A method and system for correcting magnetic resonance reconstruction errors based on multi-frequency priors

Abstract

This invention proposes a method and system for magnetic resonance reconstruction error correction based on multi-frequency priors, comprising: Step S1: acquiring undersampled k-space data and preliminary reconstructed k-space data; Step S2: constructing a multi-frequency prior constraint model including frequency equalization weighting and frequency partitioning masking, and establishing a coil joint modeling strategy for joint processing of real and imaginary parts; Step S3: using a self-supervised learning mechanism to train a neural network in the self-calibration signal region to predict k-space errors; Step S4: compensating the predicted errors into the preliminary reconstruction results to obtain the corrected magnetic resonance image. This invention solves the problem of detail loss caused by uneven error distribution across different frequency bands through multi-frequency priors, significantly reduces computational overhead through coil joint modeling, and achieves efficient and high-precision scanning-specific error correction without requiring complete sampling of training data.

Description

Technical Field

[0001] This invention proposes a method and system for correcting magnetic resonance reconstruction errors based on multi-frequency priors, which relates to the field of magnetic resonance imaging technology. Background Technology

[0002] Magnetic Resonance Imaging (MRI) is a non-invasive medical imaging technique based on the principle of nuclear magnetic resonance. With its excellent soft tissue contrast and high spatial resolution, it has irreplaceable value in fields such as clinical diagnosis, nervous system research, and tumor detection.

[0003] In magnetic resonance imaging (MRI) scans, raw data is acquired in the form of a frequency-coded space (k-space). To shorten the lengthy scan time and reduce artifacts caused by the patient's physiological movements (such as breathing and heartbeat), undersampling is often used in practice, acquiring a portion of the k-space data at a rate lower than the Nyquist sampling rate. However, this undersampling leads to severe aliasing artifacts in the images, necessitating advanced reconstruction algorithms to recover high-quality images from incomplete data.

[0004] In existing technologies, a common method is the SPARK correction method. Based on the Scan-Specific Artifact Reduction Network (SPARK) of deep learning, SPARK uses self-calibration signal (ACS) region information to predict k-space errors, demonstrating significant application value in the field of magnetic resonance reconstruction error correction. As a self-supervised correction strategy, this method can be applied to various reconstruction schemes without additional training data. However, existing SPARK methods do not fully consider the differentiated impact of errors in different k-space frequency bands on image quality: low-frequency components mainly determine image contrast, while high-frequency components encode fine structural details, and the amplitude difference between the two can reach three orders of magnitude. Directly using all sampled values ​​for supervision can easily lead to the optimization process being biased towards low frequencies, resulting in the loss of high-frequency details. Furthermore, SPARK uses channel-independent modeling for multi-coil data, failing to fully utilize the correlation between coils, leading to model redundancy and, in some cases, even causing image quality degradation. Summary of the Invention

[0005] In view of this, in order to fill the gaps and deficiencies in the prior art, the present invention proposes a magnetic resonance reconstruction error correction method and system based on multi-frequency priors to solve the problems that have occurred in the background art.

[0006] This invention proposes a method and system for correcting magnetic resonance reconstruction errors based on multi-frequency priors, comprising the following:

[0007] According to a first aspect of the present invention, a magnetic resonance reconstruction error correction method based on multi-frequency priors is proposed, characterized by comprising the following:

[0008] Step S1: Obtain undersampled k-space data and preliminary reconstructed k-space data;

[0009] Step S2: Construct a multi-frequency prior constraint model that includes frequency equalization weighting and frequency partitioning mask, and establish a coil joint modeling strategy that combines real and imaginary parts;

[0010] Step S3: Use a self-supervised learning mechanism to train a neural network in the self-calibration signal region to predict k-space errors;

[0011] Step S4: Compensate the predicted error into the preliminary reconstruction result to obtain the corrected magnetic resonance image.

[0012] Further, step S1 includes the following:

[0013] Step S1: The initially reconstructed k-space data is as follows:

[0014] ;

[0015] in, For the reconstructed k-space data, This refers to the k-space data acquired by the coil under undersampling conditions; Let k represent the k-space obtained by reconstructing undersampled data using any MRI reconstruction method.

[0016] Further, step S2 includes the following:

[0017] Step S21: Use a multi-frequency prior constraint model to solve the problem of uneven amplitude distribution in k-space. The multi-frequency prior constraint model is composed of frequency equalization weighting and frequency partitioning masking strategy.

[0018] The frequency equalization weighting includes: designing a smooth weighting function W to suppress low-frequency energy and increase high-frequency weights, so that the network can learn the features of each frequency band in a balanced way;

[0019] The frequency partitioning mask includes: constructing a binary mask. The k-space is divided into multiple frequency bands, and differentiated constraints are applied to the errors of different frequency bands to prevent high-frequency noise from interfering with the recovery of low-frequency structures.

[0020] Furthermore, step S2 also includes the following:

[0021] Step S22: During the training and inference process of the multi-frequency prior constraint model, the coil-by-coil independent processing mode is changed, and the real part data of all coils in the k-space data of the reconstructed image are concatenated. The imaginary part data is concatenated as The data are input into the network separately, and the method of simultaneously training and inferring multi-coil data is used to reduce the number of model parameters and training and inference time.

[0022] During the training phase, for the k-space data of the input reconstructed image, the real or imaginary parts of all coils are used. First, a frequency-balanced weighting strategy is used to weight it, resulting in:

[0023] ;

[0024] Where W is a smoothing weighting function, which uses a smoothly varying weighting matrix to apply progressive weight adjustments to the data at different locations in the k-space, including the following:

[0025] ;

[0026] Where a and b are k-space coordinates.

[0027] in For amplitude adjustment parameters, Let P be the smoothness parameter, where the expression for P is as follows:

[0028] ;

[0029] Where P is the normalization factor; P is used to suppress low-frequency energy and increase high-frequency weights.

[0030] Furthermore, step S2 also includes the following:

[0031] Step S23: Input the weighted result into n identical networks respectively to predict the reconstruction error, and obtain:

[0032] ;

[0033] in, The weighted k-space error of the network prediction; For the improved ResNet network, These are the parameters to be learned for the nth network.

[0034] The network input consists of weighted k-space data after frequency equalization weighting.

[0035] Furthermore, the network contains several residual blocks, each containing two convolutional layers and a nonlinear activation function, and retains linear features through skip connections; the network output layer is the predicted weighted k-space error.

[0036] Step S24: Obtain different frequency bands in the k-space using a frequency partition mask, weight the low-frequency path and the weighted full-frequency path, to obtain:

[0037] ;

[0038] in To obtain different frequency bands of the weighted k-space error of the prediction using frequency partitioning masks; binary mask. By designing a binary mask matrix, frequency-selective truncation of k-space data can be achieved. The expression for the binary mask matrix includes the following:

[0039] ;

[0040] Where L represents the range of the area to be preserved.

[0041] Further, step S3 includes the following:

[0042] Step S31: Using a self-supervised learning model, train the neural network in the self-calibration signal region and backpropagate, where the loss function of the self-supervised learning model is:

[0043] ;

[0044] Where A is the ACS central region extraction operator.

[0045] Further, step S4 includes the following:

[0046] Step S41: After training, infer the reconstructed k-space results and calculate the weighted error of the prediction. Perform inverse weighting to obtain the original scale error:

[0047] ;

[0048] Step S42: Gradually segment and fuse the errors output from different paths:

[0049] ;

[0050] in, To extract the mask for each frequency band step by step, The expressions include:

[0051] ;

[0052] in, and This determines the range of the extracted frequency band;

[0053] Step S43: Compensate for the correction error to the initial reconstruction result. From:

[0054] ;

[0055] The corrected magnetic resonance image is obtained.

[0056] According to a second aspect of the present invention, a magnetic resonance reconstruction error correction system based on multi-frequency priors is proposed for performing a magnetic resonance reconstruction error correction method based on multi-frequency priors as described in any one of the present invention, characterized in that the magnetic resonance reconstruction error correction system based on multi-frequency priors includes the following:

[0057] The data acquisition module is used to acquire undersampled k-space data and preliminary reconstructed k-space data;

[0058] The multi-frequency prior constraint module is used to construct a multi-frequency prior constraint model that includes frequency equalization weighting and frequency partitioning masking, and to establish a coil joint modeling strategy that combines real and imaginary parts.

[0059] The self-supervised learning mechanism module is used to train a neural network in the self-calibration signal region to predict k-space errors using a self-supervised learning mechanism.

[0060] The error correction module is used to compensate for the predicted errors in the preliminary reconstruction results to obtain the corrected magnetic resonance image.

[0061] According to a third aspect of the present invention, the present invention provides a magnetic resonance reconstruction error correction system based on multi-frequency priors, comprising an electronic device, wherein the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements a magnetic resonance reconstruction error correction method based on multi-frequency priors as described in any one of the present invention.

[0062] According to a fourth aspect of the present invention, the present invention provides a magnetic resonance reconstruction error correction system based on multi-frequency priors, comprising a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements a magnetic resonance reconstruction error correction method based on multi-frequency priors as described in any one of the present invention.

[0063] The present invention has the following advantages:

[0064] The core advantages of this invention are reflected in the following aspects:

[0065] This invention introduces a multi-frequency prior and applies differentiated constraints to k-space reconstruction errors through frequency-balanced weighting and selective masking, effectively preserving high-frequency details. Simultaneously, it proposes a coil joint strategy for unified modeling of multi-coil complex-valued data and improves the network architecture by incorporating residual connections. Compared to existing zero-sample scan-specific reconstruction methods, this method does not require complete sampling of training data, significantly reducing the number of parameters and computational costs. Furthermore, it achieves higher robustness and competitive reconstruction quality under high acceleration factors, making it suitable for reconstructed image correction in clinical MRI. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of the steps of the present invention.

[0067] Figure 2 This is a schematic diagram of the image reconstruction and correction process of the present invention.

[0068] Figure 3 This is a schematic diagram of the network used in this invention.

[0069] Figure 4 This is a schematic diagram of the experimental results of the reconstructed MRI brain map of the present invention. Detailed Implementation

[0070] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0071] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0072] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0073] like Figures 1 to 4 As shown, this invention proposes a method and system for correcting magnetic resonance reconstruction errors based on multi-frequency priors, including the following:

[0074] According to a first aspect of the present invention, a magnetic resonance reconstruction error correction method based on multi-frequency priors is proposed, characterized by comprising the following:

[0075] Step S1: Obtain undersampled k-space data and preliminary reconstructed k-space data;

[0076] Step S2: Construct a multi-frequency prior constraint model that includes frequency equalization weighting and frequency partitioning mask, and establish a coil joint modeling strategy that combines real and imaginary parts;

[0077] Step S3: Use a self-supervised learning mechanism to train a neural network in the self-calibration signal region to predict k-space errors;

[0078] Step S4: Compensate the predicted error into the preliminary reconstruction result to obtain the corrected magnetic resonance image.

[0079] Further, step S1 includes the following:

[0080] Step S1: The initially reconstructed k-space data is as follows:

[0081] ;

[0082] in, For the reconstructed k-space data, This refers to the k-space data acquired by the coil under undersampling conditions; Let k represent the k-space obtained by reconstructing undersampled data using any MRI reconstruction method.

[0083] Further, step S2 includes the following:

[0084] Step S21: Use a multi-frequency prior constraint model to solve the problem of uneven amplitude distribution in k-space. The multi-frequency prior constraint model is composed of frequency equalization weighting and frequency partitioning masking strategy.

[0085] The frequency equalization weighting includes: designing a smooth weighting function W to suppress low-frequency energy and increase high-frequency weights, so that the network can learn the features of each frequency band in a balanced way;

[0086] The frequency partitioning mask includes: constructing a binary mask. The k-space is divided into multiple frequency bands, and differentiated constraints are applied to the errors of different frequency bands to prevent high-frequency noise from interfering with the recovery of low-frequency structures.

[0087] Furthermore, step S2 also includes the following:

[0088] Step S22: During the training and inference process of the multi-frequency prior constraint model, the coil-by-coil independent processing mode is changed, and the real part data of all coils in the k-space data of the reconstructed image are concatenated. The imaginary part data is concatenated as The data are input into the network separately, and the method of simultaneously training and inferring multi-coil data is used to reduce the number of model parameters and training and inference time.

[0089] During the training phase, for the k-space data of the input reconstructed image, the real or imaginary parts of all coils are used. First, a frequency-balanced weighting strategy is used to weight it, resulting in:

[0090] ;

[0091] Where W is a smoothing weighting function, which uses a smoothly varying weighting matrix to apply progressive weight adjustments to the data at different locations in the k-space, including the following:

[0092] ;

[0093] Where a and b are k-space coordinates.

[0094] in For amplitude adjustment parameters, Let P be the smoothness parameter, where the expression for P is as follows:

[0095] ;

[0096] Where P is the normalization factor; P is used to suppress low-frequency energy and increase high-frequency weights.

[0097] Furthermore, step S2 also includes the following:

[0098] Step S23: Input the weighted result into n identical networks respectively to predict the reconstruction error, and obtain:

[0099] ;

[0100] in, The weighted k-space error of the network prediction; For the improved ResNet network, These are the parameters to be learned for the nth network.

[0101] The network input consists of weighted k-space data after frequency equalization weighting.

[0102] Furthermore, the network contains several residual blocks, each containing two convolutional layers and a nonlinear activation function, and retains linear features through skip connections; the network output layer is the predicted weighted k-space error.

[0103] Step S24: Obtain different frequency bands in the k-space using a frequency partition mask, weight the low-frequency path and the weighted full-frequency path, to obtain:

[0104] ;

[0105] in To obtain different frequency bands of the weighted k-space error of the prediction using frequency partitioning masks; binary mask. By designing a binary mask matrix, frequency-selective truncation of k-space data can be achieved. The expression for the binary mask matrix includes the following:

[0106] ;

[0107] Where L represents the range of the area to be preserved.

[0108] Further, step S3 includes the following:

[0109] Step S31: Using a self-supervised learning model, train the neural network in the self-calibration signal region and backpropagate, where the loss function of the self-supervised learning model is:

[0110] ;

[0111] Where A is the ACS central region extraction operator.

[0112] Further, step S4 includes the following:

[0113] Step S41: After training, infer the reconstructed k-space results and calculate the weighted error of the prediction. Perform inverse weighting to obtain the original scale error:

[0114] ;

[0115] Step S42: Gradually segment and fuse the errors output from different paths:

[0116] ;

[0117] in, To extract the mask for each frequency band step by step, The expressions include:

[0118] ;

[0119] in, and This determines the range of the extracted frequency band;

[0120] Step S43: Compensate for the correction error to the initial reconstruction result. From:

[0121] ;

[0122] The corrected magnetic resonance image is obtained.

[0123] The process of step S4 corresponds to Figure 2 The reasoning process in part (B) of the document.

[0124] According to a second aspect of the present invention, a magnetic resonance reconstruction error correction system based on multi-frequency priors is proposed for performing a magnetic resonance reconstruction error correction method based on multi-frequency priors as described in any one of the present invention, characterized in that the magnetic resonance reconstruction error correction system based on multi-frequency priors includes the following:

[0125] The data acquisition module is used to acquire undersampled k-space data and preliminary reconstructed k-space data;

[0126] The multi-frequency prior constraint module is used to construct a multi-frequency prior constraint model that includes frequency equalization weighting and frequency partitioning masking, and to establish a coil joint modeling strategy that combines real and imaginary parts.

[0127] The self-supervised learning mechanism module is used to train a neural network in the self-calibration signal region to predict k-space errors using a self-supervised learning mechanism.

[0128] The error correction module is used to compensate for the predicted errors in the preliminary reconstruction results to obtain the corrected magnetic resonance image.

[0129] According to a third aspect of the present invention, the present invention provides a magnetic resonance reconstruction error correction system based on multi-frequency priors, comprising an electronic device, wherein the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements a magnetic resonance reconstruction error correction method based on multi-frequency priors as described in any one of the present invention.

[0130] According to a fourth aspect of the present invention, the present invention provides a magnetic resonance reconstruction error correction system based on multi-frequency priors, comprising a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements a magnetic resonance reconstruction error correction method based on multi-frequency priors as described in any one of the present invention.

[0131] In addition to the above, the present invention also has related embodiments, including the following:

[0132] The image correction process in this embodiment of the invention.

[0133] like Figure 2 As stated, in which Figure 2 (A) represents the training process; Figure 2 (B) represents the reasoning process; Figure 2 (C) Select a mask for the frequency band; Figure 2(D) is the frequency partition mask module.

[0134] Figure 3 This is the network used in the embodiments of the present invention.

[0135] Figure 4 This is the experimental result of the corrected and reconstructed MRI brain map in an embodiment of the present invention. A 16-channel two-dimensional magnetic resonance image acquired in a clinical setting was used. Figure 4 (A) is the original image; Figure 4 (B) is the result image corrected by the method of the present invention; (c) is the sampling template used, with an acceleration factor of 4 and an automatic calibration area of ​​30.

[0136] As shown in Table 1, the comparison results between the method of the present invention and the method using SPARK correction are as follows:

[0137] Table 1 is a comparative table of the method of the present invention and the method using SPARK correction.

[0138] Magnetic resonance reconstruction image correction methods Relative L2-norm error Peak signal-to-noise ratio Reconstruction time (in seconds) SPARK calibration method 0.0828 36.61 36.6 The correction method of the present invention 0.0730 37.69 9.1

[0139] Table 1 compares the relative L2-norm error, peak signal-to-noise ratio, and correction time of the method of the present invention with those of the SPARK correction method. It can be seen that the method of the present invention can achieve the purpose of rapid correction.

[0140] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A magnetic resonance reconstruction error correction method based on multi-frequency priors, characterized in that, Includes the following: Step S1: Obtain undersampled k-space data and preliminary reconstructed k-space data; Step S2: Construct a multi-frequency prior constraint model that includes frequency equalization weighting and frequency partitioning mask, and establish a coil joint modeling strategy that combines real and imaginary parts; Step S3: Use a self-supervised learning mechanism to train a neural network in the self-calibration signal region to predict k-space errors; Step S4: Compensate the predicted error into the preliminary reconstruction result to obtain the corrected magnetic resonance image.

2. The magnetic resonance reconstruction error correction method based on multi-frequency priors according to claim 1, characterized in that, Step S1 includes the following: Step S1: The initially reconstructed k-space data is as follows: ； in, For the reconstructed k-space data, This refers to the k-space data acquired by the coil under undersampling conditions; Let k represent the k-space obtained by reconstructing undersampled data using any MRI reconstruction method.

3. The magnetic resonance reconstruction error correction method based on multi-frequency priors according to claim 2, characterized in that, Step S2 includes the following: Step S21: Use a multi-frequency prior constraint model to solve the problem of uneven amplitude distribution in k-space. The multi-frequency prior constraint model is composed of frequency equalization weighting and frequency partitioning masking strategy. The frequency equalization weighting includes: designing a smooth weighting function W to suppress low-frequency energy and increase high-frequency weights, so that the network can learn the features of each frequency band in a balanced way; The frequency partition mask includes: Constructing a binary mask The k-space is divided into multiple frequency bands, and differentiated constraints are applied to the errors of different frequency bands to prevent high-frequency noise from interfering with the recovery of low-frequency structures.

4. The magnetic resonance reconstruction error correction method based on multi-frequency priors according to claim 3, characterized in that, Step S2 also includes the following: Step S22: During the training and inference process of the multi-frequency prior constraint model, the coil-by-coil independent processing mode is changed, and the real part data of all coils in the k-space data of the reconstructed image are concatenated. The imaginary part data is concatenated as The data are input into the network separately, and the method of simultaneously training and inferring multi-coil data is used to reduce the number of model parameters and training and inference time. During the training phase, for the k-space data of the input reconstructed image, the real or imaginary parts of all coils are used. First, a frequency-balanced weighting strategy is used to weight it, resulting in: ； Where W is a smoothing weighting function, which uses a smoothly varying weighting matrix to apply progressive weight adjustments to the data at different locations in the k-space, including the following: ； Where a and b are k-space coordinates. in For amplitude adjustment parameters, Let P be the smoothness parameter, where the expression for P is as follows: ； Where P is the normalization factor; P is used to suppress low-frequency energy and increase high-frequency weights.

5. The magnetic resonance reconstruction error correction method based on multi-frequency priors according to claim 4, characterized in that, Step S2 also includes the following: Step S23: Input the weighted result into n identical networks respectively to predict the reconstruction error, and obtain: ； in, The weighted k-space error of the network prediction; For the improved ResNet network, These are the parameters to be learned for the nth network. The network input consists of weighted k-space data after frequency equalization weighting. Furthermore, the network contains several residual blocks, each containing two convolutional layers and a nonlinear activation function, and retains linear features through skip connections; the network output layer is the predicted weighted k-space error. Step S24: Obtain different frequency bands in the k-space using a frequency partition mask, weight the low-frequency path and the weighted full-frequency path, to obtain: ； in To obtain different frequency bands of the weighted k-space error of the prediction using frequency partitioning masks; binary mask. By designing a binary mask matrix, frequency-selective truncation of k-space data can be achieved. The expression for the binary mask matrix includes the following: ； Where L represents the range of the area to be preserved.

6. The magnetic resonance reconstruction error correction method based on multi-frequency priors according to claim 5, characterized in that, Step S3 includes the following: Step S31: Using a self-supervised learning model, train the neural network in the self-calibration signal region and backpropagate, where the loss function of the self-supervised learning model is: ； Where A is the ACS central region extraction operator.

7. The magnetic resonance reconstruction error correction method based on multi-frequency priors according to claim 6, characterized in that, Step S4 includes the following: Step S41: After training, infer the reconstructed k-space results and calculate the weighted error of the prediction. Perform inverse weighting to obtain the original scale error: ； Step S42: Gradually segment and fuse the errors output from different paths: ； in, To extract the mask for each frequency band step by step, The expressions include: ； in, and This determines the range of the extracted frequency band; Step S43: Compensate for the correction error to the initial reconstruction result. From: ； The corrected magnetic resonance image is obtained.

8. A magnetic resonance reconstruction error correction system based on multi-frequency priors, used to execute a magnetic resonance reconstruction error correction method based on multi-frequency priors as described in any one of claims 1 to 7, characterized in that, The aforementioned magnetic resonance reconstruction error correction system based on multi-frequency priors includes the following: The data acquisition module is used to acquire undersampled k-space data and preliminary reconstructed k-space data; The multi-frequency prior constraint module is used to construct a multi-frequency prior constraint model that includes frequency equalization weighting and frequency partitioning masking, and to establish a coil joint modeling strategy that combines real and imaginary parts. The self-supervised learning mechanism module is used to train a neural network in the self-calibration signal region to predict k-space errors using a self-supervised learning mechanism. The error correction module is used to compensate for the predicted errors in the preliminary reconstruction results to obtain the corrected magnetic resonance image.

9. A magnetic resonance reconstruction error correction system based on multi-frequency priors, comprising an electronic device, wherein the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a magnetic resonance reconstruction error correction method based on multi-frequency priors as described in any one of claims 1 to 7.

10. A magnetic resonance reconstruction error correction system based on multi-frequency priors, comprising a computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a magnetic resonance reconstruction error correction method based on multi-frequency priors as described in any one of claims 1 to 7.

Images