Magnetic Resonance CEST Imaging Method, Apparatus, Media, and Equipment Based on Automatic Image Artifact Correction

Abstract

This invention discloses a magnetic resonance imaging (MRI) CEST method, apparatus, medium, and device based on automatic image artifact calibration. The method includes: acquiring undersampled k-space data of the MRI CEST of the object to be imaged; acquiring a trained deep neural network, which consists of an image reconstruction module and an image artifact suppression module; and using the neural network to reconstruct and suppress artifacts in the CEST source image. This invention integrates an image artifact suppression strategy into a deep learning image reconstruction architecture. By leveraging image artifact priors, it significantly reduces the MRI CEST imaging time while ensuring image quality, thus solving the problem of reduced actual image resolution caused by data undersampling and T2 attenuation during MRI CEST imaging.

Description

Technical Field

[0001] This invention belongs to the field of magnetic resonance imaging, and in particular relates to a rapid chemical exchange saturation transfer imaging technique that combines deep neural networks and an automatic image artifact calibration strategy. Background Technology

[0002] Chemical Exchange Saturation Transfer (CEST) imaging selectively saturates specific endogenous protons and utilizes the exchange between water protons and endogenous protons to accumulate and amplify the magnetic resonance (MR) signals generated by endogenous protons. Therefore, it can sensitively detect the spatial distribution of low-concentration metabolites in vivo at the molecular level, enabling imaging of metabolites and important physiological parameters such as pH. This extends MR imaging from the structural level to the metabolic and functional levels. Currently, CEST imaging technology has demonstrated unique advantages and great potential in the diagnosis and treatment of many diseases. However, the high richness of CEST imaging information comes at the cost of long imaging times. To ensure the quality and reliability of CEST imaging, multiple frames often need to be acquired over a wide saturation shift frequency range. Therefore, although CEST imaging has a promising application prospect, the long imaging time severely restricts its development and clinical application.

[0003] To address the aforementioned issues, Compressed Sensing (CS) reconstruction methods based on the sparsity of the image transform domain have been used to accelerate CEST imaging. However, drawbacks such as time-consuming nonlinear iterative steps during reconstruction and the loss of detailed information in the reconstructed image limit the application of CS algorithms in CEST imaging. Parallel imaging methods based on multi-channel coil acquisition effectively reduce CEST imaging time by modeling and mining the spatial sensitivity differences between different channel coils. Among them, the representative Variable Speed ​​Sensitivity Coding (vSENSE) method provides an effective image artifact suppression strategy. The fast imaging (KIPI) method, which combines k-space and image space reconstruction, further demonstrates that this artifact suppression strategy can effectively solve the problem of reduced actual image resolution caused by T2 decay during data acquisition. However, the above artifact suppression strategies are entirely based on a pixel-by-pixel linear model, failing to accurately model the complex relationship between undersampled and fully sampled images, and ignoring the correlation between pixels.

[0004] In recent years, deep learning-based image reconstruction algorithms have demonstrated unique advantages in mining and utilizing implicit information in data, providing a novel approach to accelerating CEST imaging. However, an important type of information—artifact information in the image to be reconstructed—has not yet been effectively utilized as a priori for image reconstruction. How to leverage the data-driven advantages of deep neural networks to fully utilize priori information on image artifacts in the image reconstruction process, thereby achieving higher quality and faster CEST imaging, is a problem worthy of further research.

[0005] Please also refer to Chinese Patent Publication No. CN 114820849 A, which discloses a deep learning-based magnetic resonance CEST image reconstruction method, apparatus, medium, and device, and also provides an efficient multi-channel CEST data simulation method. This work effectively accelerates CEST imaging in clinical scenarios by combining deep neural networks and coil sensitivity encoding. Summary of the Invention

[0006] The purpose of this invention is to provide a rapid magnetic resonance imaging (CEST) method, device, medium, and equipment with short imaging time, good reconstruction effect, and high clinical and commercial value, based on the concept of automatic image artifact calibration, in order to solve the problem of long CEST imaging time.

[0007] The specific technical solution adopted in this invention is as follows:

[0008] In a first aspect, the present invention provides a magnetic resonance imaging (CEST) method based on automatic image artifact calibration, comprising:

[0009] S1. For the target object to be subjected to magnetic resonance CEST imaging, acquire multi-channel k-space data of the target object and the corresponding coil sensitivity map; the multi-channel k-space data consists of all k-space data frames acquired at all CEST saturation offset frequencies, including one frame of fully sampled k-space data of calibration frame and the remaining unsampled k-space data of non-calibration frames.

[0010] S2. The full sample k-space data of the calibration frame is subjected to z-spectrum modulation, and retrospective undersampling is performed, with the undersampling trajectory being consistent with the undersampled k-space data of the non-calibration frame, to obtain the undersampled k-space data of the calibration frame. At the same time, the full sample k-space data of the calibration frame after z-spectrum modulation is subjected to inverse Fourier transform and multi-channel merging to obtain the merged full sampled image of the calibration frame.

[0011] S3. Obtain a trained deep neural network, which is composed of an image reconstruction module and an image artifact suppression module cascaded together. The inputs of the image reconstruction module are uncalibrated frame undersampled k-space data, calibrated frame undersampled k-space data, and coil sensitivity map. After image reconstruction, the three inputs output an uncalibrated frame merged image with residual artifacts and a calibrated frame merged image with residual artifacts. The inputs of the image artifact suppression module are the calibrated frame merged full sampled image, the uncalibrated frame merged image with residual artifacts, and the calibrated frame merged image with residual artifacts. The calibrated frame merged full sampled image and the calibrated frame merged image with residual artifacts are first subjected to artifact estimation to obtain an image artifact calibration spectrum. Then, the image artifact calibration spectrum is used to perform artifact calibration on the uncalibrated frame merged image with residual artifacts. The final output is the reconstructed CEST source image.

[0012] S4. Input the uncalibrated frame undersampled k-space data, calibrated frame undersampled k-space data, coil sensitivity map, and calibrated frame merged full-sampled image obtained in S1 and S2 into the trained deep neural network to obtain the reconstructed and artifact-suppressed CEST source image.

[0013] As a preferred embodiment of the first aspect above, the coil sensitivity map in S1 is obtained by direct acquisition or by calculation using the full sample k-space data of the calibration frame.

[0014] As a preferred embodiment of the first aspect above, the fully sampled k-space data of the calibration frame can be obtained by full sampling or by downsampling and reconstruction through a parallel imaging algorithm.

[0015] As a preferred embodiment of the first aspect above, the specific operation steps of the z-spectrum modulation in S2 are as follows:

[0016] S11. Perform inverse Fourier transform and multi-channel merging on the undersampled k-space data of the uncalibrated frames to obtain N frames. F Uncalibrated frame undersampled merged image;

[0017] S12. Take the average value of each frame in the uncalibrated frame undersampled merged image, and take the pixel average value of each frame as a point in the z-spectrum in turn, thereby obtaining a spectrum composed of N F The z-spectrum is composed of points;

[0018] S13, Copy and stack the full-sample k-space data of the calibration frame N F Next, the stacked calibration frame full-sample k-space data is obtained;

[0019] S14. Perform channel-wise and pixel-wise dot product between the z-spectrum and the stacked calibration frame full-sample k-space data along the frame direction to obtain the calibration frame full-sample k-space data modulated by the z-spectrum.

[0020] As a preferred embodiment of the first aspect above, the image reconstruction module is implemented by an image reconstruction algorithm that incorporates parallel imaging.

[0021] As a preferred embodiment of the first aspect mentioned above, the image reconstruction module is replaced by inverse Fourier transform and multi-channel merging.

[0022] As a preferred embodiment of the first aspect above, the image artifact suppression module comprises an artifact estimation unit and an artifact calibration unit, and the artifact estimation unit and the artifact calibration unit are each implemented by a convolutional neural network based on an encoder-decoder structure; the input of the artifact estimation unit is a fully sampled image of calibration frames merged and a merged image of calibration frames containing residual artifacts, and the output of the artifact estimation unit is an image artifact calibration spectrum; the input of the artifact calibration unit is a merged image of uncalibrated frames containing residual artifacts and the image artifact calibration spectrum, and the output of the artifact calibration unit is the reconstructed CEST source image.

[0023] As a preferred embodiment of the first aspect above, the convolutional neural network based on the encoder-decoder structure includes an encoder and a decoder;

[0024] The encoder comprises two end-fused coding branches. The two input images of the convolutional neural network are input into different coding branches. Each coding branch's input image first undergoes two 3×3 convolutions to obtain its corresponding first intermediate feature. The first intermediate feature of each coding branch is then subjected to 2×2 mean pooling to obtain its corresponding second intermediate feature. The second intermediate feature of each coding branch is then concatenated with the second intermediate feature of the other coding branch along the channel direction, followed by two 3×3 convolutions to obtain its corresponding third intermediate feature. The third intermediate feature of each coding branch is then subjected to 2×2 mean pooling to obtain its corresponding fourth intermediate feature. The fourth intermediate feature of each coding branch... The first intermediate feature is concatenated with the fourth intermediate feature of another coding branch along the channel direction and then passed through two 3×3 convolutions to obtain the corresponding fifth intermediate feature. The fifth intermediate feature of each coding branch is then passed through 2×2 mean pooling to obtain the corresponding sixth intermediate feature. The sixth intermediate feature of each coding branch is then concatenated with the sixth intermediate feature of another coding branch along the channel direction and then passed through two 3×3 convolutions to obtain the corresponding seventh intermediate feature. The seventh intermediate feature of each coding branch is then passed through 2×2 mean pooling to obtain the corresponding eighth intermediate feature. The eighth intermediate features of the two coding branches are then concatenated along the channel direction and then passed through two 3×3 convolutions and one 2×2 upsampling layer before being input into the decoder.

[0025] In the decoder, the features input from the encoder are first concatenated with the seventh intermediate features from the two coding branches along the channel direction, and then sequentially passed through two layers of 3×3 convolution and one layer of 2×2 upsampling to obtain the first upsampling result. The first upsampling result is then concatenated with the fifth intermediate features from the two coding branches along the channel direction, and then sequentially passed through two layers of 3×3 convolution and one layer of 2×2 upsampling to obtain the second upsampling result. The second upsampling result is then concatenated with the third intermediate features from the two coding branches along the channel direction, and then sequentially passed through two layers of 3×3 convolution and one layer of 2×2 upsampling to obtain the third upsampling result. The third upsampling result is then concatenated with the first intermediate features from the two coding branches along the channel direction, and then sequentially passed through two layers of 3×3 convolution and one layer of 1×1 convolution to output the final result image.

[0026] In a second aspect, the present invention provides a fast magnetic resonance imaging (CEST) data processing device based on deep learning, which includes a memory and a processor;

[0027] The memory is used to store computer programs;

[0028] The processor is configured to, when executing the computer program, implement the magnetic resonance CEST imaging method based on automatic calibration of image artifacts as described in any of the embodiments of the first aspect.

[0029] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements magnetic resonance CEST imaging based on automatic calibration of image artifacts as described in any of the embodiments of the first aspect.

[0030] Fourthly, the present invention provides a magnetic resonance imaging device, which includes a magnetic resonance scanner and a control unit;

[0031] The magnetic resonance scanner is used to obtain CEST multi-channel undersampled k-space data of the target object through a parallel imaging method;

[0032] The control unit stores a computer program that, when executed, is used to implement magnetic resonance CEST imaging based on automatic image artifact calibration as described in any of the schemes in the first aspect.

[0033] The beneficial effects of this invention are as follows: This invention integrates an image artifact suppression strategy into a deep learning image reconstruction architecture. By leveraging image artifact priors, it achieves higher reconstruction results than existing methods. Simultaneously, it solves the problem of reduced actual image resolution caused by data undersampling and T2 attenuation during magnetic resonance imaging (CEST). It can achieve at least a 6-fold acceleration of CEST while maintaining image quality. Furthermore, the automatic image artifact calibration mechanism involved in this invention has high versatility and flexibility, and can be used in conjunction with various existing image reconstruction algorithms, possessing high clinical and commercial value. Attached Figure Description

[0034] Figure 1 The flowchart shows the steps of the magnetic resonance CEST imaging method based on automatic calibration of image artifacts.

[0035] Figure 2 This is a flowchart illustrating the overall implementation of Embodiments 1 and 2 of the present invention;

[0036] Figure 3 Diagram of a deep neural network architecture;

[0037] Figure 4 This is a structural diagram of the artifact suppression module in a deep neural network.

[0038] Figure 5 The network structure diagrams of the artifact estimation unit and the artifact suppression unit in the artifact suppression module are shown in Examples 1 and 2.

[0039] Figure 6 The results of artifact suppression at 2.5 ppm in the source image of a brain tumor patient in Example 1 are shown. Figure (a) from left to right shows the full-sampled source image, the effect before artifact suppression with acceleration factor R=3, the result obtained using the traditional artifact suppression algorithm, and the result obtained using the method of the present invention. Figure (b) shows the k-space undersampling mask used in this experiment with acceleration factor 3. Figure (c) from left to right shows the error graphs of the results before artifact suppression, the traditional artifact suppression algorithm, and the results obtained by the present invention relative to the full-sampled reference image, and the corresponding normalized root mean square error (nRMSE).

[0040] Figure 7Figure 2 shows the reconstruction results of APTw images from a brain tumor patient in Example 2. Figure (a) from left to right shows the fully sampled APTw reference image, the reconstruction results from the GRAPPA algorithm, the BCS algorithm, the CEST-VN algorithm, and the reconstruction result of the method of this invention. Figure (b) shows the k-space undersampling mask used in this experiment, with an acceleration factor of 5. Figure (c) from left to right shows the error graphs of the reconstruction results from the GRAPPA algorithm, BCS algorithm, CEST-VN algorithm, and the method of this invention relative to the fully sampled reference image, along with the corresponding normalized root mean square error (nRMSE).

[0041] Figure 8 This is a statistical analysis of the image reconstruction results of 5 brain tumor patients in Example 2. Figure (a) depicts the change in the normalized root mean square error (nRMSE) value of the APTw image obtained by the GRAPPA algorithm, BCS algorithm, CEST-VN algorithm and the method of the present invention as the acceleration factor R increases from 4 to 6. Figure (b) reflects the change in the normalized root mean square error (nRMSE) value of the CEST source image obtained by the GRAPPA algorithm, BCS algorithm, CEST-VN algorithm and the method of the present invention as the acceleration factor R increases from 4 to 6. Detailed Implementation

[0042] The present invention will be further described and illustrated below with reference to the accompanying drawings and specific embodiments. The technical features of each embodiment of the present invention can be combined accordingly, provided that there is no mutual conflict.

[0043] like Figure 1 As shown, as a preferred embodiment of the present invention, a magnetic resonance CEST imaging method based on automatic image artifact calibration is provided, which includes the following steps:

[0044] S1. For the target object to be subjected to CEST magnetic resonance imaging, acquire multi-channel k-space data of the target object and the corresponding coil sensitivity map. The multi-channel k-space data consists of all acquired k-space data frames at all CEST saturation offset frequencies, including one fully sampled k-space data frame as a calibration frame and the remaining undersampled k-space data frames as uncalibrated frames. For ease of description, in this invention, the aforementioned fully sampled k-space data frame is defined as the calibration frame fully sampled k-space data, and the remaining undersampled k-space data frames are defined as the uncalibrated frame undersampled k-space data.

[0045] It should be noted that, in the embodiments of the present invention, in step S1, the aforementioned coil sensitivity map can be directly acquired by the device, or it can be calculated using the full-sample k-space data of the calibration frame. Furthermore, the full-sample k-space data of the calibration frame can be obtained either through full sampling or by downsampling and reconstructing it using a parallel imaging algorithm.

[0046] It should be noted that, in the embodiments of the present invention, in step S1, the acquisition of multi-channel k-space data of the target object can be either offline or online. For the offline method, it is only necessary to read the data from the storage device containing it; while for the online method, it is necessary to use a magnetic resonance imaging device to perform CEST imaging scans on the target object to acquire this data.

[0047] It should be noted that, in the embodiments of the present invention, in step S1, the target object can be any object that can be imaged by magnetic resonance CEST, such as the brain of a patient.

[0048] S2. The full-sample k-space data of the calibration frame is subjected to z-spectrum modulation, and retrospective undersampling is performed, with the trajectory of the retrospective undersampling being consistent with the undersampled k-space data of the non-calibration frame, to obtain the undersampled k-space data of the calibration frame. At the same time, the full-sampled k-space data of the calibration frame after z-spectrum modulation is subjected to inverse Fourier transform and multi-channel merging to obtain the merged full-sampled image of the calibration frame.

[0049] It should be noted that, in the embodiments of the present invention, in step S2, the trajectory used for retrospective undersampling must be consistent with the uncalibrated frame undersampling k-space data.

[0050] It should be noted that, in the embodiments of the present invention, the specific operation steps of z-spectrum modulation in step S2 are as follows:

[0051] S11. Perform inverse Fourier transform and multi-channel merging on the uncalibrated frame undersampled k-space data to obtain an uncalibrated frame undersampled merged image. The number of frames in the uncalibrated frame undersampled merged image is defined as N. F ;

[0052] S12. Take the average value of each frame in the uncalibrated frame undersampled merged image, and use the pixel average value of each frame as a point in the z-spectrum, thereby obtaining a spectrum composed of N F The z-spectrum is composed of points;

[0053] S13, Copy and stack the full-sample k-space data of the calibration frame N F This process yields the stacked calibration frame full-sample k-space data.

[0054] S14. Perform channel-wise and pixel-wise dot product between the z-spectrum and the stacked calibration frame full-sample k-space data along the frame direction to obtain the calibration frame full-sample k-space data modulated by the z-spectrum.

[0055] S3. Obtain the trained deep neural network to reconstruct images from the data obtained in S1 and S2, thereby achieving fast magnetic resonance imaging (CEST). The inputs to this deep neural network are uncalibrated frame undersampled k-space data, calibrated frame undersampled k-space data, coil sensitivity maps, and calibrated frame merged fully sampled images. The output of this deep neural network is the CEST source image reconstructed by the network and with artifact suppression.

[0056] The deep neural network model consists of a cascaded image reconstruction module and an image artifact suppression module. The image reconstruction module takes as input undersampled k-space data of uncalibrated frames, undersampled k-space data of calibrated frames, and a coil sensitivity map. After image reconstruction, the outputs are a merged image of uncalibrated frames containing residual artifacts (i.e., the reconstructed uncalibrated frame image) and a merged image of calibrated frames containing residual artifacts (i.e., the reconstructed calibrated frame image). The image artifact suppression module takes as input the merged calibrated frame fully sampled image, the merged uncalibrated frame image containing residual artifacts, and the merged calibrated frame image containing residual artifacts. The merged calibrated frame fully sampled image and the merged calibrated frame image containing residual artifacts are first subjected to artifact estimation to obtain an image artifact calibration spectrum. Then, the image artifact calibration spectrum is used to perform artifact calibration on the merged uncalibrated frame image containing residual artifacts. The final output is the reconstructed CEST source image.

[0057] In embodiments of the present invention, the image reconstruction module can be implemented by any image reconstruction algorithm combining parallel imaging, or it can be replaced by inverse Fourier transform and multi-channel merging; while the image artifact suppression module consists of an artifact estimation unit and an artifact calibration unit, and the aforementioned artifact estimation unit and the aforementioned artifact calibration unit are each implemented by a convolutional neural network based on an encoder-decoder structure. The input of the artifact estimation unit is the fully sampled image of the calibration frame merging and the calibration frame merging image containing residual artifacts, and the output of the artifact estimation unit is the image artifact calibration spectrum; the input of the artifact calibration unit is the uncalibrated frame merging image containing residual artifacts and the image artifact calibration spectrum, and the output of the artifact calibration unit is the reconstructed CEST source image.

[0058] The aforementioned convolutional neural network based on an encoder-decoder structure includes an encoder and a decoder, and theoretically, any convolutional neural network capable of performing the corresponding function can be used to implement it. In the embodiments of the present invention, the preferred structure of this convolutional neural network based on the encoder-decoder structure is as follows: Figure 5 As shown.

[0059] It should be noted that before using the aforementioned deep neural network for image reconstruction, network parameter optimization is required. The specific method for network parameter optimization can refer to the conventional training methods for deep neural networks, i.e., setting a loss function and using an optimizer to optimize the set of trainable parameters θ until the optimal parameter set is obtained.

[0060] The loss function used for network training can be adjusted and optimized according to actual conditions, and the optimal set of network parameters is obtained. Then, the actual undersampled CEST source image can be reconstructed using a network containing these optimal parameters.

[0061] S4. Input the uncalibrated frame undersampled k-space data, calibrated frame fully sampled k-space data, coil sensitivity map, and calibrated frame merged fully sampled image obtained in S1 and S2 into the trained deep neural network to obtain the reconstructed and artifact-suppressed CEST source image.

[0062] The fully sampled CEST source images obtained from the network reconstruction described above can be further processed and analyzed using CEST according to actual needs.

[0063] The following describes the specific technical effects of the methods described in S1-S4 and S11-S14, combined with specific embodiments, so that those skilled in the art can better understand the essence of the present invention. To demonstrate the flexibility of the present invention in practical use, Embodiment 1 uses inverse Fourier transform and multi-channel merging to implement the function of the image reconstruction module, while Embodiment 2 uses a deep neural network to implement the function of the image reconstruction module.

[0064] Example 1

[0065] The magnetic resonance CEST imaging method based on automatic image artifact calibration described in S1-S4 above will be applied to a specific embodiment to demonstrate its technical effects. The specific framework and flow of the methods described in S1-S4 and S11-S14 above are as previously stated and will not be repeated in full in this embodiment. The focus below is on demonstrating the specific implementation details and technical effects of each step. The complete flow of this embodiment is as follows: Figure 2 As shown.

[0066] 1. Data Preparation

[0067] 1.1 Training Data Simulation

[0068] A multi-channel magnetic resonance CEST source image dataset was generated through simulation. The detailed simulation process can be found in Chinese Patent Publication No. CN114820849A. This embodiment simulated and generated a total of 9250 sets of multi-channel fully sampled k-space data samples.

[0069] 1.2 MRI Data Acquisition

[0070] In order to monitor the model training status and test the model's generalization performance on real clinical data, this embodiment obtained measured CEST data through MR experiments as samples in the test set.

[0071] Specifically, the brains of two glioma patients were scanned using a 3-Tesla Siemens scanner (MAGNETOM Prisma, Siemens Healthcare, Erlangen, Germany) with a 16-channel head coil. The sequence used in the scanning experiment was the 2D fast spin echo (TSE) CEST imaging sequence, with the following acquisition parameters: saturation pulse duration of 1.0 s, intensity of 2 μT, flip angle FA = 90°; echo time (TE) = 6.7 ms; repetition time (TR) = 3 s; field of view (FOV) = 212 × 186 mm. 2 Resolution = 2.2 × 2.2 mm 2 Layer thickness = 5mm; Turbine coefficient = 96. A total of 54 frequency offset frames were collected, including unsaturated frame S0 and saturated frames saturated at frequencies of 0, ±0.25, ±0.5, ±0.75, ±1, ±1.5, ±2(2), ±2.5(2), ±3(2), ±3.25(2), ±3.5(6), ±3.75(2), ±4(2), ±4.5, ±5, ±6ppm (the numbers in parentheses represent the number of times the corresponding frequency frame was repeatedly collected).

[0072] To correct for B0 field inhomogeneity, this embodiment calculates the B0 spectrum using the direct water saturation shift referencing (WASSR) method. The WASSR sequence used has a TR of 2 s and a saturation pulse intensity of 0.5 μT. A total of 26 frequency points equidistantly distributed within the range of -1.5 to 1.5 ppm were acquired, and other parameters are consistent with the CEST imaging sequence described above.

[0073] 1.3 Data Preprocessing and Dataset Construction

[0074] In this embodiment, one frame at -3.75 ppm is set as the calibration frame, and the remaining 53 frames are non-calibration frames. To construct the network's training, validation, and test sets, the fully sampled k-space data samples acquired from simulations and MRI experiments need to be preprocessed. The specific preprocessing method is as follows:

[0075] Step 1: For the full sample k-space data of the calibration frame, use the ESPIRiT algorithm to calculate the corresponding coil sensitivity map.

[0076] Step 2: Based on the coil sensitivity map obtained in Step 1, perform inverse Fourier transform and multi-channel merging on the 53 frames of uncalibrated frame full-sample k-space data to obtain the 53 frames of uncalibrated frame full-sample images.

[0077] Step 3: For the 53 frames of uncalibrated frame fully sampled k-space data, use as follows: Figure 6 (b) shows the 1D variable density Cartesian undersampling mode, which performs 3 times retrospective undersampling on 53 frames of uncalibrated frame full sampled k-space data to obtain 53 frames of uncalibrated frame undersampled k-space data. Then, based on the coil sensitivity map obtained in Step 1, the 53 frames of uncalibrated frame undersampled k-space data are subjected to inverse Fourier transform and multi-channel merging to obtain 53 frames of uncalibrated frame undersampled images.

[0078] Step 4: For the full sample k-space data of the calibration frame, copy and stack the full sample k-space data of the calibration frame 53 times to obtain 53 full sample k-space data of the calibration frame. Then, based on the coil sensitivity map obtained in Step 1, perform inverse Fourier transform and multi-channel merging on the 53 full sample k-space data of the calibration frame to obtain 53 full sample images of the calibration frame.

[0079] Step 5: Use the 1D variable density Cartesian undersampling mode described in Step 3 to retrospectively undersample the 53-frame calibration frame full-sample k-space data described in Step 4 to obtain 53-frame calibration frame undersampled k-space data. Then, based on the coil sensitivity map obtained in Step 1, perform inverse Fourier transform and multi-channel merging on the 53-frame calibration frame undersampled k-space data to obtain 53-frame calibration frame undersampled images.

[0080] Each independent data sample thus consists of 1) 53 undersampled calibration frames, 2) 53 fully sampled calibration frames, 3) 53 undersampled uncalibrated frames, and 4) 53 fully sampled uncalibrated frames. The first three will serve as input to the neural network, while the 53 fully sampled uncalibrated frames will serve as the sample's label (Ground Truth). Of the 9250 sets of samples generated in the simulation, 7000 sets were randomly selected as the training set, and the remainder as the validation set. The model's test set was constructed from samples acquired during MRI experiments.

[0081] 2. Model building and training

[0082] Build using the deep learning framework PyTorch, such as Figure 3 The deep neural network shown comprises an image reconstruction module and an image artifact suppression module. The image reconstruction module is replaced by inverse Fourier transform and multi-channel merging. The architecture of the image artifact suppression module is as follows: Figure 4As shown, the image artifact suppression module used in this embodiment consists of an artifact estimation unit and an artifact calibration unit. Both units employ convolutional neural networks based on an encoder-decoder structure, as detailed below. Figure 5 As shown. The structure of this convolutional neural network based on the encoder-decoder structure is as follows:

[0083] The encoder contains two end-fused coding branches. The two input images of the convolutional neural network are input into different coding branches. Each coding branch's input image first undergoes two 3×3 convolutions to obtain its corresponding first intermediate feature. The first intermediate feature of each coding branch is then subjected to 2×2 mean pooling to obtain its corresponding second intermediate feature. The second intermediate feature of each coding branch is then concatenated with the second intermediate feature of the other coding branch along the channel direction, followed by two 3×3 convolutions to obtain its corresponding third intermediate feature. The third intermediate feature of each coding branch is then subjected to 2×2 mean pooling to obtain its corresponding fourth intermediate feature. The fourth intermediate feature of each coding branch... The feature is then concatenated with the fourth intermediate feature of another coding branch along the channel direction and then passed through two 3×3 convolutions to obtain the corresponding fifth intermediate feature. The fifth intermediate feature of each coding branch is then passed through 2×2 mean pooling to obtain the corresponding sixth intermediate feature. The sixth intermediate feature of each coding branch is then concatenated with the sixth intermediate feature of another coding branch along the channel direction and then passed through two 3×3 convolutions to obtain the corresponding seventh intermediate feature. The seventh intermediate feature of each coding branch is then passed through 2×2 mean pooling to obtain the corresponding eighth intermediate feature. The eighth intermediate features of the two coding branches are then concatenated along the channel direction and then passed through two 3×3 convolutions and one 2×2 upsampling layer before being input into the decoder.

[0084] In the decoder, the features input from the encoder are first concatenated with the seventh intermediate features from the two coding branches along the channel direction, and then passed through two layers of 3×3 convolution and one layer of 2×2 upsampling to obtain the first upsampling result. The first upsampling result is then concatenated with the fifth intermediate features from the two coding branches along the channel direction, and then passed through two layers of 3×3 convolution and one layer of 2×2 upsampling to obtain the second upsampling result. The second upsampling result is then concatenated with the third intermediate features from the two coding branches along the channel direction, and then passed through two layers of 3×3 convolution and one layer of 2×2 upsampling to obtain the third upsampling result. The third upsampling result is then concatenated with the first intermediate features from the two coding branches along the channel direction, and then passed through two layers of 3×3 convolution and one layer of 1×1 convolution to output the final image.

[0085] It should be noted that for the artifact estimation unit, the inputs to its two coding branches are the fully sampled image of the calibration frame merge and the calibration frame merge image containing residual artifacts, respectively. For the artifact calibration unit, the inputs to its two coding branches are the uncalibrated frame merge image containing residual artifacts and the image artifact calibration spectrum output by the artifact estimation unit, respectively.

[0086] The network was trained end-to-end using the multi-channel CEST dataset generated by simulation. During training, the Adam optimizer was used to optimize the network parameter set θ by minimizing the loss function value shown in Equation (1) below, thus obtaining the relatively optimal network parameters. The weight coefficient μ(e) in the loss function is set to a constant of 10, and all neural networks are trained on two NVIDIA RTX 3080Ti GPUs for 50 iterations. During training, the validation set is used to test the generalization ability of the current model and determine whether the network has reached the convergence condition.

[0087] The loss function used in this embodiment is:

[0088]

[0089] Where θ represents the set of all learnable parameters of the network, n is used to count the samples in the training set, and N n The total number of training samples in the training set; S represents the nth CEST source image output by a deep neural network. n This represents the label of the nth group of fully sampled CEST source images; Indicates based on The intensity spectrum of the magnetization transfer rate asymmetry effect corresponding to the m-th offset frequency was calculated. Indicates based on S n The intensity spectrum of the magnetization transfer rate asymmetry effect corresponding to the m-th offset frequency is calculated, N M denoted as the total number of positive and negative frequency pairs; μ(e) represents the weight coefficient related to the number of training rounds e of the network.

[0090] 3. Image Reconstruction and Post-processing

[0091] After training, the network described above can be used for image artifact suppression based on actual data acquisition. Specifically, in this embodiment, undersampled calibration frame images, fully sampled calibration frame images, and undersampled uncalibrated frame images from brain tumor patients are input into the network, and the network output is the CEST source image after artifact suppression. Simultaneously, to compare and evaluate the performance of this algorithm, the traditional artifact suppression strategy in the KIPI algorithm is used to suppress artifacts on the same undersampled images, and this result is used as a control group.

[0092] In addition, to eliminate the influence of B0 field inhomogeneity, the WASSR method was used to perform B0 correction on the CEST source images reconstructed by various algorithms, and then the APTw image was calculated using the following formula:

[0093]

[0094] We will use APTw images as an example to perform CEST analysis.

[0095] 4. Results Analysis

[0096] Figure 7 Using a source image at 2.5 ppm from a brain tumor patient as an example, the image artifact suppression performance of this algorithm is demonstrated. The experimental results show that, under the same 3x undersampling mode, this algorithm can effectively suppress image artifacts caused by data undersampling, and compared to traditional algorithms, it can obtain a higher quality source image. The difference between the image after artifact suppression using this method and the fully sampled image is minimal.

[0097] Example 2

[0098] Unlike Example 1, this example uses a deep neural network to implement the image reconstruction module. The complete process of this example is as follows: Figure 2 As shown. The training data simulation and MRI data acquisition process in this embodiment are the same as in Embodiment 1, and will not be described again here.

[0099] 1. Data Preprocessing and Dataset Construction

[0100] In this embodiment, one frame at -3.75 ppm is set as the calibration frame, and the remaining 53 frames are non-calibration frames. To construct the network's training, validation, and test sets, the fully sampled k-space data samples acquired from simulations and MRI experiments need to be preprocessed. The specific preprocessing method is as follows:

[0101] Step 1: For the full sample k-space data of the calibration frame, use the ESPIRiT algorithm to calculate the corresponding coil sensitivity map.

[0102] Step 2: For the 53 frames of uncalibrated frame full-sample k-space data, based on the coil sensitivity map obtained in Step 1, the Sensitivity Encoding (SENSE) algorithm is used to reconstruct the image of the 53 frames of uncalibrated frame full-sample k-space data, and obtain the 53 frames of uncalibrated frame full-sample image.

[0103] Step 3: For the 53 frames of uncalibrated frame fully sampled k-space data, use as follows: Figure 7 (b) shows a 1D variable density Cartesian undersampling mode, which performs 5 times retrospective undersampling on 53 frames of uncalibrated frame fully sampled k-space data to obtain 53 frames of uncalibrated frame undersampled k-space data.

[0104] Step 4: Based on the 53 frames of uncalibrated frame undersampled k-space data generated in Step 3, Fourier transform and multi-channel merging are used to obtain the corresponding merged image. Then, the mean of each merged image is taken to obtain a z-spectrum consisting of 53 points.

[0105] Step 5: For the full sample k-space data of the calibration frame, use the z-spectrum obtained in Step 4 to perform z-spectrum modulation on the full sample k-space data of the calibration frame to obtain the modulated full sample k-space data of the calibration frame of 53 frames. Then, based on the coil sensitivity map obtained in Step 1, use the sensitivity coding (SENSE) algorithm to reconstruct the image of the full sample k-space data of the calibration frame of 53 frames to obtain the full sample image of the calibration frame of 53 frames.

[0106] Step 6: Based on the modulated 53-frame calibration frame full-sample k-space data obtained in Step 5, the 1D variable density Cartesian undersampling mode described in Step 3 is used to undersample the modulated 53-frame calibration frame full-sample k-space data to obtain the 53-frame calibration frame undersampled k-space data.

[0107] Each independent data sample thus consists of: 1) a coil sensitivity map, 2) 53 frames of undersampled k-space data from calibration frames, 3) 53 frames of undersampled k-space data from uncalibrated frames, 4) 53 frames of fully sampled images from calibration frames, and 5) 53 frames of fully sampled images from uncalibrated frames. The first four will serve as inputs to the neural network, while the 53 frames of fully sampled CEST source images from uncalibrated frames will serve as the sample's label (Ground Truth). Of the 9250 sets of samples generated in the simulation, 7000 sets were randomly selected as the training set, and the remainder as the validation set. The model's test set was constructed from samples acquired during MRI experiments.

[0108] 2. Model building and training

[0109] Build using the deep learning framework PyTorch, such as Figure 3 The deep neural network shown comprises an image reconstruction module and an image artifact suppression module. The image reconstruction module uses the CEST-VN algorithm (see the deep learning-based magnetic resonance CEST image reconstruction method in Chinese patent publication CN 114820849 A) ​​to achieve image reconstruction. The number of iteration modules K is set to 10, and the time-frequency convolution kernel size in each iteration module is set to 7×7×7. The number of basis functions and convolution groups N is... v The value is 48, and the Gaussian radial basis function N is... w The number is 31. Furthermore, the image artifact suppression module structure used in this embodiment is the same as in Embodiment 1, and will not be repeated here.

[0110] The network was trained end-to-end using the multi-channel CEST dataset generated by simulation. During training, the Adam optimizer was used to optimize the network parameter set θ by minimizing the loss function value shown in Equation (1) above, thus obtaining the relatively optimal network parameters. The weight coefficient μ(e) in the loss function is set to a constant of 10, and all neural networks are run for 40 iterations on two NVIDIA RTX 3080Ti GPUs. During training, the validation set is used to test the generalization ability of the current model and determine whether the network has reached the convergence condition.

[0111] 3. Image Reconstruction and Post-processing

[0112] After training, the network described above can be used for image reconstruction based on actual acquired data. Specifically, in this embodiment, the undersampled k-space data of calibration frames, undersampled k-space data of uncalibrated frames, fully sampled images of calibration frames, and coil sensitivity maps are used as inputs to the network. The network output is the fully sampled CEST source image after reconstruction and artifact suppression. Simultaneously, to compare and evaluate the performance of this algorithm, the GRAPPA algorithm, the BCS algorithm combined with parallel imaging, and the CEST-VN algorithm are used to reconstruct images from the same undersampled data.

[0113] In addition, to eliminate the influence of B0 field inhomogeneity, the WASSR method was used to perform B0 correction on the CEST source images reconstructed by various algorithms, and then the APTw image was calculated using the following formula:

[0114]

[0115] We will use APTw images as an example to perform CEST analysis.

[0116] 4. Results Analysis

[0117] Figure 7 This paper demonstrates the reconstruction results of an APTw image from a brain tumor patient using the proposed algorithm. The experimental results show that, under the same 5x undersampling mode, compared to advanced image reconstruction algorithms such as GRAPPA, BCS, and CEST-VN, this reconstruction algorithm achieves significantly higher CEST image quality, with minimal difference between the reconstructed image and the fully sampled image.

[0118] Figure 8 This paper presents statistical results on the reconstruction quality of APTw images and corresponding source images of five brain tumor patients using the proposed algorithm at different acceleration levels. The results demonstrate that, compared to advanced image reconstruction algorithms such as GRAPPA, BCS, and CEST-VN, this algorithm achieves statistically higher image reconstruction quality and exhibits stronger robustness across different acceleration scenarios.

[0119] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a fast magnetic resonance CEST imaging data processing device based on deep learning, corresponding to the magnetic resonance CEST imaging method based on automatic image artifact calibration provided in the above embodiments, which includes a memory and a processor.

[0120] The memory is used to store computer programs;

[0121] The processor is configured to implement the magnetic resonance CEST imaging method based on automatic calibration of image artifacts as described above when executing the computer program.

[0122] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the magnetic resonance CEST imaging method based on automatic calibration of image artifacts provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, implements the magnetic resonance CEST imaging method based on automatic calibration of image artifacts as described above.

[0123] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media can also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs. Of course, with the widespread use of cloud servers, the aforementioned software programs can also be hosted on cloud platforms to provide corresponding services; therefore, computer-readable storage media are not limited to the form of local hardware.

[0124] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0125] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the device and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0126] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0127] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a magnetic resonance imaging device corresponding to the magnetic resonance CEST imaging method based on automatic image artifact calibration provided in the above embodiments, which includes a magnetic resonance scanner and a control unit. Wherein:

[0128] Magnetic resonance scanners are used to obtain k-space undersampled frames, k-space undersampled masks, and corresponding coil sensitivity maps of a target object at all frequency offsets through parallel imaging methods.

[0129] The control unit stores a computer program that, when executed, is used to implement the magnetic resonance CEST imaging method based on automatic calibration of image artifacts as described above.

[0130] It should be noted that the magnetic resonance imaging (MRI) device can be any MRI scanner capable of implementing parallel imaging methods. Its structure is existing technology, and mature commercial products can be used; the specific model is not limited. Furthermore, in addition to storing the aforementioned computer programs, the control unit of the MRI device should also contain the imaging sequences and other software programs necessary for implementing CEST imaging.

[0131] Of course, the aforementioned control unit can be an independent control unit or a control unit built into the magnetic resonance scanner. That is, the magnetic resonance CEST imaging method based on automatic calibration of image artifacts can be integrated into the control unit of the magnetic resonance imaging device in the form of a data processing program, so that the reconstruction results can be directly output by the magnetic resonance scanner without the need for additional control units.

[0132] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method of magnetic resonance CEST imaging based on automatic calibration of image artifacts, characterized in that, The method comprises the following steps: S1, acquiring multi-channel k-space data of a target object to be subjected to magnetic resonance CEST imaging, and a corresponding coil sensitivity map; the multi-channel k-space data is composed of k-space data frames acquired at all CEST saturation shift frequencies, including a calibration frame full-sampling k-space data and the rest of the non-calibration frame undersampling k-space data; S2, z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation is performed on the calibration frame full-sampling k-space data, and the z-spectrum modulation ​ ​ 2. The method of magnetic resonance CEST imaging based on automatic calibration of image artifacts as claimed in claim 1, wherein, ​ 3. The method of automatically calibrating magnetic resonance CEST imaging based on image artifacts as claimed in claim 1, wherein, ​ S11, inverse Fourier transform and multi-channel merging are performed on the non-calibration frame undersampled k-space data to obtain non-calibration frame undersampled merged images with N frames F ​ S12, averaging each frame in the non-calibration frame undersampling merged image, the pixel mean value of each frame image is sequentially taken as a point in the z spectrum, thereby obtaining a z spectrum composed of N F points. S13, copying and stacking N the calibration frame full sampling k-space data F stacked calibration frame full sampling k-space data; ​ 4. The method of automatically calibrating magnetic resonance CEST imaging based on image artifacts as claimed in claim 1, wherein, ​ 5. The method of automatically calibrating magnetic resonance CEST imaging based on image artifacts as claimed in claim 1, wherein, ​ 6. The method for magnetic resonance CEST imaging based on automatic calibration of image artifacts as claimed in claim 1, wherein, ​ An input of the artifact calibration unit is a non-calibration frame combined image containing remaining artifacts and the image artifact calibration spectrum, and an output of the artifact calibration unit is a reconstructed CEST source image.

7. The method of magnetic resonance CEST imaging based on automatic calibration of image artifacts as claimed in claim 6, wherein, The convolutional neural network based on an encoding-decoding structure comprises an encoder and a decoder; The encoder comprises two encoding branches fused at the end, and two input images of the convolutional neural network are respectively input into different encoding branches; Each input image of each encoding branch is respectively subjected to two layers of 3*3 convolution to obtain corresponding first intermediate features, and each first intermediate feature of each encoding branch is subjected to 2*2 mean value pooling to obtain corresponding second intermediate features, and the second intermediate features of each encoding branch are respectively spliced with the second intermediate features of the other encoding branch along the channel direction and then subjected to two layers of 3*3 convolution to obtain corresponding third intermediate features, and each third intermediate feature of each encoding branch is subjected to 2*2 mean value pooling to obtain corresponding fourth intermediate features, and the fourth intermediate features of each encoding branch are respectively spliced with the fourth intermediate features of the other encoding branch along the channel direction and then subjected to two layers of 3*3 convolution to obtain corresponding fifth intermediate features, and each fifth intermediate feature of each encoding branch is subjected to 2*2 mean value pooling to obtain corresponding sixth intermediate features, and the sixth intermediate features of each encoding branch are respectively spliced with the sixth intermediate features of the other encoding branch along the channel direction and then subjected to two layers of 3*3 convolution to obtain corresponding seventh intermediate features, and each seventh intermediate feature of each encoding branch is subjected to 2*2 mean value pooling to obtain corresponding eighth intermediate features, and the eighth intermediate features of the two encoding branches are spliced along the channel direction and then subjected to two layers of 3*3 convolution and one layer of 2*2 up-sampling to be input into the decoder; In the decoder, the features input from the encoder are spliced with the seventh intermediate features of the two encoding branches along the channel direction and then subjected to two layers of 3*3 convolution and one layer of 2*2 up-sampling in sequence to obtain a first up-sampling result, the first up-sampling result is spliced with the fifth intermediate features of the two encoding branches along the channel direction and then subjected to two layers of 3*3 convolution and one layer of 2*2 up-sampling in sequence to obtain a second up-sampling result, the second up-sampling result is spliced with the third intermediate features of the two encoding branches along the channel direction and then subjected to two layers of 3*3 convolution and one layer of 2*2 up-sampling in sequence to obtain a third up-sampling result, and the third up-sampling result is spliced with the first intermediate features of the two encoding branches along the channel direction and then subjected to two layers of 3*3 convolution and one layer of 1*1 convolution in sequence to output a final result image.

8. A deep learning based fast magnetic resonance CEST imaging data processing apparatus, characterized in that, comprise a memory and a processor; The memory is configured to store a computer program; The processor is configured to implement the magnetic resonance CEST imaging method based on image artifact automatic calibration according to any one of claims 1-7 when executing the computer program.

9. A computer-readable storage medium, characterized in that, The storage medium has a computer program stored thereon, and the computer program, when executed by a processor, implements the magnetic resonance CEST imaging method based on image artifact automatic calibration according to any one of claims 1-7.

10. A magnetic resonance imaging apparatus, characterized by comprise a magnetic resonance scanner and a control unit; The magnetic resonance scanner is used to obtain CEST multi-channel k-space data of a target object by a parallel imaging method; The control unit stores a computer program which, when executed, implements the image-artifact-based automatic calibration method for magnetic resonance CEST imaging according to any one of claims 1-7.

Images