A method and system for denoising nuclear magnetic resonance localized spectrum signals based on deep learning

Abstract

The application discloses a kind of nuclear magnetic resonance localized spectrum signal denoising method and system based on deep learning, it is related to spectrum signal denoising field of technique, first, nuclear magnetic resonance localized spectrum signal with noise is acquired, and training-target pair is constructed by zero-mean noise;Introduce self-attention mechanism, embed attention module AttentionBlock into deep neural network model;Based on training-target pair, the deep neural network model is trained, and nuclear magnetic resonance localized spectrum signal denoising model is obtained;With noisy target nuclear magnetic resonance localized spectrum signal is input into the nuclear magnetic resonance localized spectrum signal denoising model, and the nuclear magnetic resonance localized spectrum signal after denoising is obtained.The application starts from the training strategy of noise to noise, can train network under the premise of only relying on paired noise data, and through self-attention mechanism, the denoising ability of neural network is enhanced, and good data recovery effect is obtained.

Description

Technical Field

[0001] This invention relates to the field of spectral signal denoising technology, and more specifically to a method and system for denoising nuclear magnetic resonance localized spectral signals based on deep learning. Background Technology

[0002] Magnetic Resonance Spectroscopy (MRS) is an analytical method based on magnetic resonance imaging (MRI) that uses magnetic resonance to detect the chemical components in biological tissues. Unlike MRI, which primarily provides anatomical information, MRS focuses on metabolic substances and the biochemical environment within tissues, providing molecular information at the tissue or cellular level. Despite its advantages such as being non-invasive and radiation-free, MRS faces several challenges in clinical application. First, the signal-to-noise ratio (SNR) of MRS signals is generally low, affecting signal quality and the accuracy of subsequent analysis. Second, spectral resolution is limited by magnetic field strength; high-intensity magnetic field equipment is expensive and not readily available. While signal averaging can increase the SNR, this is often impractical for clinical research due to the increased number of scans and sensitivity to motion artifacts.

[0003] In traditional NMR localized spectral denoising, the HSVD method, based on singular value decomposition, is used to extract the frequency, attenuation, and phase information of the signal for denoising. This method first converts the time-domain signal into a Hankel matrix, then decomposes the matrix using SVD and separates the signal from noise through threshold filtering, reconstructing the denoised signal using only the signal components. While HSVD has advantages in signal processing, it also has drawbacks: it is sensitive to threshold selection, which may lead to the accidental deletion of signal or the retention of noise; its denoising effect is poor when the signal-to-noise ratio is low or the signal is complex; and it is computationally complex, especially inefficient when processing large datasets.

[0004] One example of deep learning-based denoising is the DnCNN model, which is based on a convolutional neural network (CNN). The DnCNN model denoises signals by learning the difference between noisy and clean signals. First, the training phase requires a large number of noisy signals and their corresponding clean signals. Then, the DnCNN model learns the mapping between these noisy and clean signals. During the testing phase, the DnCNN model can take a noisy signal as input and output a denoised signal that is as close as possible to the clean signal. This is because the DnCNN model has learned how to recover a clean signal from a noisy signal. However, a major drawback of this method is that it requires a large amount of clean signals as a label spectrum for training. This means that without sufficient noisy signals and their corresponding clean signals, it is impossible to train the DnCNN model effectively. Furthermore, this also means that if a new type of noise is encountered in a real-world application, the DnCNN model may not be able to effectively remove this new noise. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for denoising localized nuclear magnetic resonance spectral signals based on deep learning.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A deep learning-based method for denoising localized nuclear magnetic resonance spectral signals includes the following steps:

[0008] Step 1: Obtain the localized nuclear magnetic resonance spectral signal with noise, and construct training-target pairs using zero-mean noise;

[0009] Step 2: Introduce a self-attention mechanism by embedding the AttentionBlock module into the deep neural network model;

[0010] Step 3: Based on the training-target pair, train the deep neural network model with the added attention module AttentionBlock to obtain the nuclear magnetic resonance local spectral signal denoising model.

[0011] Step 4: Input the noisy target NMR localized spectrum signal into the NMR localized spectrum signal denoising model to obtain the denoised NMR localized spectrum signal.

[0012] Optionally, in step 1, the task of minimizing the empirical risk of the training-target pair is: in All are observed spectra with zero-mean noise interference, f θ (·) represents the network mapping relationship under a specific loss function L;

[0013] With clean spectrum y i As a condition, such that: That is, training-target pairs The expectation is a clean spectrum y i .

[0014] Optionally, the deep neural network model adopts a U-Net network structure or a DnCNN network structure.

[0015] Optionally, in step 2, the AttentionBlock is embedded between the DoubleConv module and the linear layer in the traditional U-Net network structure.

[0016] Optionally, the attention module AttentionBlock consists of a multi-head self-attention module (MSA) and a multilayer perceptron (MLP). The MSA is composed of several self-attention modules arranged in parallel, which can extract signal features in multiple dimensions. The MLP includes an input layer, a hidden layer, and an output layer, enabling the network to learn more complex function mappings.

[0017] Optionally, during training in step 3, the noise2noise denoising method is used to train the deep neural network model with the added attention module AttentionBlock.

[0018] Optionally, after obtaining the NMR localized spectrum signal denoising model in step 3, a test dataset is also used to test the NMR localized spectrum signal denoising model.

[0019] A deep learning-based nuclear magnetic resonance localized spectral signal denoising system includes:

[0020] The training data acquisition unit is used to introduce a self-attention mechanism, embedding the attention module AttentionBlock into the deep neural network model;

[0021] The model improvement unit is used to add a Self-Attention structure, embedding the AttentionBlock into the deep neural network model;

[0022] The model training unit is used to train the deep neural network model with the added attention module AttentionBlock based on the training-target pair to obtain the nuclear magnetic resonance local spectral signal denoising model.

[0023] The signal denoising unit is used to input the noisy target NMR localized spectrum signal into the NMR localized spectrum signal denoising model to obtain the denoised NMR localized spectrum signal.

[0024] As can be seen from the above technical solution, the present invention provides a method and system for denoising nuclear magnetic resonance localized spectral signals based on deep learning, which has the following advantages compared with the prior art:

[0025] Existing traditional denoising algorithms cannot effectively recover signals and remove noise, and their computation speed is extremely slow. Furthermore, existing deep learning algorithms rely on large amounts of labeled data to train the network. In contrast, this invention starts with a noise-to-noise training strategy, enabling the network to be trained using only pairs of noisy data. Moreover, through a self-attention mechanism, it enhances the denoising capabilities of traditional neural networks, achieving better data recovery results. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the method steps of the present invention;

[0028] Figure 2 A schematic diagram of the U-Net network structure with self-attention mechanism introduced;

[0029] Figure 3(a) is a schematic diagram of the mean spectrum;

[0030] Figure 3(b) is a schematic diagram of the noise spectrum;

[0031] Figure 3(c) is a schematic diagram of the denoised spectrum. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] This invention discloses a method for denoising localized nuclear magnetic resonance spectral signals based on deep learning. See [link to relevant documentation]. Figure 1 This includes the following steps:

[0034] Step 1: Obtain the localized nuclear magnetic resonance spectrum signal with noise, and construct the training-target pair using zero-mean noise.

[0035] In common denoising tasks, for noisy spectral maps x i and clean spectrum y i The following empirical risks need to be minimized:

[0036]

[0037] Where f θ (·) indicates the network mapping relationship under a specific loss function L. For any training pair (xi|yi), it can be regarded as a noisy input xi~p(xi|yi) is a random variable according to the clean spectrum distribution.

[0038] To estimate the true data from unreliable observations (y1, y2…), a number z with the minimum average bias can be found using a certain loss strategy:

[0039]

[0040] For L2 loss L(z, y) = (zy) 2 The minimum value can be found at the arithmetic mean of the observations:

[0041]

[0042] L1 loss, which is the sum of absolute deviations L(z, y) = |zy|, has a minimum value located at the median of the observations.

[0043] Equation (1) is the basic formula for feedforward network learning. If the point estimation process of equation (2) is extended to the neural network estimation process of lines and surfaces, and the network function f is... θ (x) parameterization, for a set of input-target pairs (xi, yi), yields the objective function for the traditional training task:

[0044]

[0045] For a complete training task, each training session can be broken down into finding a minimization target, i.e.:

[0046]

[0047] In theory, a network can minimize its loss by solving the point estimation problem for each input sample, which lays a foundation for addressing data challenges during network training. However, although each training iteration deduces the mapping model for the current training-target pair (xi, yi), the final result is not a 1:1 mapping between the input and target; in reality, the mapping is multi-valued. For example, in the super-resolution task of MRS, a low-resolution spectral image x can be interpreted using many different high-resolution spectral images y. This leads to uncertainty in the results and a certain error rate because prior knowledge about the exact location and orientation of edges and textures is lost during extraction and cannot be known in the task.

[0048] Therefore, the results of network training are closely related to the utilization of data. Analysis of equation (1) shows that the simplified minimization task and input-target training pair (x) are... i y i The mapping logic is related to the high-resolution spectrum y; the output probability depends on the training logic of the low-resolution spectrum x, where p(y|x) is a highly complex distribution of the natural spectrum consistent with the low-resolution spectrum. As mentioned earlier, using L2 loss yields a minimization task close to the arithmetic mean. For tasks involving denoising MRS using deep learning, based on the p(y|x) mapping logic, the network can learn to output all reasonable super-resolution spectra y. i The average value depends on p(y|x). If there are many reasonable super-resolution spectra, the network output will lead to the blurring of the predicted values, which is not conducive to the super-resolution task.

[0049] Many previous studies have focused on addressing the blurring trend in spectral super-resolution tasks. However, the benefit of L2 loss is that if the target is replaced with a random number that matches the expected value of the target, the network output remains unchanged. In this case, the network output no longer depends on the mapping p(y|x). Therefore, if the input conditional target distribution p(y|x) is replaced with any distribution with the same conditional expected value, the optimal network parameters in equation (5) also remain unchanged.

[0050] This is the foundation for the implementation of noise2noise denoising theory. In principle, zero-mean noise is used to construct the training-target pairs of the neural network without changing the network's learning content. The empirical risk minimization task of equation (1) is then transformed into:

[0051]

[0052] in All observed spectra are subject to zero-mean noise interference, with a potential, unobserved clean spectrum y. i As a condition, such that:

[0053]

[0054] That is, training-target pairs A clean spectrum is desired.

[0055] It is worth noting that the above network training task no longer focuses on the model damaged by noise interference, nor does it require clean data as labels as before noise2clean. It only requires that the distribution of noise data meets the requirements of equation (7).

[0056] Step 2: Introduce a self-attention mechanism by embedding the AttentionBlock module into the deep neural network model. The AttentionBlock module consists of a multi-head self-attention module (MSA) and a multilayer perceptron (MLP). The MSA module is composed of several self-attention modules arranged in parallel. The MLP module includes an input layer, a hidden layer, and an output layer.

[0057] The deep neural network model employs either the U-Net or DnCNN network architecture. Taking the U-Net network architecture as an example, the AttentionBlock module is embedded between the Double Conv module and the linear layer in the traditional U-Net network structure. (See [link to documentation]). Figure 2 The leftmost image shows the original noisy data, the middle image shows the overall structure of the network, and the rightmost image shows the AttentionBlock and Double Conv modules.

[0058] The AttentionBlock module has the following functions:

[0059] The AttentionBlock module is mainly used to enhance the model's attention to and understanding of the input data.

[0060] Selective focus on important frequency components: In local spectral data, different frequency components contribute differently to the characteristics of the signal. The AttentionBlock module helps the model learn the importance of different frequency components, enabling the model to focus on the frequency components that are more critical to the current task, while ignoring irrelevant frequency components.

[0061] Dynamically Adjusting Spectral Feature Weights: Local spectral data typically contains multiple spectral features, the importance of which may vary across different signals. By introducing an attention module (AttentionBlock), the weights of each spectral feature can be dynamically adjusted, allowing the model to adaptively select and combine spectral features based on the characteristics of the current signal. Combining Global and Local Information: Local spectral analysis usually focuses on the local frequency characteristics of a signal, but global information is also important for some tasks.

[0062] The AttentionBlock module can combine global and local information, enabling the model to capture both the overall and local features of the signal, thereby improving the model's denoising capabilities.

[0063] Step 3: Based on the training-target pair, use the noise2noise denoising method to train the deep neural network model with the added attention module AttentionBlock to obtain the nuclear magnetic resonance local spectral signal denoising model.

[0064] Furthermore, after obtaining the NMR localized spectrum signal denoising model in step 3, a test dataset can be acquired and the NMR localized spectrum signal denoising model can be tested to determine whether the NMR localized spectrum signal denoising model achieves the expected denoising effect.

[0065] Step 4: Input the noisy target NMR localized spectrum signal into the NMR localized spectrum signal denoising model to obtain the denoised NMR localized spectrum signal.

[0066] See Figures 3(a)-3(c) This is a schematic diagram of the mean spectrum, noise spectrum, and denoised spectrum of the spectral signal data.

[0067] See Table 1 below for the RE and SD values ​​of Glx, tNAA, tCho, and Ins in three cases: simulated spectrum, in vivo spectrum, and semi-simulated spectrum, under three datasets: denoised spectrum, noisy spectrum, and reference spectrum.

[0068] Table 1

[0069]

[0070]

[0071] In Table 1, the parameters are explained as follows:

[0072] Simulated spectrum: An MRS spectrum artificially simulated based on actual physical parameters. In the simulated spectrum, the directly generated spectrum is the reference spectrum. Then, Gaussian noise is added to achieve a signal-to-noise ratio of approximately 16, resulting in a noisy spectrum. After denoising via a network, the resulting spectrum is denoised.

[0073] In-vivo spectrum: The noisy spectrum is directly acquired by the instrument. After 64 superpositions, the reference spectrum is obtained. The spectrum with a signal-to-noise ratio of about 16 for a single noisy spectrum is called the noise spectrum. After network denoising, the denoised spectrum is obtained.

[0074] Semi-simulated spectrum: The reference spectrum of the semi-simulated spectrum is the same as that of the living organism, while the noise spectrum is artificially added to the spectrum with a signal-to-noise ratio of about 16. After being denoised by the network, the denoised spectrum is obtained.

[0075] Glx refers to the complex of glutamate (Glu) and glutamine (Gln) in brain spectra. In MRS, because the peak positions of Glu and Gln are very close, they are often analyzed together and referred to as Glx.

[0076] tNAA: The full name is "Total N-acetylaspartate," which typically includes N-acetylaspartate (NAA) and some of its metabolites such as N-acetylaspartylglutamate (NAAG). tNAA is used in MRS as a marker of neuronal integrity and density.

[0077] tCho: The full name is "Total Choline," which in MRS refers to the sum of choline (Cho) and its metabolites. tCho is used as an indicator of cell membrane metabolism.

[0078] Ins refers to inositol, specifically a form of inositol in the MRS—myo-inositol monophosphate. It is often associated with cell permeability or glial cell activity.

[0079] Relative Error (RE): This refers to the ratio of the difference between the measured value and the true value to the true value. It is an indicator used to represent measurement accuracy.

[0080]

[0081] SD (Standard Deviation): Standard deviation is a statistical measure of how widely a set of values ​​are distributed around the mean. In experimental data analysis such as MRS, standard deviation can be used to indicate the repeatability or consistency of the data.

[0082]

[0083] Where N is the number of samples, x i These are the values ​​for each sample. It is the sample mean. The smaller the standard deviation, the better the data consistency; conversely, the larger the standard deviation, the greater the fluctuation.

[0084] As shown in Table 1, the denoising method of the present invention has achieved significant denoising effect in all datasets, including simulated spectrum, live spectrum, and semi-simulated spectrum datasets.

[0085] Another embodiment of the present invention provides a deep learning-based nuclear magnetic resonance localized spectral signal denoising system, comprising:

[0086] The training data acquisition unit is used to acquire noisy nuclear magnetic resonance localized spectral signals and construct training-target pairs using zero-mean noise.

[0087] The model improvement unit is used to introduce a self-attention mechanism, embedding the attention module AttentionBlock into the deep neural network model;

[0088] The model training unit is used to train the deep neural network model with the added attention module AttentionBlock based on the training-target pair to obtain the nuclear magnetic resonance local spectral signal denoising model.

[0089] The signal denoising unit is used to input the noisy target NMR localized spectrum signal into the NMR localized spectrum signal denoising model to obtain the denoised NMR localized spectrum signal.

[0090] The system units disclosed in the embodiments are described in a relatively simple manner since they correspond to the methods disclosed in the embodiments. For relevant details, please refer to the description in the method section.

[0091] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0092] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for denoising nuclear magnetic resonance localized spectral signals based on deep learning, characterized in that, Includes the following steps: Step 1: Obtain the localized nuclear magnetic resonance spectral signal with noise, and construct training-target pairs using zero-mean noise; Step 2: Introduce a self-attention mechanism by embedding the AttentionBlock module into a deep neural network model, wherein the deep neural network model adopts a U-Net network structure or a DnCNN network structure. An attention module, AttentionBlock, is embedded between the Double Conv module and the linear layer in the traditional U-Net network structure. The attention module, AttentionBlock, consists of a multi-head self-attention module (MSA) and a multilayer perceptron (MLP). The multi-head self-attention module (MSA) is composed of several self-attention modules arranged in parallel. The multilayer perceptron (MLP) includes an input layer, a hidden layer, and an output layer. Step 3: Based on the training-target pair, train the deep neural network model with the added attention module AttentionBlock to obtain the nuclear magnetic resonance local spectral signal denoising model. During training, noise2noise denoising was used to train the deep neural network model with the added AttentionBlock module; Step 4: Input the noisy target NMR localized spectrum signal into the NMR localized spectrum signal denoising model to obtain the denoised NMR localized spectrum signal.

2. The method for denoising nuclear magnetic resonance localized spectral signals based on deep learning according to claim 1, characterized in that, In step 1: The task of minimizing the empirical risk of the training-target pair is: ,in The observed spectra are all subject to zero-mean noise interference, representing the network mapping relationship under a specific loss function L; With clean spectrum As a condition, such that: That is, training-target pairs A clean spectrum is expected. y i .

3. The method for denoising nuclear magnetic resonance localized spectral signals based on deep learning according to claim 1, characterized in that, After obtaining the NMR localized spectrum signal denoising model in step 3, a test dataset is used to test the NMR localized spectrum signal denoising model.

4. A system for denoising nuclear magnetic resonance localized spectral signals based on deep learning as described in any one of claims 1-3, characterized in that, include: The training data acquisition unit is used to acquire noisy nuclear magnetic resonance localized spectral signals and construct training-target pairs using zero-mean noise. The model improvement unit is used to introduce a self-attention mechanism, embedding the attention module AttentionBlock into the deep neural network model; The model training unit is used to train the deep neural network model with the added attention module AttentionBlock based on the training-target pair to obtain the nuclear magnetic resonance local spectral signal denoising model. The signal denoising unit is used to input the noisy target NMR localized spectrum signal into the NMR localized spectrum signal denoising model to obtain the denoised NMR localized spectrum signal.

Images