Nuclear magnetic resonance two-dimensional diffusion ordered spectroscopy sparse reconstruction method based on deep learning

Abstract

The present application relates to the technical field of nuclear magnetic resonance, and provides a nuclear magnetic resonance two-dimensional diffusion ordering spectrum sparse reconstruction method based on deep learning. First, a training data set is constructed based on an exponential decay model of DOSY signals, wherein the decay signals sparsely sampled in the gradient dimension and the corresponding gradient sequence are taken as input, and the ideal two-dimensional diffusion ordering reconstruction spectrum is taken as a label; a sparse reconstruction network model is built by fusing a GRU decay feature extraction module and a Transformer spectrum reconstruction module, and is trained by using a simulation data set; the sparse undersampled two-dimensional diffusion ordering data to be reconstructed is input into the trained network, so that the two-dimensional diffusion ordering spectrum is quickly and accurately reconstructed. The present application can fully utilize the time sequence characteristics of the diffusion decay signals and the global correlation information of the spectrum, reduce the data acquisition amount and the experimental time, realize high-precision and high-efficiency spectrum reconstruction, and improve the spectrum quality and the accuracy of subsequent quantitative analysis.

Description

Technical Field

[0001] This invention relates to the field of nuclear magnetic resonance (NMR) technology, and more particularly to a sparse reconstruction method for two-dimensional diffusion sorting spectra of NMR based on deep learning. Background Technology

[0002] Two-dimensional diffusion-ordered spectroscopy (DOSY) is an important analytical technique for separating and characterizing components in mixtures based on differences in molecular diffusion coefficients. By acquiring a series of NMR spectra under different gradient intensities, signal attenuation information as a function of the gradient is obtained. Based on this, the diffusion coefficient distribution is calculated through inversion, thus constructing a two-dimensional spectrum with both chemical shift and diffusion coefficients. Introducing the diffusion dimension allows signals that previously overlapped in the chemical shift dimension to be effectively separated based on differences in diffusion behavior. This significantly improves the resolvability and analytical capability of components in complex systems, making it valuable for the analysis of complex systems such as natural products, polymers, and biomacromolecules.

[0003] In DOSY experiments, obtaining diffusion information typically relies on multiple samplings at different gradient intensities. To ensure the diffusion decay curve can adequately characterize the data and improve the stability of the inversion calculation, a large number of gradient sampling points are usually required. Generally, the number of gradient dimension sampling points is set to 10–20, and this number needs to be further increased for high-resolution applications. However, as the number of gradient sampling points increases, the data acquisition time is significantly prolonged, thus limiting the widespread application of DOSY technology in rapid detection and practical applications. Conversely, when the number of gradient dimension sampling points decreases, the diffusion decay curve is undersampled, leading to a significant increase in the ill-posedness of the diffusion coefficient inversion process based on the inverse Laplace transform. Specifically, this manifests as a decrease in diffusion coefficient resolution, easy aliasing between spectral peaks of different components, and increased sensitivity of the reconstruction results to noise, resulting in phenomena such as peak broadening, positional shifts, and even spurious peaks, which seriously affect the accuracy and stability of diffusion spectrum reconstruction. Therefore, how to reduce the number of gradient dimension sampling points while ensuring the quality of spectrum reconstruction has become an important research direction for the development of DOSY technology.

[0004] Currently, the mainstream methods for DOSY data processing mainly include parameter fitting methods and gridded reconstruction methods based on regularization constraints. Parameter fitting methods, such as NMR component analysis (CORE), accelerated component analysis algorithm (SCORE), multi-parameter curve analysis (MCR), and direct exponential curve analysis algorithm (DECRA), typically require pre-setting prior information such as the number of components. Therefore, their application in systems with complex or unknown components is limited. Furthermore, their accuracy is sensitive to the quality of the attenuated signal, and their performance degrades significantly with sparse sampling. Gridded reconstruction methods based on regularization constraints, such as constrained regularization (CONTIN), maximum entropy method (MaxEnt), and low-rank sparse Laplace inversion (LRSpILT), stabilize the solution process by introducing constraints such as nonnegativity, sparsity, or low rank. However, their performance is sensitive to the regularization parameter, and they are prone to resolution degradation and reconstruction instability when the number of sampling points is insufficient.

[0005] In recent years, deep learning methods have been introduced into the field of DOSY spectrum reconstruction, improving reconstruction performance to some extent. For example, the Data-Driven Exponential Curve Translation (DRECT) method and the Joint Deep Reconstruction and Uncertainty Estimation Model (DREAM) have achieved good peak alignment and component separation capabilities without explicit physical prior constraints, and have shown strong robustness. However, when the number of sampling points in the gradient dimension is small, existing methods still face problems such as insufficient extraction of sparse features and insufficient reconstruction stability, making it difficult to achieve stable and high-quality diffusion spectrum reconstruction under sparse input conditions. Chinese invention application CN119104967A discloses a method and apparatus for fast NMR spectrum reconstruction using compressed sensing. It constructs an attention-based encoder-decoder neural network to process undersampled NMR spectra. Through multiple feature extraction and fitting processes, it gradually enhances the key features of the signal from the undersampled NMR spectrum, ultimately achieving high-precision reconstruction output and obtaining a high-quality reconstructed NMR spectrum. However, this approach primarily targets compressed sensing reconstruction of conventional NMR one-dimensional FID signals or two-dimensional spectra under sparse sampling conditions. Its core objective is to recover the time-domain information lost due to non-uniform sampling, without considering signal loss caused by gradient coding undersampling in the DOSY experiment. The undersampled DOSY reconstruction process not only requires accurate characterization of the diffusion attenuation process of signal loss but also needs to address the severe ill-conditioned inversion problem caused by the inverse Laplace transform. Therefore, high-quality reconstruction of the DOSY spectrum requires both gradient-dimensional undersampling recovery capability and effective modeling capability for the diffusion-dimensional inverse Laplace map. Existing methods primarily designed for compressed sensing reconstruction of FID signals cannot be directly applied to DOSY data analysis.

[0006] In summary, under the condition of limited sampling points in the gradient dimension, how to achieve high-resolution and stable DOSY spectrum reconstruction using limited attenuation information remains a key technical problem that urgently needs to be solved in the field of two-dimensional diffusion sorting spectrum data processing. Summary of the Invention

[0007] The purpose of this invention is to solve the problems in the prior art.

[0008] The technical solution adopted by this invention to solve its technical problem is: to provide a sparse reconstruction method for two-dimensional diffusion sorting spectra of nuclear magnetic resonance based on deep learning, comprising the following steps:

[0009] Construct a sparse undersampling simulation training dataset;

[0010] A sparse reconstruction network model was built, and the network was trained using a sparse undersampling simulation training dataset.

[0011] Sparse reconstruction of two-dimensional diffusion sorting spectrum of nuclear magnetic resonance is achieved by using a trained sparse reconstruction network model;

[0012] The sparse reconstruction network model includes a gradient dimension decay modeling module and a DOSY spectrum reconstruction module. The gradient dimension decay modeling module uses a two-layer GRU to recursively model the input decay signal to obtain decay features. The DOSY spectrum reconstruction module maps the decay features to a two-dimensional diffusion sorting spectrum.

[0013] Preferably, constructing the sparse undersampling simulation training dataset includes the following steps:

[0014] The gradient sequence is generated parametrically, and is represented as follows:

[0015] ;

[0016] ;

[0017] Among them, normalized variables ; This represents the gradient distribution adjustment coefficient, used to control the distribution form of the gradient sequence; The maximum value of the gradient sequence value; ,in, For the gradient sequence of the th Each sampling point index This represents the total number of gradient sampling points.

[0018] Construct a two-dimensional DOSY signal, represented as:

[0019] ;

[0020] ;

[0021] in, This represents the spectral peak distribution along the chemical shift dimension. For the first Chemical shift positions of each spectral peak The peak width is... For the first The amplitude of each peak For the first The diffusion coefficient of each component It is Gaussian noise; Represents chemical shift coordinates, This represents the total number of peaks in the chemical shift dimension. This indicates the total number of diffusing components.

[0022] Supervised learning labels are constructed to obtain the diffusion spectrum representation, which is expressed as:

[0023] ;

[0024] in, The broadening factor indicating the direction of diffusion. To normalize the peak intensity, The diffusion coefficient distribution grid is shown.

[0025] Preferably, the sparse reconstruction network model includes:

[0026] The gradient dimension decay modeling module adopts a two-layer stacked structure, with each layer containing an update gate and a reset gate. The input of the gradient dimension decay modeling module includes two dimensions. In the gradient dimension, the signal and gradient value corresponding to each gradient point are input into the network as a time step. The current hidden state is determined by the current input and the historical state, thereby realizing the modeling of the overall change characteristics of the decay curve and obtaining the decay characteristics.

[0027] The DOSY spectral reconstruction module first transforms the attenuation features to a unified feature space through linear mapping, and then inputs them into an attention-based encoding network for global feature modeling. The attention-based encoding network includes a multi-head attention unit, a feedforward neural network, and a normalization structure connected in sequence. The multi-head attention mechanism models the input features by weighting them in multiple subspaces. The feedforward neural network is used to perform nonlinear transformations on the features. The normalization structure is used to stabilize the training process.

[0028] Preferably, the network training includes the following steps:

[0029] The input is sparse, undersampled DOSY data, and the label is the ideal DOSY reconstructed spectrum.

[0030] The sparsely undersampled DOSY data is normalized, and the gradient sequence is used as an explicit feature to be paired point by point with the normalized signal to obtain the preprocessed data.

[0031] The preprocessed data is input into the gradient dimension decay modeling module and the DOSY spectrum reconstruction module to obtain the DOSY reconstruction spectrum result output by the model.

[0032] The DOSY reconstructed spectrum results are compared with the labels, and the reconstruction error is calculated as a loss function.

[0033] An adaptive moment estimation optimization algorithm is used to update the model parameters. By adaptively estimating the first and second moments of the gradient, the parameter update step size is dynamically adjusted, thereby improving the optimization performance of the model under sparse data distribution conditions and increasing the convergence speed and stability of the training process.

[0034] Preferably, the sparse reconstruction of two-dimensional diffusion-ordered NMR spectra using a trained sparse reconstruction network model includes the following steps:

[0035] The original attenuation signal is acquired and normalized along the gradient dimension;

[0036] The corresponding gradient sequence is used as an explicit input feature and paired point by point with the normalized decay signal to construct a joint input feature.

[0037] The joint features are input into the gradient dimension decay modeling module to perform sequence modeling in the gradient dimension and extract decay features that can characterize the overall change trend of the sparse decay curve.

[0038] The sparse attenuation features are input into the DOSY spectrum reconstruction module, and the corresponding two-dimensional DOSY reconstruction spectrum is obtained through feature mapping and global modeling.

[0039] This invention also provides a deep learning-based sparse reconstruction system for two-dimensional diffusion ordering spectra of nuclear magnetic resonance, comprising:

[0040] The dataset construction module builds a sparse undersampling simulation training dataset.

[0041] The model building module constructs a sparse reconstruction network model and trains the network using a sparse undersampling simulation training dataset.

[0042] The DOSY reconstruction module utilizes a pre-trained sparse reconstruction network model to achieve sparse reconstruction of two-dimensional diffusion sorting spectra of nuclear magnetic resonance.

[0043] The sparse reconstruction network model includes a gradient dimension decay modeling module and a DOSY spectrum reconstruction module. The gradient dimension decay modeling module uses a two-layer GRU to recursively model the input decay signal to obtain decay features. The DOSY spectrum reconstruction module maps the decay features to a two-dimensional diffusion sorting spectrum.

[0044] The present invention also provides an electronic device, comprising:

[0045] One or more processors;

[0046] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to perform any of the methods described above.

[0047] The present invention also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements any of the methods described above.

[0048] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the methods described above.

[0049] The present invention has the following beneficial effects:

[0050] (1) In practical applications, this invention can significantly reduce the data acquisition time of the DOSY experiment and improve experimental efficiency by reducing the number of gradient sampling points;

[0051] (2) This invention utilizes sequence modeling and global attention mechanism to fully characterize the decay process, and maintains good reconstruction stability and robustness under low sampling conditions. In particular, it still has resolution and analytical reliability in complex systems with overlapping spectral peaks or similar diffusion coefficients.

[0052] (3) This invention avoids the complex optimization solution process and manual parameter adjustment in traditional inversion methods, reduces the overall implementation difficulty and usage threshold, and has strong engineering practicality and promotion application value.

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a deep learning-based sparse reconstruction method for two-dimensional diffusion sorting spectra of nuclear magnetic resonance according to an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the overall structure of the sparse reconstruction network model according to an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of the reconstruction inference process of the sparse reconstruction network model in an embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of the two-dimensional diffusion sorting spectrum obtained after reconstructing real sparse DOSY data according to an embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram of the structure of a sparse reconstruction system for two-dimensional diffusion sorting spectrum of nuclear magnetic resonance based on deep learning, according to an embodiment of the present invention.

[0059] Figure 6 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0061] To address the problems of existing technologies, this invention proposes a deep learning reconstruction method that integrates a gated recurrent unit (GRU) and a multi-head attention mechanism, taking into account the limited number of sampling points in the gradient dimension of DOSY spectra. By jointly modeling the diffusion attenuation features in the gradient dimension and the spectral peak features in the chemical shift dimension, the reconstruction quality and diffusion component resolution of DOSY spectra under limited gradient sampling conditions are improved.

[0062] See Figure 1 As shown in the figure, an embodiment of the present invention provides a sparse reconstruction method for two-dimensional diffusion sorting spectra of nuclear magnetic resonance based on deep learning, which includes the following steps:

[0063] S101, Construct a sparse undersampling simulation training dataset;

[0064] S102, Build a sparse reconstruction network model and train the network using a sparse undersampling simulation training dataset;

[0065] S103, using a trained sparse reconstruction network model to achieve sparse reconstruction of two-dimensional diffusion sorting spectrum of nuclear magnetic resonance;

[0066] The sparse reconstruction network model includes a gradient dimension decay modeling module and a DOSY spectrum reconstruction module. The gradient dimension decay modeling module uses a two-layer GRU to recursively model the input decay signal to obtain decay features. The DOSY spectrum reconstruction module maps the decay features to a two-dimensional diffusion sorting spectrum.

[0067] Specifically, in step S101, a sparse simulation dataset for model training is constructed based on the physical model of signal decay with gradient in the DOSY experiment. This embodiment focuses on the problem of a small number of gradient dimension sampling points and inconsistent gradient distribution in actual DOSY experiments. It constructs a sparse dataset that closely resembles the real experimental conditions through parameterization, ensuring that the generated data can reflect the characteristics of sparse DOSY signals in actual experiments.

[0068] Because the numerical range and distribution of gradient sequences differ under different experimental conditions, and the number of sampling points in the gradient dimension is usually small, this embodiment generates an independent gradient sequence for each sample during the dataset construction process. This allows different samples to have different distribution characteristics in the gradient dimension, simulating the non-fixed distribution of b values ​​in real experiments, thereby improving the model's adaptability to situations where gradient sequences are not fixed and the number of points is limited in actual data. For the first... For each sample, its gradient sequence is represented as:

[0069]

[0070] Among them, the number of gradient dimension points Set to 6. The gradient sequence is generated parametrically: first in the interval Internally generated normalized variables Then, the gradient distribution is adjusted through nonlinear mapping and scaling parameters, and finally normalized to... Within the range. Among them, The range of values ​​is The scale parameter takes values ​​in the range of: The above process can generate gradient sequences with different densities and distributions to simulate the non-fixed gradient distribution in real DOSY experiments. For example, gradient sequences can be generated as follows:

[0071]

[0072] in, This represents the gradient distribution adjustment coefficient, used to control the distribution form of the gradient sequence (such as uniform distribution or exponential distribution).

[0073] Based on this, a two-dimensional sparse DOSY signal is constructed according to the exponential decay model. For the first... The two-dimensional sparse DOSY attenuation signal of each sample is represented as follows:

[0074]

[0075]

[0076] in This represents the peak distribution along the chemical shift dimension. The relevant parameter settings are as follows: diffusion coefficient. , group number 3; Peak position Peak number ; Peak width The value is 0.06; For the first The amplitude of each peak is taken as a value. Gaussian noise In this embodiment, the signal-to-noise ratio is set. Simultaneously with signal generation, a Gaussian function distribution is used to construct supervised learning labels along the diffusion coefficient dimension, resulting in a two-dimensional DOSY reconstructed spectrum with dimensions of [missing information]. , which correspond to the chemical shift dimension and the diffusion coefficient dimension, respectively.

[0077] The final training data includes: input signal dimension Dimension of gradient sequence Tag Dimension The generated simulated dataset has a total of 30,000 samples, including 27,000 samples in the training set, 2,000 samples in the validation set, and 1,000 samples in the test set.

[0078] Specifically, in step S102, to address the difficulty in reconstructing sparse DOSY signals due to insufficient gradient dimension information, a sparse reconstruction network model is constructed and its parameters are set. In this embodiment, the network model consists of two parts: a gradient dimension attenuation modeling module and a diffusion spectrum reconstruction module. Its overall structure is as follows: Figure 2 As shown. The gradient dimension decay modeling module adopts a two-layer GRU. The GRU network uses a two-layer stacked structure, with each layer containing an update gate and a reset gate. The input feature dimension of the GRU is 2, corresponding to the normalized decay signal value and the corresponding gradient value, respectively; the output hidden dimension is 512; and the number of network layers is 2. In the gradient dimension, each gradient sampling point and the decay signal value are input into the network as a time step. The hidden state of the current time step is determined by the current input and the hidden state of the previous time step, thereby realizing the modeling of the dynamic change process of the decay curve; the modeling process is represented as follows:

[0079] ;

[0080] ;

[0081] ;

[0082] ;

[0083] in, It's an update door. It's a door reset. yes Activation function It is the hyperbolic tangent activation function. This represents the hidden state that is passed on to the next time step. This indicates the hidden state at the previous moment. Indicates the candidate hidden state. This indicates the information to be entered at the current moment.

[0084] The DOSY spectral reconstruction module includes a multi-head attention unit, a feedforward neural network, and a normalization (Add&Norm) structure. In this example, the overall parameters are as follows: linear mapping layer parameters are 512-140; encoder layers are 10; multi-head attention units are 7; feedforward neural network dimension is 4096; the parameters of the two linear layers in the feedforward module are set to 140-4096 and 4096-140 respectively; label dimension is 140; Dropout coefficient is 0.1; and GRU hidden dimension is 512.

[0085] During model training, a sparse DOSY decay signal with fewer sampling points in the gradient dimension is used as input, and the corresponding DOSY reconstructed spectrum is used as a supervision signal to learn the mapping relationship between the sparse DOSY signal and the DOSY reconstructed spectrum. In the input processing stage, the original decay signal is first normalized in the gradient dimension to eliminate amplitude differences between different samples and improve the stability of model training. Simultaneously, the corresponding gradient sequence is used as an explicit input feature and paired point-by-point with the normalized signal to construct joint input features, enabling the model to utilize both signal amplitude information and gradient numerical information. Subsequently, the joint features are input to the gradient dimension decay modeling module for sequence modeling in the gradient dimension, extracting features that characterize the overall trend of the sparse decay curve. Finally, the sparse decay features are input to the DOSY spectrum reconstruction module, and the corresponding two-dimensional DOSY reconstructed spectrum is obtained through feature mapping and global modeling.

[0086] In the loss calculation phase, mean squared error (MSE) is used as the loss function to measure the difference between the model output and the label. To ensure training stability, the labels are normalized and scaled accordingly, and the Adam optimization algorithm is used to update the model parameters. By adaptively estimating the first and second moments of the gradient, the parameter update step size is dynamically adjusted, thereby improving the optimization performance of the model under sparse data distribution conditions and increasing the convergence speed and stability of the training process. The initial learning rate is set to... Furthermore, a learning rate scheduling strategy is introduced to adaptively adjust the learning rate based on changes in the validation set loss, thereby improving the model's convergence performance. In this embodiment, the number of training epochs is set to 40, ultimately obtaining a deep learning model for sparse DOSY spectrum reconstruction. The network's loss function is Mean Squared Error (MSE).

[0087]

[0088] in, This represents the reconstructed two-dimensional DOSY reconstruction spectrum. Indicates a label.

[0089] In this embodiment, the sparse DOSY attenuation signal with a small number of gradient dimension sampling points is used as input during the testing phase to reconstruct the DOSY spectrum. First, the sparse attenuation curves obtained during the testing phase are normalized, and the corresponding gradient sequences are paired point-by-point with the attenuation signal as explicit input features to construct the model input. Based on this, the input data is fed into the trained reconstruction model, outputting the corresponding two-dimensional DOSY reconstruction spectrum, thus achieving direct reconstruction from the sparse DOSY signal to the DOSY reconstruction spectrum. Figure 4 The results show the reconstruction results of real sparse DOSY data. The results demonstrate that, under actual experimental conditions, the method of this invention can still obtain DOSY reconstructed spectra with clear structure and good peak separation, while maintaining good diffusion coefficient resolution. This indicates that the constructed model can adapt to situations such as unstable gradient distribution, limited number of sampling points, and noise interference in real experiments, exhibiting good generalization ability. Experimental results show that the sparse reconstruction method proposed in this invention can achieve high-quality two-dimensional DOSY spectrum reconstruction without adding extra sampling points or using preprocessing operations such as interpolation completion, even with a small number of sampling points in the gradient dimension. This method can fully utilize limited attenuation information to recover the diffusion coefficient distribution, effectively alleviating the resolution degradation and reconstruction instability problems of traditional methods under sparse sampling conditions, and providing an efficient and reliable data processing scheme for component analysis, peak separation, and molecular diffusion characteristic characterization of complex mixtures.

[0090] See Figure 5 As shown in the figure, an embodiment of the present invention provides a sparse reconstruction system for two-dimensional diffusion sorting spectra of nuclear magnetic resonance based on deep learning, comprising:

[0091] Dataset construction module 501 constructs a sparse undersampling simulation training dataset;

[0092] Model building module 502 builds a sparse reconstruction network model and trains the network using a sparse undersampling simulation training dataset.

[0093] DOSY reconstruction module 503 uses a trained sparse reconstruction network model to achieve sparse reconstruction of two-dimensional diffusion sorting spectrum of nuclear magnetic resonance.

[0094] The sparse reconstruction network model includes a gradient dimension decay modeling module and a DOSY spectrum reconstruction module. The gradient dimension decay modeling module uses a two-layer GRU to recursively model the input decay signal to obtain decay features. The DOSY spectrum reconstruction module maps the decay features to a two-dimensional diffusion sorting spectrum.

[0095] Figure 6 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. For example... Figure 6 As shown, the electronic device of this embodiment includes a processor 601 and a memory 602; wherein the memory 602 is used to store computer execution instructions; and the processor 601 is used to execute the computer execution instructions stored in the memory to implement the various steps performed by the electronic device in the above embodiment. For details, please refer to the relevant descriptions in the foregoing method embodiments.

[0096] Alternatively, the memory 602 can be either standalone or integrated with the processor 601.

[0097] When the memory 602 is set up independently, the electronic device also includes a bus 603 for connecting the memory 602 and the processor 601.

[0098] This invention also provides a computer storage medium storing computer execution instructions, which, when executed by a processor, implement the method described above.

[0099] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0100] In the embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0101] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0102] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0103] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0104] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0105] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0106] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0107] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0108] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic device or host device.

[0109] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A sparse reconstruction method for two-dimensional diffusion sorting spectra of nuclear magnetic resonance based on deep learning, characterized in that, Includes the following steps: Construct a sparse undersampling simulation training dataset; A sparse reconstruction network model was built, and the network was trained using a sparse undersampling simulation training dataset. Sparse reconstruction of two-dimensional diffusion sorting spectrum of nuclear magnetic resonance is achieved by using a trained sparse reconstruction network model; The sparse reconstruction network model includes a gradient dimension decay modeling module and a DOSY spectrum reconstruction module. The gradient dimension decay modeling module uses a two-layer GRU to recursively model the input decay signal to obtain decay features. The DOSY spectrum reconstruction module maps the attenuation features into a two-dimensional diffusion sorting spectrum.

2. The sparse reconstruction method for two-dimensional diffusion sorting spectra of nuclear magnetic resonance based on deep learning according to claim 1, characterized in that, The construction of the sparse undersampling simulation training dataset includes the following steps: The gradient sequence is generated parametrically, and is represented as follows: ； ； Among them, normalized variables ; This represents the gradient distribution adjustment coefficient, used to control the distribution form of the gradient sequence; The maximum value of the gradient sequence value; ,in, For the gradient sequence of the th Each sampling point index This represents the total number of gradient sampling points; Construct a two-dimensional DOSY signal, represented as: ； ； in, This represents the spectral peak distribution along the chemical shift dimension. For the first Chemical shift positions of each spectral peak The peak width is spectral width. For the first The amplitude of each peak For the first The diffusion coefficient of each component It is Gaussian noise; Represents chemical shift coordinates, This represents the total number of peaks in the chemical shift dimension. Indicates the total number of diffusing components; Supervised learning labels are constructed to obtain the diffusion spectrum representation, which is expressed as: ； in, The broadening factor indicating the direction of diffusion. To normalize the peak intensity, The diffusion coefficient distribution grid is shown.

3. The sparse reconstruction method for two-dimensional diffusion sorting spectra of nuclear magnetic resonance based on deep learning according to claim 1, characterized in that, The sparse reconstruction network model includes: The gradient dimension decay modeling module adopts a two-layer stacked structure, with each layer containing an update gate and a reset gate. The input of the gradient dimension decay modeling module includes two dimensions. In the gradient dimension, the signal and gradient value corresponding to each gradient point are input into the network as a time step. The current hidden state is determined by the current input and the historical state, thereby realizing the modeling of the overall change characteristics of the decay curve and obtaining the decay characteristics. The DOSY spectral reconstruction module first transforms the attenuation features to a unified feature space through linear mapping, and then inputs them into an attention-based encoding network for global feature modeling. The attention-based encoding network includes a multi-head attention unit, a feedforward neural network, and a normalization structure connected in sequence. The multi-head attention mechanism models the input features by weighting them in multiple subspaces. The feedforward neural network is used to perform nonlinear transformations on the features. The normalization structure is used to stabilize the training process.

4. The sparse reconstruction method for two-dimensional diffusion sorting spectra of nuclear magnetic resonance based on deep learning according to claim 1, characterized in that, The network training includes the following steps: The input is sparse, undersampled DOSY data, and the label is the ideal DOSY reconstructed spectrum. The sparsely undersampled DOSY data is normalized, and the gradient sequence is used as an explicit feature to be paired point by point with the normalized signal to obtain the preprocessed data. The preprocessed data is input into the gradient dimension decay modeling module and the DOSY spectrum reconstruction module to obtain the DOSY reconstruction spectrum result output by the model. The DOSY reconstructed spectrum results are compared with the labels, and the reconstruction error is calculated as a loss function. An adaptive moment estimation optimization algorithm is used to update the model parameters. By adaptively estimating the first and second moments of the gradient, the parameter update step size is dynamically adjusted, thereby improving the optimization performance of the model under sparse data distribution conditions and increasing the convergence speed and stability of the training process.

5. The sparse reconstruction method for two-dimensional diffusion sorting spectra of nuclear magnetic resonance based on deep learning according to claim 1, characterized in that, The method of using a trained sparse reconstruction network model to achieve sparse reconstruction of two-dimensional diffusion-ordered NMR spectra includes the following steps: The original attenuation signal is acquired and normalized along the gradient dimension; The corresponding gradient sequence is used as an explicit input feature and paired point by point with the normalized decay signal to construct a joint input feature. The joint features are input into the gradient dimension decay modeling module to perform sequence modeling in the gradient dimension and extract decay features that can characterize the overall change trend of the sparse decay curve. The sparse attenuation features are input into the DOSY spectrum reconstruction module, and the corresponding two-dimensional DOSY reconstruction spectrum is obtained through feature mapping and global modeling.

6. A sparse reconstruction system for two-dimensional diffusion-ordered nuclear magnetic resonance spectra based on deep learning, characterized in that, include: The dataset construction module builds a sparse undersampling simulation training dataset. The model building module constructs a sparse reconstruction network model and trains the network using a sparse undersampling simulation training dataset. The DOSY reconstruction module utilizes a pre-trained sparse reconstruction network model to achieve sparse reconstruction of two-dimensional diffusion sorting spectra of nuclear magnetic resonance. The sparse reconstruction network model includes a gradient dimension decay modeling module and a DOSY spectrum reconstruction module. The gradient dimension decay modeling module uses a two-layer GRU to recursively model the input decay signal to obtain decay features. The DOSY spectrum reconstruction module maps the attenuation features into a two-dimensional diffusion sorting spectrum.

7. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to perform the method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-5.

Images