A method of magnetic resonance fingerprinting reconstruction with phase compensated dual signal contrast encoding

Abstract

The application discloses a phase compensation double signal contrast encoding magnetic resonance fingerprint reconstruction method, which is guided by a Gridding reconstruction technology based on Fourier transform, and introduces a phase compensation module for double signal contrast encoding magnetic resonance fingerprint imaging, and dimension reduction processing is performed on the Gridding reconstruction result by introducing singular value decomposition, so that the complexity of network calculation is effectively reduced. Meanwhile, a neural network is used for denoising in the reconstruction process. The denoising network is based on an improved UNet structure, the input of the network contains the real part and the imaginary part of the subspace signal, and the denoised subspace signal is output after network processing. The signal after denoising processing is recovered to the original time sequence space through singular value decomposition inverse transformation for subsequent processing steps. Through the introduction of the Gridding guide module and the phase compensation module, the present application makes full use of part of the prior information, thereby providing a more accurate and efficient self-supervised magnetic resonance image reconstruction method without additional data sets.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging technology, and in particular to a magnetic resonance fingerprint reconstruction method based on phase-compensated dual-signal contrast coding. Background Technology

[0002] Magnetic resonance fingerprinting (MRF) technology has attracted widespread attention due to its novelty and versatility, as it has been proposed for simultaneous multi-parameter mapping. From an informatics perspective, MRF utilizes theoretical predictions of the magnetic resonance signals that biological tissues may produce as prior knowledge to establish a correlation between the observed signal and quantitative parameters. Therefore, multiple quantitative parameters, such as T1, T2, PD, and B1, can be acquired simultaneously in a single scan.

[0003] The MRF imaging process mainly includes data acquisition, Fourier transform-based image reconstruction, and pattern matching. Notably, the acquisition process employed a variable density spiral (VDS) trajectory. Initially, Madan et al. used a Fourier transform-based Gridding algorithm for MRF reconstruction. Furthermore, the dual-signal contrast-coded MRF sequence proposed by Huihui, Shizhuo et al., presents new challenges to image reconstruction due to the phase difference between the two codes.

[0004] Meanwhile, various reconstruction algorithms have been applied to MRF reconstruction, each with its own advantages and disadvantages. The Gridding reconstruction technique proposed by Ma et al. offers a relatively fast reconstruction speed, but there is still significant room for improvement in accuracy. Model-based reconstruction methods, such as low-rank methods including the FLOR algorithm, employ conjugate gradient descent for iterative reconstruction. The FLOR algorithm achieved good results in simulation experiments. However, when applied to real-world experiments with multiple coils, the reconstruction speed significantly slowed down, and the overall reconstruction quality was poor. Furthermore, deep learning methods have been increasingly applied to MRF reconstruction in recent years, attracting widespread attention due to their superior denoising capabilities and fast computation speed. However, obtaining magnetic resonance fingerprint datasets remains a significant challenge, greatly limiting the application of existing supervised learning networks. Against this backdrop, self-supervised models have important practical significance. The DIP (Depth Image Prior) reconstruction algorithm proposed by Jesse et al. requires no additional dataset training. This method combines a physical model with a DIP network to achieve MRF reconstruction. Compared with traditional iterative reconstruction methods, the DIP algorithm shows a significant improvement in quantitative accuracy. However, DIP does not consider the phase issue between different frames of MRF, and there is still room for improvement in the application of double contrast-coded MRF. The accuracy of MRF quantification depends on the reconstruction quality of the signal, and accurate K-space signal reconstruction remains an important task. Summary of the Invention

[0005] In view of the above, this invention provides a magnetic resonance fingerprint reconstruction method based on phase-compensated dual-signal contrast coding. This method first preprocesses the image using the Gridding algorithm, then reduces the dimensionality of the Gridding reconstruction result by introducing singular value decomposition (SVD). Next, a neural network is used for denoising during the reconstruction process. The denoising network is based on an improved UNet structure; its input includes the real and imaginary parts of the subspace signal, and the output is a denoised subspace signal. The denoised signal is then restored to its original temporal space using inverse SVD transform. This reconstruction method further improves the clarity and accuracy of the reconstructed image. This invention is achieved through the following technical solutions:

[0006] This invention discloses a magnetic resonance fingerprint reconstruction method based on phase-compensated dual-signal contrast coding, characterized by comprising the following steps:

[0007] 1) The preprocessing of the traditional Fourier transform-based gridding reconstruction algorithm is used to integrate multi-coil data to obtain a spatial domain image. In addition, phase information is provided for the dual-signal encoded MRF.

[0008] 2) Construct a fingerprint dictionary using the Bloch equation and sequence parameters of magnetic resonance imaging, extract the basis of principal components using the dictionary, and extract principal components from the preprocessing results of step 1) using the basis of principal components to achieve data compression, reduce computational complexity, and obtain the initial subspace data;

[0009] 3) Extract the real and imaginary parts of the initial subspace data obtained in step 2) and input them into the neural network for denoising, and output the denoised subspace data;

[0010] 4) Perform magnetic resonance forward physical sampling on the denoised subspace data from step 3), including expanding the multi-coil data using coil sensitivity maps and non-uniform Fourier transform, in order to predict the K-space data.

[0011] 5) Calculate the loss function by combining the K-space data predicted in step 4) with the actual sampled K-space data, and then backpropagate to update the neural network in step 3).

[0012] 6) Match the temporal variation signal corresponding to each pixel of the subspace data in steps 3) and 4) with the dictionary in step 2) to index the specific parameter and obtain the quantitative image of the parameter.

[0013] As a further improvement, in step 1) of this invention, the traditional Gridding reconstruction algorithm based on Fourier transform reconstructs the non-Cartesian sampling trajectory and processes multi-coil K-space data. If it is a traditional MRF signal, there is no need to extract phase information. If it is a double contrast-coded MRF signal, the phase difference information of the double signal contrast coding is extracted.

[0014] As a further improvement, in step 2) of the present invention, the basis of the principal components is obtained by performing singular value decomposition on the established dictionary. The basis of the principal components is used to map the time series data onto the subspace, which facilitates subsequent denoising and other processing.

[0015] As a further improvement, in step 3) of this invention, the neural network uses the UNet structure and is not pre-trained, but only initialized and updated through backpropagation of the loss function.

[0016] As a further improvement, the magnetic resonance physical sampling process in step 4) of the present invention is as follows:

[0017]

[0018] in, For the K-space data sampled by the c-th coil at time point t, Let A be the image reconstructed at time point t. Let A be the system matrix.

[0019] A = TrajFMΔΦ;

[0020] Where Traj is the sampling trajectory, F is the Fourier transform, and M is the coil sensitivity map.

[0021] As a further improvement, the ΔΦ described in this invention is extracted from the result obtained in step 1) and used to compensate for the phase difference between the two signals in the reconstruction of dual-signal contrast coding. This parameter is not required for traditional MRF.

[0022] As a further improvement, the loss function in step 4) of this invention adopts the spiral trajectory weighted MSE function:

[0023]

[0024] As a further improvement, the weight w described in this invention is generated based on a spiral trajectory, with a smaller weight used in areas with dense sampling points and a larger weight used in areas with sparse sampling points.

[0025] As a further improvement, the dictionary matching method in step 6) of the present invention is as follows: calculate the inner product of the reconstructed dynamic signal and the dynamic signal in the dictionary, and then sum their absolute values. The largest result indicates that the dynamic signal and the reconstructed signal are the best match, and the corresponding parameter is the physical parameter of the pixel.

[0026] The beneficial effects of adopting the technical solution of this invention are as follows:

[0027] 1. This invention innovatively uses the Gridding reconstruction algorithm as a preprocessing step. Gridding preprocessing extracts sampling trajectory weights and phase compensation information from k-space data, providing structured prior knowledge for subsequent neural networks and avoiding artifacts caused by the lack of physical constraints in purely data-driven methods. The initial reconstruction results from Gridding's Fast Fourier Transform (FFT) are fused with the residuals from the neural network output, forming a hybrid reconstruction framework of "physical model + data-driven," improving the signal-to-noise ratio (SNR) and relative errors. More accurate parametric maps (such as T1 / T2 relaxation times) can provide reliable evidence for quantitative imaging diagnosis (such as tumor grading and ischemic stroke staging), reducing the risk of subjective misjudgment by physicians.

[0028] 2. This invention is a self-supervised neural network that uses an untrained UNet network, eliminating the need for pre-training with additional datasets. By employing k-space data consistency loss, it avoids dependence on external datasets, addressing the pain points of medical data privacy and high annotation costs. It provides a "out-of-the-box" solution for resource-constrained medical institutions, requiring no additional GPU clusters or large-scale annotation datasets.

[0029] 3. This invention performs Singular Value Decomposition (SVD) on K-space data, retaining the top k singular values ​​(energy percentage > 95%), and compressing the data volume to 3% of the original size. This significantly reduces memory consumption during computation and decreases the hardware requirements of medical centers.

[0030] 4. This invention, by introducing phase compensation, solves the phase difference problem of double-contrast encoded data to a certain extent, further improving accuracy by approximately 4% compared to without phase difference compensation. It can be used as a processing method for double-signal encoding, extracting information between the two signals through preprocessing, thereby providing conditions for subsequent processing and improving accuracy. Attached Figure Description

[0031] Figure 1 This is an overall flowchart of the present invention;

[0032] Figure 2 It mainly includes the Gridding algorithm preprocessing diagram (top left), the DIP denoising network structure diagram (bottom), and the physical imaging process diagram (top right), which are schematic diagrams of the reconstruction algorithm of this invention;

[0033] Figure 3 It includes multi-coil data generation and non-uniform Fourier transform, and is a schematic diagram of the physical imaging process of the present invention;

[0034] Figure 4 The ablation experiment results of each component of this invention are shown in the figure.

[0035] Figure 5 A comparison chart showing the results with the existing reconstruction method FLOR;

[0036] Figure 6 The image shows the reconstruction results of different brain layers using this invention. Detailed Implementation

[0037] To describe the present invention in more detail, the counting scheme of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] This invention provides a magnetic resonance fingerprint imaging reconstruction method. Figure 1 The flowchart of the method is shown. Figure 2 The overall structure of the reconstruction algorithm is demonstrated. It mainly consists of four parts: sampled data preprocessing, neural network denoising, simulated physical sampling and loss function calculation, and parameter quantification.

[0039] 1. Sample data preprocessing:

[0040] (1) Data acquisition: In the MRI instrument, double-contrast encoded MRF sequences and pre-set VDS sampling trajectories were used. t K-space data y were collected. t,c Where t represents the time point used in the MRF sequence, and c represents the number of coils. The acquired data is time-series, multi-coil frequency domain data. Simultaneously, it is necessary to use an instrument to scan the coil sensitivity map M. c .

[0041] (2) K-space reconstruction: Reconstruct the coil sensitivity map M c Sampling trajectory Traj t and K-space data y t,c The data is input into a traditional Fourier transform-based Gridding reconstruction algorithm, integrates multi-coil data, and outputs a spatial domain image. Represented as:

[0042]

[0043] (3) For dual-signal contrast coding, such as alternating sampling of FISP and PSIF sequences, the signal in the FISP module mainly comes from the FID signal excited by the RF of this module; the rewinder gradient of the PSIF module is located at the end of the module, and the readout gradient is the time-reverse form of the readout gradient in the FISP module. The signal mainly comes from the echo signal obtained after the signal in the previous module is refocused by the RF of this module. There is a phase difference between the two modules. The phase difference between the dual contrast codes is extracted using the processing results of the Gridding algorithm. The average phase difference is:

[0044]

[0045] in, for The average phase of the FISP module in the middle, for The average phase of the PSIF module.

[0046] 2. Neural Network Denoising:

[0047] (1) Principal component extraction: For the reconstruction obtained above Its time duration t is typically several hundred, which greatly increases the computational burden on subsequent neural network processing. Therefore, principal component analysis is chosen for dimensionality reduction:

[0048] A. By establishing a fingerprint database containing all possible MRF signals, called a dictionary, the actual acquired signal is matched with the signal in the fingerprint database to achieve quantization imaging. The dictionary establishment process includes: (1) limiting the variation range of each parameter, such as (100:20:200ms); (2) based on the Bloch equation, calculating all possible time evolution curves according to the dynamic range of different parameters to generate the dictionary dict.

[0049] B. Perform singular value decomposition on the dictionary established above:

[0050] [U,S,V]=SVD(dict)

[0051] Singular Value Decomposition (SVD) is used here. The first k principal components are selected; that is, for a parameter U, the first k columns are taken, denoted as Uk. k This is how it can be done. Principal component extraction:

[0052]

[0053] in, for The subspace projection of contains the first k principal components.

[0054] C. In the experiment, t was set to 468 and k to 6. Principal component analysis can achieve a data compression of 78 times while preserving [data quality / value]. The vast majority of information.

[0055] (2) DIP Neural Network:

[0056] A. This network mainly targets Denoising is performed to obtain higher quality subspace data. During input, the real and imaginary parts need to be separated; therefore, the input dimension is N×N×2k, where N×N is the image size and 2k represents k real parts and k imaginary parts. The network output is the denoised subspace data. The size remains N×N×2k.

[0057] B. The network structure is the UNet network structure, which includes 3×3 convolutional layers, five 2×2 pooling layers, and five 2×2 upsampling layers. The network weight parameters do not require additional datasets for training, and denoising is achieved through iterative self-supervised network training.

[0058] 3. Physical forward sampling and loss function calculation:

[0059] A. Subspace data output by the network Perform physical forward sampling. First, use U... k transpose conjugate U k * Restore it to its original temporal space:

[0060]

[0061] like Figure 3 Predicting K-space data using prior information:

[0062]

[0063] Where A is the system matrix:

[0064] A=TrajFMΔΦ

[0065] Where Traj is the sampling trajectory, F is the Fourier transform, and M is the coil sensitivity map.

[0066] B. K-space data predicted by the physical sampling process K-space data y obtained by instrument sampling t,c Calculate the loss function, which is set as a spiral density-weighted MSE loss function, where the spiral density weight w is generated based on the sampling trajectory in order to balance the density transformation of the sampling trajectory.

[0067]

[0068] The weight w is generated based on the spiral trajectory and is used to compensate for the non-uniformity of the sampling trajectory. It is reflected in the fact that the inside of the trajectory represents low-frequency information and the weight is small when the sampling density is large, while the outside of the trajectory represents high-frequency information and the weight is large when the sampling density is small.

[0069] Using a loss function, the network weight parameters are updated through backpropagation. The update stops once the network converges or reaches a preset number of iterations, and the final time series signal is output.

[0070] 4. Parameter quantification:

[0071] For timing signals The dot product is calculated by multiplying the dynamic signal in the dictionary with the reconstructed signal. The largest dot product indicates that the dynamic signal best matches the reconstructed signal, and its corresponding parameters are the physical parameters of that pixel. This yields the distribution map of the quantitative parameters (T1, T2) for the entire image.

[0072] The experiment was divided into two parts: a simulation experiment and an actual scanning experiment.

[0073] (1) Simulation Experiment Section

[0074] like Figure 4 Experimental results and error maps were used to construct a set of brain simulation data as experimental reference values. Simulated ablation experiments were conducted to test the proposed method using a simulated spiral sampling trajectory. Simultaneously, a mask was selected to extract the region of interest, and the relative error between this region and the reference value was calculated. Where I and These represent the true values ​​of the quantitative parameters and the simulation results, respectively. The upper part shows the T1 quantitative result graph and error graph, and the lower part shows the T2 quantitative result graph and error graph. The relative error is displayed in the lower left corner of the error graph.

[0075] The first column represents the experimental reference values. The second column represents the experimental results obtained using Gridding reconstruction. The third column represents the results after removing phase alignment, the fourth column represents the results after removing Gridding preprocessing, and the fifth column represents the results of this invention.

[0076] (2) Real Experiment Section

[0077] In vivo experiments were conducted using a Siemens Prisma 3.0T 20-channel head coil scanner with the written consent of healthy volunteers. Figure 5The experiment comparing the present invention with the existing reconstruction method FLOR is shown. The reference is the result of reconstruction with 12 interleaves and Gridding. The two reconstruction methods were applied to traditional MRF and dual-contrast encoded MRF, respectively.

[0078] like Figure 6 Table 1 shows the reconstruction results of different brain layers using the reconstruction algorithm of this invention, including T1 and T2 quantitative results of traditional MRF and dual contrastive MRF (Dual). Table 1 shows the normalized standard deviation of each quantitative map in this experiment.

[0079] Table 1 shows the normalized standard deviation results for different brain layers.

[0080]

[0081] The normalized standard deviation reflects the signal-to-noise ratio of the image to some extent. The table shows that the quantitative effect of T2 is slightly worse than that of T1. By introducing dual-signal contrast coding and the reconstruction algorithm mentioned in this invention, T2 is improved in all five layers.

[0082] The parameters are defined as follows in the embodiment:

[0083] T1: Longitudinal relaxation time, which refers to the time required for the longitudinal magnetization vector to recover from zero to 67% of the total signal strength;

[0084] T2: Transverse relaxation time, which refers to the time required for the transverse magnetization vector to decay from 100% to 37%;

[0085] Average phase difference refers to the average phase difference between the FISP signal and the PSIF signal in dual-signal contrast coding (MRF).

[0086] Image domain data obtained by preliminary reconstruction of K-space data.

[0087] right The subspace image obtained after principal component extraction.

[0088] Subspace image after DIP denoising.

[0089] right The recovered image domain data is used as the network output.

[0090] U k The basis of the principal components is obtained by performing singular value decomposition on the dictionary and is used for data compression.

[0091] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. It will be apparent to those skilled in the art that various modifications can be made to the above embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.

Claims

1. A magnetic resonance fingerprint reconstruction method based on phase-compensated dual-signal contrast coding, characterized in that, Includes the following steps: 1) Preprocessing is performed using the traditional Fourier transform-based Gridding reconstruction algorithm. During the preprocessing process, multi-coil data is integrated to obtain a spatial domain image. In addition, phase information is provided for the dual-signal coded MRF. In step 1), the traditional Fourier transform-based Gridding reconstruction algorithm reconstructs the non-Cartesian sampling trajectory and processes the multi-coil K-space data. If it is a traditional MRF signal, there is no need to extract phase information. If it is a dual-contrast coded MRF signal, the phase difference information of the dual-signal contrast coding is extracted. 2) Construct a fingerprint dictionary using the Bloch equation and sequence parameters of magnetic resonance imaging, extract the basis of principal components using the dictionary, and extract principal components from the preprocessing results of step 1) using the basis of principal components to achieve data compression, reduce computational complexity, and obtain the initial subspace data; 3) Extract the real and imaginary parts of the initial subspace data obtained in step 2) and input them into the neural network for denoising, and output the denoised subspace data; 4) Perform magnetic resonance forward physical sampling on the denoised subspace data from step 3), including expanding the multi-coil data using coil sensitivity maps and non-uniform Fourier transform, in order to predict the K-space data. 5) Calculate the loss function by combining the K-space data predicted in step 4) with the actual sampled K-space data, and then backpropagate to update the neural network in step 3); 6) Match the temporal variation signal corresponding to each pixel of the subspace data in steps 3) and 4) with the dictionary in step 2) to index the specific parameter and obtain the quantitative image of the parameter.

2. The method according to claim 1, characterized in that: In step 2), the basis of the principal components is obtained by performing singular value decomposition on the established dictionary. The basis of the principal components is used to map the time series data onto the subspace, which facilitates subsequent denoising and other processing.

3. The method according to claim 1 or 2, characterized in that: In step 3), the neural network uses the UNet structure and is not pre-trained. It is only initialized and updated through backpropagation of the loss function.

4. The method according to claim 3, characterized in that: The magnetic resonance physical sampling process in step 4) is as follows: ； in, For the K-space data sampled by the c-th coil at time point t, The image reconstructed at time point t. For the system matrix: ； Where Traj is the sampling trajectory, F is the Fourier transform, and M is the coil sensitivity map.

5. The method according to claim 4, characterized in that, The aforementioned The result obtained from step 1) is extracted and used to compensate for the phase difference between the two signals in the reconstruction of dual-signal contrast coding. This parameter is not required for traditional MRF.

6. The method according to claim 1, 2, 4, or 5, characterized in that, The loss function in step 4) is the spiral trajectory weighted MSE function: 。 7. The method according to claim 6, characterized in that, w represents the weight, which is generated based on the spiral trajectory. Smaller weights are used in areas with dense sampling points, while larger weights are used in areas with sparse sampling points.

8. The method according to claim 7, characterized in that, The dictionary matching method in step 6) is as follows: calculate the inner product of the reconstructed dynamic signal and the dynamic signal in the dictionary, and then sum their absolute values. The largest result indicates that the dynamic signal and the reconstructed signal are the best match, and the corresponding parameter is the physical parameter of the pixel.

Images