Magnetic resonance image reconstruction method, device, equipment, medium and product

Abstract

The invention discloses a magnetic resonance image reconstruction method, device and equipment, a medium and a product. The method comprises the following steps: determining a coil sensitivity mapping image set according to k space data to be reconstructed corresponding to a magnetic resonance image set to be reconstructed; according to the coil sensitivity mapping image set, performing zero-filling reconstruction on the to-be-reconstructed k space data to obtain an initial magnetic resonance reconstruction image set; determining a target magnetic resonance reconstruction image set corresponding to the magnetic resonance image set to be reconstructed according to the initial magnetic resonance reconstruction image set, the coil sensitivity mapping image set and a pre-trained target image reconstruction model; wherein the target image reconstruction model comprises a space-time convolutional network and a data consistency layer, the space-time convolutional network comprises a space-time deformation convolutional layer, and the embodiment of the invention improves the reconstruction effect of the magnetic resonance image.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging technology, and in particular to a method, apparatus, device, medium and product for reconstructing magnetic resonance images. Background Technology

[0002] Cardiac cine magnetic resonance imaging (CMRI) is a non-invasive medical imaging technique used to observe and evaluate the structure and function of the heart. It provides detailed information about cardiac anatomy, such as heart size, shape, ventricular and atrium volumes, and valvular movement. Furthermore, it can be used to assess cardiac function parameters, such as systolic function (ejection fraction), diastolic function, and ventricular myocardial motility abnormalities. Cardiac cMRI is crucial for diagnosing and assessing cardiac lesions and developing personalized treatment plans.

[0003] In clinical practice, cardiac cine MRI scans are lengthy, increasing the risk of image artifacts caused by cardiac motion, necessitating MRI image reconstruction. However, traditional image reconstruction methods are ill-suited for irregular anatomical structures and struggle to model non-rigid motions and deformations present in the images. Summary of the Invention

[0004] This invention provides a method, apparatus, device, medium, and product for reconstructing magnetic resonance images, which solves the problem that traditional image reconstruction methods cannot adapt to irregular anatomical structures, enabling the image reconstruction model to model non-rigid motion and deformation, thereby improving the reconstruction effect of magnetic resonance images.

[0005] According to an embodiment of the present invention, a method for reconstructing a magnetic resonance image is provided, the method comprising:

[0006] Based on the k-space data to be reconstructed corresponding to the set of magnetic resonance images to be reconstructed, determine the coil sensitivity mapping set;

[0007] Based on the coil sensitivity mapping atlas, zero-filling reconstruction is performed on the k-space data to be reconstructed to obtain an initial magnetic resonance reconstructed image set;

[0008] Based on the initial magnetic resonance reconstructed image set, the coil sensitivity mapping set, and the pre-trained target image reconstruction model, the target magnetic resonance reconstructed image set corresponding to the magnetic resonance image set to be reconstructed is determined.

[0009] The target image reconstruction model includes a spatiotemporal convolutional network and a data consistency layer, wherein the spatiotemporal convolutional network contains a spatiotemporal deformation convolutional layer.

[0010] According to another embodiment of the present invention, a magnetic resonance image reconstruction apparatus is provided, the apparatus comprising:

[0011] The coil sensitivity mapping set determination module is used to determine the coil sensitivity mapping set based on the k-space data to be reconstructed corresponding to the magnetic resonance image set to be reconstructed.

[0012] The initial magnetic resonance reconstruction image set determination module is used to perform zero-filling reconstruction on the k-space data to be reconstructed according to the coil sensitivity mapping atlas to obtain the initial magnetic resonance reconstruction image set;

[0013] The target magnetic resonance reconstruction image set determination module is used to determine the target magnetic resonance reconstruction image set corresponding to the magnetic resonance image set to be reconstructed based on the initial magnetic resonance reconstruction image set, the coil sensitivity mapping map set, and the pre-trained target image reconstruction model.

[0014] The target image reconstruction model includes a spatiotemporal convolutional network and a data consistency layer, wherein the spatiotemporal convolutional network contains a spatiotemporal deformation convolutional layer.

[0015] According to another embodiment of the present invention, an electronic device is provided, the electronic device comprising:

[0016] At least one processor; and

[0017] A memory communicatively connected to the at least one processor; wherein,

[0018] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the magnetic resonance image reconstruction method according to any embodiment of the present invention.

[0019] According to another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the magnetic resonance image reconstruction method according to any embodiment of the present invention.

[0020] According to another embodiment of the present invention, a computer program product is provided, including a computer program that, when executed by a processor, implements the magnetic resonance image reconstruction method according to any embodiment of the present invention.

[0021] The technical solution of this invention sets up a data consistency layer and a spatiotemporal convolutional network with a spatiotemporal deformation convolutional layer in the image reconstruction model. The data consistency layer combined with the coil sensitivity mapping atlas can reduce imaging artifacts in the target magnetic resonance reconstruction image. The spatiotemporal deformation convolutional layer solves the problem that traditional image reconstruction methods cannot adapt to irregular anatomical structures, enabling the image reconstruction model to model non-rigid motion and deformation, thereby improving the reconstruction effect of magnetic resonance images.

[0022] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart illustrating a method for reconstructing a magnetic resonance image according to an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram illustrating the training process of a target image reconstruction model provided in one embodiment of the present invention;

[0026] Figure 3 A model architecture diagram of a target image reconstruction model provided in one embodiment of the present invention;

[0027] Figure 4 This is a schematic diagram of a spatial deformation convolution kernel provided in one embodiment of the present invention;

[0028] Figure 5 This is a schematic diagram illustrating the principle of a spatial deformation convolution unit provided in one embodiment of the present invention;

[0029] Figure 6 This is a schematic diagram of a time-deformed convolution kernel provided in one embodiment of the present invention;

[0030] Figure 7 A flowchart illustrating another method for reconstructing magnetic resonance images according to an embodiment of the present invention;

[0031] Figure 8 This is a schematic diagram of a spatially rotated convolution kernel provided in one embodiment of the present invention;

[0032] Figure 9 This is a schematic diagram of the architecture of a rotating prior fusion layer provided in one embodiment of the present invention;

[0033] Figure 10 A model architecture diagram of a target image reconstruction model provided in one embodiment of the present invention;

[0034] Figure 11 A comparative schematic diagram of a magnetic resonance reconstructed image provided in one embodiment of the present invention;

[0035] Figure 12 This is a schematic diagram of the structure of a magnetic resonance image reconstruction device provided in one embodiment of the present invention;

[0036] Figure 13 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0037] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0038] It should be noted that the terms "first," "second," "initial," "target," "reference," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0039] Currently, commonly used fast magnetic resonance imaging (MRI) techniques include parallel imaging and compressed sensing. Parallel imaging accelerates imaging by leveraging the correlation of multi-channel coils, while compressed sensing accelerates it by reducing k-space sampling points and utilizing the sparsity of the imaged object. However, due to hardware limitations, the acceleration capability of parallel imaging is limited, while compressed sensing has a longer image reconstruction time and involves complex selection of sparse transforms and reconstruction parameters. In recent years, the use of Convolutional Neural Networks (CNNs) for cardiac cinema MRI image reconstruction has shown great potential. CNNs extract image features from MRI images through multi-layer convolution and pooling operations, learn the optimal parameters for reconstruction from a large amount of training data, or directly learn the mapping relationship from undersampled k-space data to the complete MRI image, thereby achieving better image quality and higher speedup than the aforementioned parallel imaging or compressed sensing techniques.

[0040] However, traditional CNN networks employ a fixed geometric structure, primarily manifested in: 1) convolutional units sampling the input image at fixed locations; 2) pooling layers reducing spatial resolution by a fixed ratio; and 3) pooling layers segmenting the region of interest (RoI) into fixed spatial units. Therefore, traditional CNN networks have significant limitations when learning cardiac motion: they cannot adapt to irregular cardiac anatomy and struggle to model the heart's non-rigid movements and deformations (including a series of contractions, relaxations, and blood pumping). Both of these issues lead to a decrease in the reconstruction accuracy of magnetic resonance images.

[0041] Figure 1 This is a flowchart illustrating a method for reconstructing a magnetic resonance image according to an embodiment of the present invention. This embodiment is applicable to the reconstruction of magnetic resonance images. The method can be executed by a magnetic resonance image reconstruction device, which can be implemented in hardware and / or software and can be configured in a terminal device. Figure 1 As shown, the method includes:

[0042] S110. Determine the coil sensitivity mapping set based on the k-space data to be reconstructed corresponding to the magnetic resonance image set to be reconstructed.

[0043] Specifically, the set of MRI images to be reconstructed contains at least two MRI images to be reconstructed, which are undersampled MRI images. For example, the MRI images to be reconstructed can be cardiac cine MRI images or chest MRI images, etc. The image type of the MRI images to be reconstructed is not limited here, and can be customized according to actual needs.

[0044] For example, the data format of the magnetic resonance image set to be reconstructed can be represented as [X, Y, T, C], where X represents the image width of the magnetic resonance image to be reconstructed, Y represents the image height of the magnetic resonance image to be reconstructed, T represents the number of image frames of the magnetic resonance image to be reconstructed in the magnetic resonance image set, and C represents the number of channels.

[0045] Specifically, the k-space data to be reconstructed represents the undersampled k-space data corresponding to the set of magnetic resonance images to be reconstructed. For example, the k-space data to be reconstructed can be acquired using an RF coil array, or it can be obtained by performing a Fourier transform on each magnetic resonance image to be reconstructed in the set of magnetic resonance images to be reconstructed. The method of acquiring the k-space data to be reconstructed is not limited here, and can be customized according to actual needs.

[0046] For example, the data format of the k-space data to be reconstructed can be represented as [k x k y [T, C], where k x k y The frequency domain scale of the k-space data to be reconstructed is represented by T, the number of image frames of the magnetic resonance images to be reconstructed in the set of magnetic resonance images to be reconstructed is represented by C, and the number of channels is represented by C.

[0047] In one optional embodiment, determining the coil sensitivity mapping set based on the k-space data to be reconstructed corresponding to the magnetic resonance image set to be reconstructed includes: inputting the k-space data to be reconstructed corresponding to the magnetic resonance image set to be reconstructed into a pre-trained sensitivity estimation model to obtain the output coil sensitivity mapping set.

[0048] Specifically, the coil sensitivity map atlas contains at least two coil sensitivity maps, which represent a visualization of the relative weights of each coil across the spatial dimensions of the image, flexibly representing the overlapping anatomical structures acquired in the field of view. For example, the data format of the coil sensitivity map atlas can be represented as [X,Y,C,M], where M represents the number of coil sensitivity maps in the atlas.

[0049] Sensitivity Encoding (SENSE) is a reconstruction technique based on parallel imaging in the image domain. Its core idea is to use the different sensitivities of coils to magnetic resonance signals at different spatial locations to dealias undersampled magnetic resonance images.

[0050] Generalized autocalibarating partially parallel acquisition (grappa) is a K-space-based reconstruction algorithm. Its principle is that each data point in the k-space of a coil can be represented by a linear combination of the data points adjacent to that data point in all coils, thereby achieving the synthesis of missing data points.

[0051] The iterative self-consistent parallel imaging reconstruction using eigenvector maps (espirit) is a reconstruction method that combines SENSE and grappa techniques.

[0052] In one optional embodiment, determining the coil sensitivity mapping set based on the k-space data to be reconstructed corresponding to the set of magnetic resonance images to be reconstructed includes: determining a calibration matrix based on the k-space data to be reconstructed corresponding to the set of magnetic resonance images to be reconstructed; determining the ESPIRiT reconstruction operator based on the singular vectors corresponding to the calibration matrix; and performing eigenvalue decomposition on the ESPIRiT reconstruction operator to obtain the coil sensitivity mapping set.

[0053] In one optional embodiment, determining the calibration matrix based on the k-space data to be reconstructed corresponding to the set of magnetic resonance images to be reconstructed includes: obtaining the current k-space data corresponding to each magnetic resonance image to be reconstructed in the k-space data to be reconstructed; calibrating the current k-space data according to the automatic calibration data of the preset calibration region in the current k-space data to obtain the current matrix; and constructing the calibration matrix based on at least two current matrices.

[0054] In one optional embodiment, the preset calibration region is the central region in k-space. Specifically, the automatic calibration data is obtained by performing an additional scan on the ACS (Auto Calibration Signal) rows in the preset calibration region during the formal scan.

[0055] Each block within the preset calibration region corresponds to a row in the current matrix, and the columns of the current matrix represent shifted versions of the preset calibration region.

[0056] In another optional embodiment, determining the calibration matrix based on the k-space data to be reconstructed corresponding to the set of magnetic resonance images to be reconstructed includes: performing time averaging on the k-space data to be reconstructed corresponding to the set of magnetic resonance images to be reconstructed to obtain average k-space data; and performing calibration processing on the k-space data to be reconstructed based on the average calibration data of a preset calibration region in the average k-space data to obtain the calibration matrix.

[0057] Specifically, the average calibration data represents the average result of the automatic calibration data of the preset calibration regions corresponding to the multiple magnetic resonance images to be reconstructed in the k-space data to be reconstructed.

[0058] Specifically, the number of eigenvectors calculated by eigenvalue decomposition represents the number of coil sensitivity maps in the coil sensitivity map set. Ideally, there is only a single eigenvector corresponding to an absolute eigenvalue of "1" at each location, and all other eigenvalues ​​are much smaller than 1. However, in practical applications, the ESPIRiT reconstruction operator can be decomposed into multiple eigenvectors corresponding to an absolute eigenvalue of "1" or additional eigenvalues ​​smaller than but close to "1".

[0059] The number of coil sensitivity maps is a preset hyperparameter. In an optional embodiment, the number of coil sensitivity map sets is two.

[0060] S120. Based on the coil sensitivity mapping atlas, zero-fill reconstruction is performed on the k-space data to be reconstructed to obtain the initial magnetic resonance reconstruction image set.

[0061] The magnetic resonance imaging process can be simulated as a linear system, which can be expressed as:

[0062] y = PFSx

[0063] Where x represents the fully sampled magnetic resonance image set or the fully sampled k-space data, P represents the k-space sampling operator, F represents the discrete Fourier transform, S represents the coil sensitivity operator, and y represents the magnetic resonance image set or the k-space data to be reconstructed.

[0064] In this embodiment, the coil sensitivity operator S can be replaced by the coil sensitivity mapping set E, that is, the initial magnetic resonance reconstruction image set is obtained by solving y = PFEx using the least squares method.

[0065] Specifically, the initial magnetic resonance reconstructed image set contains a sequence of initial magnetic resonance reconstructed images corresponding to each magnetic resonance image to be reconstructed. The number of initial magnetic resonance reconstructed images in the initial magnetic resonance reconstructed image sequence is the same as the number of coil sensitivity maps in the coil sensitivity map set. For example, the data format of the initial magnetic resonance reconstructed image set can be represented as [X, Y, T, M].

[0066] S130. Based on the initial magnetic resonance reconstructed image set, the coil sensitivity mapping set, and the pre-trained target image reconstruction model, determine the target magnetic resonance reconstructed image set corresponding to the magnetic resonance image set to be reconstructed.

[0067] In this embodiment, the target image reconstruction model includes a spatiotemporal convolutional network and a data consistency layer, wherein the spatiotemporal convolutional network contains a spatiotemporal deformation convolutional layer.

[0068] Among them, the Spatiotemporal Deformable Convolutional Network (Deformable CNN) is a neural network with a sampled grid that can freely deform. It adapts to the shape and deformation of objects by learning the bias of the convolutional kernel from the input feature map.

[0069] In one optional embodiment, the spatiotemporal deformable convolutional layer includes a three-dimensional deformable convolution module and a nonlinear activation module. For example, the size of the spatiotemporal convolution kernel in the spatiotemporal deformable convolutional network is 3×3×1.

[0070] In one optional embodiment, the target magnetic resonance reconstructed image set corresponding to the magnetic resonance image set to be reconstructed is determined based on the initial magnetic resonance reconstructed image set, the coil sensitivity mapping set, and the pre-trained target image reconstruction model. This includes: performing deformation convolution processing on the initial magnetic resonance reconstructed image set through a spatiotemporal deformation convolution layer to obtain a spatiotemporal magnetic resonance image set; and performing data consistency processing on the spatiotemporal magnetic resonance image set based on the coil sensitivity mapping set through a data consistency layer to obtain the target magnetic resonance reconstructed image set.

[0071] Since the initial MRI reconstructed image amplifies noise and residual aliasing artifacts, prior information from the MRI image is incorporated into the ESPIRiT reconstruction through regularization. MRI image reconstruction itself is a nonlinear inverse problem, which can be expressed as:

[0072]

[0073] in, Let A represent the magnetic resonance reconstructed image, A represent the encoding matrix, which may include k-space sampling operators, discrete Fourier transform, and coil sensitivity mapping set, x represent the fully sampled magnetic resonance image set or fully sampled k-space data, y represent the k-space data to be reconstructed, and R represent sparse transform, such as wavelet transform and total variation.

[0074] Specifically, the data consistency layer uses a coil sensitivity mapping atlas to project back and forth between the k-space and the image domain, thereby enhancing the consistency between the k-space data to be reconstructed and the intermediate iterative outputs of the spatiotemporal convolutional network.

[0075] For example, when the sparse transformation R is a suitable convex function, an approximate gradient descent algorithm can be used to solve the optimization problem of the above formula through alternating updates and iterations:

[0076] x k+1 =x k -2tA H (y-Axk )

[0077] Where t is the step size of the approximate gradient descent algorithm.

[0078] Figure 2 This is a schematic diagram illustrating the training process of a target image reconstruction model according to an embodiment of the present invention. Specifically, the training k-space data corresponding to the training magnetic resonance image set is time-averaged to obtain average k-space data. Based on the average calibration data corresponding to the average k-space data, the training k-space data is calibrated and its eigenvalues ​​are decomposed to obtain a coil sensitivity mapping set. Based on the coil sensitivity mapping set, zero-padding reconstruction is performed on the training k-space data to obtain an initial magnetic resonance reconstructed image set. The initial magnetic resonance reconstructed image set and the coil sensitivity mapping set are input into the untrained initial image reconstruction model to obtain the output predicted magnetic resonance reconstructed image set.

[0079] Specifically, during the training of the target image reconstruction model, the spatiotemporal convolutional network and the data consistency layer undergo a certain number of iterations, such as 10 or 20. The number of iterations is not limited here and can be customized according to actual needs. In each iteration, two iterative update steps—the data consistency layer and the spatiotemporal convolutional network—alternate. The prior information in each iteration is enforced by the spatiotemporal convolutional network, performing supervised end-to-end training so that the predicted magnetic resonance reconstructed image set output by the initial image reconstruction model gradually approximates the fully sampled magnetic resonance image.

[0080] In an optional embodiment, the loss function L of the target image reconstruction model is expressed as:

[0081]

[0082] Where, N train It is the number of training magnetic resonance images. It is the predicted magnetic resonance reconstructed image set of the i-th training magnetic resonance image set in the k-th iteration, x i This represents the fully sampled magnetic resonance image set corresponding to the i-th training magnetic resonance image set.

[0083] The technical solution of this embodiment sets a data consistency layer and a spatiotemporal convolutional network with a spatiotemporal deformation convolutional layer in the image reconstruction model. The data consistency layer combined with the coil sensitivity mapping atlas can reduce imaging artifacts in the target magnetic resonance reconstruction image. The spatiotemporal deformation convolutional layer solves the problem that traditional image reconstruction methods cannot adapt to irregular anatomical structures, enabling the image reconstruction model to model non-rigid motion and deformation, thereby improving the reconstruction effect of magnetic resonance images.

[0084] Based on the above embodiments, optionally, the spatiotemporal deformation convolutional layer includes at least one spatial deformation convolutional module and at least one temporal deformation convolutional module connected in series. The spatial deformation convolutional module includes spatial deformation convolutional units and nonlinear activation units, and the temporal deformation convolutional module includes temporal deformation convolutional units and nonlinear activation units.

[0085] Specifically, the number of spatial deformation convolutional modules can be the same as or different from the number of temporal deformation convolutional modules. For example, there can be two spatial deformation convolutional modules and one temporal deformation convolutional module.

[0086] Specifically, at least one spatial deformation convolutional module and at least one temporal deformation convolutional module can be connected sequentially in any order. Taking the example above, the arrangement order can be spatial deformation convolutional module A, spatial deformation convolutional module B, and temporal deformation convolutional module A; alternatively, it can be spatial deformation convolutional module A, temporal deformation convolutional module A, and spatial deformation convolutional module B; or it can be temporal deformation convolutional module A, spatial deformation convolutional module A, and spatial deformation convolutional module B. The arrangement order of the spatial and temporal deformation convolutional modules in the spatiotemporal deformation convolutional layer is not limited here; it can be customized according to actual needs.

[0087] Figure 3 This is a model architecture diagram of a target image reconstruction model provided in one embodiment of the present invention. Specifically, in the spatiotemporal convolutional network of the target image reconstruction model, two (2+1)D deformation convolutional layers are set, and each (2+1)D deformation convolutional layer contains a spatial deformation convolutional module and a temporal deformation convolutional module.

[0088] Spatial graph convolution operation consists of the following two steps: 1) using a square grid 1) Sample the input feature map x; 2) Use a square grid The weights w in the algorithm are used to sum all the sampled values ​​in turn to obtain the output feature map y.

[0089] Specifically, square grid This can be called a spatial graph convolution kernel, which defines the size and expansion of the receptive field. For example, a spatial graph convolution kernel of size 3×3 with weights of 1 can be defined:

[0090] For each sampling point p0 on the output feature map y, its value can be expressed in the following form:

[0091]

[0092] Among them, pn enumerates square grids. All sampling locations.

[0093] In spatial deformation convolution operations, square grids A bias Δpn was added, where, Therefore, each sampling point p0 on the output feature map y can be represented in the following form:

[0094]

[0095] Due to square grid An offset Δpn was added, and its sampling position changed from pn to p. n +Δp n Therefore, sampling is performed on an irregular grid. This was performed on top. However, due to the bias Δp n The values ​​are usually non-integer values, and the biased sampling position may not fall on the sampling position of the input feature map x. Therefore, bilinear interpolation is needed to obtain the optimized sampling position.

[0096] For example, the bilinear interpolation method can be expressed as:

[0097]

[0098] Where p represents an arbitrary sampling position that may contain a score value after bias, q enumerates all sampling positions in the input feature map x, and G represents the bilinear interpolation matrix.

[0099] G can be decomposed into two one-dimensional matrices, as shown below:

[0100] G(q,p)=g(q x p x )·g(q y ,p y )

[0101] Where g(a, b) = max(0, 1-|ab|).

[0102] Specifically, spatial graphical convolution kernels with learned bias are called spatial deformation convolution kernels. Figure 4 This is a schematic diagram of a spatial deformation convolution kernel provided in one embodiment of the present invention. Specifically, Figure 4 The nine dots in the middle left image form a 3×3 spatial graphic convolution kernel. Figure 4 The nine blue dots in the middle right figure represent the learned bias Δp. n A 3×3 spatial graphical convolution kernel, with arrows connecting the learning bias Δp. n Sampling positions before and after.

[0103] Figure 5This is a schematic diagram illustrating the principle of a spatial deformation convolutional unit according to an embodiment of the present invention. Specifically, the input feature map has N channels, and "3×3" represents the size of the spatial graphic convolutional kernel, used to learn the bias field from the input feature map. Since each sampling position is represented by (x, y), when the input feature map has N channels, the bias field has 2N channels. The bias field represents the representation of the set of x and y coordinates; for example, the first N channels are used to record the x-coordinate of the bias, and the last N channels are used to record the y-coordinate of the bias.

[0104] The sampling grid in the spatial deformation convolution module possesses the ability to deform freely, enabling more accurate and adaptive fitting of object structures, especially irregular structures such as curved or uneven surfaces. During the training of the target image reconstruction model, the spatial graph convolution kernel and the bias field are learned simultaneously. To learn the bias field, the gradient is backpropagated through the calculation process of the aforementioned bilinear interpolation method.

[0105] Specifically, the time-varying convolution kernel that learns the bias is called the time-varying convolution kernel. Figure 6 This is a schematic diagram of a time-deformable convolution kernel provided in one embodiment of the present invention. Specifically, Figure 6 In the coordinate system, the horizontal axis represents time, and the vertical axis represents the pixel values ​​of the input feature map. Figure 6 The image above shows a sampling diagram of a 3×1×1 temporal convolutional kernel, which indicates that the temporal convolutional kernel samples the input feature map at the same and fixed sampling interval. Figure 6 The image below shows a sampling diagram of a 3×1×1 temporal deformation convolution kernel, demonstrating that the sampling interval of the temporal deformation convolution kernel can change according to the pixel information of the input feature map.

[0106] The technical solution of this embodiment decomposes the spatiotemporal deformation convolutional layer into a spatial deformation convolutional module and a temporal deformation convolutional module. The spatial deformation convolutional module can more accurately and adaptively fit object structures, especially irregular structures such as curved or uneven surfaces. The temporal deformation convolutional module can sample from non-grid pixels of the input feature map. Compared to three-dimensional spatiotemporal deformation convolutional kernels, the (2+1)D deformation convolutional kernel results in fewer model parameters, easier training, and higher reconstruction accuracy.

[0107] Figure 7 This is a flowchart illustrating another magnetic resonance image reconstruction method provided in an embodiment of the present invention. This embodiment further refines the model architecture of the "target image reconstruction model" in the above embodiment. Specifically, the target image reconstruction model includes a spatiotemporal convolutional network and a data consistency layer. The spatiotemporal convolutional network includes a spatiotemporal deformation convolutional layer, a rotation prior fusion layer, and an output layer. For example... Figure 7 As shown, the method includes:

[0108] S210. Determine the coil sensitivity mapping set based on the k-space data to be reconstructed corresponding to the magnetic resonance image set to be reconstructed.

[0109] S220. Based on the coil sensitivity mapping atlas, zero-fill reconstruction is performed on the k-space data to be reconstructed to obtain the initial magnetic resonance reconstruction image set.

[0110] In this embodiment, S210-S220 are the same as those in the above embodiments. Figure 1 The S110-S120 shown are the same or similar, and will not be described again in this embodiment.

[0111] S230. The initial magnetic resonance reconstructed image set is processed by deformation convolution through a spatiotemporal deformation convolution layer to obtain a convolutional magnetic resonance image set.

[0112] In one optional embodiment, the spatiotemporal deformable convolutional layer includes a three-dimensional deformable convolution module and a nonlinear activation module. In another optional embodiment, the spatiotemporal deformable convolutional layer includes at least one spatial deformable convolution module and at least one temporal deformable convolution module connected in series.

[0113] S240. By rotating the prior fusion layer, a prior fusion process is performed on the convolutional magnetic resonance image set according to at least two spatially rotated convolution kernels to obtain a fused magnetic resonance image set.

[0114] In this embodiment, each spatial rotation convolution kernel is obtained by rotating a shared spatial deformation convolution kernel, and each spatial rotation convolution kernel has a different rotation angle.

[0115] Since cardiac motion is a complex, quasi-circular, non-rigid scaling motion, it can be decomposed into multiple local scaling motions that are identical in multiple directions. Specifically, by rotating the angle, more possibilities can be derived from the bias learned by the shared spatial deformation convolution kernel, and the sharing of the spatial deformation convolution kernel means that adding a rotation prior fusion layer will not increase the number of model parameters excessively.

[0116] Specifically, the definition of the shared spatial deformation convolution kernel in the rotation prior fusion layer is the same as or similar to the definition of the spatial deformation convolution kernel in the above embodiment, and will not be repeated here.

[0117] In one optional embodiment, the rotation prior fusion layer includes a rotation pooling module and at least two rotation deformation convolution modules. The rotation prior fusion layer performs prior fusion processing on the convolutional magnetic resonance image set based on at least two spatial rotation convolution kernels to obtain a fused magnetic resonance image set. This includes: using the rotation deformation convolution module to rotate a shared spatial deformation convolution kernel according to a rotation angle to obtain a spatial rotation convolution kernel, and performing deformation convolution processing on the convolutional magnetic resonance image set based on the spatial rotation convolution kernel to obtain a rotated magnetic resonance image set; and using the rotation pooling module to perform rotation pooling processing on the rotated magnetic resonance image sets corresponding to the at least two rotation deformation convolution modules respectively to obtain the fused magnetic resonance image set.

[0118] Specifically, the number of rotational deformation convolutional modules is the same as the number of rotation angles, or in other words, the number of rotational deformation convolutional modules is the same as the number of spatial rotational convolutional kernels.

[0119] In one optional embodiment, the multiple rotation angles are preset. In another optional embodiment, the angular interval between the sorted multiple rotation angles is determined based on the number of rotational deformation convolution modules. Specifically, the angular interval = 360° / N, where N represents the number of rotational deformation convolution modules.

[0120] Figure 8 This is a schematic diagram of a spatially rotated convolution kernel provided in one embodiment of the present invention. Figure 8 Taking the number of rotational deformation convolutional modules N=4 as an example, the rotation angles corresponding to the four rotational deformation convolutional modules are 0°, 90°, 180° and 270° respectively. Figure 8 The top left image shows the convolutional magnetic resonance image output by the spatiotemporal deformation convolutional layer. The red, green, blue, and pink arrows represent the rotation directions corresponding to the four rotation angles mentioned above.

[0121] Figure 8 The image below shows the four spatial rotation convolution kernels obtained by rotating the shared spatial deformation convolution kernel using the four rotation angles mentioned above. The four colors represent the four different spatial rotation convolution kernels.

[0122] In one optional embodiment, the rotational deformation convolution module includes rotational deformation convolution units and a nonlinear activation layer. For example, the nonlinear activation layer can be a SiLU activation layer. The advantage of including nonlinear activation units is that it can further improve the expressive power of the target image reconstruction model.

[0123] Figure 9This is a schematic diagram of the architecture of a rotational prior fusion layer provided in one embodiment of the present invention. Specifically, the rotational prior fusion layer includes a rotational pooling module and at least two rotational deformation convolutional modules. The input feature map is a convolutional magnetic resonance image F from the convolutional magnetic resonance image set output by the spatiotemporal deformation convolutional layer. The spatial rotational convolutional kernel group includes a shared spatial deformation convolutional kernel and spatial rotational convolutional kernels obtained at four separate rotation angles. The rotating magnetic resonance image sequence includes rotating magnetic resonance images output by each of the four rotating deformation convolution modules.

[0124] Specifically, the rotating magnetic resonance image set contains rotating magnetic resonance images corresponding to each magnetic resonance image to be reconstructed, and the fused magnetic resonance image set contains fused magnetic resonance images corresponding to each magnetic resonance image to be reconstructed.

[0125] In one optional embodiment, the rotation pooling layer is an average pooling layer. Specifically, the average pooling layer is used to obtain a rotation magnetic resonance image corresponding to the magnetic resonance image to be reconstructed from each set of rotation magnetic resonance images for each magnetic resonance image to be reconstructed, and to perform average pooling processing on at least two rotation magnetic resonance images to obtain a fused magnetic resonance image corresponding to the magnetic resonance image to be reconstructed.

[0126] In another optional embodiment, the rotation pooling layer is a max pooling layer. Specifically, the max pooling layer is used to obtain a rotation magnetic resonance image corresponding to the magnetic resonance image to be reconstructed from each set of rotation magnetic resonance images for each magnetic resonance image to be reconstructed, and to perform max pooling on at least two rotation magnetic resonance images to obtain at least two feature values; the rotation magnetic resonance image corresponding to the largest feature value is used as the fused magnetic resonance image corresponding to the magnetic resonance image to be reconstructed.

[0127] S250: Through the output layer, output a spatiotemporal magnetic resonance image set based on the fused magnetic resonance image set.

[0128] In one alternative embodiment, the output layer includes a nonlinear activation module and a spatiotemporal graph convolution module. Figure 10 This is a model architecture diagram of a target image reconstruction model provided in one embodiment of the present invention. Specifically, the target image reconstruction model includes a spatiotemporal convolutional network and a data consistency layer. The spatiotemporal convolutional network includes a spatiotemporal deformation convolutional layer, a rotation prior fusion layer, and an output layer. The output layer includes a nonlinear activation module and a spatiotemporal graph convolutional module.

[0129] In one optional embodiment, the spatiotemporal graphics convolution module is a three-dimensional graphics convolution module, or the spatiotemporal graphics convolution module is a series of spatial graphics convolution units and temporal conventional convolution units.

[0130] S260. Through the data consistency layer, based on the coil sensitivity mapping atlas, the spatiotemporal magnetic resonance image set is processed to obtain the target magnetic resonance reconstruction image set.

[0131] In this embodiment, S260 is the same as or similar to the data consistency layer in the above embodiments, and will not be described again in this embodiment.

[0132] Setting a spatiotemporal deformation convolutional layer in the target image reconstruction model requires the target image reconstruction model to additionally learn the bias of the deformation convolution kernel during training, thereby increasing the training parameters of the target image reconstruction model and reducing the training efficiency of the model.

[0133] The technical solution of this embodiment adds the rotation prior, which is common in cardiac operation, to the model framework by setting a rotation prior fusion layer and an output layer in the spatiotemporal convolutional network. Specifically, the rotation prior fusion layer performs prior fusion processing on the convolutional magnetic resonance image set output by the spatiotemporal deformation convolutional layer based on at least two spatial rotation convolutional kernels to obtain a fused magnetic resonance image set. The output layer outputs the spatiotemporal magnetic resonance image set based on the fused magnetic resonance image set. Each spatial rotation convolutional kernel is obtained by rotating a shared spatial deformation convolutional kernel, and each spatial rotation convolutional kernel has a different rotation angle. This solves the problem of low training efficiency of the target image reconstruction model and further improves the reconstruction effect of magnetic resonance images.

[0134] The spatiotemporal convolutional network provided in the above embodiments of the present invention is named the Rotational-prior-fused Spatiotemporal Deformable Convolutional Neural Network (RSD-CNN).

[0135] The magnetic resonance imaging reconstruction method provided in the above embodiments of the present invention, along with three other cardiac cine magnetic resonance imaging reconstruction methods, were subjected to retrospective undersampling experiments on a 3T magnetic resonance imaging system with speedups of 12x, 16x, and 20x. The three cardiac cine magnetic resonance imaging reconstruction methods are a deep learning network-based reconstruction method (MoDL), a deep learning and convolutional neural network-based reconstruction method (R2plus 1D), and a deep learning and rotationally equivariant network-based reconstruction method (SRE-CNN). Specifically, the Adam optimizer was used to update the network parameters, the network learning rate was set to 0.001, the batch size was set to 1, and the number of training iterations was 30. All experiments were performed on a single NVIDIA Tesla V100 GPU and implemented using the PyTorch framework.

[0136] Table 1 shows the experimental results of the test metrics for the four reconstruction methods. The test metrics include Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM).

[0137]

[0138]

[0139] As can be seen from Table 1, the magnetic resonance image reconstruction method provided by the embodiments of the present invention improves the peak signal-to-noise ratio and structural similarity of the magnetic resonance reconstructed image compared with the above three traditional reconstruction methods.

[0140] Figure 11 This is a comparative schematic diagram of a magnetic resonance reconstruction image provided in an embodiment of the present invention. Specifically, the first set of image sequences, from top to bottom, consists of a fully sampled magnetic resonance image under 16x acceleration, a k-space mask image, a yt image corresponding to the fully sampled magnetic resonance image, and a yt image corresponding to the k-space mask image. The second set of image sequences, from top to bottom, consists of a magnetic resonance reconstruction image under 16x acceleration using the MoDL method, a k-space mask image, a yt image corresponding to the magnetic resonance reconstruction image, and a yt image corresponding to the k-space mask image. The third set of image sequences, from top to bottom, consists of a magnetic resonance reconstruction image under 16x acceleration using the R2plus1D method, a k-space mask image, and a yt image corresponding to the k-space mask image. The first set of images consists of three image sequences: a magnetic resonance reconstructed image (YT) corresponding to the magnetic resonance reconstructed image and a YT image corresponding to the k-space mask image. The second set of images consists of three image sequences: a magnetic resonance reconstructed image (YT) corresponding to the magnetic resonance reconstructed image and a YT image corresponding to the k-space mask image, obtained by using the magnetic resonance image method provided in this embodiment of the invention; ...

[0141] Among them, the k-space mask image represents the difference image between the magnetic resonance reconstructed image and its corresponding fully sampled magnetic resonance image, and the yt image represents the image formed by taking all y and t values ​​along a line along the x-axis for the fully sampled magnetic resonance image, the magnetic resonance reconstructed image, or the k-space mask image, and is used to represent the reconstruction effect along the time direction.

[0142] from Figure 11As can be seen, the magnetic resonance image reconstruction method provided by the embodiments of the present invention has a better representation of the detailed features of the heart compared with the above three traditional reconstruction methods, the k-space mask image shows less difference, and the image quality of the magnetic resonance reconstructed image shown by the yt image is also higher.

[0143] The above results demonstrate that the magnetic resonance image reconstruction method provided by the embodiments of the present invention can more accurately and adaptively fit the complex non-rigid motion and deformation in cardiac motion.

[0144] The following are embodiments of the magnetic resonance image reconstruction apparatus provided in this invention. This apparatus and the magnetic resonance image reconstruction method described above belong to the same inventive concept. For details not described in detail in the embodiments of the magnetic resonance image reconstruction apparatus, please refer to the content of the magnetic resonance image reconstruction method in the above embodiments.

[0145] Figure 12 This is a schematic diagram of a magnetic resonance image reconstruction apparatus provided in one embodiment of the present invention. Figure 12 As shown, the device includes: a coil sensitivity mapping set determination module 310, an initial magnetic resonance reconstruction image set determination module 320, and a target magnetic resonance reconstruction image set determination module 330.

[0146] The coil sensitivity mapping set determination module 310 is used to determine the coil sensitivity mapping set based on the k-space data to be reconstructed corresponding to the magnetic resonance image set to be reconstructed.

[0147] The initial magnetic resonance reconstruction image set determination module 320 is used to perform zero-filling reconstruction on the k-space data to be reconstructed based on the coil sensitivity mapping atlas to obtain the initial magnetic resonance reconstruction image set.

[0148] The target magnetic resonance reconstruction image set determination module 330 is used to determine the target magnetic resonance reconstruction image set corresponding to the magnetic resonance image set to be reconstructed based on the initial magnetic resonance reconstruction image set, the coil sensitivity mapping map set, and the pre-trained target image reconstruction model.

[0149] The target image reconstruction model includes a spatiotemporal convolutional network and a data consistency layer. The spatiotemporal convolutional network contains a spatiotemporal deformation convolutional layer.

[0150] The technical solution of this embodiment sets a data consistency layer and a spatiotemporal convolutional network with a spatiotemporal deformation convolutional layer in the image reconstruction model. The data consistency layer combined with the coil sensitivity mapping atlas can reduce imaging artifacts in the target magnetic resonance reconstruction image. The spatiotemporal deformation convolutional layer solves the problem that traditional image reconstruction methods cannot adapt to irregular anatomical structures, enabling the image reconstruction model to model non-rigid motion and deformation, thereby improving the reconstruction effect of magnetic resonance images.

[0151] In an optional embodiment, the spatiotemporal convolutional network further includes a rotation prior fusion layer and an output layer;

[0152] Target magnetic resonance reconstruction image set determination module 330 includes:

[0153] The convolutional magnetic resonance image set determination unit is used to perform deformation convolution processing on the initial magnetic resonance reconstructed image set through a spatiotemporal deformation convolution layer to obtain a convolutional magnetic resonance image set.

[0154] The fused magnetic resonance image set determination unit is used to perform prior fusion processing on the convolutional magnetic resonance image set by rotating the prior fusion layer according to at least two spatial rotation convolution kernels to obtain the fused magnetic resonance image set; wherein, each spatial rotation convolution kernel is obtained by rotating a shared spatial deformation convolution kernel, and each spatial rotation convolution kernel has a different rotation angle.

[0155] The spatiotemporal magnetic resonance image set determination unit is used to output a spatiotemporal magnetic resonance image set through the output layer based on the fused magnetic resonance image set;

[0156] The target magnetic resonance reconstruction image set determination unit is used to perform data consistency processing on the spatiotemporal magnetic resonance image set through the data consistency layer based on the coil sensitivity mapping atlas to obtain the target magnetic resonance reconstruction image set.

[0157] In one alternative embodiment, the rotating prior fusion layer includes a rotating pooling module and at least two rotating deformable convolutional modules;

[0158] The fused magnetic resonance image set determination unit is specifically used for:

[0159] The rotation deformation convolution module rotates the shared spatial deformation convolution kernel according to the rotation angle to obtain the spatial rotation convolution kernel, and performs deformation convolution processing on the convolutional magnetic resonance image set according to the spatial rotation convolution kernel to obtain the rotational magnetic resonance image set.

[0160] By using a rotation pooling module, rotation pooling is performed on the rotation magnetic resonance image sets corresponding to at least two rotation deformation convolution modules to obtain a fused magnetic resonance image set.

[0161] In one optional embodiment, the spatiotemporal deformation convolutional layer includes at least one spatial deformation convolutional module and at least one temporal deformation convolutional module connected in series. The spatial deformation convolutional module includes spatial deformation convolutional units and nonlinear activation units, and the temporal deformation convolutional module includes temporal deformation convolutional units and nonlinear activation units.

[0162] In an optional embodiment, the coil sensitivity mapping atlas determination module 310 includes:

[0163] The calibration matrix determination unit is used to determine the calibration matrix based on the k-space data to be reconstructed corresponding to the magnetic resonance image set to be reconstructed.

[0164] The coil sensitivity mapping set determination unit is used to determine the ESPIRiT reconstruction operator based on the singular vectors corresponding to the calibration matrix, and to perform eigenvalue decomposition on the ESPIRiT reconstruction operator to obtain the coil sensitivity mapping set.

[0165] In one optional embodiment, the calibration matrix determination unit is specifically used for:

[0166] The average k-space data corresponding to the set of magnetic resonance images to be reconstructed is calculated by time averaging to obtain the average k-space data.

[0167] The calibration matrix is ​​obtained by calibrating the average k-space data based on the average calibration data of the preset calibration region in the average k-space data.

[0168] The magnetic resonance image reconstruction apparatus provided in this embodiment of the invention can execute the magnetic resonance image reconstruction method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of executing the method.

[0169] Figure 13 This is a schematic diagram of an electronic device provided according to one embodiment of the present invention. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0170] like Figure 13As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor 11. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0171] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information or data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0172] Processor 11 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the magnetic resonance image reconstruction method provided in the above embodiments.

[0173] In some embodiments, the magnetic resonance image reconstruction method provided in the above embodiments can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the magnetic resonance image reconstruction method described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the magnetic resonance image reconstruction method by any other suitable means (e.g., by means of firmware).

[0174] Various embodiments of the systems and techniques described above herein can be implemented in the following systems or combinations thereof: digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0175] Computer programs for implementing the magnetic resonance image reconstruction method of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0176] In the context of this application, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium can be a machine-readable storage medium. Examples of machine-readable storage media include, based on an electrical connection of at least one wire, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0177] To provide interaction with a user, the systems and techniques described herein can be implemented on a terminal device having: a display device for displaying information to the user (e.g., a cathode-ray tube (CRT) or liquid crystal display (LCD) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the terminal device. Other types of devices can also provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0178] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0179] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. It addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0180] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0181] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for reconstructing magnetic resonance images, characterized in that, include: Based on the k-space data to be reconstructed corresponding to the set of magnetic resonance images to be reconstructed, determine the coil sensitivity mapping set; Based on the coil sensitivity mapping atlas, zero-filling reconstruction is performed on the k-space data to be reconstructed to obtain an initial magnetic resonance reconstructed image set; Based on the initial magnetic resonance reconstructed image set, the coil sensitivity mapping set, and the pre-trained target image reconstruction model, the target magnetic resonance reconstructed image set corresponding to the magnetic resonance image set to be reconstructed is determined. The target image reconstruction model includes a spatiotemporal convolutional network and a data consistency layer, wherein the spatiotemporal convolutional network contains a spatiotemporal deformation convolutional layer.

2. The method according to claim 1, characterized in that, The spatiotemporal convolutional network also includes a rotation prior fusion layer and an output layer; The step of determining the target magnetic resonance reconstructed image set corresponding to the magnetic resonance image set to be reconstructed based on the initial magnetic resonance reconstructed image set, the coil sensitivity mapping set, and the pre-trained target image reconstruction model includes: The initial magnetic resonance reconstructed image set is processed by deformation convolution through the spatiotemporal deformation convolution layer to obtain a convolutional magnetic resonance image set; Through the rotation prior fusion layer, the convolutional magnetic resonance image set is subjected to prior fusion processing based on at least two spatial rotation convolution kernels to obtain a fused magnetic resonance image set; wherein, each spatial rotation convolution kernel is obtained by rotating a shared spatial deformation convolution kernel, and each spatial rotation convolution kernel has a different rotation angle. Through the output layer, a spatiotemporal magnetic resonance image set is output based on the fused magnetic resonance image set; The target magnetic resonance reconstructed image set is obtained by performing data consistency processing on the spatiotemporal magnetic resonance image set through the data consistency layer based on the coil sensitivity mapping atlas.

3. The method according to claim 2, characterized in that, The rotating prior fusion layer includes a rotating pooling module and at least two rotating deformable convolution modules; The step of performing prior fusion processing on the convolutional magnetic resonance image set using the rotation prior fusion layer, based on at least two spatially rotated convolutional kernels, to obtain a fused magnetic resonance image set includes: The rotation deformation convolution module rotates the shared spatial deformation convolution kernel according to the rotation angle to obtain a spatial rotation convolution kernel, and performs deformation convolution processing on the convolutional magnetic resonance image set according to the spatial rotation convolution kernel to obtain a rotational magnetic resonance image set. The rotation pooling module performs rotation pooling processing on the rotation magnetic resonance image sets corresponding to at least two rotation deformation convolution modules to obtain a fused magnetic resonance image set.

4. The method according to any one of claims 1-3, characterized in that, The spatiotemporal deformation convolutional layer includes at least one spatial deformation convolutional module and at least one temporal deformation convolutional module connected in series. The spatial deformation convolutional module includes spatial deformation convolutional units and nonlinear activation units, and the temporal deformation convolutional module includes temporal deformation convolutional units and nonlinear activation units.

5. The method according to claim 1, characterized in that, The step of determining the coil sensitivity mapping set based on the k-space data to be reconstructed corresponding to the magnetic resonance image set to be reconstructed includes: The calibration matrix is ​​determined based on the k-space data to be reconstructed corresponding to the set of magnetic resonance images to be reconstructed; Based on the singular vectors corresponding to the calibration matrix, the ESPIRiT reconstruction operator is determined, and the ESPIRiT reconstruction operator is decomposed into eigenvalues ​​to obtain the coil sensitivity mapping set.

6. The method according to claim 5, characterized in that, The step of determining the calibration matrix based on the k-space data to be reconstructed corresponding to the set of magnetic resonance images to be reconstructed includes: The average k-space data corresponding to the set of magnetic resonance images to be reconstructed is calculated by time averaging to obtain the average k-space data. The calibration matrix is ​​obtained by calibrating the average k-space data based on the average calibration data of the preset calibration region in the average k-space data.

7. A magnetic resonance image reconstruction apparatus, characterized in that, include: The coil sensitivity mapping set determination module is used to determine the coil sensitivity mapping set based on the k-space data to be reconstructed corresponding to the magnetic resonance image set to be reconstructed. The initial magnetic resonance reconstruction image set determination module is used to perform zero-filling reconstruction on the k-space data to be reconstructed according to the coil sensitivity mapping atlas to obtain the initial magnetic resonance reconstruction image set; The target magnetic resonance reconstruction image set determination module is used to determine the target magnetic resonance reconstruction image set corresponding to the magnetic resonance image set to be reconstructed based on the initial magnetic resonance reconstruction image set, the coil sensitivity mapping map set, and the pre-trained target image reconstruction model. The target image reconstruction model includes a spatiotemporal convolutional network and a data consistency layer, wherein the spatiotemporal convolutional network contains a spatiotemporal deformation convolutional layer.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method for reconstructing magnetic resonance images according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method for reconstructing a magnetic resonance image according to any one of claims 1-6.

10. A computer program product comprising a computer program that, when executed by a processor, implements a method for reconstructing magnetic resonance images according to any one of claims 1-6.

Images