Dynamic magnetic resonance image reconstruction method, apparatus, and device

Abstract

Embodiments of the present application provide a dynamic magnetic resonance image reconstruction method, device and equipment, the method comprising: acquiring dynamic magnetic resonance raw data of a target object by using a center-periphery alternating acquisition mode; performing channel compression on the raw data to obtain compressed data; extracting a time basis function, a coil sensitivity map and down-sampling data based on the compressed data; constructing an optimization expression of a spatial basis function according to the time basis function, the coil sensitivity map, the down-sampling data and a down-sampling mask, taking a weighted sum of a reconstruction error and a difference between adjacent frames as an optimization target; eliminating the time dimension in the process of solving the optimization expression of the spatial basis function based on a conjugate gradient descent algorithm to obtain an optimal solution of the spatial basis function; and reconstructing a dynamic magnetic resonance image of the target object according to the optimal solution of the spatial basis function and the time basis function. Not only the reconstruction speed is improved, but also the reconstruction quality is improved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of image processing, and particularly relate to a dynamic magnetic resonance image reconstruction method, device and equipment. BACKGROUND

[0002] Nuclear Magnetic Resonance Imaging (MRI) technology is a commonly used imaging technology in the clinic, which uses a highly integrated instrument that integrates physical and chemical technologies, and can provide images containing anatomical structure and pathological information of a target object in a non-invasive manner. Dynamic magnetic resonance imaging technology is to sample magnetic resonance at a certain frame rate, so as to realize dynamic imaging of the target object, which has important value in the fields of medical examination and diagnosis of heart diseases, etc. Due to the limitations of physical laws, hardware devices and subject safety, the sampling frequency of the magnetic resonance device is limited, and the time resolution (equivalent to the frame rate) of the dynamic magnetic resonance imaging is also limited. The dynamic magnetic resonance imaging technology for fast moving objects has become a difficult and hot issue in the field.

[0003] Under the requirement of high time resolution, the downsampling multiple of the dynamic magnetic resonance imaging is very high, which brings great difficulty to the reconstruction of the dynamic magnetic resonance image under high time resolution. The skilled person in the art tries to mine the unique structure and properties of the dynamic magnetic resonance data to reconstruct a high-quality image as much as possible without distortion from less sampling data. Low-rank subspace technology is a representative method, which expresses the dynamic magnetic resonance data in the form of linear combination of a plurality of basis vectors, mines the correlation and redundancy of the dynamic data in space and time, and greatly reduces the degree of freedom of reconstruction, so as to realize the reconstruction of the dynamic magnetic resonance image with high time resolution. However, the expression of low-rank usually requires a long acquisition time and a large total data volume, resulting in a very long reconstruction time, which can be tens of minutes to several hours. The slow reconstruction process seriously limits its application in the clinic. Therefore, there is an urgent need for a fast dynamic magnetic resonance image reconstruction method that can meet the needs of clinical application. SUMMARY

[0004] Embodiments of the present application provide a dynamic magnetic resonance image reconstruction method, device and equipment to solve the problem of long time consumption of the existing reconstruction method.

[0005] In a first aspect, embodiments of the present application provide a dynamic magnetic resonance image reconstruction method, comprising:

[0006] acquiring dynamic magnetic resonance raw data of a target object by using a center-periphery alternating acquisition mode;

[0007] performing channel compression on the raw data to obtain compressed data;

[0008] extracting a temporal basis function, a coil sensitivity map and down-sampled data based on the compressed data;

[0009] constructing an optimization expression of a spatial basis function according to the temporal basis function, the coil sensitivity map, the down-sampled data and a down-sampling mask, with an optimization objective of minimizing a weighted sum of reconstruction errors and differences between adjacent frames;

[0010] eliminating the time dimension in a process of solving the optimization expression of the spatial basis function based on a conjugate gradient descent algorithm to obtain an optimal solution of the spatial basis function;

[0011] reconstructing a dynamic magnetic resonance image of the target object according to the optimal solution of the spatial basis function and the temporal basis function.

[0012] In an embodiment, the dynamic magnetic resonance raw data of the target object is acquired by using a center-periphery alternating acquisition mode, including:

[0013] alternately performing center acquisition and periphery acquisition on the target object at a preset frequency to obtain center acquisition data used for motion navigation, denoted as motion navigation raw data, and periphery acquisition data used for image reconstruction, denoted as image reconstruction raw data.

[0014] In an embodiment, the raw data is channel compressed to obtain compressed data, including:

[0015] The motion navigation raw data and the image reconstruction raw data are respectively channel compressed by using a geometric decomposition coil channel compression algorithm to obtain motion navigation compressed data and image reconstruction compressed data.

[0016] In an embodiment, the temporal basis function, the coil sensitivity map and the down-sampled data are extracted based on the compressed data, including:

[0017] The motion navigation compressed data is singular value decomposed to obtain the temporal basis function;

[0018] The image reconstruction compressed data is averaged along the time dimension and is two-dimensionally inverse Fourier reconstructed with full sampling to obtain a time-averaged multi-channel image, and the image is normalized to obtain the coil sensitivity map;

[0019] The positions not acquired in the image reconstruction compressed data are zero-padded to obtain the down-sampled data.

[0020] In an embodiment, the optimization expression of the spatial basis function is:

[0021]

[0022] Wherein, M represents a down-sampling mask, S represents a coil sensitivity map, V represents a temporal basis function, y represents down-sampled data, D represents a first-order difference operator along a time dimension, U represents a spatial basis function, F represents a two-dimensional discrete Fourier transform operator, and λ represents a regularization weight;

[0023] Then the optimal solution of the spatial basis function U * is:

[0024] U * = (V H S H F H MFSV + λV H D H DV) -1 ( H S H F H y).

[0025] In one embodiment, the elimination of the time dimension in the process of solving the optimization expression based on the spatial basis function using the conjugate gradient descent algorithm comprises:

[0026] The operator A in the conjugate gradient descent algorithm is expressed as:

[0027] A = V H S H F H MFSV + λV H D H DV.

[0028] The time basis function V in the operator A is exchanged in order with the coil sensitivity map S and the two-dimensional discrete Fourier transform operator F, to obtain:

[0029] A = S H F H V H MVFS + λV H D H DV.

[0030] In one embodiment, the method further comprises: combining V H , M and V in the operator A into one low-dimensional operator without the time dimension.

[0031] In a second aspect, an embodiment of the present application provides a dynamic magnetic resonance image reconstruction device, comprising:

[0032] An acquisition module configured to acquire dynamic magnetic resonance raw data of a target object using a center-periphery alternating acquisition mode;

[0033] A compression module configured to perform channel compression on the raw data to obtain compressed data;

[0034] extracting a temporal basis function, a coil sensitivity map and down-sampling data based on the compressed data;

[0035] an optimization module configured to construct an optimization expression of the spatial basis function according to the temporal basis function, the coil sensitivity map, the down-sampling data and a down-sampling mask, with a weighted sum of reconstruction errors and differences between adjacent frames as an optimization target;

[0036] a dimension reduction module configured to eliminate a time dimension in a process of solving the optimization expression of the spatial basis function based on a conjugate gradient descent algorithm, to obtain an optimal solution of the spatial basis function;

[0037] a reconstruction module configured to reconstruct a dynamic magnetic resonance image of the target object according to the optimal solution of the spatial basis function and the temporal basis function.

[0038] In a third aspect, an embodiment of the present application provides an electronic device, comprising:

[0039] at least one processor and a memory;

[0040] the memory stores computer-executed instructions;

[0041] the at least one processor executes the computer-executed instructions stored in the memory, so that the at least one processor executes the dynamic magnetic resonance image reconstruction method according to any one of the first aspect.

[0042] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores computer-executed instructions, and the computer-executed instructions are executed by a processor to implement the dynamic magnetic resonance image reconstruction method according to any one of the first aspect.

[0043] The dynamic magnetic resonance image reconstruction method, device and electronic device provided by the embodiments of the present application acquire dynamic magnetic resonance raw data of a target object by using a center-periphery alternating acquisition mode; perform channel compression on the raw data to obtain compressed data; extract a temporal basis function, a coil sensitivity map and down-sampling data based on the compressed data; construct an optimization expression of a spatial basis function according to the temporal basis function, the coil sensitivity map, the down-sampling data and a down-sampling mask, with a weighted sum of reconstruction errors and differences between adjacent frames as an optimization target; eliminate a time dimension in a process of solving the optimization expression of the spatial basis function based on a conjugate gradient descent algorithm, to obtain an optimal solution of the spatial basis function; and reconstruct a dynamic magnetic resonance image of the target object according to the optimal solution of the spatial basis function and the temporal basis function. The reconstruction speed is improved, and the reconstruction quality is improved. BRIEF DESCRIPTION OF DRAWINGS

[0044] The accompanying drawings, which are incorporated herein and constitute part of this specification, illustrate embodiments consistent with the application and, together with the description, further serve to explain the principles of the application.

[0045] Figure 1 A flow chart of a dynamic magnetic resonance image reconstruction method provided by an embodiment of the application;

[0046] Figure 2 A schematic diagram of a proof process of a merging operator provided by an embodiment of the application;

[0047] Figure 3 A flow chart of a whole process of dynamic magnetic resonance image reconstruction on a heart provided by an embodiment of the application;

[0048] Figure 4 A comparison schematic diagram of reconstruction effects provided by an embodiment of the application;

[0049] Figure 5 A structural schematic diagram of a dynamic magnetic resonance image reconstruction device provided by an embodiment of the application;

[0050] Figure 6 A structural schematic diagram of an electronic device provided by an embodiment of the application.

[0051] The specific embodiments of the application have been shown by the above-described drawings, and will be described in more detail hereinafter. These drawings and the written description are not intended to restrict the scope of the inventive concept in any way, but to explain the inventive concept to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0052] The application will be described in further detail by way of specific embodiments with reference to the accompanying drawings. Like reference numerals in different embodiments designate like elements. In the following embodiments, many specific details are described to provide a thorough understanding of the application. However, one skilled in the relevant art will readily recognize that the application can be practiced without one or more of the specific details, or with other elements, materials, methods, components, materials, and / or operations. In some instances, well-known structures, materials, and / or operations have not been described in detail in order to avoid obscuring the application.

[0053] In addition, features described in the specification, operations or characteristics can be combined in any appropriate manner to form various embodiments. At the same time, the steps or actions in the method description can also be sequentially exchanged or adjusted in a manner obvious to those skilled in the art. Therefore, the various sequences in the specification and drawings are only for the purpose of clearly describing a certain embodiment, and do not mean that the sequence is necessary, unless otherwise stated that a certain sequence must be followed.

[0054] The serial numbers of the components in the specification, such as "first", "second", etc., are only used to distinguish the described objects, and do not have any order or technical meaning. The "connection" and "coupling" in the present application include direct and indirect connection (coupling) unless otherwise specified.

[0055] Under the requirement of high temporal resolution, the downsampling ratio of dynamic magnetic resonance imaging is very high, which makes the dynamic magnetic resonance image reconstruction under high temporal resolution an inverse problem of a large underdetermined linear system. Under the requirement of high temporal resolution, the downsampling ratio of dynamic magnetic resonance imaging is very high, and the ill-conditioned nature of directly inverting the linear equation system is very strong, and the solving process is extremely sensitive to noise and error, and reliable image quality cannot be obtained. In recent decades, researchers have tried to mine the unique structure and properties of dynamic magnetic resonance data to provide additional prior information for the solution of this inverse problem, so as to reconstruct the real image as much as possible without distortion from less sampling data.

[0056] Low-rank subspace technology is a representative method, which mines the correlation and redundancy of dynamic data in space and time by expressing dynamic magnetic resonance data as a linear combination of several basis vectors, greatly reduces the degree of freedom of reconstruction, and can realize dynamic magnetic resonance imaging with high temporal resolution (10 ms and below). For example, the reconstruction technology of partial separable model combined with sparse prior (abbreviated as PS+Sparse), due to the application of sparse transformation in high-dimensional time domain space, it is necessary to use costly nonlinear truncated optimization in the optimization process, resulting in a long time-consuming; the reconstruction technology of partial separable model combined with locally low-rank prior (abbreviated as PS+LLR), the optimization of the regular term needs to calculate the time-consuming singular value decomposition, also leading to the extension of the reconstruction time. In summary, the existing dynamic magnetic resonance image reconstruction technology based on low-rank subspace, the regular term design and optimization algorithm all unintentionally or intentionally destroy the dimensionality reduction property of the subspace, resulting in a long reconstruction process and an unstable optimization process. High computational complexity, long reconstruction time; sensitive to hyperparameters, unstable optimization process. These all seriously restrict the application of dynamic magnetic resonance imaging technology in clinical.

[0057] In order to solve at least one problem existing in the prior art, the low-rank subspace technology is fully tapped in the dimension reduction characteristics in the iteration process of the optimization algorithm, and the entire iteration algorithm is innovatively converted into a low-dimensional projection space for implementation, so that the high-cost calculation process in the high-dimensional space is avoided, the calculation complexity is effectively reduced, and the optimization process is more stable. The following will be described in detail through specific embodiments.

[0058] Figure 1 A flowchart of a dynamic magnetic resonance image reconstruction method provided by an embodiment of the present application is shown in FIG. 1. As shown in the figure, the dynamic magnetic resonance image reconstruction method provided by the embodiment can include the following steps. Figure 1

[0059] S101, acquiring dynamic magnetic resonance raw data of a target object by using a center-periphery alternating acquisition mode.

[0060] The target object in the embodiment may, for example, be a heart. The target object is sampled for dynamic magnetic resonance data by using a (k-t) space center-periphery alternating acquisition mode. The dynamic magnetic resonance raw data has a dimension of 3, which is kx, ky and t respectively, representing the frequency encoding direction, the phase encoding direction and the time dimension. In the embodiment, a Cartesian rectangular coordinate system is used for sampling trajectory, so that the kx direction is fully sampled (satisfying the Nyquist sampling rate). The down-sampling process occurs in the ky-t two dimensions, so that the acquisition mode is actually a (ky-t) space sampling mode. The center-periphery alternating acquisition mode means that the value of ky is periodically and alternately collected at the center (near the position of ky=0) and the periphery (far positions of ky>0 and ky<0) as the time t moves.

[0061] In an alternative embodiment, the dynamic magnetic resonance raw data of the target object is acquired by using a center-periphery alternating acquisition mode, which can specifically include: alternately collecting the target object at the center and the periphery at a preset frequency, respectively obtaining center collection data for motion navigation, denoted as motion navigation raw data, and periphery collection data for image reconstruction, denoted as image reconstruction raw data.

[0062] Specifically, the center collection part is fixed at the position of ky=0, and the periphery collection part is randomly traversed in the position of ky, and the two are alternated at a frequency of 1:3, that is, once ky=0 is collected, and then three random ky values are collected. Correspondingly, the collected data will also be divided into two parts: 1. Center collection data for motion navigation, denoted as y Nav ​; 2. peripheral data acquisition, used for image reconstruction, the image reconstruction raw data is denoted as y Img .

[0063] S102, channel compression is performed on the raw data to obtain compressed data.

[0064] To reduce the redundancy coil channel dimension and reduce the reconstruction calculation amount, after the data acquisition is completed, channel compression can be performed on the data first. Considering that the Cartesian sampling trajectory is used in the present application, a geometric-decomposition coil compression (GCC) method can be used. In an optional implementation, channel compression is performed on the raw data to obtain compressed data, which can specifically include: using a geometric-decomposition coil compression algorithm to perform channel compression on the motion navigation raw data and the image reconstruction raw data respectively to obtain motion navigation compressed data and image reconstruction compressed data. Specifically, 24-channel data can be compressed into 8-channel data. Coil compression is performed on y Nav and y Img respectively, and the compressed data can be denoted as motion navigation compressed data and image reconstruction compressed data

[0065] Optionally, a direct compression method based on singular-value decomposition (SVD) can also be used to perform channel compression on the raw data.

[0066] S103, time basis functions, coil sensitivity maps, and down-sampling data are extracted based on the compressed data.

[0067] For ease of description, the following variables are defined in the present embodiment: the frequency encoding direction dimension size Nx, the phase encoding direction dimension size Ny, the time dimension size Nt, the coil channel number Nc, and the rank L of the low-rank approximation.

[0068] The specific way of extracting time basis functions, coil sensitivity maps, and down-sampling data based on the compressed data is as follows:

[0069] (k-t) spatial down-sampling mask M: a [Nx, Ny, Nt] matrix. It is determined by the center-peripheral acquisition trajectory, and specifically, in the (k-y-t) space, the acquired positions are marked as 1, and the unsampled positions are marked as 0.

[0070] Time basis function V: a [L, Nt] matrix. It is extracted from the motion navigation compressed data SVD is performed to obtain V by taking the first L eigenvectors of the time-dimension feature space.

[0071] coil sensitivity map S: a matrix of [Nx, Ny, Nc]. The image reconstruction compressed data The data is averaged along the time dimension and a full-sampling 2D inverse Fourier reconstruction is performed to obtain a time-averaged multi-channel image. The image is normalized using the SOS (sum-of-square) algorithm to obtain the estimated coil sensitivity map S.

[0072] (k-t) spatial down-sampling data y: a matrix of [Nx, Ny, Nc, Nt]. The image reconstruction compressed data The data is zero-padded, i.e., the acquired positions are padded with The data is filled with data corresponding to the acquired positions and zero-padded for the positions not acquired.

[0073] That is, in an optional embodiment, the extraction of the time basis function, the coil sensitivity map and the down-sampling data based on the compressed data can specifically include: performing SVD on the motion navigation compressed data to obtain the time basis function; averaging the image reconstruction compressed data along the time dimension and performing a full-sampling 2D inverse Fourier reconstruction to obtain a time-averaged multi-channel image, and normalizing the image to obtain the coil sensitivity map; and zero-padded the positions not acquired in the image reconstruction compressed data to obtain the down-sampling data.

[0074] S104, according to the time basis function, the coil sensitivity map, the down-sampling data and the down-sampling mask, constructing an optimization expression of the spatial basis function with minimizing the weighted sum of the reconstruction error and the difference between adjacent frames as the optimization target.

[0075] In an optional embodiment, the optimization expression of the spatial basis function is:

[0076]

[0077] wherein M represents the down-sampling mask, S represents the coil sensitivity map, V represents the time basis function, y represents the down-sampling data, D represents a first-order difference operator along the time dimension, U represents the spatial basis function, F represents a 2D discrete Fourier transform operator, and λ represents a regularization weight. The first term in the above formula is used to represent the reconstruction error, and the second term is used to measure the difference between adjacent frames. λ can be used to weigh the contributions of the two terms to the target, and its value is related to the absolute value of the down-sampling data. To better weigh the two terms, λ can be positively correlated with the energy size of the down-sampling data. Optionally, if the maximum modulus value of y is normalized to 1.0, the value of the corresponding λ is about 0.05.

[0078] It should be noted that the shape information of each operator is deliberately ignored in the optimized expression of the spatial basis function in this embodiment, because M, S, V, F and D are all linear transformations, although the implementation methods are different (for example, M and S are point-by-point multiplication, V and D are right-multiplication matrix multiplication, and F is a function library that uses fast Fourier transform FFT to achieve), but their essence can all be converted into the form of left-multiplication two-dimensional matrix through appropriate form transformation. In the algorithm implementation process, the shape of the intermediate variable will be frequently changed to meet the needs of each operator calculation, so as to achieve the fastest calculation speed. Therefore, the shape information of each operator is directly omitted in this embodiment, and it is considered as an abstract linear transformation operator, and only the calculation order is concerned. The above optimization objective is a generalized Tikhonov regularization problem. By taking the partial derivative and setting the gradient to zero, the optimal solution U of the spatial basis function can be obtained * * H H H H H -1 H H H

[0079] S105, in the process of solving the optimization expression of the spatial basis function based on the conjugate gradient descent algorithm, the time dimension is removed to obtain the optimal solution of the spatial basis function.

[0080] The iteration steps of the conjugate gradient descent algorithm (CG) are as follows:

[0081] Ap=A·p (k)

[0082] α=(r H r) (k-1) / 〈p (k) ,Ap>

[0083] x (k) =x (k-1) +α·Ap

[0084] r (k) =r (k-1) -α·Ap

[0085]

[0086] β=(r H r) (k) / (r H r) (k-1) ​​​​​​​​​​​

[0087] p (k+1) = r ( k ) + a · p (k)

[0088] where p (k) represents the descent direction of the kth step, r (k) represents the residual of the kth step, a represents the step size of descent, and b represents the conjugate correction coefficient of the descent direction. “·” represents matrix multiplication.

[0089] In general, the initial value of the CG algorithm is x (0) = 0, r (0) = b-Ax (0) , p (0) = r (0)

[0090] In this embodiment, each element of the CG algorithm is obtained in the following manner:

[0091] A = V H S H F H MFSV + aV H D H DV, r (0) = V H S H F H y

[0092] The greatest contribution of this application lies in the design of the optimization target, which can very efficiently use the CG algorithm for optimization, while achieving very high image reconstruction quality and extremely fast reconstruction speed. The core lies in the optimization idea of dimension reduction. Specifically, for the term aV H D H DV in A, since V and D can be calculated in advance and combined, the term aV H D H DV can actually be explicitly written as a small matrix of [L, L], which can be calculated efficiently. Since V and D both contain the time dimension Nt, but aV H D H DV does not contain the time dimension. Therefore, after combining, the calculation efficiency is obviously improved (because the number of frames Nt is much larger than the rank L). For the term V H S H F H MFAV in A, its matrix form has a very high dimension and cannot be explicitly written. Directly calculating and combining in advance is not realistic. However, there is still a method to project this term into a low-dimensional space in theory.

[0093] Specifically, the time basis function V in the operator A is exchanged with the coil sensitivity map S and the two-dimensional discrete Fourier transform F, respectively, by swapping. After the swapping, the operator A is rewritten as:

[0094] A = S H F H V H MVFS + λV H D H DV

[0095] After the swapping, the coil sensitivity map S and the two-dimensional discrete Fourier transform F only need to operate on the low-dimensional spatial basis function U, and no longer need to operate on each frame of the high-dimensional dynamic image, thus greatly reducing the computational amount. According to the associative law of matrix multiplication, considering that the coil sensitivity map S and the two-dimensional discrete Fourier transform F are linear transformations of the spatial dimension and are independent of the time dimension, the dimensions of the V are independent of each other, and the commutativity is possessed.

[0096] After the above swapping, the V H , M, and V in the operator A can be further combined into an operator φ, which is a complete low-dimensional operator (a [L, L] matrix at each spatial position, not containing the time dimension), and the operation complexity thereof is independent of the frame number, and high efficiency of low-dimensional optimization can be achieved. The proof process of the operator can be referred to as shown in Figure 2 .

[0097] That is, in an optional implementation, eliminating the time dimension in the process of solving the optimization expression based on the spatial basis function by using the conjugate gradient descent algorithm includes:

[0098] The operator A in the conjugate gradient descent algorithm is expressed as:

[0099] A = V H S H F H MFSV + λV H D H DV

[0100] The time basis function V in the operator A is exchanged with the coil sensitivity map S and the two-dimensional discrete Fourier transform operator F, respectively, to obtain:

[0101] A = S H F H V H MVFS + λV H D H DV

[0102] Further, the V HM and V are merged into one low-dimensional operator without time dimension.

[0103] S106, reconstruct the dynamic magnetic resonance image of the target object according to the optimal solution of the spatial basis function and the time basis function.

[0104] After obtaining the optimal solution of the spatial basis function and the time basis function, the high-time-resolution dynamic image of the target object can be obtained through low-rank recovery (X = UV).

[0105] The dynamic magnetic resonance image reconstruction method provided by the embodiment obtains the dynamic magnetic resonance raw data of the target object by using the center-periphery alternating acquisition mode; performs channel compression on the raw data to obtain compressed data; extracts the time basis function, the coil sensitivity map, and the down-sampling data based on the compressed data; constructs an optimization expression of the spatial basis function according to the time basis function, the coil sensitivity map, the down-sampling data, and the down-sampling mask, with the weighted sum of the reconstruction error and the difference between adjacent frames being minimized as the optimization target; eliminates the time dimension in the solving process based on the optimization expression of the spatial basis function by using the conjugate gradient descent algorithm to obtain the optimal solution of the spatial basis function; and reconstructs the dynamic magnetic resonance image of the target object according to the optimal solution of the spatial basis function and the time basis function. Not only the reconstruction speed is improved, but also the reconstruction quality is improved.

[0106] The method provided in the present application is used to perform high-time-resolution dynamic magnetic resonance reconstruction on the heart. The overall flow is shown in Figure 3 , Figure 3 The overall flow of the dynamic magnetic resonance image reconstruction of the heart provided by an embodiment of the present application is shown in Figure 3 Before imaging, the positioning method of the conventional cardiac magnetic resonance scan can be used to complete the positioning of the cardiac magnetic resonance imaging. The reconstruction method provided in the present application can reconstruct a high-quality dynamic image of the heart in a very short time. The reconstruction speed and the reconstruction quality are significantly superior to the current mainstream methods. The comparison of the reconstruction effects of the methods is shown in Figure 4 , Figure 4 The comparison of the reconstruction effects provided by an embodiment of the present application is shown in Figure 4 As shown in the figure, from top to bottom, there are a reference image (Reference), a zero-filling reconstruction method (zero-filling, ZF), a partial separable-least square reconstruction method (Partial Separable-Least Square, PS-LSQR), a low rank plus sparse reconstruction method (low rank plus sparse, L+S), a tensor nuclear norm low rank constraint reconstruction method (Tensor Nuclear Norm, TNN), and the reconstruction method provided in the present application (Proposed). Figure 4The first column is a panoramic view of the reconstruction result; the second, third and fourth columns are enlarged images corresponding to the end-diastolic, systolic and end-systolic time instants of the region of interest marked with a box in the first column; the fifth column is an error map representing the difference between the reconstruction result and the reference image, and the most ideal case is no difference, that is, the error map should be all black (i.e., all zero); the sixth column is a one-dimensional dynamic image drawn along a vertical line through the left ventricle of the heart along the time dimension, which is often used by dynamic magnetic resonance to depict the reconstruction effect in the time dimension, and is generally referred to as an M-mode image; and the seventh column is an error map of the M-mode image, which is also better the lower the value, and the ideal case is all zero.

[0107] The performance of the above methods is statistically analyzed in terms of normalized mean square error (nRMSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and reconstruction time (recon time), and the statistical results are shown in Table 1:

[0108] Table 1

[0109]

[0110] As can be seen from Table 1, the present application has significantly lower error, significantly higher signal-to-noise ratio and structural similarity index, and the reconstruction time is only in the order of seconds, which can better meet the needs of clinical diagnosis.

[0111] In summary, the reconstruction method provided by the present application can effectively solve the problems of long reconstruction time and sensitivity to hyperparameters of the existing method, and unstable reconstruction quality. By fully exploiting the dimension reduction characteristics of the low-rank subspace in the iteration process of the optimization algorithm, the entire iteration algorithm is innovatively converted into a low-dimensional projection space, thereby avoiding the costly calculation process in a high-dimensional space, effectively reducing the computational complexity, and having at least 20 times the acceleration effect compared with the existing reconstruction method, so that the reconstruction result can be obtained within a few seconds, thereby solving the core problem of slow reconstruction speed based on the low-rank subspace. Moreover, the method is less sensitive to hyperparameters, and has relatively stable reconstruction quality under various parameter settings. Therefore, the method provided by the present application has important value for real-time magnetic resonance dynamic imaging with high temporal resolution, and has high potential for arrhythmia imaging and diagnosis.

[0112] Figure 5 A structural schematic diagram of a dynamic magnetic resonance image reconstruction device provided by an embodiment of the present application is shown in FIG. 1. Figure 5 As shown in FIG. 1, the dynamic magnetic resonance image reconstruction device 50 provided by the present embodiment can include an acquisition module 501, a compression module 502, an extraction module 503, an optimization module 504, a dimension reduction module 505 and a reconstruction module 506.

[0113] The acquisition module 501 is configured to acquire dynamic magnetic resonance raw data of the target object by using a center-periphery alternating acquisition mode.

[0114] The compression module 502 is configured to perform channel compression on the raw data to obtain compressed data.

[0115] The extraction module 503 is configured to extract a temporal basis function, a coil sensitivity map, and down-sampling data based on the compressed data.

[0116] The optimization module 504 is configured to construct an optimization expression of a spatial basis function, taking a weighted sum of a reconstruction error and a difference between adjacent frames as an optimization target.

[0117] The dimension reduction module 505 is configured to eliminate a time dimension in a process of solving the optimization expression of the spatial basis function based on a conjugate gradient descent algorithm to obtain an optimal solution of the spatial basis function.

[0118] The reconstruction module 506 is configured to reconstruct a dynamic magnetic resonance image of the target object according to the optimal solution of the spatial basis function and the temporal basis function.

[0119] The device of the embodiment can be used to perform the method of the embodiment. Figure 1 The technical solutions of the method embodiments shown in the drawings have similar implementation principles and technical effects, and will not be described here.

[0120] In an alternative implementation, the acquisition module 501 is configured to acquire dynamic magnetic resonance raw data of the target object by using a center-periphery alternating acquisition mode, and can specifically include the following steps.

[0121] The target object is alternately subjected to center acquisition and periphery acquisition at a preset frequency to obtain center acquisition data for motion navigation, denoted as motion navigation raw data, and periphery acquisition data for image reconstruction, denoted as image reconstruction raw data.

[0122] In an alternative implementation, the compression module 502 is configured to perform channel compression on the raw data to obtain compressed data, and can specifically include the following steps.

[0123] The motion navigation raw data and the image reconstruction raw data are respectively subjected to channel compression by using a geometric decomposition coil channel compression algorithm to obtain motion navigation compressed data and image reconstruction compressed data.

[0124] In an alternative implementation, the extraction module 503 is configured to extract a temporal basis function, a coil sensitivity map, and down-sampling data based on the compressed data, and can specifically include the following steps.

[0125] The motion navigation compressed data is subjected to singular value decomposition to obtain a temporal basis function.

[0126] The image reconstruction compressed data is averaged along the time dimension and full-sampling two-dimensional inverse Fourier reconstruction is performed to obtain a time-averaged multi-channel image, and the image is normalized to obtain a coil sensitivity map;

[0127] Zero padding is performed on the positions not acquired in the image reconstruction compressed data to obtain down-sampling data.

[0128] In an optional implementation, the optimized expression of the spatial basis function is:

[0129]

[0130] wherein M represents a down-sampling mask, S represents a coil sensitivity map, V represents a time basis function, y represents down-sampling data, D represents a first-order difference operator along the time dimension, U represents a spatial basis function, F represents a two-dimensional discrete Fourier transform operator, and λ represents a regularization weight;

[0131] The optimal solution U of the spatial basis function is * :

[0132] U * = (V H S H F H MFSV+λV H D H DV) -1 (V H S H F H y).

[0133] In an optional implementation, the dimension reduction module 505 is configured to eliminate the time dimension in the process of solving the optimized expression of the spatial basis function based on the conjugate gradient descent algorithm, and can specifically include:

[0134] The operator A in the conjugate gradient descent algorithm is represented as:

[0135] A = V H S H F H MFSV+λV H D H DV;

[0136] The time basis function V in the operator A is exchanged in sequence with the coil sensitivity map S and the two-dimensional discrete Fourier transform operator F, respectively, to obtain:

[0137] A = S H F H V H MVFS+λV H D H DV.

[0138] In an alternative embodiment, the dimension reduction module 505 is further configured to combine V H , M and V into one low-dimensional operator without time dimension.

[0139] The embodiment of the present application also provides an electronic device, referring to FIG. 6, Figure 6 The embodiment of the present application is only taken as an example for illustration, and the present application is not limited to this. Figure 6 The embodiment of the present application is only taken as an example for illustration, and the present application is not limited to this. Figure 6 The structure schematic diagram of the electronic device provided by the embodiment of the present application is shown in FIG. 6. As shown in the figure, Figure 6 The electronic device 60 provided by the embodiment can include a memory 601, a processor 602 and a bus 603. The bus 603 is used to realize the connection between the elements.

[0140] The computer program stored in the memory 601 can realize the technical solution of any method embodiment when executed by the processor 602.

[0141] The memory 601 and the processor 602 are directly or indirectly electrically connected to realize the transmission or interaction of data. For example, the elements can be electrically connected through one or more communication buses or signal lines, such as through the bus 603. The memory 601 stores the computer program for realizing the dynamic magnetic resonance image reconstruction method, including at least one software function module stored in the memory 601 in the form of software or firmware. The processor 602 executes the software program and module stored in the memory 601 to perform various functional applications and data processing.

[0142] The memory 601 may be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The memory 601 is used to store programs, and the processor 602 executes the programs after receiving execution instructions. Furthermore, the software programs and modules in the memory 601 may also include an operating system, which may include various software components and / or drivers for managing system tasks (e.g., memory management, storage device control, power management, etc.), and may communicate with various hardware or software components to provide an operating environment for other software components.

[0143] The processor 602 may be an integrated circuit chip with signal processing capabilities. The processor 602 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. The methods, steps, and logic diagrams disclosed in the embodiments of the present invention may be implemented or executed. The general-purpose processor may be a microprocessor or any conventional processor. It is understood that Figure 6 The structure is only for illustration and may also include Figure 6 More or fewer components than shown, or with Figure 6 Different configurations shown. Figure 6 Each component shown in the figure may be implemented using hardware and / or software.

[0144] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. The computer program is executed by a processor to implement the technical solution of any of the above method embodiments.

[0145] The various embodiments in the present disclosure are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0146] The scope of protection of the present disclosure is not limited to the above-described embodiments, and it is obvious to those skilled in the art that various modifications and changes can be made to the present disclosure without departing from the scope and spirit of the present disclosure. If these modifications and changes belong to the scope of the claims of the present disclosure and their equivalent technologies, the intention of the present disclosure also includes these modifications and changes.

Claims

1. A dynamic magnetic resonance image reconstruction method, characterized in that: include: The dynamic magnetic resonance raw data of the target object is acquired using a center-periphery alternating acquisition mode; Performing channel compression on the original data to obtain compressed data; extracting a time basis function, a coil sensitivity map, and downsampled data based on the compressed data; Constructing an optimization expression for a spatial basis function based on the temporal basis function, the coil sensitivity map, the downsampling data, and the downsampling mask, with minimizing a weighted sum of a reconstruction error and a difference between adjacent frames as an optimization goal; Eliminating the time dimension in the process of solving the optimization expression based on the spatial basis function using the conjugate gradient descent algorithm to obtain the optimal solution of the spatial basis function; The dynamic magnetic resonance image of the target object is reconstructed according to the optimal solution of the spatial basis function and the temporal basis function.

2. The method according to claim 1, characterized in that The method of acquiring dynamic magnetic resonance raw data of a target object using a center-periphery alternating acquisition mode includes: The target object is alternately subjected to central acquisition and peripheral acquisition at a preset frequency to obtain central acquisition data for motion navigation, recorded as motion navigation raw data, and peripheral acquisition data for image reconstruction, recorded as image reconstruction raw data.

3. The method according to claim 2, characterized in that The performing channel compression on the original data to obtain compressed data includes: A geometric decomposition coil channel compression algorithm is used to perform channel compression on the motion navigation raw data and the image reconstruction raw data respectively, to obtain motion navigation compressed data and image reconstruction compressed data.

4. The method according to claim 3, characterized in that The extracting of time basis functions, coil sensitivity maps, and downsampled data based on the compressed data includes: Performing singular value decomposition on the motion navigation compressed data to obtain the time basis function; Averaging the image reconstruction compressed data along the time dimension and performing full-sample two-dimensional inverse Fourier reconstruction to obtain a time-averaged multi-channel image and normalizing the image to obtain the coil sensitivity map; Zero padding is performed on uncollected positions in the image reconstruction compressed data to obtain downsampled data.

5. The method according to any one of claims 1 to 4, characterized in that The optimization expression of the spatial basis function is: Where M represents the downsampled mask, S represents the coil sensitivity map, V represents the time basis function, y represents the downsampled data, D represents the first-order difference operator along the time dimension, U represents the spatial basis function, F represents the two-dimensional discrete Fourier transform operator, and λ represents the regularization weight; Then the optimal solution U of the spatial basis function is * for: U * =(V H S H F H MFSV+λV H D H DV) -1 (V H S H F H y)。 6. The method according to claim 5, characterized in that Eliminating the time dimension in the process of solving the optimization expression based on the spatial basis function using the conjugate gradient descent algorithm includes: The operator A in the conjugate gradient descent algorithm is expressed as: A=V H S H F H MFSV+λV H D H DV; The time basis function V in operator A is exchanged with the coil sensitivity map S and the two-dimensional discrete Fourier transform operator F to obtain: A=S H F H V H MVFS+λV H D H DV。 7. The method according to claim 6, characterized in that The method further includes: H , M and V are merged into a low-dimensional operator that does not include the time dimension.

8. A dynamic magnetic resonance image reconstruction device, characterized in that: include: An acquisition module, configured to acquire dynamic magnetic resonance raw data of a target object using a center-periphery alternating acquisition mode; A compression module, configured to perform channel compression on the original data to obtain compressed data; an extraction module, configured to extract a time basis function, a coil sensitivity map, and downsampled data based on the compressed data; an optimization module, configured to construct an optimization expression for a spatial basis function based on the temporal basis function, the coil sensitivity map, the downsampling data, and the downsampling mask, with minimization of a weighted sum of a reconstruction error and a difference between adjacent frames as an optimization objective; A dimensionality reduction module is used to eliminate the time dimension in the process of solving the optimization expression based on the spatial basis function using the conjugate gradient descent algorithm to obtain the optimal solution of the spatial basis function; A reconstruction module is used to reconstruct the dynamic magnetic resonance image of the target object according to the optimal solution of the spatial basis function and the time basis function.

9. An electronic device, characterized in that: include: at least one processor and memory; The memory stores computer-executable instructions; The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs the dynamic magnetic resonance image reconstruction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the dynamic magnetic resonance image reconstruction method according to any one of claims 1 to 7.

Images