Method for accelerating chemical exchange saturation transfer contrast imaging

Through an unsupervised learning algorithm combining non-Cartesian sequences and implicit neural expression, the problems of long CEST imaging time and low image quality are solved, and high-multiple acceleration and high-quality CEST image reconstruction are achieved, which is suitable for clinical applications.

CN119001566APending Publication Date: 2024-11-22SHANGHAI JIAOTONG UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411163909.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-05-22
Filing Date
2024-08-22
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing chemical exchange saturation transfer imaging (CEST) technology has problems such as long imaging time and low image quality, especially in the case of high acceleration, the reconstruction quality is difficult to guarantee, and the lack of effective sampling strategies and deep learning models limits it. Extensiveness in clinical applications.

Method used

Multi-dimensional sparse collection is performed using non-Cartesian sequences, combined with implicit neural expression and explicit prior information, and the unsupervised learning algorithm INRESP is carried out on the Z-spectrum dimension through stacked star trajectories or three-dimensional spherical radial trajectories with golden angle rotation. Randomize undersampling and use mixed feature hash encoding for image reconstruction to achieve high-multiple acceleration and high-quality CEST image reconstruction.

Benefits of technology

Significantly shortens the CEST imaging time, improves the reconstruction quality and efficiency of the image, reduces motion artifacts, achieves high plural acceleration, and obtains results close to the full sampling gold standard without full sampling parameter data, which is noise-robust and strong generalization ability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119001566A_ABST
    Figure CN119001566A_ABST
Patent Text Reader

Abstract

The invention discloses a method for accelerating chemical exchange saturation transfer contrast imaging, which relates to the field of magnetic resonance imaging, and comprises the following steps of: acquiring k space data with different frequency offsets by adopting a non-Cartesian sequence, inputting the k space data with different frequency offsets into a deep learning algorithm, iterating to obtain a CEST image reconstruction model, and obtaining a CEST image according to the reconstruction model. According to the method, multi-dimensional sparse sampling can be realized based on the improved non-Cartesian acquisition sequence, so that high-multiple acceleration is realized, and the scanning time is shortened; the deep learning algorithm can realize high-quality reconstruction of high-power under-sampling data of specific scanning under the condition of no full-sampling data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of magnetic resonance imaging, and in particular to a method for accelerating chemical exchange saturation transfer contrast imaging. Background Art

[0002] Chemical exchange saturation transfer imaging (CEST) is a novel magnetic resonance imaging (MRI) technique that provides a powerful sensitivity enhancement mechanism for quantitative imaging of certain low-concentration metabolites and protein molecules that are difficult to observe with conventional MRI methods. By providing molecular-level information and non-invasive diagnostic capabilities, CEST imaging significantly enhances clinicians' ability to diagnose a variety of diseases. Despite its many advantages, CEST imaging typically requires the acquisition of data at different saturation frequencies (i.e., frequency offsets), and its long acquisition time may lead to treatment delays, increased patient discomfort, and an increased likelihood of scan failure. Therefore, the implementation of accelerated imaging techniques is critical to promoting the wider clinical application of CEST.

[0003] Conventional CEST imaging sequences consist of a saturation module and a readout module. The readout module usually uses fast spin echo (FSE) or echo planar imaging (EPI) technology. The FSE readout module provides higher-quality CEST images, but requires a longer scan time to collect data at multiple saturation frequencies, which severely limits its application. The EPI readout module can save scanning time, but the obtained CEST images may be distorted, thereby interfering with the diagnostic results.

[0004] In recent years, a variety of techniques have been developed to accelerate CEST imaging. Parallel imaging (PI) is a mature MRI acceleration technique that has been integrated into conventional CEST sequences, especially in three-dimensional applications. Compressed sensing (CS) has become another popular technique for MRI acceleration. Specifically, sensitivity coding technology (CS-SENSE-CEST) significantly improves the CEST imaging speed through a variable density k-space undersampling scheme while maintaining image integrity and quantitative accuracy. In addition, advanced methods such as parallel blind compressed sensing (PBCS), SLAM-CEST (Spectroscopy with Linear Algebraic Modeling), and kt-SLR (kt Sparsity and Low-Rank) effectively utilize the inherent spatiotemporal sparsity priors of CEST data and greatly shorten the scan time of CEST imaging. Among multi-task techniques, low-rank tensor space (LRTES) modeling is used to exploit the correlation between images acquired at different saturation frequencies to reduce the scan time of CEST imaging. Deep learning algorithms have become a powerful tool to improve the efficiency of MRI. Supervised learning techniques are effectively used to convert low-quality images into high-quality images. However, obtaining a sufficient amount of fully sampled reference data in MRI is a major challenge, especially in CEST imaging. To address these limitations, unsupervised deep learning algorithms have been developed. Among them, TDDIP (Time-Dependent Deep Image Prior) technology has been proposed to accelerate MRI reconstruction, which uses a one-dimensional manifold to learn the time dependency of MRI data. However, when the acceleration factor of the above methods is large, the reconstruction results often have certain artifacts, resulting in errors in the diagnostic results.

[0005] In summary, CEST imaging technology currently has the following problems:

[0006] 1) The biggest problem is the long imaging time. Since data needs to be collected at multiple saturation frequencies, the conventional imaging sequence scanning process takes a long time. Long-term scanning will result in continuous RF energy input, which increases the energy absorption of the tissue, thereby increasing the SAR value. This high SAR value brings safety and technical challenges, which not only increases the discomfort of patients, but also limits its wide application in clinical settings. And because of the long imaging time, the slightest movement of the patient will cause motion artifacts, which will reduce the clarity and accuracy of the image and affect the diagnosis. Especially for patients who need to remain still, such as children, the elderly or patients with certain diseases. It may even affect the patient's compliance and even make it impossible to complete the scan.

[0007] 2) CEST imaging is inherently limited by long scan times and high SAR, which makes it difficult to obtain large, fully sampled datasets. This limitation in data acquisition poses a significant challenge to the application of supervised deep learning methods, as supervised learning relies on large amounts of high-quality annotated data to train models. In CEST imaging, the problems of insufficient data and incomplete sampling make it difficult for traditional supervised deep learning models to learn and generalize effectively, thus limiting the practicality and effectiveness of these methods in CEST imaging.

[0008] 3) Although existing compressed sensing algorithms and unsupervised learning algorithms have shown potential in accelerating CEST imaging, further technological innovation and optimization are still needed to achieve high-fold acceleration to overcome the current problems of reconstruction quality and image accuracy. This requires the development of more efficient sampling strategies, reconstruction algorithms, and deep learning models to significantly improve imaging speed while ensuring image quality.

[0009] How to ensure efficient acquisition of CEST imaging and the quality of reconstruction under high acceleration is a hot issue that needs to be studied urgently.

[0010] Therefore, those skilled in the art have devoted themselves to developing a method to accelerate chemical exchange saturation transfer contrast imaging. Summary of the invention

[0011] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to accelerate the acquisition of CEST imaging data and how to ensure the quality of reconstructed CEST images under high acceleration.

[0012] To achieve the above object, the present invention provides a method for accelerating chemical exchange saturation transfer contrast imaging, comprising the following steps:

[0013] Step 1, setting radio frequency saturation pulses with different frequency offsets, and after each radio frequency saturation pulse with different frequency offsets, using a non-Cartesian sequence to collect data to obtain collected k-space data with different frequency offsets;

[0014] The non-Cartesian sequence adopts a randomized undersampling method in the Z spectrum dimension, and the Z spectrum dimension is the sequence acquisition dimension corresponding to the radio frequency saturation pulses with different frequency offsets;

[0015] The non-Cartesian sequence is a stacked star trajectory sequence of golden angle rotation or a three-dimensional spherical radial trajectory sequence of golden angle rotation;

[0016] Step 2: Input the collected k-space data with different frequency offsets into a deep learning algorithm to reconstruct a CEST image.

[0017] Furthermore, the deep learning algorithm described in step 2 is an unsupervised learning algorithm INRESP that combines implicit neural expression with explicit prior information, and specifically includes the following steps:

[0018] Step 2.1, generating the corresponding uniformly distributed initialization coordinates through a coordinate generation function according to the acquired k-space data with different frequency offsets;

[0019] Step 2.2, inputting the initialization coordinates into a trainable encoding module, the trainable encoding module encodes the initialization coordinates into a transformation domain, extracts and calculates the feature vector of the image pixel point, and obtains a CEST feature map;

[0020] Step 2.3, inputting the CEST feature map into a feature conversion module for calculation to obtain the real part and the imaginary part of the CEST image, wherein the feature conversion module is composed of an MLP network;

[0021] Step 2.4: Input the real and imaginary parts of the CEST image and the acquired k-space data of different frequency offsets into the INR reconstruction data consistency model, and input the explicit prior information into the explicit sparse prior regularization model to obtain the data consistency term R DC Loss function, weighted joint sparse regularization term R WJS Loss function, time domain TV regularization term R TV Loss function and low-rank regularization term R LR The loss function is as follows:

[0022] The real and imaginary parts of the CEST image and the acquired k-space data of different frequency offsets are input into the INR reconstruction data consistency model to obtain the following data consistency formula:

[0023]

[0024] in, represents the acquired k-space data of different frequency offsets, represents the k-space data of different frequency offsets fitted, where is the complex CEST image obtained by combining the real and imaginary parts of the CEST image, Φ = F nc S represents a multi-coil undersampling operator, the multi-coil undersampling operator adopts the non-Cartesian sequence, F nc is a non-uniform fast Fourier transform operator, S represents a coil sensitivity map, and θ represents a weight to be optimized in the INR reconstruction data consistency model;

[0025] The explicit prior information is weighted joint sparsity, time domain TV, and low rank. The explicit prior information is brought into the explicit sparse prior regularization model to obtain the following explicit prior information regularization formula:

[0026]

[0027] in, is the complex CEST image obtained by combining the real and imaginary parts of the CEST image,

[0028] is the regularization term of the weighted joint sparsity,

[0029] is the regularization term of the time-domain TV,

[0030] is the low-rank regularization term;

[0031] D n (n=1,2) and D3 are the first-order difference matrices of spatial dimension and frequency offset dimension, is the weighted tensor of the weighted joint sparse terms, ‖‖ 21 Indicates L 21 paradigm, ‖‖1 represents L1 paradigm, ‖‖ * represents the kernel normal form, λ1, λ2 and λ3 are regularization parameters;

[0032] The data consistency formula and the explicit prior information regularization formula are combined according to the optimization problem model in CS theory to obtain the unsupervised learning algorithm INRESP as follows:

[0033]

[0034] in, is the data consistency item R DC Loss function,

[0035] is the weighted joint sparse regularization term R WJS Loss function,

[0036] is the time domain TV regularization term R TV Loss function,

[0037] is the low-rank regularization term R LR Loss function;

[0038] Step 2.5: Calculate the data consistency term R by chain rule DC Loss function, the weighted joint sparse regularization term R WJS Loss function, the time domain TV regularization term R TV The loss function and the low-rank regularization term R LRThe gradient of the loss function with respect to model parameters, wherein the model parameters include the image pixel feature vector and the weight parameters of the MLP network;

[0039] And use the gradient to update the model parameters through the Adam optimization algorithm to minimize the data consistency term R DC Loss function, the weighted joint sparse regularization term R WJS Loss function, the time domain TV regularization term R TV The loss function and the low-rank regularization term R LR Loss function, and obtain the CEST image reconstruction model;

[0040] Step 2.6: input the initialization coordinates into the CEST image reconstruction model to reconstruct the CEST image.

[0041] Furthermore, the randomized undersampling method in the Z spectrum dimension in step 1 is golden angle rotation sampling.

[0042] Furthermore, the stacked star trajectory sequence of the golden angle rotation in step 1 adopts radial sampling of the golden angle rotation in the kx-ky plane and Cartesian sampling along the kz dimension.

[0043] Furthermore, the stacked star trajectory sequence of the golden angle rotation in step 1 uses a center-outward sampling method in the kz direction.

[0044] Furthermore, the three-dimensional spherical radial trajectory sequence of the golden angle rotation in step 1 is to perform golden angle rotation sampling between each interlacing.

[0045] Furthermore, the trainable encoding module in step 2 is a hybrid feature hash encoding module, the transform domain is a hybrid hash table, and the specific steps of the encoding are as follows:

[0046] In the hybrid feature hash coding module, M grids Mi (i=1, 2, ..., M) of different resolutions and M hybrid hash tables Hi (i=1, 2, ..., M) are constructed;

[0047] Each of the grids Mi (i=1, 2, ..., M) is further divided into voxel grids of different resolutions through a voxel window, the number of the voxel windows is W, and L=M·W voxel grids are obtained, each of the voxel grids has 8 corner points, each of the corner points corresponds to a corner point index, each of the corner points stores F corner point feature vectors, and L of the voxel grids store 8×L×F corner point feature vectors;

[0048] The 8×L×F corner point feature vectors are stored in the M hybrid hash tables Hi (i=1, 2, …, M), and a hybrid hash table link index of each corner point index is generated according to the spatial hash function h(x). The formula of the hash function h(x) is as follows:

[0049]

[0050] Among them, x i is the corner point index, π1=1, π2 and π3 are two large prime numbers, represents a bitwise XOR operator, mod represents a modulo operator, and T is the size of the hybrid hash table;

[0051] Trilinear interpolation is performed on the mixed hash table link indexes of the eight corner point indexes of each voxel grid to obtain L image pixel feature vectors, and the L image pixel feature vectors are combined to obtain the CEST feature map.

[0052] Furthermore, the method further comprises performing B0 correction and / or B1 correction on the CEST image reconstructed in step 2.6.

[0053] Furthermore, the method also includes the calculation of the CEST quantitative image, and the calculation formula is as follows:

[0054] (S(ω)-S(-ω)) / S0,

[0055] Wherein, S(ω) is a CEST image obtained by the method of accelerating chemical exchange saturation transfer contrast imaging after applying a specific frequency offset radio frequency saturation pulse ω,

[0056] S(-ω) is a CEST image obtained by the method of accelerating chemical exchange saturation transfer contrast imaging after applying a symmetrical frequency-shifted radio frequency saturation pulse of -ω,

[0057] S0 is a CEST image obtained by the method of accelerating chemical exchange saturation transfer contrast imaging without applying a radio frequency saturation pulse.

[0058] Furthermore, the method also includes fitting of the Z spectrum and fitting of the MTR spectrum.

[0059] The fitting of the Z spectrum is specifically as follows:

[0060] The different frequency deviations are used as the abscissa and the average magnetization intensity of the selected area of ​​the reconstructed CEST image is used as the ordinate, and a graph is drawn to obtain a fitted Z spectrum;

[0061] The fitting of the MTR spectrum is specifically:

[0062] Based on the Z spectrum, the saturated CEST signal intensity at the different frequency offsets is subtracted from the unsaturated CEST signal intensity, and the result is normalized with the baseline CEST signal intensity in the absence of the RF saturation pulse to obtain a fitted MTR spectrum.

[0063] The technical effects of the present invention are as follows:

[0064] 1) The present invention achieves high-multiple acceleration based on multi-dimensional sparse sampling of improved non-Cartesian acquisition sequences, shortens scanning time, and further shortens total CEST imaging time. Use stacked star tracks rotated at the golden angle or three-dimensional spherical radial tracks rotated at the golden angle, that is, after each saturation pulse, collect data along the rotating radial path in three-dimensional space, and use the golden angle (about 137.5 degrees) for continuous angle increment sampling to achieve more uniform under-sampling acquisition, which is conducive to higher-multiple accelerated imaging in the spatial dimension. In the acquisition of data with different frequency deviations, the present invention initializes the angle of the first radial track with the angle increment of the golden angle. This rotation angle method increases the incoherence of the k-space data collected in the frequency deviation direction, and achieves acceleration in the frequency deviation dimension.

[0065] Since the non-Cartesian acquisition strategy can accelerate in both the phase encoding direction (spatial dimension) and the frequency encoding direction (frequency offset dimension), it can achieve acceleration in more dimensions. In contrast, the traditional Cartesian acquisition method can only achieve acceleration in the phase encoding direction. Specifically, CEST imaging based on stack-of-stars sequence acquisition can be sparsely sampled from three directions: X, Y, and frequency offset dimensions, thereby achieving a 20-fold or even 30-fold acceleration. CEST imaging based on three-dimensional spherical radial trajectory (3Dkooshball) sequence acquisition can be sparsely sampled from four directions: X, Y, Z, and frequency offset dimensions, thereby achieving a 30-fold or even 50-fold acceleration.

[0066] In the spatial dimension, the non-Cartesian acquisition sequence based on golden angle rotation optimizes k-space filling, improves the uniformity and coverage of data acquisition, and reduces acquisition redundancy and motion artifacts. Specifically, using stacked star trajectories or three-dimensional spherical radial trajectories (3D kooshball) trajectories with golden angle rotation can still achieve uniform acquisition even under high-fold undersampling conditions. Its unique golden angle increment ensures that the radial path of each acquisition fills the k-space in a non-repetitive and non-periodic manner. This golden angle rotation avoids overlaps and gaps in data sampling, evenly distributes sampling points, and maximizes k-space coverage, thereby improving the quality and efficiency of image reconstruction. This method is particularly effective in reducing motion artifacts and enhancing temporal resolution.

[0067] In the frequency deviation dimension, the golden angle rotation method increases the incoherence of CEST data in the frequency deviation direction, so that the sampling of each frequency deviation point is evenly distributed and non-overlapping, thereby ensuring the diversity and integrity of the sampling and reducing data redundancy. This also achieves acceleration in the frequency deviation dimension. As a result, the data sparsity in the frequency deviation direction is higher, achieving higher-multiple accelerated acquisition.

[0068] 2) A center-out acquisition method is used in the kz direction, so that after each RF saturation pulse, the lower frequency signals are acquired first, making the CEST effect more obvious and the imaging quality further improved.

[0069] 3) The unsupervised learning algorithm INRESP, which combines implicit neural representation (INR) with explicit sparse prior (ESP), can achieve high-quality reconstruction of highly undersampled data of specific scans in the absence of full sampling parameter data.

[0070] The algorithm takes advantage of implicit neural expression to learn the intrinsic features of CEST data through continuous expression functions and trainable encoding methods, and combines explicit prior information in the image domain into the unsupervised learning algorithm of implicit neural expression through a composite loss function. Since the algorithm makes full use of the inherent features of the data and the explicit prior information of the CEST image, in practical applications, we only need to collect a small amount of effective k-space data to reconstruct a better CEST image and realize accelerated imaging of CEST images. The algorithm is universal and general, and has the advantages of strong noise robustness and generalization ability. It can reconstruct highly undersampled non-Cartesian data and obtain results close to the full sampling gold standard.

[0071] 4) Implicit neural representation combined with trainable hybrid-feature hashencoding can improve the overall performance of the network and reduce the size and number of parameters of the network. By exploiting the redundancy of different feature grids in the multi-resolution hybrid hash table, higher quality CEST image reconstruction can be achieved with fewer coding parameters.

[0072] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flow chart of a method for accelerating chemical exchange saturation transfer contrast imaging according to a preferred embodiment of the present invention;

[0074] Figure 2 It is a schematic diagram of a steady-state pulse data acquisition sequence based on a golden angle stack-of-stars trajectory according to an embodiment of the present invention;

[0075] Figure 3 is a schematic diagram of a steady-state pulse data acquisition sequence based on a three-dimensional spherical radial trajectory (3D kooshball) according to an embodiment of the present invention;

[0076] Figure 4 is a schematic diagram of a stack-of-stars trajectory of golden angle rotation according to an embodiment of the present invention;

[0077] Figure 5 is a schematic diagram of a three-dimensional spherical radial trajectory (3D kooshball) trajectory of an embodiment of the present invention;

[0078] Figure 6 is an algorithm flow chart of an embodiment of the present invention;

[0079] Figure 7 It is a schematic diagram of the overall framework structure of an algorithm of an embodiment of the present invention;

[0080] Figure 8 It is a visual comparison diagram of hybrid feature hash coding and traditional hash coding according to an embodiment of the present invention;

[0081] Fig. 9 This is a prospective experimental result diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0082] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the present invention is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0083] like Figure 1 FIG. 1 is a flow chart of a method for accelerating chemical exchange saturation transfer contrast imaging of the present invention, the method comprising:

[0084] S10: Using a non-Cartesian sequence to acquire data, specifically, setting radio frequency saturation pulses with different frequency offsets, and using a non-Cartesian sequence to acquire data after each radio frequency saturation pulse with different frequency offsets, to obtain acquired k-space data with different frequency offsets, which can be expressed as follows:

[0085]

[0086] in, is the corresponding CEST image sequence to be reconstructed, Φ = F ncS represents the multi-coil undersampling operator, using a non-Cartesian sequence, F nc is the non-uniform fast Fourier transform (NUFFT) operator, S represents the coil sensitivity map, and ε is the measurement error;

[0087] The non-Cartesian sequence uses a randomized undersampling method in the Z spectrum dimension, which is the sequence acquisition dimension corresponding to the RF saturation pulses with different frequency offsets;

[0088] The non-Cartesian sequence is a golden angle rotated stacked star trajectory sequence or a golden angle rotated three-dimensional spherical radial trajectory sequence.

[0089] S20: Input the acquired k-space data with different frequency offsets into the deep learning algorithm to reconstruct the CEST image.

[0090] In some embodiments, the stack-of-stars sequence of golden angle rotation or the three-dimensional spherical radial trajectory sequence (3D kooshball) of golden angle rotation in S10 uses golden angle rotation in the Z spectral dimension to achieve randomized undersampling of the Z spectral dimension. CEST imaging usually requires the acquisition of data at multiple frequency offsets, while the traditional CEST sequence uses the same sampling trajectory at different frequency offsets, and does not fully utilize the information of the data in the frequency offset dimension. The present application uses golden angle rotation on the radial acquisition sequence corresponding to the RF saturation pulses at different frequency offsets to avoid overlapping of data sampling, evenly distribute the sampling points, and maximize the k-space coverage, thereby improving the quality and efficiency of image reconstruction.

[0091] In some embodiments, the golden angle rotated stacked star trail sequence in S10 employs radial sampling in the kx-ky plane and golden angle rotated Cartesian sampling along the kz dimension.

[0092] In some embodiments, the three-dimensional spherical radial trajectory sequence of the golden angle rotation in S10: each radial spoke or a group of radial spokes in the sequence is called an interleaf, and golden angle rotation sampling is performed between each interleaf. The trajectory of the golden angle rotation can be obtained by calculating the azimuth angle and the polar angle.

[0093] Since the non-Cartesian acquisition strategy can accelerate in both the spatial dimension and the frequency offset dimension, an additional dimension of acceleration is achieved. In contrast, the traditional Cartesian acquisition method can only achieve acceleration in the spatial dimension. Specifically, CEST imaging based on stack-of-stars sequence acquisition can sparsely sample from three directions, X, Y, and frequency offset, and can achieve a 20-fold or even 30-fold acceleration. CEST imaging based on three-dimensional spherical radial trajectory (3D kooshball) sequence acquisition can sparsely sample from four directions, X, Y, Z, and frequency offset, and can achieve a 30-fold or even 50-fold acceleration.

[0094] In the spatial dimension, the non-Cartesian acquisition sequence based on golden angle rotation optimizes k-space filling, improves the uniformity and coverage of data acquisition, and reduces acquisition redundancy and motion artifacts. Specifically, using stacked star trajectories or three-dimensional spherical radial trajectories (3D kooshball) trajectories with golden angle rotation can still achieve uniform acquisition even under high-multiple undersampling conditions. Its unique golden angle increment ensures that the radial path of each acquisition fills the k-space in a non-repetitive and non-periodic manner. This rotation angle avoids overlaps and gaps in data sampling, evenly distributes sampling points, and maximizes k-space coverage, thereby improving the quality and efficiency of image reconstruction. This method is particularly effective in reducing motion artifacts and enhancing temporal resolution.

[0095] In the frequency deviation dimension, the rotation angle method increases the incoherence of CEST data in the frequency deviation direction, so that the sampling of each frequency deviation point is evenly distributed and non-overlapping, thereby ensuring the diversity and integrity of the sampling and reducing data redundancy. This also achieves acceleration in the frequency deviation dimension, thereby achieving higher quality image reconstruction and faster scanning speed.

[0096] In summary, the non-Cartesian sequential imaging of golden angle rotation achieves undersampling in multi-dimensional directions through multi-dimensional sparse sampling, thereby significantly improving the acquisition efficiency.

[0097] In some embodiments, the deep learning algorithm in S20 is an unsupervised learning algorithm INRESP that combines implicit neural expression with explicit prior information. The k-space data of different frequency offsets acquired by a non-Cartesian sequence are input into the unsupervised learning algorithm INRESP that combines implicit neural expression with explicit prior information to reconstruct a CEST image.

[0098] like Figure 6 The flowchart of the unsupervised learning algorithm INRESP is shown in Figure 7 The figure shows the overall framework structure diagram of the unsupervised learning algorithm INRESP, and its specific steps include:

[0099] S201: Generate the corresponding uniformly distributed initialization coordinates through a coordinate generation function according to the collected k-space data with different frequency offsets; the value of the coordinate in each dimension can be uniformly distributed from 0 to 1, and the size of the coordinate point is consistent with the size of the CEST image to be reconstructed.

[0100] The initialization coordinates generated by the data collected by the stacked star trajectory sequence based on the golden angle rotation are the three-dimensional coordinates (x k ,y k ,z).

[0101] The initialization coordinates generated by the data collected by the three-dimensional spherical radial trajectory sequence based on the golden angle rotation are the four-dimensional coordinates (x k ,y k ,z k ,z).

[0102] In some embodiments the coordinate generation function is the torch.meshgrid function.

[0103] S202: Input the initialization coordinates into the trainable coding module, extract and calculate the features, and obtain the CEST feature map. The trainable coding module encodes the continuous initialization coordinates into a transformation domain. During the training process, the trainable coding module gradually learns the inherent features of the CEST image carried in the initialization coordinates, calculates and obtains the CEST feature map.

[0104] In some embodiments, the training encoding module in S202 is a hybrid feature hash encoding module, and the transform domain is a hybrid hash table, such as Figure 8 The figure shows a schematic diagram of hybrid feature hash coding. The specific coding steps are as follows: in the hybrid feature hash coding module, M grids of different resolutions are constructed. and M hybrid hash tables Hi (i = 1, 2, ..., M), the maximum size of the hybrid hash table Hi (i = 1, 2, ..., M) is T, and each grid Mi (i = 1, 2, ..., M) is further divided into different resolutions through the voxel window, so that the original M grids of different resolutions Increase to L = M·W voxel grids of different resolutions, where W is the number of voxel windows. The number of voxel grids L and the number of hybrid hash tables M are adjustable hyperparameters, and higher reconstruction quality can be obtained by adjusting these two hyperparameters.

[0105] Each voxel grid has 8 corner points (vertices), each corner point stores F corner point feature vectors, and L voxel grids store 8×L×F corner point feature vectors;

[0106] The 8×L×F corner point feature vectors are stored in M ​​hybrid hash tables Hi, and the link relationship between each corner point index and its corresponding position in the hybrid hash table is generated according to the spatial hash function h(x). That is, the hash table link index of each corner point index can be obtained according to the hash function h(x). The hash function h(x) formula is as follows:

[0107]

[0108] Among them, x i is the corner index, π1=1, π2 and π3 are two large prime numbers, The XOR operator represents the bit orientation, mod represents the modulo operator, and T is the size of the hybrid hash table.

[0109] To calculate the encoding of a certain pixel point (sampling point), trilinear interpolation needs to be performed on the 8 hash table link indexes of the voxel grid containing this pixel point, so as to obtain the encoding of this pixel point, that is, the hash table link index of this pixel point in the hybrid hash table. The feature vector corresponding to this link index is the feature vector of the pixel point (denoted as the image pixel point). By using the above method to obtain the image pixel point feature vectors of each pixel point in the CEST image and combining them, the CEST feature map is obtained.

[0110] The hybrid hash table can adaptively mix the voxel grid into a small number of hybrid hash tables. For example, Figure 8 in (b), the image pixel point feature vectors obtained by trilinear interpolation of the corner points of the voxel grids with L different resolutions are adaptively stored in M hybrid hash tables, where M < L. This reduces the number of hash tables, thereby achieving higher-quality reconstruction with fewer encoding parameters. In addition, the hybrid hash table uses a voxel window instead of an ultra-large-scale grid to simulate obtaining feature vectors from the missing-resolution grid, which helps to avoid feature loss caused by the reduction in the number of hash tables. For example, Figure 8 in (b), the same pixel point is selected by different voxel windows, so that the same image pixel point feature vector can be obtained by trilinear interpolation of voxel grids with different resolutions and then stored in the same hybrid hash table. The integration of the hybrid feature hash table enables some feature grids to be adaptively mixed, effectively reducing the number of hash tables and reducing the size of the encoding parameters.

[0111] The hybrid feature hash encoding can adaptively merge some feature grids into one hash table. Feature vectors are obtained after trilinear interpolation of the eight corner features, and all feature vectors are combined to obtain the feature map.

[0112] Generally speaking, the hybrid hash table makes full use of the redundancy of different feature grids in the multi-resolution hash table and can achieve higher-quality reconstruction with fewer encoding parameters.

[0113] Since the same human tissue region in the image contains both the information of the large-resolution grid and the information of the small-resolution grid, putting the grid information carrying the same tissue information into the same hash table can enable these redundant information to learn from each other adaptively to capture more local information and make up for the lack of features.

[0114] S203: Input the CEST feature map into the feature conversion module for calculation to obtain the real part and the imaginary part of the CEST image. The feature conversion module is composed of an MLP network.

[0115] S204: Input the real and imaginary parts of the CEST image and the acquired k-space data of different frequency offsets into the INR reconstruction data consistency model, input the prior information into the explicit sparse prior regularization model, and obtain the data consistency term R according to the optimization problem model in the CS theory. DC Loss function, weighted joint sparse regularization term R WJS Loss function, time domain TV regularization term R TV Loss function and low-rank regularization term R LR Loss function, the specific process is as follows:

[0116] The real and imaginary parts of the CEST image are converted into fitted k-space data with different frequency offsets through the multi-coil undersampling operator, coil sensitivity map, and Fourier transform operator, and then input into the INR reconstruction data consistency model together with the acquired k-space data with different frequency offsets to obtain the following data consistency formula:

[0117]

[0118] in, represents the k-space data acquired at different frequency offsets, represents the k-space data of different frequency offsets fitted, where is the complex CEST image composed of the real and imaginary parts of the CEST image, Φ = F nc S represents the multi-coil undersampling operator, using the same undersampling non-Cartesian sequence as in actual sampling, F nc is the non-uniform fast Fourier transform operator (NUFFT), S represents the coil sensitivity map, and θ represents the weight that needs to be optimized in the INR reconstruction data consistency model; the data consistency formula indicates that the smaller the difference between the fitted k-space data with different frequency offsets and the acquired k-space data with different frequency offsets, the more accurate the simulation result.

[0119] In CEST imaging, due to the similarity of anatomical structures and biological tissue types, adjacent pixels usually have highly similar features. In addition, due to the constant change in the chemical exchange process between water molecules in biological tissues and the surrounding environment, the CEST signals of adjacent frequency points also show a high degree of similarity. Therefore, the CEST signal shows significant continuity in both the spatial domain (i.e., the spatial dimension) and the saturation frequency offset domain (i.e., the Z spectrum domain or Z spectrum dimension). Implicit neural representation (INR) learns and represents continuous functions through a neural network model, which can capture and reconstruct the complex structure and details of the target object. Compared with traditional discrete representations, INR has better continuity expression and generalization capabilities, and can effectively process and reconstruct complex data structures. Based on the continuity of CEST signals, INR can effectively learn the continuity and similarity within CEST data, thereby better reconstructing CEST images.

[0120] The explicit prior information is weighted joint sparsity, time domain TV, and low rank. The explicit prior information is brought into the explicit sparse prior regularization model, and the following explicit prior information regularization formula is obtained:

[0121]

[0122] in, It is the complex CEST image composed of the real and imaginary parts of the CEST image.

[0123] is the regularization term of the weighted joint sparsity,

[0124] is the regularization term of time-domain TV,

[0125] is a low-rank regularization term;

[0126] D n (n=1,2) and D3 are the first-order difference matrices of spatial dimension and frequency offset dimension, is the weighted tensor of the weighted joint sparse terms, ‖‖ 21 represents L21 normal form, ‖‖1 represents L1 normal form, ‖‖ * represents the kernel normal form, λ1, λ2 and λ3 are regularization parameters;

[0127] In CEST imaging, voxels in the same interval have similar Z spectra and show high correlation in the spatial and Z spectrum domains. This correlation provides strong explicit prior information, which is combined into the loss function of the unsupervised learning model to further constrain CEST imaging with explicit prior information. By regularizing the explicit prior information, the images generated by the model are ensured to be highly consistent with the real images in terms of structural features, so that the model not only minimizes the reconstruction error during training, but also retains and enhances these explicit prior features, thereby obtaining higher quality CEST image reconstruction results.

[0128] According to the optimization problem model in CS theory, the data consistency formula and the explicit prior information regularization formula are combined to obtain the unsupervised learning algorithm INRESP.

[0129] The optimization problem model in the above CS theory is as follows:

[0130]

[0131] in, represents the fitted data, represents the measured data, R(·) represents the sparsity operator, and λ is the regularization parameter; Part of it is the data consistency (DC) term, which ensures the data fidelity between the fitted data and the measured data; It is the regularization part, which helps to obtain the best results in underdetermined situations;

[0132] The obtained unsupervised learning algorithm INRESP is as follows:

[0133]

[0134] in, is the data consistency term R DC Loss function,

[0135] is the weighted joint sparse regularization term R WJS Loss function,

[0136] is the time domain TV regularization term R TV Loss function,

[0137] is the low-rank regularization term R LR Loss function.

[0138] S205: Calculate the data consistency term R by using the chain rule DC Loss function, the weighted joint sparse regularization term R WJS Loss function, the time domain TV regularization term R TV The loss function and the low-rank regularization term R LR The gradient of the loss function with respect to the model parameters;

[0139] The model parameters are updated by using the Adam optimization algorithm using the gradient to minimize the data consistency term R DC Loss function, weighted joint sparse regularization term R WJS Loss function, time domain TV regularization term R TV Loss function and low-rank regularization term R LR After about 500 iterations, the loss function tends to be stable, the model parameters gradually converge, and a CEST image reconstruction model is obtained. The model parameters include the image pixel feature vector stored in the hybrid hash table Hi and the weight parameters of the MLP network.

[0140] S206: Input the initialization coordinates into the CEST image reconstruction model to reconstruct the CEST image.

[0141] In some embodiments, the stacked star trail sequence of the golden angle rotation in S10 is as follows: Figure 2The steady-state pulse sequence shown, that is, the CEST saturation module adopts a steady-state pulse sequence, which only applies a short RF saturation pulse. Using the stacked star trajectory with golden angle rotation, the acquisition mode of golden angle rotation is adopted in the frequency deviation direction. The cross section of the acquisition trajectory is as follows Figure 4 As shown in (a), in each radial stack, the acquisition is scanned from the center to the outside, and the final acquisition trajectory is as follows: Figure 4 (b) as shown.

[0142] In some embodiments, S10 adopts Figure 2 The stacked star trajectory sequence of the golden angle rotation of the steady-state pulse shown in the figure was used to carry out data acquisition experiments on a 3.0T medical magnetic resonance imaging system (uMR790, Shanghai United Imaging Medical). The acquisition parameters were as follows: the single RF saturation pulse time was 50ms, the imaging field of view (FOV) was 220mm×220mm×42mm; the imaging resolution was 1.35mm×1.35mm×3.5mm; the TR was 4.71ms; the TE was 2.34ms; a total of 30 frequency deviations were collected: ±6, ±5.5, ±5, ±4.5, ±4, ±3.5, ±3.5, ±3.5, ±3, ±2.5, ±2.0, ±1.5, ±1, ±0.5, 0 and -300. Before the data acquisition of each frequency deviation began, there was a 1-second pre-RF saturation pulse to achieve steady state. A total of four experiments were conducted under different acceleration factors: full sampling with a total acquisition time of 16 minutes and 31 seconds, collecting 300 radial stacks for each frequency offset; 10 times accelerated sampling, collecting 30 radial stacks for each frequency offset, with a total acquisition time of 2 minutes and 8 seconds; 20 times accelerated sampling, collecting 15 radial stacks for each frequency offset, with a total acquisition time of 1 minute and 36 seconds; 30 times accelerated sampling, collecting 10 radial stacks for each frequency offset, with a total acquisition time of 1 minute and 4 seconds.

[0143] In some embodiments, the three-dimensional spherical radial trajectory sequence of the golden angle rotation in S10 is as follows: Figure 3 The steady-state pulse sequence shown in the figure, that is, the CEST saturation module adopts a steady-state pulse sequence, which only applies a short RF saturation pulse. The kooshball trajectory with golden angle rotation is used, and the acquisition mode with golden angle rotation is adopted in the frequency deviation direction. The acquisition trajectory is as follows Figure 5 shown.

[0144] In some embodiments, a small MLP network is used in S202, such as an MLP network with a width of 64 and a depth of 3. Because the trainable encoding module shares most of the data calculation, the small MLP network can meet the computing requirements.

[0145] In some embodiments, the following steps may be further performed based on the reconstructed CEST image:

[0146] S30: B0 and B1 correction:

[0147] B0 correction is used to correct the frequency offset caused by magnetic field drift. B1 correction is used to adjust the saturation effect caused by uneven RF field strength. These correction steps ensure accurate measurement of the CEST effect, thereby improving the reliability and effectiveness of the reconstructed CEST image in quantitative analysis.

[0148] In some embodiments, B0 correction is implemented by the CEST-minZ method, specifically, the CEST signal intensity values ​​corresponding to each pixel point of the reconstructed CEST image along the Z spectral dimension are taken to form the Z spectral domain, and the minimum CEST signal intensity value is found after applying smoothing spline interpolation in the Z spectral domain. The minimum CEST signal intensity value is the B0 value at the pixel point in the reconstructed CEST image. All pixels are traversed by the CEST-minZ method to obtain a B0 map. The CEST signal intensity value at the required saturation frequency is extracted according to the denser image in the Z spectral domain after smoothing spline interpolation, and then the CEST image after B0 correction is obtained.

[0149] S40: Fitting of Z spectrum and MTR spectrum:

[0150] The offset frequency deviation is used as the horizontal coordinate, and the average magnetization intensity of the selected area of ​​the reconstructed CEST image is used as the vertical coordinate. These data points are plotted into a graph to obtain a fitted Z spectrum to show how the saturation effect changes with the frequency deviation. Based on the fitted Z spectrum, the saturated CEST signal intensity under RF saturation pulses with different frequency deviations is subtracted from the unsaturated CEST signal intensity, and normalized with the baseline CEST signal intensity when there is no RF saturation to obtain the fitted MTR spectrum.

[0151] S50: Calculation of CEST quantitative images:

[0152] The CEST quantitative image (APTw quantitative image) is obtained according to (S(ω)-S(-ω)) / S0.

[0153] Wherein, S(ω) is the k-space data of different frequency offsets acquired by the above method after applying a specific frequency offset radio frequency saturation pulse ω, and the CEST image obtained according to the above CEST image reconstruction model,

[0154] S(-ω) is the k-space data with different frequency offsets acquired by the above method after applying a symmetrical frequency offset RF saturation pulse of -ω, and the CEST image obtained according to the above CEST image reconstruction model.

[0155] S0 is the CEST image obtained according to the above CEST image reconstruction model without applying the RF saturation pulse and with k-space data of different frequency offsets acquired by the above method.

[0156] In some embodiments, ω=3.5ppm, -ω=-3.5ppm, and the result after B0 correction is as follows Fig. 9 As shown, the first row of the figure shows the comparison results of the CEST image with a frequency deviation of 3.5ppm reconstructed by the method of accelerated chemical exchange saturation transfer contrast imaging of the invention and the ordinary full sampling image in the forward-looking experiment with acceleration factors of 10 times, 20 times, and 30 times, and the second row shows the error diagram of the image reconstructed by accelerated sampling and the ordinary full sampling image, as well as the size of the standard root mean square error (nRMSE) with the ordinary full sampling image. The third row of the figure shows the comparison results of the APTw image calculated by the CEST image reconstructed by the method of accelerated chemical exchange saturation transfer contrast imaging of the present invention and the full sampling APTw image in the forward-looking experiment with acceleration factors of 10 times, 20 times, and 30 times. The fourth row shows the error diagram of the APTw image calculated by the CEST image reconstructed by the method of accelerated chemical exchange saturation transfer contrast imaging of the present invention and the full sampling APTw image after being magnified twice, as well as the size of the standard root mean square error (nRMSE) with the full sampling APTw image. It can be seen from the error diagram that the error between the reconstruction result diagram of the INRESP algorithm and the full sampling diagram is small.

[0157] In summary, the present invention has the advantage of three-dimensional fast high-resolution imaging: CEST imaging based on stack-of-stars sequence acquisition can achieve a resolution of 1.35mm×1.35mm×3.5mm, and CEST imaging can be obtained under 2min conditions. CEST imaging based on three-dimensional spherical radial trajectory (3D kooshball) sequence acquisition can achieve isotropic high-resolution imaging, and obtain whole-brain high-resolution CEST imaging in a shorter time. The present invention has the advantage of high accuracy: compared with the quantitative results of the full sampling gold standard, the CEST images obtained by the present invention have good consistency with them. The present invention has the advantage of stability: it has good consistency with the full sampling gold standard on different objects. The present invention has the advantage of easy implementation: the present invention mainly includes two parts: data acquisition sequence and image reconstruction, wherein the data acquisition sequence can be set on the magnetic resonance system, and the image reconstruction part can be set on the server, or it can be set and integrated into the magnetic resonance imaging system.

[0158] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for accelerating chemical exchange saturation transfer contrast imaging, characterized in that: The method comprises the following steps: Step 1, setting radio frequency saturation pulses with different frequency offsets, and after each radio frequency saturation pulse with different frequency offsets, using a non-Cartesian sequence to collect data to obtain collected k-space data with different frequency offsets; The non-Cartesian sequence adopts a randomized undersampling method in the Z spectral dimension, and the Z spectral dimension is The sequence acquisition dimensions corresponding to the radio frequency saturation pulses with different frequency offsets; The non-Cartesian sequence is a stacked star trajectory sequence of golden angle rotation or a three-dimensional spherical radial trajectory sequence of golden angle rotation; Step 2: Input the collected k-space data with different frequency offsets into a deep learning algorithm to reconstruct a CEST image.

2. The method for accelerating chemical exchange saturation transfer contrast imaging according to claim 1, characterized in that: The deep learning algorithm described in step 2 is an unsupervised learning algorithm INRESP that combines implicit neural expression with explicit prior information, and specifically includes the following steps: Step 2.1, generating the corresponding uniformly distributed initialization coordinates through a coordinate generation function according to the acquired k-space data with different frequency offsets; Step 2.2, inputting the initialization coordinates into a trainable encoding module, the trainable encoding module encodes the initialization coordinates into a transformation domain, extracts and calculates the feature vector of the image pixel point, and obtains a CEST feature map; Step 2.3, inputting the CEST feature map into a feature conversion module for calculation to obtain the real part and the imaginary part of the CEST image, wherein the feature conversion module is composed of an MLP network; Step 2.4: Input the real and imaginary parts of the CEST image and the acquired k-space data of different frequency offsets into the INR reconstruction data consistency model, and input the explicit prior information into the explicit sparse prior regularization model to obtain the data consistency term R DC Loss function, weighted joint sparse regularization term R WJS Loss function, time domain TV regularization term R TV Loss function and low-rank regularization term R LR The loss function is as follows: The real and imaginary parts of the CEST image and the acquired k-space data of different frequency offsets are input into the INR reconstruction data consistency model to obtain the following data consistency formula: in, represents the acquired k-space data of different frequency offsets, Φ(xθ) represents the fitted k-space data of different frequency offsets, wherein is the complex CEST image obtained by combining the real and imaginary parts of the CEST image, Φ = F nc S represents a multi-coil undersampling operator, the multi-coil undersampling operator adopts the non-Cartesian sequence, F nc is a non-uniform fast Fourier transform operator, S represents a coil sensitivity map, and θ represents a weight to be optimized in the INR reconstruction data consistency model; The explicit prior information is weighted joint sparsity, time domain TV, and low rank. The explicit prior information is brought into the explicit sparse prior regularization model to obtain the following explicit prior information regularization formula: in, is the complex CEST image obtained by combining the real and imaginary parts of the CEST image, is the regularization term of the weighted joint sparsity, is the regularization term of the time-domain TV, is the low-rank regularization term; D n (n=1, 2) and D3 are the first-order difference matrices of spatial dimension and frequency offset dimension, is the weighted tensor of weighted joint sparse terms, || || 21 Indicates L 21 paradigm, || ||1 represents L1 paradigm, || || * represents the kernel normal form, λ1, λ2 and λ3 are regularization parameters; According to the optimization problem model in CS theory, the data consistency formula and the explicit prior information regularization formula are combined to obtain the unsupervised learning algorithm INRESP as follows: in, is the data consistency item R DC Loss function, is the weighted joint sparse regularization term R WJS Loss function, is the time domain TV regularization term R TV Loss function, is the low-rank regularization term R LR Loss function; Step 2.5: Calculate the data consistency term R by chain rule DC Loss function, the weighted joint sparse regularization term R W JS loss function, the time domain TV regularization term R TV The loss function and the low-rank regularization term R LR The gradient of the loss function with respect to model parameters, wherein the model parameters include the image pixel feature vector and the weight parameters of the MLP network; And use the gradient to update the model parameters through the Adam optimization algorithm to minimize the data consistency term R DC Loss function, the weighted joint sparse regularization term R W JS loss function, the time domain TV regularization term R TV The loss function and the low-rank regularization term R LR Loss function, and obtain the CEST image reconstruction model; Step 2.6: input the initialization coordinates into the CEST image reconstruction model to reconstruct the CEST image.

3. The method for accelerating chemical exchange saturation transfer contrast imaging according to claim 1, characterized in that: The random undersampling method in the Z spectrum dimension described in step 1 is to use golden angle rotation sampling.

4. The method for accelerating chemical exchange saturation transfer contrast imaging according to claim 1, characterized in that: The stacked star trajectory sequence of the golden angle rotation in step 1 adopts radial sampling of the golden angle rotation in the kx-ky plane and Cartesian sampling along the kz direction.

5. The method for accelerating chemical exchange saturation transfer contrast imaging according to claim 1, characterized in that: The stacked star trajectory sequence of the golden angle rotation described in step 1 uses a center-out sampling method in the kz direction.

6. The method for accelerating chemical exchange saturation transfer contrast imaging according to claim 1, characterized in that: The three-dimensional spherical radial trajectory sequence of the golden angle rotation in step 1 is to perform golden angle rotation sampling between each interlacing.

7. The method for accelerating chemical exchange saturation transfer contrast imaging according to claim 1, characterized in that: In step 2, the trainable encoding module is a hybrid feature hash encoding module, the transform domain is a hybrid hash table, and the specific steps of the encoding are as follows: In the hybrid feature hash coding module, M grids Mi (i=1, 2, ..., M) of different resolutions and M hybrid hash tables Hi (i=1, 2, ..., M) are constructed; Each of the grids Mi (i=1, 2, ..., M) is further divided into voxel grids of different resolutions through a voxel window, the number of the voxel windows is W, and L=M·W voxel grids are obtained, each of the voxel grids has 8 corner points, each of the corner points corresponds to a corner point index, each of the corner points stores F corner point feature vectors, and L of the voxel grids store 8×L×F corner point feature vectors; The 8×L×F corner point feature vectors are stored in M ​​hybrid hash tables Hi (i=1, 2, ..., M), and a hybrid hash table link index of each corner point index is generated according to the spatial hash function h(x). The formula of the hash function h(x) is as follows: Among them, x i is the corner point index, π1=1, π2 and π3 are two large prime numbers, represents a bitwise XOR operator, mod represents a modulo operator, and T is the size of the hybrid hash table; Trilinear interpolation is performed on the mixed hash table link indexes of the 8 corner point indexes of each voxel grid to obtain L image pixel feature vectors, and the L image pixel feature vectors are combined to obtain the CEST feature map.

8. The method for accelerating chemical exchange saturation transfer contrast imaging according to claim 1, characterized in that: The method further comprises performing B0 correction and / or B1 correction on the CEST image reconstructed in step 2.

6.

9. The method for accelerating chemical exchange saturation transfer contrast imaging according to claim 1, characterized in that: The method also includes the calculation of the CEST quantitative image, and the calculation formula is as follows: (S(ω)-S(-ω)) / S0, Wherein, S(ω) is a CEST image obtained by the method of accelerating chemical exchange saturation transfer contrast imaging after applying a specific frequency offset radio frequency saturation pulse ω, S(-ω) is a CEST image obtained by the method of accelerating chemical exchange saturation transfer contrast imaging after applying a symmetrical frequency-shifted radio frequency saturation pulse of -ω, S0 is a CEST image obtained by the method of accelerating chemical exchange saturation transfer contrast imaging without applying a radio frequency saturation pulse.

10. The method for accelerating chemical exchange saturation transfer contrast imaging according to claim 1, characterized in that: The method also includes fitting of the Z spectrum and fitting of the MTR spectrum. The fitting of the Z spectrum is specifically as follows: The different frequency deviations are used as the abscissa and the average magnetization intensity of the selected area of ​​the reconstructed CEST image is used as the ordinate, and a graph is drawn to obtain a fitted Z spectrum; The fitting of the MTR spectrum is specifically: Based on the Z spectrum, the saturated CEST signal intensity at the different frequency offsets is subtracted from the unsaturated CEST signal intensity, and the result is normalized with the baseline CEST signal intensity in the absence of the RF saturation pulse to obtain a fitted MTR spectrum.

Citation Information

Cited By

  • Magnetic resonance chemical exchange saturation transfer imaging method, device, equipment, product and storage medium

    CN121995285A