Nuclear magnetic resonance heteronuclear spectrum and pure chemical shift spectrum noise reduction method based on Gaussian adaptive iterative network

Abstract

The invention discloses a noise reduction method for a nuclear magnetic resonance heteronuclear spectrum and a pure chemical shift spectrum based on a Gaussian adaptive iterative network, and the method comprises the steps: 1, constructing a pure chemical shift training data set, and carrying out the preprocessing; 2, constructing a Gaussian adaptive iterative network, wherein the specific network architecture comprises a data preprocessing module and N iterative denoising modules; and a third step of establishing a back propagation loss function and training the established network to obtain an optimal nuclear magnetic resonance spectrum noise reduction model which is used for processing the noisy nuclear magnetic resonance spectrum and outputting the nuclear magnetic resonance spectrum after noise reduction. According to the method, the noise-containing magnetic resonance spectrum can be input into the trained deep neural network to be subjected to adaptive noise reduction processing, so that the noise reduction effect of the nuclear magnetic resonance heteronuclear spectrum and the pure chemical shift spectrum can be effectively improved.

Description

Technical Field

[0001] The invention relates to the field of magnetic resonance spectrum noise reduction, and in particular to a nuclear magnetic resonance spectrum noise reduction method based on a Gaussian adaptive iterative network. Background Art

[0002] Nuclear magnetic resonance (NMR) technology is an important tool for analyzing molecular structure, chemical environment and dynamic behavior by measuring the signal generated by the resonance of atomic nuclei in an external magnetic field. NMR spectra can provide accurate information inside molecules, including chemical shift, coupling constants, and the relative positions and quantities of different atoms. ) represents the resonant frequency shift caused by the change of the local magnetic field inside the sample, usually measured in parts per million (ppm). By analyzing the chemical shift, nuclei in different chemical environments in the sample can be distinguished, and then the molecular structure can be analyzed. However, NMR spectra are often accompanied by noise, which comes from factors such as instrument inherent noise, environmental noise, and thermal noise. The presence of noise not only affects the clarity of the spectrum, but may also interfere with the identification and quantitative analysis of important signals. Therefore, how to effectively remove or reduce noise and improve the signal-to-noise ratio (SNR) of the spectrum is a key issue in the application of NMR technology. Spectra with high signal-to-noise ratios can more accurately reflect the true characteristics of molecules, while spectra with low signal-to-noise ratios may cause confusion between signals and noise, affecting the judgment of the chemical environment and the analysis of molecular structures.

[0003] Traditionally, there are a variety of methods to improve the signal-to-noise ratio of NMR spectra. In terms of hardware, the use of magnets with higher magnetic field strength, low-noise amplifiers, and optimized receiving coil design can significantly enhance signal strength and thus improve signal-to-noise ratio. However, these hardware upgrades are often costly, complex to operate, and require regular maintenance. In addition, as the magnetic field strength increases, sample handling and safety requirements become more stringent. On the other hand, data acquisition optimization is also a common method, such as optimizing signal acquisition efficiency by adjusting the parameters of the pulse sequence (such as relaxation time, duty cycle, repetition time, etc.). However, this method requires precise control of experimental conditions and is usually accompanied by a long experimental time. This not only reduces the efficiency of the experiment, but may also make it difficult to meet experimental requirements when high throughput is required. In terms of signal processing, commonly used noise reduction methods include low-pass filtering after Fourier transform, waveform denoising, wavelet transform, etc. These methods can suppress noise to a certain extent, but may also cause signal loss, especially when the noise is close to the signal frequency component. In addition, these methods are usually based on fixed noise models and cannot cope with complex or non-Gaussian noise, resulting in limited effectiveness when dealing with dynamic or complex systems.

[0004] In recent years, with the development of deep learning, neural networks have been used in NMR spectra denoising. Traditional deep learning networks learn the patterns of extracting signals from noise through large-scale training data. These networks, such as convolutional neural networks (CNNs) and autoencoders, have shown certain noise reduction effects. However, these networks usually rely on large-scale data sets, and the network structure is fixed, making it difficult to adapt to different noise types or sample environments. In addition, traditional networks often lack the ability to adapt to the characteristics of the signal itself, which may lead to overfitting or signal loss when processing complex or low signal-to-noise ratio data.

[0005] In summary, although traditional methods and deep learning networks have made some progress in NMR spectrum denoising, there are still many limitations. In order to better cope with these challenges, it is necessary to develop more flexible and adaptive denoising technologies to further improve the quality and analytical ability of NMR spectra. Summary of the invention

[0006] Based on this, in order to address the problems that traditional nuclear magnetic resonance spectrum denoising methods are insufficient in adaptability and have limited effects when processing complex signals, and traditional deep learning networks perform poorly when dealing with different types of noise, the present invention proposes a nuclear magnetic resonance heteronuclear spectrum and pure chemical shift spectrum denoising method based on a Gaussian adaptive iterative network, in order to effectively improve the denoising effect of the nuclear magnetic resonance spectrum by adaptively adjusting the noise processing strategy.

[0007] The present invention is implemented by the following technical solutions:

[0008] The invention discloses a method for reducing noise of nuclear magnetic resonance heteronuclear spectrum and pure chemical shift spectrum based on Gaussian adaptive iterative network, which comprises the following steps:

[0009] Step 1: Construct the NMR spectroscopy training input dataset And the ideal data set , and generate the preprocessing matrix ,in, represents transpose, The dimension is The projection matrix;

[0010] Step 2: construct a Gaussian adaptive iterative network, including: data preprocessing module, Iterative denoising modules, and The dth data of the mth group of data Input the Gaussian adaptive iterative network and Processing is performed to obtain the feature map after noise reduction ;

[0011] Step 2.1, according to R, Input the data into the data preprocessing module for processing to obtain the initial feature map ;

[0012] Step 2.2, Iterative denoising module Processing is performed to obtain the feature map after noise reduction ;

[0013] Step 3: Use formula (1) to establish the dth normalized root mean square error loss function of the mth group of data: :

[0014] (1)

[0015] In formula (1), express The i-th data point in ;

[0016] Step 4: Input the NMR spectrum training data set The data of all groups in are input into the Gaussian adaptive iterative network for training, and the network parameters are updated using the gradient descent method until the normalized root mean square error loss function converges, thereby obtaining the optimal nuclear magnetic resonance spectrum denoising model, which is used to process the noisy nuclear magnetic resonance spectrum and output the denoised nuclear magnetic resonance spectrum.

[0017] The method for denoising the nuclear magnetic resonance heteronuclear spectrum and pure chemical shift spectrum based on a Gaussian adaptive iterative network described in the present invention is also characterized in that the step 1 comprises:

[0018] Step 1.1, Acquire a Pure Chemical Shift NMR Spectral Dataset and ideal pure chemical shift NMR spectroscopy data sets ,in, express The mth group of nuclear pure chemical shift magnetic resonance spectroscopy data, express The mth group of ideal pure chemical shift NMR spectroscopy data, M represents the total number of data sets;

[0019] in, , , express The dth pure chemical shift NMR spectrum data in, express The dth ideal pure chemical shift NMR spectrum data in As Label data of

[0020] Step 1.2, After adding different noises to the D data in The corresponding data are D pure chemical shift NMR spectra plus noise data. ,in, express The corresponding d-th pure chemical shift NMR spectrum noise data in;

[0021] Step 1.3, and After Fourier transformation, normalization is performed respectively, and the normalized mth group of pure chemical shift nuclear magnetic resonance spectrum data is obtained accordingly. And the normalized mth group of ideal pure chemical shift NMR spectroscopy data , thus obtaining the normalized pure chemical shift NMR spectroscopy training input data set and normalized ideal pure chemical shift NMR spectroscopy data set ; in, express The corresponding normalized pure chemical shift NMR spectrum data is , express The corresponding normalized dth ideal pure chemical shift NMR spectrum data, and , express The i-th data point in express The i-th data point in , I represents the total number of data points.

[0022] The step 2.1 comprises the following steps:

[0023] Step 2.1.1, using the projection matrix right The dth data of the mth group of data Projection is performed to obtain the dth observation matrix of the mth group of data ;

[0024] Step 2.1.2, calculate the dth initial feature map of the mth group of data .

[0025] Each iterative denoising module in step 2.2 includes: a feature optimization processing unit, a feature enhancement processing unit, a Gaussian adaptive adjustment unit GAU, and a data reconstruction output unit;

[0026] Step 2.2.1, when n=1, The reconstructed feature map output by the n-1th iterative denoising module is denoted as ;

[0027] In the nth iterative denoising module, the feature optimization processing unit Perform feature optimization to obtain the dth optimized feature map of the mth group of data in the nth iterative denoising module :

[0028] Step 2.2.2, the feature enhancement processing unit in the nth iteration denoising module is Perform feature enhancement processing to obtain the d-th enhanced feature map of the m-th group of data in the n-th iterative denoising module ;

[0029] Step 2.2.3, the Gaussian adaptive adjustment unit GAU in the nth iteration denoising module is Processing is performed to obtain the dth adjusted adaptive signal of the mth group of data in the nth iterative denoising module ;

[0030] Step 2.2.4, the data reconstruction output unit in the nth iterative denoising module includes: L one-dimensional convolution layers with a convolution kernel of k, Y ReLU activation functions, and Perform reconstruction processing to generate the d-th reconstructed feature map of the m-th group of data in the n-th iterative denoising module ;

[0031] Step 2.2.5, when n=2,3…N, Input into the nth iterative denoising module for processing to generate the dth reconstructed feature map of the mth group of data in the nth iterative denoising module , so that the Nth iterative denoising module outputs the dth reconstructed feature map of the mth group of data in the Nth iterative denoising module, and is recorded as the denoised feature map .

[0032] The step 2.2.3 includes:

[0033] Step 2.2.3.1, use equation (2) and equation (3) to calculate The mean With variance :

[0034] (2)

[0035] (3)

[0036] In formula (2) and formula (3), for The i-th data point in ;

[0037] Step 2.2.3.2, use formula (4) to generate the dth standard deviation estimate of the mth group of data in the Gaussian adaptive adjustment unit GAU of the nth iterative denoising module: :

[0038] (4)

[0039] In formula (4), 𝜀 is a positive number;

[0040] Step 2.2.3.3, use formula (5) to calculate and generate the d-th Gaussian weight map of the m-th group of data in the Gaussian adaptive adjustment unit GAU of the n-th iterative denoising module: :

[0041] (5)

[0042] In formula (5), is the dth width parameter to be learned for the mth group of data in the nth iterative denoising module;

[0043] Step 2.2.3.4, using formula (6), generate the d-th adaptive signal of the m-th group of data in the Gaussian adaptive adjustment unit GAU of the n-th iterative denoising module :

[0044] (6)

[0045] In formula (6), is the dth offset parameter to be learned of the mth group of data in the Gaussian adaptive adjustment unit GAU of the nth iterative denoising module, It is the d-th scaling coefficient to be learned of the m-th group of data in the Gaussian adaptive adjustment unit GAU of the n-th iterative denoising module.

[0046] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the nuclear magnetic resonance spectrum noise reduction method, and the processor is configured to execute the program stored in the memory.

[0047] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the nuclear magnetic resonance spectrum noise reduction method when the computer program is executed by a processor.

[0048] Compared with the prior art, the method for improving the signal-to-noise ratio of nuclear magnetic resonance spectroscopy of the present invention has the following beneficial effects:

[0049] 1. The present invention provides a Gaussian adaptive iterative network, which introduces a Gaussian adaptive unit (GAU) module into the iterative shrinkage threshold network structure. The module can be adaptively adjusted according to the noise characteristics of the input data, thereby optimizing the signal processing strategy under different noise environments. The network combines the structural advantages of traditional optimization algorithms and the computational efficiency of deep learning networks. Compared with traditional optimization methods, the Gaussian adaptive iterative network significantly improves the accuracy and robustness of signal reconstruction, enhances the effect of the network in processing complex and low signal-to-noise ratio signals, and further improves the noise reduction performance.

[0050] 2. The GAU module can adaptively enhance the ability to capture weak signals at low sampling rates and effectively suppress noise, thereby shortening the experimental acquisition time, reducing equipment requirements and data processing costs. In addition, the GAU module can dynamically adjust parameters according to the statistical characteristics of the input signal, further improving the generalization ability of the model and the reconstruction accuracy of complex signals.

[0051] 3. All network parameters in the present invention (including threshold parameters, step size, and adaptive parameters of the GAU module, etc.) are automatically adjusted through a joint training strategy, so that the network can flexibly adapt to different types of input data, thereby enhancing the robustness and noise reduction performance of the model. The introduction of the GAU module enhances the network's ability to handle different noise types and signal structures, and broadens the scope of application of this method.

[0052] 4. When generating training data sets, one ideal data corresponds to multiple inputs with different noise levels. Through this method, the network can learn and adapt to the signal characteristics under different noise conditions during the training process, thereby significantly improving the ability to suppress different noise levels and the ability to recover weak peak signals. This diversified noise processing strategy improves the robustness and accuracy of the network in practical applications.

[0053] 5. The present invention uses pure chemical shift data to train the network, which helps the model to better learn and understand the signal characteristics of different nuclear spectrum types. Task, this training method strengthens the network's adaptability to different signal types and further broadens its application scope. Through this training method, the network not only performs well in processing heteronuclear spectra, but also can be used for denoising of pure chemical shift spectra. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flow chart of the nuclear magnetic resonance spectrum noise reduction method of the present invention;

[0055] Figure 2 It is a functional module diagram of the nuclear magnetic resonance spectrum noise reduction method in the present invention;

[0056] Figure 3 A schematic diagram of generating a training set used in the present invention;

[0057] Figure 4 Flowchart of the deep neural network training process;

[0058] Figure 5 It is the flowchart of the data preprocessing module;

[0059] Figure 6 It is the flowchart of the nth iteration process;

[0060] Figure 7 is a flowchart of the Gaussian adaptive unit module;

[0061] Figure 8 It is a schematic diagram of the functional modules of the deep learning network model adopted in the present invention;

[0062] Fig. 9 A schematic diagram of the experimental results of applying a deep learning network model to perform denoising on a pure chemical shift simulation spectrum in the present invention;

[0063] Fig.10 Schematic diagram of the experimental results of applying the deep learning network model to denoise the 2D HSQC spectrum of azithromycin in the present invention. DETAILED DESCRIPTION

[0064] In this embodiment, a method for reducing noise of nuclear magnetic resonance heteronuclear spectrum and pure chemical shift spectrum based on Gaussian adaptive iterative network is shown in the specific flow chart. Figure 1 As shown, the following steps are included:

[0065] Step 1: Construct the NMR spectroscopy training input dataset And the ideal data set , and generate the preprocessing matrix ,in, represents transpose, The dimension is The projection matrix;

[0066] Step 1.1, Acquire a Pure Chemical Shift NMR Spectral Dataset and ideal pure chemical shift NMR spectroscopy data sets ,in, express The mth group of nuclear pure chemical shift magnetic resonance spectroscopy data, express The mth group of ideal pure chemical shift NMR spectroscopy data, M represents the total number of data sets;

[0067] in, , , express The dth pure chemical shift NMR spectrum data in, express The dth ideal pure chemical shift NMR spectrum data in As Label data of

[0068] Pure chemical shift NMR spectroscopy data And ideal pure chemical shift NMR spectroscopy data All are simulated pure chemical displacement data, calculated by equation (7) and equation (8) respectively:

[0069] (7)

[0070] (8)

[0071] In formula (7) and formula (8), is the indirect spectral width, is the direct dimension sampling time, A is the amplitude, f is the frequency, is the relaxation time, and J is the scalar coupling constant. In this embodiment, Set to 50, the range of A is , f range is , The range is , the range of J is Pure chemical shift NMR spectroscopy data has a scalar coupling effect and exhibits a more complex oscillation pattern, which is represented by a cosine function. This causes the signal to be modulated by the scalar coupling constant J as it decays over time. Ideal pure chemical shift NMR spectroscopy data only considers the effects of chemical shift and relaxation, exhibiting clear frequency and amplitude characteristics, with clean signals and no noise.

[0072] In this embodiment, a simulated pure chemical displacement data set generated by MATLAB is used for training. The specific generation method is shown in Figure 3 , where the total number of training data groups M is 6000, the number of data in each group D is 4, there are a total of 24000 data, and the total number of points in each data I is 8192.

[0073] Step 1.2, After adding different noises to the D data in The corresponding data are D pure chemical shift NMR spectra plus noise data. ,in, express The corresponding d-th pure chemical shift NMR spectrum noise data in .

[0074] Step 1.3, and After Fourier transformation, normalization is performed respectively, and the normalized mth group of pure chemical shift nuclear magnetic resonance spectrum data is obtained accordingly. And the normalized mth group of ideal pure chemical shift NMR spectroscopy data , thus obtaining the normalized pure chemical shift NMR spectroscopy training input data set and normalized ideal pure chemical shift NMR spectroscopy data set ;

[0075] in, express The corresponding normalized pure chemical shift NMR spectrum data is , express The corresponding normalized dth ideal pure chemical shift NMR spectrum data, and , express The i-th data point in express The i-th data point in , I represents the total number of data points.

[0076] Step 2: construct a Gaussian adaptive iterative network, including: data preprocessing module, An iterative denoising module, whose functional module diagram is as follows Figure 2 as shown; and The dth data of the mth group of data Input Gaussian adaptive iterative network and Processing is performed to obtain the feature map after noise reduction ;

[0077] In this embodiment, the design method of the Gaussian adaptive iterative network includes the following steps: First, the input noisy nuclear magnetic resonance data is preprocessed, and then N iterations are performed in sequence. In each iteration, the input data is first optimized, and then feature enhancement is achieved through forward convolution and soft threshold operation to generate an intermediate feature map. Next, the intermediate feature map is adaptively modeled with Gaussian distribution using the GAU adaptive adjustment module. On this basis, high-quality signal data after denoising is generated by a signal reconstruction module composed of reverse convolution. After N iterations, a nuclear magnetic resonance signal with a high signal-to-noise ratio is finally obtained. Each step is described in detail below.

[0078] Step 2.1, The input data is processed in the preprocessing module to obtain the initial feature map ;

[0079] In this embodiment, the data preprocessing module projects and initializes the input data. Figure 5 , the specific process is as follows:

[0080] Step 2.1.1, using the projection matrix right The dth data of the mth group of data Projection is performed to obtain the dth observation matrix of the mth group of data ;

[0081] Step 2.1.2, calculate the dth initial feature map of the mth group of data .

[0082] Through the preprocessing of this step, the convergence direction of the data in the iterative process is adjusted, thereby enhancing the model's ability to identify and retain signal features. Secondly, by performing matrix operations on noisy nuclear magnetic resonance spectrum data, the irregular noise in the data is initially smoothed, effectively improving the stability of the data and reducing numerical fluctuations in subsequent iterations. In addition, this step provides an optimized starting point for the model, allowing the model to more quickly approach the global optimal solution and avoid falling into the local optimum.

[0083] Step 2.2, each iterative denoising module includes: a feature optimization processing unit, a feature enhancement processing unit, a Gaussian adaptive adjustment unit GAU, and a data reconstruction output unit;

[0084] In this embodiment, the initial feature map is iterated 9 times in total. The nth iteration process is shown in FIG. Figure 6 , whose main purpose is to optimize the initial feature map through multiple iterations , gradually reducing noise and improving feature map quality. The specific process is as follows:

[0085] Step 2.2.1, when n=1, The reconstructed feature map output by the n-1th iterative denoising module is denoted as ;

[0086] In the nth iterative denoising module, the feature optimization processing unit uses formula (9) to Perform feature optimization to obtain the dth optimized feature map of the mth group of data in the nth iterative denoising module :

[0087] (9)

[0088] In this embodiment, the learning rate parameter is set , which is used to control the amplitude and direction of the gradient adjustment. Through this operation, the errors in the signal are adjusted and corrected, which initially suppresses the noise interference and enhances the main components of the signal, ensuring the overall stability and clarity of the signal, thereby achieving a preliminary improvement in signal quality.

[0089] Step 2.2.2, the feature enhancement processing unit in the nth iteration denoising module is Perform feature enhancement processing to obtain the d-th enhanced feature map of the m-th group of data in the n-th iterative denoising module ;

[0090] This processing further focuses on enhancing the key features in the signal while effectively removing the remaining low-amplitude noise, so that the main information of the signal is more prominently retained and optimized.

[0091] Step 2.2.2.1, the feature enhancement processing unit includes: a forward convolution layer and a feature screening layer. The forward convolution layer includes L one-dimensional convolution layers with a convolution kernel of k and Y ReLU activation functions;

[0092] In this embodiment, the forward convolution consists of two one-dimensional convolution layers with a convolution kernel of 25 and one ReLU activation function; First, the first one-dimensional convolutional layer extracts preliminary features, then the ReLU activation function performs nonlinear processing, and then the second one-dimensional convolutional layer further extracts and enhances features to obtain an intermediate feature map. , then the feature screening processing layer uses formula (10) to Perform feature screening to generate the enhanced nth feature map :

[0093] (10)

[0094] In formula (10), sign is a sign function, and 𝛼 represents a screening parameter. In this embodiment, the screening parameter 𝛼=0.1 is used to highlight the key features of the signal and suppress noise.

[0095] Step 2.2.3, the Gaussian adaptive adjustment unit GAU in the nth iteration denoising module is Processing is performed to obtain the dth adjusted adaptive signal of the mth group of data in the nth iterative denoising module ;

[0096] The innovation of the present invention is the introduction of the GAU module. This module can be dynamically adjusted according to the statistical characteristics of the input signal, thereby significantly improving the accuracy and noise resistance of signal reconstruction. Through layer-by-layer adaptive optimization, the present invention effectively improves the signal-to-noise ratio of the signal and restores more detailed information. Figure 7 , specifically including the following steps:

[0097] Step 2.2.3.1, use equation (2) and equation (3) to calculate The mean With variance :

[0098] (2)

[0099] (3)

[0100] In formula (2) and formula (3), for The i-th data point in ;

[0101] Step 2.2.3.2, use formula (4) to generate the dth standard deviation estimate of the mth group of data in the Gaussian adaptive adjustment unit GAU of the nth iterative denoising module: :

[0102] (4)

[0103] In formula (4), 𝜀 is a positive number;

[0104] Step 2.2.3.3, use formula (5) to calculate and generate the d-th Gaussian weight map of the m-th group of data in the Gaussian adaptive adjustment unit GAU of the n-th iterative denoising module: :

[0105] (5)

[0106] In formula (5), It is the dth width parameter to be learned for the mth group of data in the nth iterative denoising module.

[0107] Step 2.2.3.4, using formula (6), generate the d-th adaptive signal of the m-th group of data in the Gaussian adaptive adjustment unit GAU of the n-th iterative denoising module :

[0108] (6)

[0109] In formula (6), is the dth offset parameter to be learned of the mth group of data in the Gaussian adaptive adjustment unit GAU of the nth iterative denoising module, It is the d-th scaling coefficient to be learned of the m-th group of data in the Gaussian adaptive adjustment unit GAU of the n-th iterative denoising module.

[0110] The Gaussian adaptive unit (GAU) dynamically optimizes the balance between signal and noise according to the noise characteristics of the input data by adaptively adjusting the signal processing strategy. This not only further suppresses complex and nonlinear noise, but also enhances the subtle features of the signal, ensuring that the model can stably output high-quality reconstruction results under different noise levels.

[0111] Step 2.2.4, the data reconstruction output unit in the nth iterative denoising module includes: L one-dimensional convolution layers with a convolution kernel of k, Y ReLU activation functions, and Perform reconstruction processing to generate the d-th reconstructed feature map of the m-th group of data in the n-th iterative denoising module ;

[0112] In this embodiment, the reconstruction process consists of two one-dimensional convolutional layers with a convolution kernel of 25 and one ReLU activation function; First, it is initially reconstructed through the first one-dimensional convolutional layer, then nonlinearly processed through the ReLU activation function, and then further restored and reconstructed through the second one-dimensional convolutional layer to obtain the nth reconstructed feature map .

[0113] Step 2.2.5, when n=2,3…N, Input into the nth iterative denoising module for processing to generate the dth reconstructed feature map of the mth group of data in the nth iterative denoising module , so that the Nth iterative denoising module outputs the dth reconstructed feature map of the mth group of data in the Nth iterative denoising module, and is recorded as the denoised feature map .

[0114] Step 3: Use formula (1) to establish the normalized root mean square error loss function for back propagation: :

[0115] (1)

[0116] In formula (1), express The i-th data point in .

[0117] Step 4: Input the NMR spectroscopy training data set The data from all groups in the training are input into the Gaussian adaptive iterative network for training, and the network parameters are updated using the gradient descent method until the normalized root mean square loss function The optimal NMR spectrum denoising model is obtained until convergence, which is used to process the noisy NMR spectrum and output the denoised NMR spectrum.

[0118] In this embodiment, the neural network model is trained. For details on the training process, please refer to Figure 4 The network layers of the deep learning model are set to 9, the learning rate is 0.003, the batch size is 32, and the Adam optimization algorithm is used. The early stopping training strategy is adopted. When the loss value does not decrease for 20 consecutive cycles, the training is stopped, and the model with the smallest verification loss value is saved as the final model.

[0119] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0120] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0121] In this example, the deep learning network model was tested using the 2D HSQC spectrum of azithromycin and the pure chemical shift simulation spectrum to verify the effect achieved by the present application. Fig. 9 , Fig.10 ,in Fig. 9 Parts (a), (b), and (c) are respectively the input and label of pure chemical displacement simulation data and the results of denoising and reconstruction using the method of the present application; Fig.10 Part (a), part (b), and part (c) are respectively the spectrum (input) of azithromycin 2D HSQC spectrum under 4 repeated scans, the spectrum (label) under 32 repeated scans, and the result of noise reduction reconstruction using the method of the present application. According to the comparison of the spectra in the figure, it can be seen that using the method provided by the present invention, the network can effectively suppress noise and pseudo-peak signals after processing nuclear magnetic resonance heteronuclear spectra and pure chemical shift spectra, and successfully recover weak peaks that are almost completely submerged by severe noise, thereby achieving a significant improvement in signal-to-noise ratio.

[0122] In summary, the nuclear magnetic resonance spectrum denoising method of the present invention adopts a deep learning model of a Gaussian adaptive iterative network that introduces a Gaussian adaptive unit module. The specific functional modules of the network model refer to Figure 8. By inputting noisy nuclear magnetic resonance spectrum data into the model and gradually reducing the noise through multiple iterations, the data quality after each iteration is significantly improved. The method uses feature optimization, feature enhancement processing, Gaussian adaptive adjustment, and data reconstruction output in each iteration to ensure the stability and consistency of the signal, thereby generating high-quality denoised spectrum data. Compared with traditional methods, the present invention introduces a Gaussian adaptive unit (GAU), which can be adaptively adjusted according to the noise characteristics of the input data and dynamically optimize the balance between signal and noise. This can not only significantly improve the signal-to-noise ratio of the signal, but also enhance the subtle features of the signal, retain the key details of the original signal, and has the advantages of significant noise reduction effect and high processing efficiency. Combined with the iterative shrinkage threshold algorithm and compressed sensing theory, the GAU module further improves the robustness and adaptability of the model, and significantly improves the accuracy of signal reconstruction and the noise reduction effect.

Claims

1. A method for denoising nuclear magnetic resonance heteronuclear spectra and pure chemical shift spectra based on a Gaussian adaptive iterative network, characterized in that: The following steps are involved: Step 1: Construct the NMR spectroscopy training input dataset And the ideal data set , and generate the preprocessing matrix ,in, represents transpose, The dimension is The projection matrix; Step 2: construct a Gaussian adaptive iterative network, including: data preprocessing module, Iterative denoising modules, and The dth data of the mth group of data Input the Gaussian adaptive iterative network and Processing is performed to obtain the feature map after noise reduction ; Step 2.1, according to R, Input the data into the data preprocessing module for processing to obtain the initial feature map ; Step 2.2, Iterative denoising module Processing is performed to obtain the feature map after noise reduction ; Step 3: Use formula (1) to establish the dth normalized root mean square error loss function of the mth group of data: : (1) In formula (1), express The i-th data point in ; Step 4: Input the NMR spectrum training data set The data of all groups in are input into the Gaussian adaptive iterative network for training, and the network parameters are updated using the gradient descent method until the normalized root mean square error loss function converges, thereby obtaining the optimal nuclear magnetic resonance spectrum denoising model, which is used to process the noisy nuclear magnetic resonance spectrum and output the denoised nuclear magnetic resonance spectrum.

2. The method for denoising NMR heteronuclear spectrum and pure chemical shift spectrum based on Gaussian adaptive iterative network according to claim 1, characterized in that: The step 1 comprises: Step 1.1, Acquire a Pure Chemical Shift NMR Spectral Dataset and ideal pure chemical shift NMR spectroscopy data sets ,in, express The mth group of nuclear pure chemical shift magnetic resonance spectroscopy data, express The mth group of ideal pure chemical shift NMR spectroscopy data, M represents the total number of data sets; in, , , express The dth pure chemical shift NMR spectrum data in, express The dth ideal pure chemical shift NMR spectrum data in As Label data of Step 1.2, After adding different noises to the D data in The corresponding data are D pure chemical shift NMR spectra plus noise data. ,in, express The corresponding d-th pure chemical shift NMR spectrum noise data in; Step 1.3, and After Fourier transformation, normalization is performed respectively, and the normalized mth group of pure chemical shift nuclear magnetic resonance spectrum data is obtained accordingly. And the normalized mth group of ideal pure chemical shift NMR spectroscopy data , thus obtaining the normalized pure chemical shift NMR spectroscopy training input data set and normalized ideal pure chemical shift NMR spectroscopy data set ; in, express The corresponding normalized pure chemical shift NMR spectrum data is , express The corresponding normalized dth ideal pure chemical shift NMR spectrum data, and , express The i-th data point in express The i-th data point in , I represents the total number of data points.

3. The method for denoising NMR heteronuclear spectrum and pure chemical shift spectrum based on Gaussian adaptive iterative network according to claim 1, characterized in that: The step 2.1 comprises the following steps: Step 2.1.1, using the projection matrix right The dth data of the mth group of data Projection is performed to obtain the dth observation matrix of the mth group of data ; Step 2.1.2, calculate the dth initial feature map of the mth group of data .

4. The method for denoising NMR heteronuclear spectrum and pure chemical shift spectrum based on Gaussian adaptive iterative network according to claim 1, characterized in that: Each iterative denoising module in step 2.2 includes: a feature optimization processing unit, a feature enhancement processing unit, a Gaussian adaptive adjustment unit GAU, and a data reconstruction output unit; Step 2.2.1, when n=1, The reconstructed feature map output by the n-1th iterative denoising module is denoted as ; In the nth iterative denoising module, the feature optimization processing unit Perform feature optimization to obtain the dth optimized feature map of the mth group of data in the nth iterative denoising module : Step 2.2.2, the feature enhancement processing unit in the nth iteration denoising module is Perform feature enhancement processing to obtain the d-th enhanced feature map of the m-th group of data in the n-th iterative denoising module ; Step 2.2.3, the Gaussian adaptive adjustment unit GAU in the nth iteration denoising module is Processing is performed to obtain the dth adjusted adaptive signal of the mth group of data in the nth iterative denoising module ; Step 2.2.4, the data reconstruction output unit in the nth iterative denoising module includes: L one-dimensional convolution layers with a convolution kernel of k, Y ReLU activation functions, and Perform reconstruction processing to generate the d-th reconstructed feature map of the m-th group of data in the n-th iterative denoising module ; Step 2.2.5, when n=2,3…N, Input into the nth iterative denoising module for processing to generate the dth reconstructed feature map of the mth group of data in the nth iterative denoising module , so that the Nth iterative denoising module outputs the dth reconstructed feature map of the mth group of data in the Nth iterative denoising module, and is recorded as the denoised feature map .

5. The method for denoising NMR heteronuclear spectrum and pure chemical shift spectrum based on Gaussian adaptive iterative network according to claim 4, characterized in that: The step 2.2.3 includes: Step 2.2.3.1, use equation (2) and equation (3) to calculate The mean With variance : (2) (3) In formula (2) and formula (3), for The i-th data point in ; Step 2.2.3.2, use formula (4) to generate the dth standard deviation estimate of the mth group of data in the Gaussian adaptive adjustment unit GAU of the nth iterative denoising module: : (4) In formula (4), 𝜀 is a positive number; Step 2.2.3.3, use formula (5) to calculate and generate the d-th Gaussian weight map of the m-th group of data in the Gaussian adaptive adjustment unit GAU of the n-th iterative denoising module: : (5) In formula (5), is the dth width parameter to be learned for the mth group of data in the nth iterative denoising module; Step 2.2.3.4, using formula (6), generate the d-th adaptive signal of the m-th group of data in the Gaussian adaptive adjustment unit GAU of the n-th iterative denoising module : (6) In formula (6), is the dth offset parameter to be learned of the mth group of data in the Gaussian adaptive adjustment unit GAU of the nth iterative denoising module, It is the d-th scaling coefficient to be learned of the m-th group of data in the Gaussian adaptive adjustment unit GAU of the n-th iterative denoising module.

6. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the nuclear magnetic resonance spectrum noise reduction method described in any one of claims 1 to 5, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the nuclear magnetic resonance spectrum noise reduction method according to any one of claims 1 to 5 are executed.

Images