High-field CEST MRI Z-spectrum synthesis method based on deep learning and singular value decomposition

Through deep learning and singular value decomposition technology, low-field CEST MRI Z-spectrum data are converted into high-field CEST MRI Z-spectrum data, solving the problems of signal overlap and high equipment costs under clinical field strength, and achieving efficient chemical substance identification and Quantitative, providing rich diagnostic data.

CN119064834APending Publication Date: 2024-12-03TSINGHUA UNIVERSITY +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410953469.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

CEST imaging under clinical field strength faces challenges such as signal overlap, high equipment costs, and risk of radio frequency energy deposition, which leads to difficulties in obtaining high-field CEST data, affecting the accurate identification and quantification of specific chemical substances.

Method used

Using the high-field CEST MRI Z-spectrum synthesis method based on deep learning and singular value decomposition, the low-field CEST MRI Z-spectrum data is converted into high-field CEST MRI Z-spectrum data through pre-trained high-field CEST MRI Z-spectrum data, using depth The neural network performs B0 correction and feature distillation, and singular value decomposition is performed for dimensionality reduction and inverse transformation to generate target high-field CEST MRI Z-spectrum data.

Benefits of technology

It provides richer and more accurate clinical diagnostic data, reduces equipment costs, improves the resolution and quantitative capabilities of specific chemicals, and overcomes the equipment and signal analysis difficulties of high-field CEST imaging.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119064834A_ABST
    Figure CN119064834A_ABST
Patent Text Reader

Abstract

The invention discloses a high-field CEST MRI Z-spectrum synthesis method based on deep learning and singular value decomposition. The method comprises the following steps: acquiring target low-field CEST MRI Z-spectrum data; a pre-trained high-field Z-spectrum data generation model is utilized, target high-field CEST MRI Z-spectrum data is generated according to the target low-field CEST MRI Z-spectrum data, the high-field Z-spectrum data generation model comprises a deep neural network DNN-based B0 correction module, a DNN-based feature distillation module and a singular value decomposition SVD truncation module, the B0 correction module is used for carrying out B0 correction on the low-field CEST MRI Z-spectrum data, the feature distillation module is used for carrying out feature distillation on the low-field CEST MRI Z-spectrum data, and the singular value decomposition SVD truncation module is used for carrying out feature distillation on the low-field CEST MRI Z-spectrum data. The feature distillation module is used for performing feature distillation on the target low-field CEST MRI Z-spectrum data subjected to B0 correction according to the dimension of the target dimension reduction feature expression to obtain a target high-field feature expression, the SVD truncation module is used for performing SVD inverse transformation on the target high-field feature expression to obtain target high-field CEST MRI Z-spectrum data, and the dimension of the target dimension reduction feature expression is determined by the SVD truncation module. The method is suitable for clinical CEST diagnosis and has the advantage of low cost.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of magnetic resonance imaging technology, and in particular to a high-field CEST MRI Z-spectrum synthesis method based on deep learning and singular value decomposition. Background Art

[0002] Chemical Exchange Saturation Transfer (CEST) imaging is an important medical imaging technology used to evaluate the molecular characteristics of biological tissues. Unlike traditional magnetic resonance imaging (MRI), which only images free hydrogen atoms, CEST imaging can selectively apply saturation pulses to the specific resonance frequencies of hydrogen atoms in different groups, and indirectly obtain multi-solute molecular information by detecting the changes in water signal intensity caused by saturation transfer, thereby achieving accurate diagnosis of diseases.

[0003] However, CEST imaging faces some challenges in clinical applications. Under clinical field strength, due to the presence of multiple chemical substances in human tissues, the attenuation of water signal intensity during the saturation process is simultaneously affected by multiple competing effects, such as direct water saturation effect (DS), magnetization transfer contrast (MTC), and nuclear Overhauser effect (NOE), so that the signal components in the Z spectrum obtained by CEST imaging usually show overlapping peaks, making it difficult to accurately identify and quantify specific chemical substances. Compared with clinical field strength, CEST imaging under high field strength can obtain better isolated signal peaks and greater signal intensity, improving the resolution and quantitative ability of specific chemical substances. However, due to the high cost of high-field CEST imaging equipment, the risk of radio frequency energy deposition, and the inhomogeneity problems associated with high magnetic fields, it becomes challenging to obtain high-quality high-field CEST data in clinical applications. Summary of the invention

[0004] The present invention aims to solve one of the technical problems in the related art at least to a certain extent. To this end, one object of the present invention is to propose a high-field CEST MRI Z-spectrum synthesis method based on deep learning and singular value decomposition, which provides richer and more accurate data for clinical diagnosis and research, and has the advantage of low cost.

[0005] A second objective of the present invention is to provide an electronic device.

[0006] To achieve the above-mentioned purpose, the first aspect of the present invention proposes a high-field CEST MRI Z spectrum synthesis method based on deep learning and singular value decomposition, the method comprising: obtaining target low-field CEST MRI Z spectrum data; using a pre-trained high-field Z spectrum data generation model to generate target high-field CESTMRI Z spectrum data according to the target low-field CEST MRI Z spectrum data, wherein the high-field Z spectrum data generation model comprises a B0 correction module based on a deep neural network DNN, a feature distillation module based on the DNN, and a singular value decomposition SVD truncation module, the B0 correction module is used to perform B0 correction on the low-field CESTMRI Z spectrum data, the feature distillation module is used to perform feature distillation on the target low-field CEST MRI Z spectrum data after B0 correction according to the dimension of the target dimensionality reduction feature expression, to obtain the target high-field feature expression, the SVD truncation module is used to perform SVD inverse transform on the target high-field feature expression, to obtain the target high-field CEST MRI Z spectrum data, and the dimension of the target dimensionality reduction feature expression is determined by the SVD truncation module.

[0007] According to the high-field CEST MRI Z spectrum synthesis method based on deep learning and singular value decomposition in an embodiment of the present invention, a pre-trained high-field Z spectrum data generation model is used to convert target low-field CEST MRI Z spectrum data acquired under clinical field strength into target high-field CEST MRI Z spectrum data, and based on the characteristics of the target high-field CEST MRI Z spectrum data, accurate CEST diagnosis of target biological tissue is performed clinically, providing richer and more accurate data and low cost advantages for clinical diagnosis and research.

[0008] In addition, the high-field CEST MRIZ spectrum synthesis method based on deep learning and singular value decomposition proposed in the above embodiment of the present invention may also have the following additional technical features:

[0009] According to one embodiment of the present invention, the training method of the high-field Z spectrum data generation model includes a training process, and the training process includes: obtaining a training data set, wherein the training data set includes a plurality of simulated low-field CESTMRI Z spectrum data and a one-to-one corresponding plurality of simulated high-field CEST MRI Z spectrum data, and a plurality of simulated low-field CEST MRI Z spectra perturbed by B0 and a one-to-one corresponding plurality of B0 perturbation actual values; establishing the high-field Z spectrum data generation model; inputting the simulated low-field CEST MRI Z spectrum data perturbed by B0 into a B0 correction module of the current training high-field Z spectrum data generation model, predicting the B0 perturbation value corresponding to the current training low-field CEST MRI Z spectrum data, and using the B0 perturbation prediction value to correct the current training low-field CEST MRI Z spectrum data, inputting the corrected current training low-field CEST MRI Z spectrum data into a feature distillation module of the current training high-field Z spectrum data generation model, and outputting the current training low-field CEST MRI A training high-field feature expression corresponding to the Z spectrum data; a first loss function is obtained according to the actual value of the B0 disturbance and the predicted value of the B0 disturbance, and the network parameters of the B0 correction module of the current training high-field Z spectrum data generation model are adjusted using the first loss function to perform the next iteration until the first loss function is less than a first preset threshold and does not decrease within the first preset iteration range; a second loss function is obtained according to the target high-field feature expression and the high-field feature expression output by the feature distillation module, and the network parameters of the feature distillation module of the current training high-field Z spectrum data generation model are adjusted using the second loss function to perform the next iteration until the second loss function is less than a second preset threshold and does not decrease within the second preset iteration range.

[0010] According to one embodiment of the present invention, the training method of the high-field Z spectrum data generation model also includes a testing process, which includes: obtaining a test data set, wherein the test data set includes a plurality of collected low-field CEST MRI Z spectrum data and their one-to-one collected high-field CEST MRI Z spectrum data; inputting the collected low-field CEST MRI Z spectrum data into the B0 correction module of the current test high-field Z spectrum data generation model, predicting the B0 disturbance value corresponding to the current test low-field CEST MRI Z spectrum data, and using the B0 disturbance prediction value to correct the current test low-field CEST MRI Z spectrum data, inputting the corrected current test low-field CEST MRI Z spectrum data into the feature distillation module of the current test high-field Z spectrum data generation model, and outputting the test high-field feature expression corresponding to the current test low-field CEST MRI Z spectrum data, wherein the first test high-field Z spectrum data generation model is a high-field Z spectrum data generation model trained using the training data set.

[0011] According to one embodiment of the present invention, the training method of the high-field Z spectrum data generation model also includes an evaluation process, which includes: performing an inverse SVD transform using the test high-field feature expression and the target dimensionality reduction feature expression to obtain predicted high-field CEST MRI Z spectrum data; obtaining a third loss function based on the predicted high-field CEST MRI Z spectrum data and its corresponding collected high-field CEST MRI Z spectrum data to evaluate the performance of the high-field Z spectrum data generation model.

[0012] According to one embodiment of the present invention, the target low-field CEST MRI Z spectrum data after B0 correction is subjected to feature distillation to obtain the target high-field feature expression, including: performing frequency offset scaling and normalization processing on the target low-field simulated CEST MRI Z spectrum data after B0 correction, and then inputting it into the deep neural network corresponding to the feature distillation module, and outputting the target high-field feature expression.

[0013] According to one embodiment of the present invention, the expression for the frequency offset scaling normalization process is:

[0014]

[0015] Among them, Z m,n represents the signal value at the nth frequency offset in the Z spectrum, Z m,n The normalized result, Z n It represents the signal value of all training data under the same frequency offset. min(·) means to find the minimum value, and max(·) means to find the maximum value.

[0016] According to one embodiment of the present invention, the SVD truncation module uses a singular value decomposition method to decompose the high-field CEST MRI Z spectrum data obtained by simulation, and uses a cumulative variance explanation ratio to determine the dimension of the target dimensionality reduction feature expression according to the singular value obtained by decomposition, wherein the expression for calculating the cumulative variance explanation ratio is:

[0017]

[0018] Among them, CEV n is the cumulative variance contribution rate of the nth eigenvalue, S i is the i-th eigenvalue.

[0019] According to one embodiment of the present invention, the acquisition process of the test data set includes: using a 3T scanner to scan a phantom test sample and a positioning sample at the same time to obtain first low-field data; using a 9.4T scanner to scan the phantom test sample respectively to obtain multiple second high-field data; processing and normalizing the first low-field data and the multiple second high-field data; extracting a region of interest from the normalized first low-field data to obtain the low-field CEST MRI Z spectrum data; down-sampling the multiple normalized second high-field data respectively, and performing B0 correction and denoising on the high-field Z spectrum data obtained by downsampling to obtain the high-field CEST MRI Z spectrum data, wherein the phantom test sample includes an egg white phantom placed in a glass tube of preset specifications and a phosphocreatine phantom of different concentrations prepared in phosphate buffered saline of different pH in a glass tube of preset specifications, and the positioning sample includes phosphate buffered saline of pH=7 placed in a glass tube of preset specifications.

[0020] According to one embodiment of the present invention, the process of acquiring the training data set includes: using a multi-pool model of the Bloch equation to simulate and generate low-field CEST MRI Z spectrum data and its one-to-one corresponding high-field CEST MRI Z spectrum data, and performing B0 perturbation on the simulated low-field CEST MRI Z spectrum data.

[0021] To achieve the above-mentioned purpose, the second aspect of the present invention proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the CEST MRI Z spectrum synthesis method proposed in the first aspect of the present invention is implemented.

[0022] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1is a flow chart of a high-field CEST MRI Z-spectrum synthesis method according to an embodiment of the present invention;

[0024] Figure 2 is a flow chart of a high-field CEST MRI Z-spectrum algorithm according to a specific embodiment of the present invention;

[0025] Figure 3 is a training flow chart of a high-field Z-spectrum data generation model according to an embodiment of the present invention;

[0026] Figure 4 is a training flow chart of a B0 correction module according to a specific embodiment of the present invention;

[0027] Figure 5 is a schematic diagram of an electronic device according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0028] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0029] The following is attached to the instruction manual Figure 1-5 And the specific implementation method describes in detail the high-field CEST MRI Z spectrum synthesis method and electronic device based on deep learning and singular value decomposition of the embodiment of the present invention.

[0030] Figure 1 FIG. 1 is a flow chart of a high-field CEST MRI Z-spectrum synthesis method according to an embodiment of the present invention. Figure 1 As shown, the high-field CEST MRI Z-spectrum synthesis method based on deep learning and singular value decomposition may include:

[0031] S101, acquiring target low-field CEST MRI Z-spectrum data;

[0032] S102, using a pre-trained high-field Z spectrum data generation model, generate target high-field CEST MRI Z spectrum data according to the target low-field CEST MRI Z spectrum data, wherein the high-field Z spectrum data generation model includes a B0 correction module based on a deep neural network DNN, a feature distillation module based on the DNN, and a singular value decomposition SVD truncation module, the B0 correction module is used to perform B0 correction on the low-field CEST MRI Z spectrum data, the feature distillation module is used to perform feature distillation on the target low-field CEST MRI Z spectrum data after B0 correction according to the dimension of the target dimensionality reduction feature expression to obtain the target high-field feature expression, the SVD truncation module is used to perform SVD inverse transformation on the target high-field feature expression to obtain the target high-field CEST MRI Z spectrum data, and the dimension of the target dimensionality reduction feature expression is determined by the SVD truncation module.

[0033] It should be noted that studies have shown that CEST MRI Z spectrum at clinical field strength contains a wealth of information from different sources, but it is difficult to extract because of the confusion. CEST MRI Z spectrum data at high field strength has good peak separation characteristics, which helps to accurately interpret CEST signals.

[0034] The embodiment of the present invention utilizes the CEST MRI Z spectrum characteristics under clinical field strength to establish and train a high-field Z spectrum data generation model, so that the trained high-field Z spectrum data generation model can quickly and accurately generate high-quality CEST MRI Z spectrum data under high field strength according to the CEST MRI Z spectrum data under clinical field strength, thereby utilizing the advantages of the CEST MRI Z spectrum data under high field strength to help clinicians perform accurate CEST diagnosis.

[0035] It is feasible to perform CEST magnetic resonance imaging on the target biological tissue under clinical field strength to obtain target low-field CEST MRI Z spectrum data. The target low-field CEST MRI Z spectrum data is input into the B0 correction module in the pre-trained high-field Z spectrum data generation model, the B0 correction module performs B0 correction on the target low-field CEST MRI Z spectrum data, the corrected target low-field CEST MRI Z spectrum data is input into the feature distillation in the pre-trained high-field Z spectrum data generation model, the feature distillation performs feature extraction on the target low-field CEST MRI Z spectrum data, and outputs a k-dimensional target high-field feature expression, the SVD truncation module performs SVD inverse transformation on the k-dimensional target high-field feature expression, and the pre-trained high-field Z spectrum data generation model outputs the target high-field CEST MRI Z spectrum data of the target biological tissue, thereby improving the accuracy of the interpretation and quantification of the low-field CEST signal.

[0036] It should be noted that the dimension k of the target dimensionality reduction feature expression is determined by the SVD truncation module, and when training the high-field Z spectrum data generation model, the high-field CEST MRI Z spectrum data in the training data set is subjected to singular value decomposition to determine it. Feature distillation generates a k-dimensional target high-field feature expression based on k and the target low-field CEST MRI Z spectrum data.

[0037] The high-field Z spectrum data generation model of the embodiment of the present invention adopts a deep learning-based framework to synthesize high-field CEST MRI Z spectrum data based on low-field CEST MRI Z spectrum data. The core of the framework is to extract key features from low-field CEST MRI Z spectrum data and convert them into high-field CEST MRI Z spectrum data through deep neural networks (Deep Neural Networks, DNN) and singular value decomposition SVD technology. The high-field Z spectrum data generation model of the embodiment of the present invention converts the task of generating a high-dimensional high-field Z spectrum into two subtasks: generating a low-dimensional high-field feature expression and restoring the low-dimensional high-field feature expression to a high-dimensional complete high-field Z spectrum through the SVD inverse transformation method.

[0038] The high-field Z spectrum data generation model established in the embodiment of the present invention includes a DNN-based B0 correction module, a DNN-based feature distillation module, and a singular value decomposition (SVD) truncation module.

[0039] Since CEST MRI Z spectrum data is very susceptible to B0 inhomogeneity, resulting in frequency mismatch and Z spectrum distortion, these interferences bring inaccuracy to any task that relies on the Z spectrum. B0 correction of related CEST MRI Z spectrum data requires the use of water saturation offset reference (WASSR) or Dixon and other methods to obtain an additional B0 offset map. In order to simplify the B0 correction process, an embodiment of the present invention uses a DNN-based B0 correction module to optimize the frequency mismatch and spectral distortion of low-field CEST MRI Z spectrum data caused by B0 inhomogeneity, ensuring the accuracy of subsequent processing.

[0040] The B0 correction module in the embodiment of the present invention is implemented by DNN. The DNN adopts a fully connected neural network structure, that is, a multi-layer perceptron (MLP). The DNN receives low-field CEST MRI Z spectrum data and outputs an accurate B0 offset prediction value of the low-field CEST MRI Z spectrum data. The B0 correction module corrects the low-field CEST MRI Z spectrum data according to the B0 offset prediction value, aligns the frequency deviation of the low-field CEST MRI Z spectrum data with the correct frequency deviation, so that the subsequent feature extraction task is more accurate.

[0041] In an embodiment of the present invention, a DNN-based feature distillation module is used to convert low-field CEST MRI Z-spectral data into a k-dimensional feature expression of high-field CEST MRI Z-spectral data.

[0042] The feature distillation module in the embodiment of the present invention is implemented using a DNN. The DNN extracts feature expressions from the low-field Z spectrum data that has been corrected for B0 inhomogeneity, and outputs a k-dimensional feature expression of the high-field CEST MRI Z spectrum data. The k-dimensional feature expression contains important information that is similar to the high-field data. The feature distillation module is also implemented using a DNN with an MLP structure.

[0043] The embodiment of the present invention uses a frequency offset scaling normalization technique before the DNN of the feature distillation module to capture subtle changes in low-field CEST MRI Z-spectral data.

[0044] In an embodiment of the present invention, the SVD truncation module is used to reduce the feature dimension of the high-field CEST MRI Z spectrum data, determine the target dimension reduction feature expression, and simplify the subsequent prediction task. And when the k-dimensional feature expression of the high-field CEST MRI Z spectrum data input by the feature distillation module is received, the k-dimensional feature expression is subjected to an SVD inverse transformation to accurately restore the complete high-field CEST MRI Z spectrum data to obtain the target high-field CEST MRI Z spectrum data.

[0045] When determining the target dimensionality reduction feature expression, the SVD truncation module uses the singular value decomposition method to reduce the feature dimension of the high-field CESTMRI Z spectrum data. Specifically, the singular value decomposition method is used to decompose the high-field CEST MRI Z spectrum data in the training data set into singular values, left singular vectors and right singular vectors. Based on the singular value decomposition results, the most important singular values ​​and corresponding singular vectors are selected to reduce the feature dimension. By selecting the retained singular values ​​and singular vectors, it is ensured that key information still exists, thereby maintaining the core structure and information integrity of the high-field CEST MRI Z spectrum data, which helps to simplify the data and reduce the difficulty of the task of regressing high-dimensional parameters faced by the subsequent generation model. Among them, the singular value decomposition of the high-field CESTMRI Z spectrum data in the training data set helps to understand the main features and structure of the high-field CEST MRI Z spectrum data.

[0046] The high-field Z spectrum data generation model of the embodiment of the present invention adopts the SVD truncation module, which avoids the challenge of regressing high-dimensional parameters faced by the neural network when directly synthesizing the complete high-field Z spectrum. The introduction of the SVD truncation module provides a simpler, low-dimensional output target for the prediction task of the subsequent feature distillation module.

[0047] Figure 2FIG. 4 is a workflow of a high-field CEST MRI Z-spectrum generation model according to an embodiment of the present invention. Figure 2 As shown, the high-field CEST MRI Z spectrum generation model goes through two stages: training and testing. The training process includes three modules: B0 correction module, feature distillation module and SVD truncation module. During the training process, all networks in the B0 correction module, feature distillation module and SVD truncation module use the multi-pool model based on the Bloch equation (Bloch-McConnell equations, BMEs) to generate simulated high- and low-field paired CEST MRI Z spectrum data for training and network hyperparameter optimization. The testing process includes two modules: B0 correction module and feature distillation module. During the testing process, the B0 correction module and the feature distillation module are tested using high- and low-field paired CEST MRI Z spectrum data collected and processed by high and low scanning equipment.

[0048] All training processes of the high-field CEST MRI Z-spectrum generation model of the embodiment of the present invention are implemented using custom Python codes within the PyTorch framework.

[0049] In one embodiment of the present invention, Figure 3 As shown, the training method of the high-field Z-spectrum data generation model includes a training process, and the training process may include:

[0050] S201, obtaining a training data set, wherein the training data set includes a plurality of simulated low-field CEST MRI Z spectrum data and a one-to-one corresponding plurality of simulated high-field CEST MRI Z spectrum data, and a plurality of simulated low-field CEST MRI Z spectrum data perturbed by B0 and a plurality of one-to-one corresponding actual values ​​of B0 perturbation.

[0051] Specifically, the parameters of the set field intensity pairing are simulated using a multi-pool model based on the Bloch-McConnell equations (BMEs) to obtain high- and low-field paired CEST MRI Z spectrum data. The simulated low-field CEST MRI Z spectrum data is subjected to B0 perturbation, wherein the B0 offset range is set to be randomly selected from -1 to 1 ppm, and the actual value of the B0 perturbation corresponding to each low-field CEST MRI Z spectrum data is recorded.

[0052] S202, establishing a high-field Z-spectrum data generation model.

[0053] Specifically, a high-field Z-spectrum data generation model is established, which includes a DNN-based B0 correction module, a DNN-based feature distillation module, and a singular value decomposition (SVD) truncation module.

[0054] S203, input the low-field CEST MRI Z spectrum data obtained by simulation of B0 perturbation into the B0 correction module of the current training high-field Z spectrum data generation model, predict the B0 perturbation value corresponding to the current training low-field CEST MRI Z spectrum data, and use the B0 perturbation prediction value to correct the current training low-field CEST MRI Z spectrum data, input the corrected current training low-field CEST MRI Z spectrum data into the feature distillation module of the current training high-field Z spectrum data generation model, and output the training high-field feature expression corresponding to the current training low-field CEST MRI Z spectrum data.

[0055] S204, obtaining a first loss function according to the actual value of the B0 disturbance and the predicted value of the B0 disturbance, and using the first loss function to adjust the network parameters of the B0 correction module of the current training high-field Z spectrum data generation model for the next iteration, until the first loss function is less than a first preset threshold and does not decrease within the first preset iteration range.

[0056] S205, obtaining a second loss function according to the target high-field feature expression and the high-field feature expression output by the feature distillation module, and using the second loss function to adjust the network parameters of the feature distillation module of the current training high-field Z spectrum data generation model for the next iteration until the second loss function is less than a second preset threshold and does not decrease within the second preset iteration range.

[0057] In one embodiment of the present invention, the SVD truncation module uses a singular value decomposition method to decompose the simulated high-field CEST MRI Z spectrum data, and uses the cumulative variance explanation ratio to determine the dimension of the target dimensionality reduction feature expression according to the singular values ​​obtained by the decomposition, wherein the expression for calculating the cumulative variance explanation ratio is:

[0058]

[0059] Among them, CEV n is the cumulative variance contribution rate of the nth eigenvalue, S i is the i-th eigenvalue.

[0060] In the embodiment of the present invention, the SVD truncation module uses the singular value decomposition method to decompose the high-field CEST MRI Z spectrum data in the training data set, and uses the cumulative variance explanation ratio to determine the dimension k of the target dimensionality reduction feature expression according to the singular value obtained by the decomposition.

[0061] Specifically, the high-field CEST MRI Z spectrum data in the training data set is represented as a matrix Z with a dimension of M×N, wherein M represents the number of samples of the high-field CEST MRI Z spectrum data in the training data set, and N corresponds to the number of frequency offsets (Δω) contained in the Z spectrum data.

[0062] The SVD truncation module uses singular value decomposition to decompose the matrix Z into the product of three matrices:

[0063] Z m×N =U M×N ×S N×N ×V N×N

[0064] Among them, U M×N and V N×N are the left and right singular vectors of the matrix Z respectively.

[0065] U M×N Map Z to the new principal component space, and V N×N These principal components are mapped back to the original feature space. The matrix S represents the singular values ​​(eigenvalues) of Z, which quantifies the importance of each principal component. The SVD truncation module sorts the eigenvalues ​​in descending order according to their contribution to the data. By retaining the first k eigenvalues ​​and their corresponding eigenvectors, it is possible to retain the core structure and information of the Z spectrum data while reducing the data dimension.

[0066] The embodiment of the present invention uses a cumulative explained variance (CEV) graph to determine the k value. In the cumulative explained variance graph, the x-axis represents the number of features retained, and the y-axis represents the cumulative variance contribution. The calculation formula of CEV is:

[0067]

[0068] Among them, CEV n is the cumulative variance contribution rate of the nth eigenvalue, S i is the i-th eigenvalue.

[0069] The embodiment of the present invention selects the value of n that reaches 99% of the cumulative variance contribution rate as k. M×N , S N×N and V N×N After truncation, the reduced-dimensional feature expression is obtained, that is, the truncated eigenvalue matrix S' and its corresponding eigenvectors U' and V'. These components can be approximately restored to Z by inverse SVD transformation M×N .

[0070] ″′

[0071] Z M×N ≈U M×k ×S k×k ×V k×N

[0072] Among them, U' is used as the target of feature extraction.

[0073] Specifically, when training the high-field Z spectrum data generation model, the low-field CEST MRI Z spectrum data that has been perturbed by B0 is input into the B0 correction module of the current training high-field Z spectrum data generation model. The B0 correction module of the current training high-field Z spectrum data generation model predicts the B0 perturbation value of the low-field CEST MRI Z spectrum data and outputs the B0 perturbation prediction value, see Figure 4 According to the actual value of B0 disturbance and the predicted value of B0 disturbance corresponding to the low-field CEST MRI Z spectrum data, a first loss function is obtained, and the first loss function is used to adjust the network parameters of the B0 correction module of the current training high-field Z spectrum data generation model for the next iteration. During the iteration process, when the first loss function is less than 0.01 and does not decrease within 30 generations, the trained B0 correction module is obtained.

[0074] Among them, when training the B0 correction module, the hidden layer of the ΔB0 network (CNN that implements the B0 correction task) uses a linear function to keep the scale of input and output in the range of [-1, 1]. In order to improve convergence, the hidden layer integrates BatchNorm1d and ReLU activation functions, and the output layer uses Sigmoid activation function to generate the final prediction. The ΔB0 network is trained by back propagation, and a series of hyperparameters such as the number of hidden layers, the number of neurons in each layer, the regularization factor, the dropout rate, and the loss function are optimized by the grid search method.

[0075] The low-field CEST MRI Z spectrum data in the training data set are corrected using the trained B0 correction module, and the corrected low-field CEST MRI Z spectrum data are input into the feature distillation module of the current training high-field Z spectrum data generation model. The feature distillation module of the current training high-field Z spectrum data generation model outputs a k-dimensional high-field feature expression. A second loss function is obtained according to the target high-field feature expression and the k-dimensional high-field feature expression output by the feature distillation module. The second loss function is used to adjust the network parameters of the feature distillation module of the current training high-field Z spectrum data generation model for the next iteration. During the iteration process, when the second loss function is less than 0.001 and does not decrease within 30 generations, the trained feature distillation module is obtained.

[0076] During the training process of the B0 correction module and the feature distillation module, the back propagation algorithm is used to adjust the weights and biases of the CNN to minimize the error. This process is repeated multiple times to gradually improve the performance of the model. At the same time, during the training process, a series of hyperparameters such as the number of hidden layers, the number of neurons per layer, the regularization print, the dropout rate, and the loss function involved in the DNN of the B0 correction module and the feature distillation module can be optimized through the grid search method.

[0077] Although low-field CEST MRI Z spectrum data contains a variety of different information, due to the low specificity of its signal, different Z spectra often show great similarity. When the DNN implementing feature distillation receives approximate low-field CEST MRI Z spectrum data input, it may produce the same output, causing the DNN implementing feature distillation to fall into the "dilemma" of mean output. In order to solve this problem, the embodiment of the present invention introduces a frequency offset scaling normalization technology to capture the fine-grained changes of low-field CEST MRI Z spectrum data at key frequency offsets.

[0078] In one embodiment of the present invention, Figure 2 As shown, feature distillation is performed on the target low-field CEST MRI Z spectrum data after B0 correction to obtain the target high-field feature expression, which may include:

[0079] The target low-field simulated CEST MRI Z spectrum data after B0 correction is subjected to frequency offset scaling and normalization, and then input into the deep neural network corresponding to the feature distillation module to output the target high-field feature expression.

[0080] Specifically, before the corrected low-field CEST MRI Z spectrum data is input into the feature distillation module of the current training high-field Z spectrum data generation model, the target low-field simulation CEST MRI Z spectrum data after B0 correction is subjected to frequency offset scaling and normalization processing, and the low-field CEST MRI Z spectrum data after frequency offset scaling and normalization processing is input into the deep neural network corresponding to the feature distillation module, and the feature distillation module outputs a k-dimensional target high-field feature expression.

[0081] In one embodiment of the present invention, the expression for frequency offset scaling normalization processing is:

[0082]

[0083] Among them, Z m,n represents the signal value at the nth frequency offset in the Z spectrum, Z m,n The normalized result, Z nIt represents the signal value of all training data under the same frequency offset. min(·) means to find the minimum value, and max(·) means to find the maximum value.

[0084] Specifically, after scaling and normalization, the DNN of the feature distillation module takes the low-field CEST MRI Z-spectral data as input in a one-dimensional vector and generates a k-dimensional high-field feature representation. The DNN network architecture and training method for implementing feature distillation are similar to the ΔB0 network.

[0085] In one embodiment of the present invention, the training method of the high-field Z-spectrum data generation model further includes a testing process, which may include:

[0086] A test data set is obtained, wherein the test data set includes a plurality of acquired low-field CEST MRI Z-spectral data and one-to-one corresponding acquired high-field CEST MRI Z-spectral data.

[0087] Specifically, the CEST MRI Z spectrum data of the phantom test sample is scanned by the high and low field scanning device, and the CEST MRI Z spectrum data of the phantom test sample scanned by the high and low field scanning device is processed to obtain the collected high and low field paired CEST MRI Z spectrum data.

[0088] The collected low-field CEST MRI Z spectrum data is input into the B0 correction module of the current test high-field Z spectrum data generation model, the B0 disturbance value corresponding to the current test low-field CEST MRI Z spectrum data is predicted, and the B0 disturbance prediction value is used to correct the current test low-field CEST MRI Z spectrum data, and the corrected current test low-field CEST MRI Z spectrum data is input into the feature distillation module of the current test high-field Z spectrum data generation model, and the test high-field feature expression corresponding to the current test low-field CEST MRI Z spectrum data is output, wherein the first test high-field Z spectrum data generation model is a high-field Z spectrum data generation model trained using the training data set.

[0089] The embodiment of the present invention uses the field intensity paired phantom experimental data to conduct actual testing on the trained high-field Z spectrum data generation model to verify the performance of the trained high-field Z spectrum data generation model in practical applications. Among them, for the DNN-based B0 prediction module, the performance of the optimized ΔB0 network is tested using the original low-field CEST MRI Z spectrum data and the corresponding WASSR graph. For the DNN-based feature distillation module, the low-field CEST MRI Z spectrum data corrected by the prediction results of the B0 prediction module is used as the network input, and the dimension reduction features extracted by SVD of the preprocessed 9.4TZ spectrum data are used as the network output target.

[0090] Specifically, when testing the trained high-field Z spectrum data generation model, the low-field CEST MRI Z spectrum data in the test data set is input into the B0 correction module of the current test high-field Z spectrum data generation model, and the B0 correction module predicts the accurate B0 perturbation prediction value corresponding to the low-field CEST MRI Z spectrum data. The B0 correction module corrects the low-field CEST MRI Z spectrum data based on the B0 perturbation prediction value, so that the trained B0 correction module adjusts the B0 offset in the low-field CESTMRI Z spectrum data in the test data set. The corrected low-field CEST MRI Z spectrum data is input into the trained feature distillation module, and the trained feature distillation module extracts K-dimensional high-field CEST MRI Z spectrum data features, and outputs the k-dimensional high-field feature expression U' for testing.

[0091] In one embodiment of the present invention, the training method of the high-field Z-spectrum data generation model further includes an evaluation process, which includes:

[0092] The high-field feature expression for testing and the target dimension reduction feature expression are used to perform SVD inverse transformation to obtain the predicted high-field CEST MRI Z-spectrum data;

[0093] A third loss function is obtained according to the predicted high-field CEST MRI Z spectrum data and its corresponding acquired high-field CEST MRI Z spectrum data to evaluate the performance of the high-field Z spectrum data generation model.

[0094] Specifically, the SVD truncation module performs an inverse SVD transform based on the k-dimensional high-field feature expression U' extracted by the feature distillation module and S' and V' retained in the training phase, and restores the k-dimensional high-field feature expression U' to the original Z spectrum feature space to generate the corresponding complete high-field CEST MRI Z spectrum data. The third loss function is obtained based on the predicted high-field CEST MRI Z spectrum data and its corresponding acquired high-field CEST MRI Z spectrum data to evaluate the performance of the high-field Z spectrum data generation model.

[0095] In order to evaluate the robustness of the high-field Z-spectral data generation model to noise, a new noisy data set was created. Considering the inherent noise in the collected low-field CEST MRI Z-spectral data, the result after principal component analysis (PCA) denoising was defined as "noise level 0". By adding Gaussian white noise with different noise standard deviations (SD) (0%, 0.25%, 0.5%, 0.75%, 1%, 1.5%, 2%, 2.5%, 3%, 4% and 5%) to the low-field CESTMRI Z-spectral data with a noise level of 0, 11 noisy PCr test sets and noisy egg white test sets were created respectively.

[0096] The embodiment of the present invention evaluates the high-field Z-spectrum data generation model from two aspects.

[0097] The generated results were quantitatively compared with the true results using conventional linear regression indicators. Specifically, the root mean square error (RMSE) and goodness of fit (R2) can be used to quantitatively evaluate the high-field Z-spectrum data generation model.

[0098]

[0099] Among them, y i and Represent the experimental collection value and the network prediction value respectively, It is an average operation.

[0100] A qualitative analysis was performed based on the commonly used CEST analysis method based on contrast diagrams. Specifically, the Lorentzian difference (LD) and magnetization transfer ratio asymmetry (MTRasym) can be used. The LD method uses a Lorentzian line to describe the DS (Direct Saturation) pool and measures the difference (Δω) between the reference DS and experimental data at a specific frequency offset:

[0101] LD(Δω)=Z ref (Δω)-Z(Δω)

[0102] Among them, Z ref In order to fit the DS value, the embodiment of the present invention adopts Z values ​​in three ranges of -10 to -8 ppm, -0.5 to 0.5 ppm and 8 to 10 ppm to fit the DS.

[0103] MTR asym Comparing the z-spectrum values ​​at symmetrical positive and negative frequency offsets provides an amount of contrast for a particular Δω:

[0104] MTR asym (Δω)=Z(-Δω)-Z(Δω)

[0105] In one embodiment of the present invention, the process of acquiring the test data set may include:

[0106] Using a 3T scanner to scan the phantom test sample and the positioning sample at the same time to obtain first low-field data;

[0107] The phantom test samples were scanned using a 9.4T scanner to obtain multiple second high field data;

[0108] performing normalization processing on the first low-field data and the plurality of second high-field data;

[0109] Extracting the region of interest from the normalized first low-field data to obtain low-field CEST MRI Z-spectrum data;

[0110] The plurality of normalized second high-field data are respectively downsampled, and the high-field Z spectrum data obtained by downsampling are subjected to B0 correction and denoising to obtain high-field CEST MRI Z spectrum data, wherein the phantom test sample comprises an egg white phantom placed in a glass tube of preset specifications and phosphocreatine phantoms of different concentrations prepared in phosphate buffered saline of different pH values ​​placed in a glass tube of preset specifications, and the positioning sample comprises phosphate buffered saline of pH=7 placed in a glass tube of preset specifications.

[0111] The embodiment of the present invention acquires CEST phantom data by performing low-field and high-field scanning on a phantom test sample, and then performs a series of preprocessing operations on the CEST phantom data to obtain high- and low-field paired CEST MRI Z spectrum data for testing.

[0112] The phantom test samples in the embodiment of the present invention include two groups of phantom samples of different types. The first group of phantom test samples consisted of creatine phosphate (PCr) solutions of different concentrations, namely 10mM, 20mM, 30mM, 40mM and 50mM, all dissolved in 1× phosphate-buffered saline (PBS) and titrated to a pH value ranging from 5.5 to 8.5. The second group of phantom test samples consisted of 3 egg white phantom samples. The 3 egg white phantom samples were all from fresh eggs. All samples were placed in 15ml glass test tubes and scanned at room temperature (21±0.5°C).

[0113] In one embodiment of the present invention, the first group of phantom test samples includes 10 mM creatine phosphate phantom prepared in phosphate buffered saline at pH = 7, 20 mM creatine phosphate phantom prepared in phosphate buffered saline at pH = 7, 30 mM creatine phosphate phantom prepared in phosphate buffered saline at pH = 7, 40 mM creatine phosphate phantom prepared in phosphate buffered saline at pH = 7, 50 mM creatine phosphate phantom prepared in phosphate buffered saline at pH = 7, 50 mM creatine phosphate phantom prepared in phosphate buffered saline at pH = 5.5, and 50 mM creatine phosphate phantom prepared in phosphate buffered saline at pH = 8.5. The positioning sample includes phosphate buffered saline at pH = 7.

[0114] It is feasible to use a 3T Philips MRI scanner to acquire low-field CEST MRI Z-spectral data. A continuous wave saturation module with a saturation power of 0.7μT and a saturation time of 2.5s was used, and a relaxation enhancement (RARE) sequence was used for rapid acquisition. The specific parameters were: RARE factor = 32, TR / TE = 6000 / 3.5ms, and field of view (FOV) of 120×120mm 2 , flip angle (FA) of 90°, slice thickness of 5 mm, matrix size of 256 × 256. Exemplarily, low-field CEST MRI Z-spectral data were acquired at 200 ppm frequency offset, and the water saturation offset reference WASSR method was used to generate the B0 field inhomogeneous distribution map. The total scanning time for both PCr and egg white phantoms was about 8 min.

[0115] It is feasible to use a 9.4T Bruker MRI scanner to acquire high-field CEST MRI Z-spectral data. Due to the limitations of the equipment itself, each sample tube was scanned individually. The scanning protocol is similar to the 3T experiment but with some adjustments. Specifically, FOV = 25×25 mm 2 , FA = 180°, matrix size 64 × 64, B0 positioning still uses WASSR technology. The total scanning time for PCr samples is 100 minutes, and the total scanning time for egg white samples is 47 minutes.

[0116] It should be noted that the 3T Philip scanner can scan multiple tubes of samples at a time. Due to the limitations of the 9.4TBruker scanner, the 9.4TBruker scanner can only collect single tubes one by one. The 3T Philip scanner directly performed CEST MRI imaging on 10 test samples and 1 positioning sample, and obtained a set of 3T CEST MRI Z spectrum data with a matrix size of 256×256. The 10 test samples were placed one by one in the scanning area of ​​the 9.4TBruker scanner, and the 9.4TBruker scanner performed CEST MRI imaging on the placed test samples, and obtained 10 sets of 9.4T CEST MRI Z spectrum data with a matrix size of 64×64.

[0117] The 3T and 9.4T CEST MRI Z spectrum data collected on the scanner are normalized to obtain the 3T CESTMRI Z spectrum data and its corresponding 9.4T CEST MRI Z spectrum data. Specifically, the saturated signal value (S sat ) was normalized to the unsaturated signal value (S0), and the Z spectrum of each voxel (CEST MRI Z spectrum data) was calculated: Z(Δω)=S sat (Δω) / S0.

[0118] A region of interest (ROI) was manually drawn on each phantom tube in the 3T CEST MRI Z-spectral data using a circular area with a radius of 11 pixels. Considering the inconsistent resolution between the 9.4T CEST MRI Z-spectral data and the 3T CESTMRI Z-spectral data, the 9.4T CEST MRI Z-spectral data was downsampled to a size of 48×48, and the ROI selected was a circular area of ​​the same size as that on the 3TCEST MRI Z-spectral data. It was assumed that the phantom was internally homogeneous, that is, the selected 3T and 9.4T data corresponded voxel by voxel. In addition, the 9.4T CEST MRI Z-spectral data were corrected for B0 inhomogeneity using the WASSR method, and denoised using a PCA-based method. Only the paired CESTMRI Z-spectral data corresponding to the voxels in the selected ROI area were used for testing.

[0119] The test data set implemented in the present invention consists of the Z spectrum in the selected ROI area. For the evaluation of the B0 correction module, the original 3T CEST MRI Z spectrum data and its corresponding B0 offset were used. For the evaluation of the feature distillation module, the 3T CEST MRI Z spectrum data corrected by the ΔB0 network and the corresponding 9T CEST MRI Z spectrum feature expression were used. The latter is the original collected 9.4TZ spectrum that is B0 corrected by the WASSR method and then feature truncated by SVD.

[0120] In the test data set of the embodiment of the present invention, there are 2618 pairs of PCr phantom data and 1122 pairs of egg white phantom data.

[0121] In one embodiment of the present invention, the process of acquiring the training data set includes:

[0122] A multi-pool model of the Bloch equation is used to simulate and generate multiple low-field CEST MRI Z spectrum data and their one-to-one corresponding high-field CEST MRI Z spectrum data, and the simulated low-field CEST MRI Z spectrum data are perturbed by B0.

[0123] The embodiment of the present invention uses a multi-pool model based on Bloch (Bloch-McConnell equations, BMEs) and multi-pool parameters of high and low field pairing to generate a CEST MRI Z spectrum of high and low field strength pairing. Specifically, the multi-pool model of the Bloch equation is used to simulate the generation process of CEST data. The generation process includes simulating the generation of Z spectrum data under different magnetic field strengths (B0), considering the physical properties and exchange effects of the sample, and when performing Bloch simulation, it is necessary to set the physical property parameters of the sample, such as T1, T2, exchange rate, etc. These parameters are used to generate simulated Z spectrum data and correspond to actual data. Through Bloch simulation, a series of simulated low-field CEST MRI Z spectrum data and corresponding high-field CEST MRI Z spectrum data are generated. These simulated data are used to train B0 correction modules and feature distillation modules to learn the effectiveness of B0 correction and feature extraction. The training part uses the simulated data corresponding to the above modules for comprehensive training to ensure its effectiveness and accuracy when processing real low-field Z spectrum data.

[0124] The magnetic field strength is set to 3T and 9.4T, B1 is 0.7μT, and Tsat is 6s. For the simulation of PCr data, a water pool and two guanidine proton pools are simulated (the frequency deviation ranges are 1.8-2.0ppm and 2.5-2.7ppm, respectively). In addition, the simulated egg white Z spectrum includes water (0ppm), amide protons (3.4-3.6ppm), amine protons (1.8-2.2ppm and 2.8-3.2ppm), NOE (-1-2ppm, -2-3ppm and -3-4ppm) and MT (0-4ppm) pools. For specific parameters, see Tables 1 and 2.

[0125] Table 1 Simulation parameters of 3T and 9.4T paired Z spectra: PCr training data

[0126]

[0127] Where, the T1 value of the *solute pool is the same as the T1 value of the water pool. Abbreviation: PCr, creatine phosphate.

[0128] Table 2 Simulation parameters of 3T and 9.4T paired Z spectra: egg white training data

[0129]

[0130] Among them, the T1 value of the solute pool is the same as the T1 value of the water pool.

[0131] It should be noted that the concentrations and exchange rates of different pools are randomly selected according to uniform distribution within the specified range, and T1w and T2w are selected in the same proportion within the specified range. In order to simulate the real scene, Gaussian white noise with a mean of 0 and a standard deviation (SD) in the range of 1% and B1 inhomogeneity of -0.02 to 0.02μT are added to the 3TZ spectrum. Bloch simulation generates 5e+4 paired PCrZ spectra and 5e+5 paired egg white Z spectra.

[0132] The B0 correction module was trained using a 3T simulated Z spectrum with B0 perturbation, where the B0 offset range was set to be randomly selected from -1 to 1ppm. In addition, the feature distillation module was trained on a dataset containing k-dimensional feature expressions of the 3T original simulated Z spectrum and the corresponding 9.4T Z spectrum. The amount of PCr and egg white data used for training is consistent with the amount of their respective original simulation data. The simulation training set was created on a personal computer configured with MATLAB (www.mathworks.com, R2020b), a 4-core CPU and 16GB RAM. It takes about 2.5 hours to simulate the PCr data to generate a low-high field paired Z spectrum, while this process takes about 16 hours for the egg white data simulation.

[0133] The high-field Z spectrum data generation model in the embodiment of the present invention includes two main DNNs. The first DNN is used to obtain the B0 inhomogeneity displacement of the low-field CEST MRI Z spectrum data so as to achieve effective alignment of the Z spectrum with the correct frequency offset. The second DNN learns from the low-field CEST MRI Z spectrum data and extracts a feature expression reflecting the high-field CEST MRI Z spectrum data.

[0134] The high-field Z spectrum data generation model in the embodiment of the present invention uses SVD technology to reduce the dimension of high-field CEST MRI Z spectrum data. The low-dimensional feature expression obtained by SVD truncation can effectively reflect the key information of high-field CEST MRI Z spectrum data. Through the inverse transformation process of SVD, these low-dimensional features are converted back to a complete high-field Z spectrum.

[0135] The high-field Z-spectrum data generation model of the embodiment of the present invention is trained based on Bloch-McConnell (BM) simulation, involving more than 50,000 sets of low-field and high-field paired Z-spectrum data. It was tested and studied on two phantom models, namely phosphocreatine (PCr) phantom and egg white phantom, and data were acquired on 3T Philips and 9.4T Bruker scanners.

[0136] The present invention compares the synthesized 9.4TZ spectrum with the real data obtained by experiment. The results show that the synthesized Z spectrum is almost completely consistent with the real data in a specific frequency range, showing a high degree of accuracy and good fit. In addition, the generative model has strong robustness to B0 inhomogeneity, noise and spectrum deviation that may occur in low-field data acquisition. At the same time, the generated high-field Z spectrum does not require complex pre-processing operations such as denoising and field inhomogeneity correction when performing subsequent quantitative analysis.

[0137] The high-field CEST MRI Z-spectrum synthesis method based on deep learning and singular value decomposition proposed in the present invention combines the different advantages of low-field strength data being easily available in clinical practice and high-field strength data being convenient for signal analysis, providing a new idea for the diagnosis and application of clinical CEST.

[0138] The high-field CEST MRI Z spectrum synthesis method based on deep learning and singular value decomposition proposed in the present invention can mine hidden signal information from low-field CEST MRI Z spectrum data without considering the inherent limitations of ultra-high field (UHF) MRI systems, provide richer and more accurate data for clinical diagnosis and research, and can be used to improve the clinical diagnosis and monitoring of various diseases using CEST technology, such as tumors, neurodegenerative diseases, etc. At the same time, the embodiment of the present invention provides a new idea for the acquisition of high-field CEST data, which can quickly and accurately generate a large amount of high-field data without expensive high-field strength equipment, complex simulation parameters and traditional simulation processes of several hours or even days, which helps to promote the migration and application of mature deep learning methods in CEST data analysis.

[0139] By combining deep learning and SVD technology, the present invention successfully developed a high-field CEST MRI Z spectrum synthesis method, overcoming the limitations of the existing technology and providing important technical support for the application of MRI technology in clinical and pre-clinical research.

[0140] The invention provides an electronic device.

[0141] In this embodiment, the electronic device includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the high-field CEST MRI Z-spectrum synthesis method as described above is implemented.

[0142] Figure 5 It is a structural block diagram of an electronic device according to an embodiment of the present invention.

[0143] like Figure 5As shown, the electronic device 500 includes: a processor 501 and a memory 503. The processor 501 and the memory 503 are connected, such as through a bus 502. Optionally, the electronic device 500 may also include a transceiver 504. It should be noted that in actual applications, the transceiver 504 is not limited to one, and the structure of the electronic device 500 does not constitute a limitation on the embodiments of the present invention.

[0144] Processor 501 may be a CPU (Central Processing Unit), a general purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of the present invention. Processor 501 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0145] The bus 502 may include a path to transmit information between the above components. The bus 502 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 502 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0146] The memory 503 is used to store a computer program corresponding to the high-field CEST MRI Z-spectrum synthesis method of the above embodiment of the present invention, and the computer program is controlled and executed by the processor 501. The processor 501 is used to execute the computer program stored in the memory 503 to implement the contents shown in the above method embodiment. Figure 5 The electronic device 500 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0147] The electronic device of the embodiment of the present invention utilizes the above high-field CEST MRI Z spectrum synthesis method to mine hidden signal information from low-field CEST MRI Z spectrum data, providing richer and more accurate data for clinical diagnosis and research, and can be used to improve the clinical diagnosis and monitoring of various diseases using CEST technology. It does not require expensive high-field strength equipment and has the advantage of low cost.

[0148] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection portion with one or more wirings (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0149] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0150] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0151] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “clockwise”, “counterclockwise”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0152] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0153] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise clearly defined. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0154] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0155] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A high-field CEST MRI Z-spectrum synthesis method based on deep learning and singular value decomposition, characterized in that: The method comprises: Acquire target low-field CEST MRI Z-spectrum data; Utilizing a pre-trained high-field Z spectrum data generation model, target high-field CEST MRI Z spectrum data is generated according to the target low-field CEST MRI Z spectrum data, wherein the high-field Z spectrum data generation model includes a B0 correction module based on a deep neural network DNN, a feature distillation module based on the DNN, and a singular value decomposition (SVD) truncation module, the B0 correction module is used to perform B0 correction on the low-field CEST MRI Z spectrum data, the feature distillation module is used to perform feature distillation on the target low-field CEST MRI Z spectrum data after B0 correction according to the dimension of the target dimensionality reduction feature expression to obtain the target high-field feature expression, the SVD truncation module is used to perform an SVD inverse transform on the target high-field feature expression to obtain the target high-field CESTMRI Z spectrum data, and the dimension of the target dimensionality reduction feature expression is determined by the SVD truncation module.

2. The CEST MRI Z spectrum synthesis method according to claim 1, characterized in that: The training method of the high-field Z-spectrum data generation model includes a training process, and the training process includes: Acquire a training data set, wherein the training data set includes a plurality of simulated low-field CEST MRI Z spectrum data and a one-to-one corresponding plurality of simulated high-field CEST MRI Z spectrum data, and a plurality of simulated low-field CEST MRI Z spectra perturbed by B0 and a one-to-one corresponding actual value of B0 perturbation; Establishing the high-field Z-spectrum data generation model; Inputting the low-field CEST MRI Z spectrum data obtained by the simulation of B0 perturbation into the B0 correction module of the current training high-field Z spectrum data generation model, predicting the B0 perturbation value corresponding to the current training low-field CEST MRI Z spectrum data, and correcting the current training low-field CEST MRI Z spectrum data using the B0 perturbation prediction value, inputting the corrected current training low-field CEST MRI Z spectrum data into the feature distillation module of the current training high-field Z spectrum data generation model, and outputting the training high-field feature expression corresponding to the current training low-field CEST MRI Z spectrum data; According to the B0 disturbance actual value and the B0 disturbance predicted value, a first loss function is obtained, and the first loss function is used to adjust the network parameters of the B0 correction module of the current training high-field Z spectrum data generation model to perform the next iteration until the first loss function is less than a first preset threshold value and does not decrease within a first preset iteration range; A second loss function is obtained based on the target high-field feature expression and the high-field feature expression output by the feature distillation module, and the second loss function is used to adjust the network parameters of the feature distillation module of the current training high-field Z spectrum data generation model for the next iteration until the second loss function is less than a second preset threshold and does not decrease within a second preset iteration range.

3. The CEST MRI Z spectrum synthesis method according to claim 2, characterized in that: The training method of the high-field Z-spectrum data generation model also includes a testing process, and the testing process includes: Acquire a test data set, wherein the test data set includes a plurality of collected low-field CEST MRI Z-spectral data and one-to-one corresponding collected high-field CEST MRI Z-spectral data; The collected low-field CEST MRI Z spectrum data is input into the B0 correction module of the current test high-field Z spectrum data generation model, the B0 disturbance value corresponding to the current test low-field CEST MRI Z spectrum data is predicted, and the B0 disturbance prediction value is used to correct the current test low-field CEST MRI Z spectrum data, the corrected current test low-field CEST MRI Z spectrum data is input into the feature distillation module of the current test high-field Z spectrum data generation model, and the test high-field feature expression corresponding to the current test low-field CEST MRI Z spectrum data is output, wherein the first test high-field Z spectrum data generation model is a high-field Z spectrum data generation model trained using the training data set.

4. The CEST MRI Z spectrum synthesis method according to claim 3, characterized in that: The training method of the high-field Z-spectrum data generation model also includes an evaluation process, and the evaluation process includes: Performing an inverse SVD transformation using the test high-field feature expression and the target dimension reduction feature expression to obtain predicted high-field CEST MRI Z-spectrum data; A third loss function is obtained according to the predicted high-field CEST MRI Z spectrum data and the corresponding acquired high-field CEST MRI Z spectrum data to evaluate the performance of the high-field Z spectrum data generation model.

5. The CEST MRI Z spectrum synthesis method according to claim 1, characterized in that: The feature distillation of the target low-field CEST MRI Z spectrum data after B0 correction is performed to obtain the target high-field feature expression, including: The target low-field simulated CEST MRI Z spectrum data after B0 correction is subjected to frequency offset scaling and normalization processing, and then input into the deep neural network corresponding to the feature distillation module to output the target high-field feature expression.

6. The CEST MRI Z spectrum synthesis method according to claim 5, characterized in that: The expression of the frequency offset scaling normalization process is: Among them, Z m,n represents the signal value at the nth frequency offset in the Z spectrum, Z m,n The normalized result, Z n It represents the signal value of all training data under the same frequency offset. min(·) means to find the minimum value, and max(·) means to find the maximum value.

7. The CEST MRI Z spectrum synthesis method according to claim 2, characterized in that: The SVD truncation module uses the singular value decomposition method to decompose the high-field CEST MRI Z spectrum data obtained by the simulation, and uses the cumulative variance explanation ratio to determine the dimension of the target dimensionality reduction feature expression according to the singular value obtained by the decomposition, wherein the expression for calculating the cumulative variance explanation ratio is: Among them, CEV n is the cumulative variance contribution rate of the nth eigenvalue, S i is the ith eigenvalue.

8. The CEST MRI Z spectrum synthesis method according to claim 3, characterized in that: The process of acquiring the test data set includes: Using a 3T scanner to scan the phantom test sample and the positioning sample at the same time to obtain first low-field data; Scanning the phantom test samples respectively using a 9.4T scanner to obtain a plurality of second high field data; Performing normalization processing on the first low-field data and the plurality of the second high-field data; Extracting a region of interest from the normalized first low-field data to obtain the low-field CEST MRI Z spectrum data; The plurality of normalized second high-field data are respectively downsampled, and the high-field Z spectrum data obtained by downsampling are subjected to B0 correction and denoising to obtain the high-field CEST MRI Z spectrum data, wherein the phantom test sample comprises an egg white phantom placed in a glass tube of preset specifications and phosphocreatine phantoms of different concentrations prepared in phosphate buffered saline of different pH values ​​placed in a glass tube of preset specifications, and the positioning sample comprises phosphate buffered saline of pH=7 placed in a glass tube of preset specifications.

9. The CEST MRI Z spectrum synthesis method according to claim 8, characterized in that: The process of obtaining the training data set includes: A multi-pool model based on the Bloch equation was used to simulate and generate multiple low-field CEST MRI Z spectrum data and their one-to-one corresponding high-field CEST MRI Z spectrum data, and the simulated low-field CEST MRI Z spectrum data were perturbed by B0.

10. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the computer program is executed by the processor, the CEST MRI Z spectrum synthesis method according to any one of claims 1 to 9 is implemented.