signals.

[0024] (5) The present invention implements multiple rounds of feature iterations in multiple iterative feature layers to ensure the stability and consistency of the signal; compared with traditional methods, the present invention can significantly improve the signal-to-noise ratio of the signal, retain the details of the original signal, and has the advantages of significant noise reduction effect and high processing efficiency. The present invention has broad application prospects, for example, in the fields of medical imaging, chemical analysis, and materials science, and can provide strong support for the acquisition of high-quality nuclear magnetic resonance spectrum data. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0026] Figure 1 A flow chart of a method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network provided in Example 1 of the present invention;

[0027] Figure 2 The overall structure diagram of the deep iterative shrinkage threshold network provided in Example 1 of the present invention;

[0028] Figure 3 for Figure 2 The structure diagram of the data initialization layer in the figure;

[0029] Figure 4 for Figure 2 The structure diagram of the nth iterative feature layer in ;

[0030] Figure 5 for Figure 2 Schematic diagram of the calculation of the loss function of the deep iterative shrinkage threshold network during training;

[0031] Figure 6 This is a graph of the azithromycin test results provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] It should be noted that when a component is referred to as being "mounted on" another component, it may be directly on the other component or there may be a central component. When a component is considered to be "set on" another component, it may be directly set on the other component or there may be a central component at the same time. When a component is considered to be "fixed to" another component, it may be directly fixed on the other component or there may be a central component at the same time.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "or / and" used herein includes any and all combinations of one or more of the related listed items.

[0035] Example 1

[0036] See also Figure 1 , Figure 1 The brief process of improving the signal-to-noise ratio of NMR spectroscopy based on a deep iterative shrinkage threshold network is presented. Figure 1 As shown, the method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network mainly includes the following steps:

[0037] Step 1, obtaining a low signal-to-noise ratio nuclear magnetic resonance spectrum;

[0038] Step 2: Input the low signal-to-noise ratio NMR spectrum into the trained deep iterative shrinkage threshold network for processing to obtain a high signal-to-noise ratio NMR spectrum.

[0039] The method relies on a deep iterative shrinking threshold network - a form of the iterative shrinking threshold algorithm unfolded into a deep neural network.

[0040] See also Figure 2 , showing the overall structure of the deep iterative shrinkage threshold network, which includes: data initialization layer, iterative feature part, and spectral output layer.

[0041] 1. The data initialization layer is used to process the low signal-to-noise ratio nuclear magnetic resonance spectrum to obtain the initial feature map X_input.

[0042] like Figure 3 As shown, the data initialization layer can be designed to include: a data mapping part, a matrix generation part, and a product part.

[0043] The data mapping unit is used to map the low signal-to-noise ratio nuclear magnetic resonance spectrum through the sensing matrix Φ to obtain the observation matrix Φ x Φ x =Y input Φ; where Y input represents a low signal-to-noise ratio NMR spectrum; Φ is an M×M identity matrix; M represents the ideal noise-free dataset Y used for deep iterative shrinkage threshold network training ideal The total number of data points. It should be noted that Y input The total number of data points is also M.

[0044] The matrix generation unit is used to provide an initialization matrix Q; Q = Yideal (ΦY ideal ) T ((ΦY ideal )(ΦY ideal ) T ) -1 .

[0045] The product part is used to convert Φ x and the transpose of Q T Perform product calculation to obtain the initial feature map X_input. In other words, X_input = Φ x Q T .

[0046] 2. The iterative feature part includes: N sequentially connected iterative feature layers, which are used to perform N feature iterations on X_input to obtain the output feature map X_output of the Nth iteration N .

[0047] The value of N can be adjusted according to actual conditions. In this embodiment 1, when N is 9, the network effect and the operating load can achieve a good balance.

[0048] Since the structure and processing of iterative feature layers are similar, the nth (n∈[1,N]) iterative feature layer is used as an example to illustrate:

[0049] See also Figure 4 , the nth iterative feature layer can be designed to include: feature update layer a n , feature extraction layer b n , forward convolution layer c n , soft threshold processing layer d n , reverse convolution layer e n ; It can also include: output convolution layer f n .

[0050] 201. Feature update layer a n Used to use gradient descent method to n Update and generate the updated feature map X_gum n .

[0051] Among them, X_input n Represents the input of the nth iterative feature layer.

[0052] When n=1, X_input n is X_input; when n>1, X_input n The output feature map X_output of the n-1th iteration n-1 .

[0053] In the feature update layer an In, X_gum n The calculation formula is:

[0054] X_gum n =X_input n -λ step (Φ T Φ*X_input n );

[0055] In the formula, λ step Represents the network learnable parameters.

[0056] In this embodiment 1, λ step The initial value of is 0.5, which is used to control the step size of the gradient descent update.

[0057] 202. Feature extraction layer b n For X_gum n Perform feature extraction and generate the extracted feature map X_fem n In the feature extraction layer b n In X_fem n The calculation formula is:

[0058] X_fem n =X_gum n +λ step (Φ T Φ).

[0059] 203. Forward convolution layer c n For X_fem n Perform forward convolution processing to obtain the forward feature map X_fcm n .

[0060] Among them, c n It includes: L one-dimensional convolutional layers with kernel k, stride j, padding p, and Y ReLU activation functions.

[0061] In this embodiment 1, L is 2, k is 25, j is 1, p is 12, and Y is 1; that is, c n It consists of two one-dimensional convolutional layers with kernel size 25, stride size 1, padding 12, and one ReLU activation function. Then, X_fem n First, the first one-dimensional convolutional layer extracts preliminary features, then the ReLU activation function is used for nonlinear processing, and then the second one-dimensional convolutional layer further extracts and enhances features to obtain the forward feature map X_fcm n .

[0062] 204. Soft threshold processing layer d n For X_fcmn Perform soft threshold processing to obtain the soft threshold feature map X_stm n .

[0063] In the soft thresholding layer d n In X_stm n The calculation formula is:

[0064] X_stm n =sign(X_fcm n )*max((X_fcm n |-μ), 0);

[0065] In the formula, sign(.) represents the sign function; max(.) represents the maximum value; μ represents the soft threshold parameter.

[0066] In this embodiment 1, μ is set to 0.1 to retain the main features of the data.

[0067] 205. Reverse convolution layer e n For X_stm n Perform reverse convolution to obtain the output feature map X_output of the nth iteration n .

[0068] e n With c n The structure is the same, including: L one-dimensional convolutional layers with kernel k, stride j, padding p, and Y ReLU activation functions.

[0069] In this embodiment 1, e n It consists of two one-dimensional convolutional layers with kernel size 25, stride size 1, padding 12, and one ReLU activation function. Then, X_stm n First, it is reconstructed by the first one-dimensional convolution layer, then nonlinearly processed by the ReLU activation function, and then further restored and reconstructed by the second one-dimensional convolution layer, so as to obtain the output feature map X_output of the nth iteration. n .

[0070] 206. Output convolution layer f n Used to convert X_output to n The output convolution is reshaped to calculate the symmetric loss loss_layer_sym of the nth iteration n .

[0071] Among them, loss_layer_sym n The calculation formula is:

[0072]

[0073] In the formula, X_est n The mth data point; X_est n X_output n The feature map generated by reshaping the output convolution; X_fem n The mth data point.

[0074] f n With c n 、e n The structure is the same, including: L one-dimensional convolutional layers with kernel k, stride j, padding p, and Y ReLU activation functions.

[0075] In this embodiment 1, f n It consists of two one-dimensional convolutional layers with kernel size 25, stride size 1, padding 12, and one ReLU activation function. Then, X_output n First, the shape is adjusted by the first one-dimensional convolution layer, then nonlinear processing is performed through the ReLU activation function, and then further shape adjustment is performed through the second one-dimensional convolution layer to obtain X_est n .

[0076] It should be noted that c n 、e n 、f n Although the structures are the same, the input and output channels of the three are different, so the corresponding operations performed by the three are also different.

[0077] 3. The spectrum output layer is used to X_output N A convolution operation is performed to obtain a high signal-to-noise ratio NMR spectrum.

[0078] The above is an introduction to the structure and function of the deep iterative shrinkage threshold network.

[0079] This method requires the use of a trained deep iterative shrinkage threshold network. The following is an introduction to the training method of the deep iterative shrinkage threshold network, which includes:

[0080] S1, obtain the ideal noise-free data set Y ideal ;

[0081] in, represents the ideal noise-free data for the mth data point.

[0082] Specifically, Y idealIt can be a simulation data set generated by MATLAB, considering only the effects of chemical shift and relaxation, and its calculation expression can be written as:

[0083]

[0084] Where A represents the amplitude; t2 represents the direct dimensional sampling time; f represents the frequency; T2 represents the relaxation time, and i represents the imaginary part.

[0085] In this embodiment 1, Y ideal The total number of data points is 8192, that is, M is 8192.

[0086] S2, for Y ideal Perform fast Fourier transform (FFT) to obtain spectrum data, calculate the real part of the spectrum data, and normalize it by taking the maximum value of the real part to obtain the normalized spectrum data set Y norm ;

[0087] Performing an inverse Fourier transform (IFFT) on the normalized spectrum data set to obtain a time domain signal, and adding complex noise with random characteristics to the time domain signal to obtain a noisy time domain signal set;

[0088] Among them, complex noise can use white noise, and its mathematical expression is:

[0089] Noise=r*[randn(1,f n )+i*randn(1,f n )];

[0090] Where Nosie represents white noise; r represents the number of sampling points within the noise intensity range; f n represents the number of direct dimension sampling points of the noise; i represents the imaginary part; randn(1,f n ) represents the generation of a white noise sequence of length fn, which is used to simulate the random noise characteristics in the environment.

[0091] Perform fast Fourier transform on the noisy time domain signal set to obtain the noisy spectrum data set Y noise ; Represents the spectrum data of the mth data point after adding noise.

[0092] S3, based on Y noise , Y ideal Training the deep iterative shrinkage threshold network to obtain a trained deep iterative shrinkage threshold network;

[0093] Among them, Y noise As the input of the deep iterative shrinkage threshold network during training; Y idealas a training goal.

[0094] In S3, the training is iterated for a set number of rounds W, and the network parameters are saved at the end of each training cycle until a trained deep iterative shrinkage threshold network is obtained.

[0095] In this embodiment 1, some training parameters are set as follows:

[0096] The training learning rate L is 0.0001; the batch size of each training is 32, and the Adam optimization algorithm is used; the number of training rounds W is 300.

[0097] See also Figure 5 In each round of iterative training, the network parameters are adjusted by taking the weighted sum of the difference loss and the constraint loss as the loss function.

[0098] Specifically, the loss function used by the deep iterative shrinkage threshold network during training is loss total for:

[0099] loss total =loss discrepancy +γloss constraint ;

[0100] In the formula, loss discrepancy Represents the difference loss of the final output; loss constraint Represents the constraint loss of the iterative feature part; γ represents the weighting constant; generally, γ is 0.01.

[0101] Among them, loss discrepancy The calculation formula is:

[0102]

[0103] In the formula, Output N Output N express The final output after input into the deep iterative shrinkage threshold network;

[0104] loss constraint The calculation formula is:

[0105]

[0106] Where, loss_layer_sym n It represents the symmetric loss of the nth iteration. Its calculation method can be found in the above 206 section and will not be repeated here.

[0107] The above method can be designed as a cloud computing service when applied in practice: users only need to upload low signal-to-noise ratio NMR spectra to the cloud, and the cloud server runs the steps of the method and returns the processed high signal-to-noise ratio NMR spectra. This method is suitable for large-scale data processing and remote data processing, and is particularly helpful for scientific research projects and clinical studies that require processing of large amounts of NMR data.

[0108] When applied specifically, the above method can also be embedded as a data processing module into a variety of analysis systems, such as drug analysis systems in the pharmaceutical industry, pollution detection systems in environmental science, material analysis systems in bioinformatics, smart home devices, etc. By improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy data, it can more accurately analyze and detect trace components in samples and improve the reliability of test results.

[0109] Simulation Verification

[0110] See also Figure 6 , this embodiment 1 also tests and verifies the above method:

[0111] 0.2 M azithromycin dissolved in CDCL3 (deuterated chloroform) 13 C was selected as the spectrum acquisition object, and the acquisition instrument was a Varian 500MHz nuclear magnetic resonance spectrometer.

[0112] Figure 6 (a) is the spectrum of azithromycin under 4 repeated scans, which is used as a low signal-to-noise ratio nuclear magnetic resonance spectrum;

[0113] Figure 6 (b) is the spectrum of azithromycin at 512 repeated scans, which is an ideal noise-free NMR spectrum;

[0114] Figure 6 (c) To use the above method Figure 6 (a) is a high signal-to-noise ratio NMR spectrum obtained after processing the spectrum.

[0115] pass Figure 6 From the comparison of the three spectra, it can be seen that the method provided by the present invention can effectively suppress noise and pseudo-peak signals when processing low signal-to-noise ratio nuclear magnetic resonance spectra, and successfully restore weak peaks that are almost completely submerged by severe noise, thereby achieving a significant improvement in the signal-to-noise ratio.

[0116] Example 2

[0117] This embodiment 2 provides a device for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network, which uses the method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network disclosed in embodiment 1.

[0118] The device for improving the signal-to-noise ratio of nuclear magnetic resonance spectrum based on deep iterative shrinkage threshold network includes: a spectrum acquisition module and a spectrum processing module.

[0119] The spectrum acquisition module is used to obtain a low signal-to-noise ratio nuclear magnetic resonance spectrum. The spectrum processing module is used to input the low signal-to-noise ratio nuclear magnetic resonance spectrum into the trained deep iterative shrinkage threshold network for processing to obtain a high signal-to-noise ratio nuclear magnetic resonance spectrum.

[0120] Since the present device uses the method of improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network in Example 1, it also has the same effect and will not be repeated here.

[0121] Example 3

[0122] This embodiment 3 discloses a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network disclosed in embodiment 1 are implemented.

[0123] This embodiment 3 also discloses a readable storage medium, in which computer program instructions are stored. When the computer program instructions are read and executed by a processor, the steps of the method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network disclosed in embodiment 1 are executed.

[0124] This embodiment 3 also discloses a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network disclosed in embodiment 1 are implemented.

[0125] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network. It is characterized in that The following steps are involved: Step 1, obtaining a low signal-to-noise ratio nuclear magnetic resonance spectrum; Step 2, inputting the low signal-to-noise ratio nuclear magnetic resonance spectrum into the trained deep iterative shrinkage threshold network for processing to obtain a high signal-to-noise ratio nuclear magnetic resonance spectrum; Among them, the deep iterative shrinkage threshold network includes: The data initialization layer is used to process the low signal-to-noise ratio nuclear magnetic resonance spectrum to obtain the initial feature map X_input; The iterative feature part includes: N sequentially connected iterative feature layers, which are used to perform N feature iterations on X_input to obtain the output feature map X_output of the Nth iteration N ; as well as Spectral output layer, which is used to X_output N Perform convolution operation to obtain a high signal-to-noise ratio nuclear magnetic resonance spectrum; Among them, all convolutions in the deep iterative shrinkage threshold network are one-dimensional convolutions.

2. The method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network according to claim 1, It is characterized in that The data initialization layer includes: The data mapping unit is used to map the low signal-to-noise ratio nuclear magnetic resonance spectrum through the sensing matrix Φ to obtain the observation matrix Φ x ; Φ x =Y input Φ; where Y input represents a low signal-to-noise ratio NMR spectrum; Φ is an M×M identity matrix; M represents the ideal noise-free dataset Y used for deep iterative shrinkage threshold network training ideal The total number of data points; A matrix generation unit, which is used to provide an initialization matrix Q; Q = Y ideal (ΦY ideal ) T ((ΦY ideal )(ΦY ideal ) T ) -1 ;as well as The product part is used to convert Φ x and the transpose of Q T Perform product calculation to obtain the initial feature map X_input; X_input = Φ x Q T .

3. The method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network according to claim 1, It is characterized in that The nth iterative feature layer includes: Feature update layer a n , which is used to use the gradient descent method to n Update and generate the updated feature map X_gum n ; Among them, X_input n Represents the input of the nth iterative feature layer; Feature extraction layer b n , which is used to n Perform feature extraction and generate the extracted feature map X_fem n ; Forward convolution layer c n , which is used to n Perform forward convolution processing to obtain the forward feature map X_fcm n ; Soft thresholding layer d n , which is used to n Perform soft threshold processing to obtain the soft threshold feature map X_stm n ; as well as Deconvolutional layer n , which is used to n Perform reverse convolution to obtain the output feature map X_output of the nth iteration n ; n∈[1,N]; when n=1, X_input n is X_input; when n>1, X_input n The output feature map X_output of the n-1th iteration n-1 .

4. The method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network according to claim 3, It is characterized in that In the feature update layer a n In, X_gum n The calculation formula is: X_gum n =X_input n -l step (F T Φ*X_input n ); In the formula, λ step Represents the network learnable parameters; In the feature extraction layer b n In X_fem n The calculation formula is: X_fem n =X_gum n +λ step (F T F); In the soft thresholding layer d n In X_stm n The calculation formula is: X_stm n =sign(X_fcm n )*max((X_fcm n |-μ),0); In the formula, sign(.) represents the sign function; max(.) represents the maximum value; μ represents the soft threshold parameter.

5. The method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network according to claim 3, It is characterized in that The nth iterative feature layer further includes: Output convolution layer f n , which is used to convert X_output n The output convolution is reshaped to calculate the symmetric loss loss_layer_sym of the nth iteration n ; Among them, loss_layer_sym n The calculation formula is: In the formula, X_est n The mth data point; X_est n X_output n The feature map generated by reshaping the output convolution; X_fem n The mth data point.

6. The method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network according to claim 5, It is characterized in that The training method of the deep iterative shrinkage threshold network includes: S1, obtain the ideal noise-free data set Y ideal ;in, represents the ideal noise-free data of the mth data point; S2, for Y ideal Perform fast Fourier transform to obtain spectrum data, calculate the real part of the spectrum data, and normalize it by taking the maximum value of the real part to obtain the normalized spectrum data set Y norm ; Performing inverse Fourier transform on the normalized spectrum data set to obtain a time domain signal, and adding complex noise with random characteristics to the time domain signal to obtain a noisy time domain signal set; Perform fast Fourier transform on the noisy time domain signal set to obtain the noisy spectrum data set Y noise ; Represents the spectrum data of the mth data point after adding noise; S3, based on Y noise , Y ideal Training the deep iterative shrinkage threshold network to obtain a trained deep iterative shrinkage threshold network; Among them, Y noise As the input of the deep iterative shrinkage threshold network during training; Y ideal as a training goal.

7. The method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network according to claim 6, It is characterized in that The loss function used by the deep iterative shrinkage threshold network during training total for: loss total =loss discrepancy +γloss constraint ; In the formula, loss discrepancy Represents the difference loss of the final output; loss constraint represents the constraint loss of the iterative feature part; γ represents the weighting constant; Among them, loss discrepancy The calculation formula is: In the formula, Output N Output N express The final output after input into the deep iterative shrinkage threshold network; loss constraint The calculation formula is:

8. The method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network according to claim 1, It is characterized in that N＝9。 9. A device for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network. It is characterized in that It uses the method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network as described in any one of claims 1 to 8; The device for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network comprises: A spectrum acquisition module, which is used to acquire a low signal-to-noise ratio nuclear magnetic resonance spectrum; as well as The spectrum processing module is used to input the low signal-to-noise ratio nuclear magnetic resonance spectrum into the trained deep iterative shrinkage threshold network for processing to obtain a high signal-to-noise ratio nuclear magnetic resonance spectrum.

10. A computer program product, It is characterized in that It comprises a computer program; when the computer program is executed by a processor, it implements the steps of the method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy based on a deep iterative shrinkage threshold network as described in any one of claims 1 to 8.

Images