Self-supervised reconstruction method for human body 5.0 T multi-core 31P magnetic resonance spectrum imaging

Abstract

The invention discloses a self-supervised reconstruction method for human body 5.0 T multi-core 31P magnetic resonance spectrum imaging, a self-supervised reconstruction network comprises a position coding module, n dense convolution modules and n data consistency modules, the position coding module fully learns the position information of metabolite chemical shift, and assists in reconstruction of a 31P-MRSI spectrum image, so as to improve the accuracy of the reconstruction of the human body 5.0 T multi-core 31P magnetic resonance spectrum imaging. Then, a dense convolutional neural network structure is constructed, information flow is effectively promoted by utilizing the advantages of residual connection and dense connection, feature multiplexing is enhanced, and the capabilities of removing false peaks and reconstructing high-quality images are enhanced; compared with an existing 31P-MRSI reconstruction method, the self-supervised reconstruction network of the invention can effectively realize training under the condition of no full sampling spectrum image, data acquisition time is shortened, reconstructed high-quality spectrum images are obtained, and the network obtained by training has stronger generalization ability.

Description

Technical Field

[0001] The present invention belongs to the field of magnetic resonance imaging technology, and specifically relates to a 5.0T multi-core human body. 31 A self-supervised reconstruction method for P magnetic resonance spectroscopy imaging, suitable for undersampling 31 P-MRSI data reconstruct high-quality spectral images. Background Art

[0002] Magnetic Resonance Spectroscopic Imaging (MRSI) is a magnetic resonance imaging technique. 31 P-MRSI can provide biochemical information about tissues by measuring high-energy phosphates (i.e., ATP, creatine phosphate PCr, and inorganic phosphate).

[0003] The inherently low signal-to-interference plus noise ratio (SNR) 31 The main limiting factor of P-MRSI is the low spatial resolution and extremely long acquisition time. The emergence of ultra-high field 5.0T magnetic resonance systems has improved the low-field magnetic resonance experimental equipment. 31 The problem of low spatial resolution of P-MRSI can better study the relationship between tissue structure and function. Compared with other ultra-high field magnetic resonance systems (such as 7.0T magnetic resonance systems), the 5.0T magnetic resonance system avoids the B1 + The problem of spectral imaging quality degradation caused by field RF inhomogeneity is B1 + The uniformity of the field helps 31 P-MRSI spectral imaging achieves more accurate chemical shift selection and metabolite concentration measurement, supporting more reliable metabolite quantification and biochemical analysis. In addition, the 5.0T magnetic resonance system has lower material costs and is more suitable for clinical application.

[0004] In addition, various acceleration methods commonly used in MRI have been introduced to reduce acquisition time while maintaining sufficient SNR for spectral analysis. 31 P-MRSI and P-MRSI have shown potential in obtaining high-quality spectra. 31 The P-MRSI method is Compressed Sensing (CS), which uses 31 The spectrum of P-MRSI has sparseness in the transform domain and the incoherence of undersampling is used for nonlinear reconstruction, which can better restore the full sampling 31P-MRSI spectral images, but the CS optimization process takes a long time and is greatly affected by parameters. At present, deep learning methods have shown great potential in the field of undersampling MRSI reconstruction. Some researchers have used convolutional neural networks (CNN) for the human B cell receptor CD79b protein. 1 H- 15 The reconstruction of N HSQC spectral images can reconstruct MRSI quickly, reliably and with high quality. However, due to the difficulty in acquiring MRSI spectral images, it is difficult for existing deep learning MRSI reconstruction methods to use a large amount of actual acquired data as training data for supervised learning. Self-supervised learning can automatically extract feature information related to imaging quality from the original MRSI spectral image data to assist in reconstruction, without the need to acquire a large amount of fully sampled MRSI data, thus solving the problem of long acquisition time. Therefore, the use of self-supervised learning methods for 31 PMRSI reconstruction, improvement 31 P MRSI spatial resolution and shortened acquisition time can obtain high-quality reconstructed spectral images when data is scarce. Summary of the invention

[0005] The purpose of the present invention is to provide a 5.0T multi-core human body to solve the above problems in the prior art. 31 A self-supervised reconstruction method for P magnetic resonance spectroscopy imaging.

[0006] The above-mentioned purpose of the present invention is achieved by the following technical means:

[0007] Human body 5.0T multi-core 31 The method for self-supervised reconstruction of P magnetic resonance spectroscopy imaging is characterized by comprising the following steps:

[0008] Step 1, obtaining a plurality of under-sampled k-space data, taking each under-sampled k-space data as a sample data, and then randomly dividing all the sample data into a training set, a validation set, and a test set according to a proportion;

[0009] Step 2, randomly divide each sample data in the training set into two parts, respectively as the first under-sampled k-space data and the second under-sampled k-space data, and obtain the first under-sampled spectral image and the first under-sampled matrix corresponding to the first under-sampled k-space data; obtain the second under-sampled spectral image and the second under-sampled matrix corresponding to the second under-sampled k-space data; and use the first under-sampled spectral image as the input sample;

[0010] Step 3: construct a self-supervised reconstruction network, the input of the self-supervised reconstruction network is the input sample, and the output of the self-supervised reconstruction network is the reconstructed full-sampled spectrum image;

[0011] Step 4, performing Fourier transform on the reconstructed fully sampled spectral image to obtain corresponding reconstructed fully sampled k-space data, then multiplying the reconstructed fully sampled k-space data by the second under-sampling matrix to obtain reconstructed under-sampled k-space data, and then performing inverse Fourier transform on the reconstructed under-sampled k-space data to obtain a reconstructed under-sampled spectral image;

[0012] Step 5: Set the total loss function;

[0013] Step 6: Input the input samples of the training set into the self-supervised reconstruction network to train the self-supervised reconstruction network. After the training is completed, save the model parameters to obtain the trained self-supervised reconstruction network.

[0014] Step 7, input the input sample corresponding to the test set obtained after executing step 2 on the test set into the trained self-supervised reconstruction network to obtain the reconstructed full-sampled spectrum image corresponding to the test set;

[0015] Step 8: Obtain the actual under-sampled spectral image and input it into the trained self-supervised reconstruction network to obtain the reconstructed full-sampled spectral image corresponding to the actual under-sampled spectral image.

[0016] As described above, the self-supervised reconstruction network in step 3 includes a position encoding module, n dense convolution modules, and n data consistency modules. The position encoding module includes an encoder. The input sample is input into the encoder. The output of the encoder and the input are input into the first dense convolution module after channel splicing operation. The output of the bth data consistency module is input into the b+1th dense convolution module, where b is 1 to n-1; the output of the ath dense convolution module and the first under-sampled k-space data are added and input into the ath data consistency module, where a is 1 to n; the output of the nth data consistency module is the output of the self-supervised reconstruction network, and the output of the self-supervised reconstruction network reconstructs the fully sampled spectral image.

[0017] As mentioned above, the dense convolution module includes multiple densely connected convolution modules. The input of the first convolution module is the input of the dense convolution module, and the inputs of the remaining convolution modules are the results of the outputs of all previous convolution modules after channel splicing operations. The outputs of all convolution modules are added together as the output of the corresponding dense convolution module.

[0018] As mentioned above, in each dense convolution module, except for the last convolution module which only includes a convolution layer, the remaining convolution modules all include a convolution layer and an activation layer, the input of the convolution layer is the input of the corresponding convolution module, the output of the last convolution layer is the output of the corresponding convolution module, and the outputs of the remaining convolution layers are all input into the activation layer of the corresponding convolution module, and the output of the activation layer is the output of the corresponding convolution module.

[0019] As described above, the total loss function of step 5 is the sum of the first loss function and the second loss function, wherein the first loss function calculates the error between the second undersampled k-space data and the reconstructed undersampled k-space data; the second loss function calculates the error between the second undersampled spectral image and the reconstructed undersampled spectral image.

[0020] As described above, the first loss function and the second loss function in step 5 are both mean square error loss functions.

[0021] As described above, step 6 specifically includes the following process: inputting the input samples in the training set into the self-supervised reconstruction network to obtain the reconstructed fully sampled spectral image corresponding to each input sample, and then executing step 4 on the reconstructed fully sampled spectral image to obtain the reconstructed under-sampled k-space data, setting the learning rate, batch size, and total number of iterations, and then performing end-to-end training on the self-supervised reconstruction network through the ADAM optimizer and the set total loss function. After the number of iterations reaches the total number of iterations, the training is stopped, and the parameters of the self-supervised reconstruction network are saved to obtain a trained self-supervised reconstruction network.

[0022] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements any one of the above-mentioned human body 5.0T multi-core 31 Steps of a self-supervised reconstruction method for P magnetic resonance spectroscopy imaging.

[0023] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any one of the above-mentioned human body 5.0T multi-core 31 Steps of a self-supervised reconstruction method for P magnetic resonance spectroscopy imaging.

[0024] A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, any one of the human body 5.0T multi-core 31 Steps of a self-supervised reconstruction method for P magnetic resonance spectroscopy imaging.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) The self-supervised reconstruction network of the present invention can utilize the magnetic resonance system B1 + The uniformity of the field is characterized by fully learning the position information of the chemical shift of metabolites through position encoding, assisting 31 The P-MRSI spectral image is reconstructed, and then a dense convolutional neural network structure is constructed. The advantages of residual connection and dense connection are utilized to effectively promote the flow of information, strengthen feature reuse, and enhance the ability to remove pseudo-peaks and reconstruct high-quality images.

[0027] (2) Compared with the existing31 P-MRSI reconstruction method, the present invention no longer relies on a large number of full sampling 31 P-MRSI spectral image, the present invention effectively utilizes 31 The intrinsic laws and characteristics of P-MRSI data are learned using self-supervised methods 31 The characteristics of P-MRSI itself can be used for training without fully sampled spectral images, reducing the need for high-quality fully sampled images. 31 The demand for P-MRSI spectral images is to obtain reconstructed high-quality spectral images while shortening the data acquisition time, so that the trained network has stronger generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the method of the present invention;

[0029] FIG2( a ) is a schematic diagram of a sampling matrix of Cartesian undersampling in Embodiment 1 of the present invention;

[0030] FIG2( b ) is a first undersampling matrix of Embodiment 1 of the present invention;

[0031] FIG2( c ) is a second undersampling matrix of Embodiment 1 of the present invention;

[0032] Figure 3 Schematic diagram of the structure of the self-supervised reconstruction network of embodiment 1 of the present invention;

[0033] Figure 4 Schematic diagram of the structure of the dense convolution module of Example 1 of the present invention, wherein Indicates channel splicing operation;

[0034] FIG. 5( a ) is a 2D full sampling 5.0T in the test set in Example 1 of the present invention. 31 P-MRSI spectrum image (where ppm is the unit of chemical shift on the abscissa and amplitude is the signal amplitude);

[0035] Figure 5(b) shows the 2D undersampling 5.0T obtained by using the zero-filling method. 31 P-MRSI spectral images;

[0036] FIG5(c) is a diagram of the 2D undersampling 5.0T in FIG5(b) using the self-supervised reconstruction network trained in Example 1 of the present invention. 31 P-MRSI spectral images were reconstructed to obtain the reconstructed 5.0T 31 P-MRSI spectral image. DETAILED DESCRIPTION

[0037] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below in conjunction with embodiments. The embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0038] Embodiment 1:

[0039] Human body 5.0T multi-core 31 The method for self-supervised reconstruction of P magnetic resonance spectroscopy imaging comprises the following steps:

[0040] Step 1, obtaining a plurality of under-sampled k-space data, taking each under-sampled k-space data as a sample data, and then randomly dividing all the sample data into a training set, a validation set, and a test set according to a proportion;

[0041] In this embodiment, LCModel spectrum quantitative analysis software is used to simulate and obtain 1024 5.0T 31 P-MRSI spectrum data (5.0T is the magnetic field strength), and then 1024 5.0T 31 The P-MRSI spectral data were Fourier transformed to obtain multiple 2D full-sampling 5.0T 31 P-MRSI k-space data, and then each 2D full sampling 5.0T according to the Cartesian undersampling sampling matrix 31 The P-MRSI k-space data were undersampled to obtain the corresponding 2D undersampled 5.0T 31 P-MRSI k-space data (corresponding to the undersampled k-space data in step 1); in this embodiment, each 2D undersampled 5.0T 31 The P-MRSI k-space data was used as a sample data, and then all sample data were randomly divided into a training set, a validation set, and a test set in a ratio of 3:1:1;

[0042] Step 2: randomly divide each sample data in the training set into two parts, respectively as the first under-sampled k-space data and the second under-sampled k-space data; perform inverse Fourier transform on the first under-sampled k-space data to obtain the corresponding spectral image as the first under-sampled spectral image, and obtain the under-sampled sampling matrix corresponding to the first under-sampled k-space data as the first under-sampled matrix; perform inverse Fourier transform on the second under-sampled k-space data to obtain the corresponding spectral image as the second under-sampled spectral image, and obtain the under-sampled sampling matrix corresponding to the second under-sampled k-space data as the second under-sampled matrix; the first under-sampled spectral image is the input sample.

[0043] The input samples are used to input into the self-supervised reconstruction network, and the second under-sampled k-space data and the corresponding second under-sampled spectral image are used to calculate the loss and train the self-supervised reconstruction network.

[0044] In this embodiment, step 2 is performed on each sample data in the training set to obtain the first undersampling 5.0T 31 P-MRSI k-space data (corresponding to the first under-sampled k-space data in step 2), the first under-sampled 5.0T 31 P-MRSI spectrum image (corresponding to the first under-sampled spectrum image in step 2), and the first under-sampled matrix; the second under-sampled 5.0T 31 P-MRSI k-space data (corresponding to the second under-sampled k-space data in step 2), the second under-sampled 5.0T 31 The P-MRSI spectrum image (corresponding to the second under-sampled spectrum image in step 2) and the second under-sampled matrix are converted into the first under-sampled 5.0T 31 P-MRSI spectral images are used as input samples.

[0045] Step 3, construct a self-supervised reconstruction network, the self-supervised reconstruction network includes a position encoding module, n dense convolution modules, and n data consistency modules, the position encoding module includes an encoder, the input sample is input into the encoder of the position encoding module, the output and input of the encoder of the position encoding module are input into the first dense convolution module after channel splicing operation, the output of the bth data consistency module is input into the b+1th dense convolution module, b is 1 to n-1; the output of the ath dense convolution module and the first under-sampled k-space data are added and input into the ath data consistency module, a is 1 to n; the output of the nth data consistency module is the output of the self-supervised reconstruction network, and the output of the self-supervised reconstruction network reconstructs the fully sampled spectral image.

[0046] The dense convolution module includes multiple densely connected convolution modules. The input of the first convolution module is the input of the dense convolution module. The input of the remaining convolution modules is the result of the channel concatenation operation of the outputs of all convolution modules before the current convolution module in the corresponding dense convolution module. The outputs of all convolution modules are added together as the output of the corresponding dense convolution module.

[0047] In each dense convolution module, except for the last convolution module which only includes convolution layers, the remaining convolution modules include convolution layers and activation layers. The input of the convolution layer is the input of the corresponding convolution module, the output of the last convolution layer is the output of the corresponding convolution module, and the output of the remaining convolution layers is input into the activation layer of the corresponding convolution module, and the output of the activation layer is the output of the corresponding convolution module.

[0048] The positional encoding module learns a uniform B1 by encoding a vector with the same length as the input content. + Field (This example is a uniform B1 under a 5.0T system +The more accurate position information of the chemical shift of metabolites reflected by the 31 Efficiency and accuracy of P-MRSI spectral image reconstruction.

[0049] Multiple densely connected convolutional modules enable the self-supervised reconstruction network to learn complex mapping relationships, so that each convolutional layer can receive rich information from the front of the entire network, which helps the network consider both global and local features during the reconstruction process;

[0050] The data consistency module is used to combine the reconstructed k-space data with the first under-sampled k-space data to ensure data consistency.

[0051] In this embodiment, the structural diagram of the self-supervised reconstruction network is as follows: Figure 3 As shown, the self-supervised reconstruction network includes a position encoding module, three dense convolution modules, and three data consistency modules, each of which includes four densely connected convolution modules.

[0052] Step 4, performing Fourier transform on the reconstructed fully sampled spectral image to obtain corresponding reconstructed fully sampled k-space data, then multiplying the reconstructed fully sampled k-space data by the second under-sampling matrix to obtain reconstructed under-sampled k-space data, and then performing inverse Fourier transform on the reconstructed under-sampled k-space data to obtain a reconstructed under-sampled spectral image;

[0053] Step 5, setting a total loss function. The total loss function is the sum of the first loss function and the second loss function, wherein the first loss function calculates the error between the second undersampled k-space data and the reconstructed undersampled k-space data; the second loss function calculates the error between the second undersampled spectral image and the reconstructed undersampled spectral image.

[0054] In this embodiment, the first loss function and the second loss function are both mean square error (L2) loss functions.

[0055] Step 6: Input the input samples of the training set into the self-supervised reconstruction network to train the self-supervised reconstruction network. After the training is completed, save the model parameters to obtain the trained self-supervised reconstruction network.

[0056] Specifically, the input samples in the training set are input into the self-supervised reconstruction network to obtain the reconstructed fully sampled spectral images corresponding to each input sample, and then step 4 is performed on the reconstructed fully sampled spectral images to obtain the reconstructed under-sampled k-space data, the learning rate, batch size, and total number of iterations are set, and then the self-supervised reconstruction network is end-to-end trained through the ADAM optimizer and the set total loss function. The training is stopped when the number of iterations reaches the total number of iterations, and the parameters of the self-supervised reconstruction network are saved to obtain the trained self-supervised reconstruction network.

[0057] In this embodiment, the first undersampling is 5.0T 31 The P-MRSI spectral image is input into the self-supervised reconstruction network to perform step 6 to obtain the reconstructed full sampling 5.0T 31 P-MRSI spectral image (corresponding to the reconstructed full-sample spectral image in step 6), reconstructed full-sample 5.0T 31 P-MRSI k-space data (corresponding to the reconstructed full-sample k-space data in step 4), reconstructed undersampled 5.0T 31 P-MRSI k-space data (corresponding to the reconstructed undersampled k-space data in step 4), reconstructed undersampled 5.0T 31 P-MRSI spectral image (corresponding to the reconstructed undersampled spectral image in step 4), the learning rate is initialized to 0.0001, the batch size is set to 10, and the total number of iterations is set to 200.

[0058] Step 7: Execute step 2 on the test set to obtain the input samples corresponding to the test set, input the input samples corresponding to the test set into the trained self-supervised reconstruction network in step 6, and obtain the reconstructed full-sampled spectrum image corresponding to the test set.

[0059] In this embodiment, the test set is executed in step 7 to obtain the reconstructed full sampling 5.0T corresponding to the test set. 31 P-MRSI spectral images and computational reconstruction of full sampling 5.0T 31 Signal-to-noise ratio (SNR) index of P-MRSI spectral image, calculation and reconstruction of full sampling 5.0T 31 P-MRSI spectral images and 2D full sampling 5.0T 31 The peak intensity correlation (R 2 ), and by using the 2D undersampling 5.0T obtained using the zero-filling method 31 P-MRSI spectral images are analyzed and compared to measure the performance of the trained self-supervised reconstruction network;

[0060] Comparing Figure 5(b) and Figure 5(a), it can be seen that the reconstruction of 5.0T obtained by the self-supervised reconstruction network of the present invention 31 P-MRSI spectral images and 2D full sampling 5.0T 31 The peak correlation of P-MRSI spectral images is higher and the reconstruction of 5.0T 31 The signal-to-noise ratio of P-MRSI spectroscopy images is also better.

[0061] In the actual application process, the undersampled k-space data collected can be directly used in step 1. In this embodiment, the 2D full sampling 5.0T 31 The P-MRSI spectral image is used to measure the performance of the trained self-supervised reconstruction network and does not participate in the training of the self-supervised reconstruction network.

[0062] The self-supervised reconstruction network of the present invention can utilize the magnetic resonance system B1 + The uniformity of the field is characterized by fully learning the position information of the chemical shift of metabolites through position encoding, assisting 31 The P-MRSI spectral image is reconstructed, and then a dense convolutional neural network structure is constructed. The advantages of residual connection and dense connection are utilized to effectively promote the flow of information, strengthen feature reuse, and enhance the ability to remove pseudo-peaks and reconstruct high-quality images.

[0063] Compared with the existing 31 P-MRSI reconstruction method, the present invention no longer relies on a large number of full sampling 31 P-MRSI spectral image, the present invention effectively utilizes 31 The intrinsic laws and characteristics of P-MRSI data are learned using self-supervised methods 31 The characteristics of P-MRSI itself can be used for training without fully sampled spectral images, reducing the need for high-quality fully sampled images. 31 The demand for P-MRSI spectral images is to obtain reconstructed high-quality spectral images while shortening the data acquisition time, so that the trained network has stronger generalization ability.

[0064] Step 8: Input the actual under-sampled spectral image into the trained self-supervised reconstruction network to obtain the reconstructed fully sampled spectral image corresponding to the actual under-sampled spectral image.

[0065] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods.

[0066] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, steps 1 to 8 in the above method embodiments are implemented.

[0067] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, steps 1 to 8 in the above method embodiments are implemented.

[0068] In one embodiment, a computer program product is provided, including a computer program, which implements steps 1 to 8 in the above method embodiments when executed by a processor.

[0069] It should be noted that the embodiments described in the present invention are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the described embodiments or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.

Claims

1. Human body 5.0T multi-core 31 A self-supervised reconstruction method for P magnetic resonance spectroscopy imaging, characterized in that The following steps are involved: Step 1, obtaining a plurality of under-sampled k-space data, taking each under-sampled k-space data as a sample data, and then randomly dividing all the sample data into a training set, a validation set, and a test set according to a proportion; Step 2, randomly divide each sample data in the training set into two parts, respectively as the first under-sampled k-space data and the second under-sampled k-space data, and obtain the first under-sampled spectral image and the first under-sampled matrix corresponding to the first under-sampled k-space data; obtain the second under-sampled spectral image and the second under-sampled matrix corresponding to the second under-sampled k-space data; and use the first under-sampled spectral image as the input sample; Step 3: construct a self-supervised reconstruction network, the input of the self-supervised reconstruction network is the input sample, and the output of the self-supervised reconstruction network is the reconstructed full-sampled spectrum image; Step 4, performing Fourier transform on the reconstructed fully sampled spectral image to obtain corresponding reconstructed fully sampled k-space data, then multiplying the reconstructed fully sampled k-space data by the second under-sampling matrix to obtain reconstructed under-sampled k-space data, and then performing inverse Fourier transform on the reconstructed under-sampled k-space data to obtain a reconstructed under-sampled spectral image; Step 5: Set the total loss function; Step 6: Input the input samples of the training set into the self-supervised reconstruction network to train the self-supervised reconstruction network. After the training is completed, save the model parameters to obtain the trained self-supervised reconstruction network. Step 7, input the input sample corresponding to the test set obtained after executing step 2 on the test set into the trained self-supervised reconstruction network to obtain the reconstructed full-sampled spectrum image corresponding to the test set; Step 8: Obtain the actual under-sampled spectral image and input it into the trained self-supervised reconstruction network to obtain the reconstructed full-sampled spectral image corresponding to the actual under-sampled spectral image.

2. The human 5.0T multi-core according to claim 1 31 A self-supervised reconstruction method for P magnetic resonance spectroscopy imaging, characterized in that The self-supervised reconstruction network in step 3 includes a position encoding module, n dense convolution modules, and n data consistency modules. The position encoding module includes an encoder. The input sample is input into the encoder. The output of the encoder and the input are input into the first dense convolution module after channel splicing operation. The output of the bth data consistency module is input into the b+1th dense convolution module, where b is 1 to n-1; the output of the ath dense convolution module and the first under-sampled k-space data are added and input into the ath data consistency module, where a is 1 to n; the output of the nth data consistency module is the output of the self-supervised reconstruction network, and the output of the self-supervised reconstruction network reconstructs the fully sampled spectral image.

3. The human 5.0T multi-core according to claim 2 31 A self-supervised reconstruction method for P magnetic resonance spectroscopy imaging, characterized in that The dense convolution module includes multiple densely connected convolution modules, the input of the first convolution module is the input of the dense convolution module, the input of the remaining convolution modules are the results of the outputs of all previous convolution modules after channel splicing operations, and the outputs of all convolution modules are added together as the output of the corresponding dense convolution module.

4. The human body 5.0T multi-core according to claim 3 31 A self-supervised reconstruction method for P magnetic resonance spectroscopy imaging, characterized in that In each dense convolution module, except for the last convolution module which only includes a convolution layer, the remaining convolution modules all include a convolution layer and an activation layer, the input of the convolution layer is the input of the corresponding convolution module, the output of the last convolution layer is the output of the corresponding convolution module, the outputs of the remaining convolution layers are all input into the activation layer of the corresponding convolution module, and the output of the activation layer is the output of the corresponding convolution module.

5. The human 5.0T multi-core according to claim 1 31 A self-supervised reconstruction method for P magnetic resonance spectroscopy imaging, characterized in that The total loss function of step 5 is the sum of the first loss function and the second loss function, wherein the first loss function calculates the error between the second undersampled k-space data and the reconstructed undersampled k-space data; the second loss function calculates the error between the second undersampled spectral image and the reconstructed undersampled spectral image.

6. The human 5.0T multi-core according to claim 5 31 A self-supervised reconstruction method for P magnetic resonance spectroscopy imaging, characterized in that In step 5, the first loss function and the second loss function are both mean square error loss functions.

7. The human 5.0T multi-core according to claim 1 31 A self-supervised reconstruction method for P magnetic resonance spectroscopy imaging, characterized in that The step 6 specifically includes the following process: inputting the input samples in the training set into the self-supervised reconstruction network to obtain the reconstructed fully sampled spectral image corresponding to each input sample, then executing step 4 on the reconstructed fully sampled spectral image to obtain the reconstructed under-sampled k-space data, setting the learning rate, batch size, and total number of iterations, and then performing end-to-end training on the self-supervised reconstruction network through the ADAM optimizer and the set total loss function, stopping the training after the number of iterations reaches the total number of iterations, and saving the parameters of the self-supervised reconstruction network to obtain a trained self-supervised reconstruction network.

8. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: The processor implements the human body 5.0T multi-core described in any one of claims 1 to 7 when executing the computer program 31 Steps of a self-supervised reconstruction method for P magnetic resonance spectroscopy imaging.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the human body 5.0T multi-core described in any one of claims 1 to 7 is implemented 31 Steps of a self-supervised reconstruction method for P magnetic resonance spectroscopy imaging.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the human body 5.0T multi-core described in any one of claims 1 to 7 is implemented 31 Steps of a self-supervised reconstruction method for P magnetic resonance spectroscopy imaging.

Images