Magnetic resonance imaging method, device and system and storage medium

[0029] Example one

[0030] figure 2 This is a schematic flow chart of a magnetic resonance imaging method provided in the first embodiment of the present invention. This embodiment is applicable to the case of performing magnetic resonance imaging based on a neural network. The method can be executed by the magnetic resonance imaging apparatus provided in the embodiments of the present application. Specifically include the following steps:

[0031] S110. Obtain an original model of magnetic resonance imaging, and establish an initial imaging model according to an iterative algorithm used to solve the original model. The iterative algorithm includes at least one of undetermined parameters, undetermined solving operators, and undetermined structural relationships.

[0032] S120. Training the initial imaging model based on the sample data to generate a magnetic resonance imaging model, where the training of the initial imaging model is used to learn the undetermined parameters, undetermined solving operators, and undetermined structural relationships in the iterative algorithm at least one.

[0033] S130: Acquire under-sampled K-space data to be processed, and input the under-sampled K-space data into the magnetic resonance imaging model to generate a magnetic resonance image.

[0034] Among them, the iterative algorithm can be any one of a basic dual algorithm, an alternating direction multiplier algorithm, and an iterative threshold shrinkage algorithm. The initial imaging model is established according to the selected iterative algorithm, and the initial imaging model can include a preset number of iterative modules. , Determine the connection relationship of each iterative module according to the iterative algorithm, connect a preset number of iterative modules to form an initial imaging model. The iteration module may be a calculation module including undetermined factors, or a network module including multiple network layers. The iterative algorithm can include any one of undetermined parameters, undetermined solving operators, and undetermined structural relations, and can also include two or three undetermined factors among undetermined parameters, undetermined solving operators, and undetermined structural relations. Correspondingly, the initial imaging model includes undetermined factors corresponding to the iterative algorithm. Illustratively, if the initial imaging model is a network model, the undetermined factors are replaced based on the undetermined network parameters in the network model. Among them, the undetermined parameter may be a parameter used to process input information in an iterative algorithm, and the undetermined structural relationship may be a calculation relationship between input information in the iterative algorithm or input information processed by the undetermined parameter, such as addition, subtraction, multiplication, and division. The undetermined solving operator may be a calculation function performed on input information or input information processed by undetermined parameters in an iterative algorithm.

[0035] In this embodiment, the initial imaging model is established according to the selected iterative algorithm, and the initial imaging model is trained based on the sample data in the sample set. The sample data may be input into the initial imaging model to obtain the output magnetic resonance imaging model. Resonance image, a loss function is determined according to the standard magnetic resonance image generated from the output magnetic resonance image and the fully sampled K-space data corresponding to the sample data, and the parameters to be learned in the initial imaging model (for example, It can be a network parameter of a network model or a parameter to be determined in a calculation model, etc.) to generate a magnetic resonance imaging model. Wherein, the number of samples may be the under-sampled K-space data collected by multiple detection targets for magnetic resonance detection. The iterative model of the preset level in the initial imaging model to be trained sequentially iteratively process the input sample data, and the output magnetic resonance image of the initial imaging model is generated based on the output magnetic resonance image and the full sampling K corresponding to the sample data The standard magnetic resonance image generated by the spatial data determines the loss function, and the parameters to be learned of the initial imaging model are adjusted backward through the loss function. The next sample data is processed based on the adjusted initial imaging model, and the above training process is repeated until the loss value obtained according to the loss function is less than the less error value, it is determined that the MRI model training is completed. Optionally, the loss function loss is determined according to the following formula: Among them, the Is the output magnetic resonance image of the magnetic resonance imaging model, the x ref A standard magnetic resonance image generated for the fully sampled K-space data corresponding to the sample data.

[0036] It should be noted that the initial imaging model based on an iterative algorithm containing different undetermined factors can be trained based on different sample data. For example, the number of sample data can be different, exemplary, and the undetermined included in the iterative algorithm The greater the number of factors, the greater the number of sample data used to train the initial imaging model established based on the iterative algorithm.

[0037] The technical solution provided by this embodiment establishes an initial imaging model based on an iterative algorithm including at least one of undetermined parameters, undetermined solving operators, and undetermined structural relationships, and learns the undetermined factors in the iterative algorithm through training of the initial imaging model. The prior art improves the degree of freedom of the magnetic resonance imaging model based on fixed solving operators, fixed parameters, and fixed structural relationships. Compared with traditional methods, the magnetic resonance imaging model obtained through learning can improve the quality of the reconstructed magnetic resonance image.

[0038] Optionally, training the initial imaging model based on the sample data to generate a magnetic resonance imaging model includes: processing the first sample data, the initial parameters, and the initial image in the sample set based on the undetermined parameters and the preset structural relationship, Obtain first input information; perform iterative training on the initial imaging model based on the first input information, learn the pending parameters, and generate a first magnetic resonance imaging model for magnetic resonance imaging, wherein the initial imaging model It includes a preset number of iterative modules, which are sequentially connected to perform iterative processing on the first input information, and the iterative module includes a preset solution operator. Wherein, the initial parameter can be 0, and the pixel data in the initial image can be 0. In this embodiment, the parameters to be determined are learned through the initial imaging model, and the iteration module in the initial imaging model includes a fixed structure relationship and a fixed solving operator. The initial imaging model is trained through the above training method, and the first magnetic resonance imaging model that can be used for magnetic resonance imaging is obtained. The above method is based on the traditional iterative algorithm and determines the undetermined parameters through learning, instead of the fixed parameters determined based on empirical values, and improves the applicability of the model. At the same time, because the above training method needs to learn only one parameter to be determined, the amount of training sample data required is small, which is suitable for a small amount of sample data.

[0039] In some embodiments, by adjusting the structure of the initial imaging model, adding undetermined factors in sequence, gradually learning the undetermined parameters, undetermined solving operators, and undetermined structural relationships in the iterative algorithm to establish and create a high-precision MRI model . Optionally, establish an initial imaging model that includes preset structural relationships, undetermined parameters, and undetermined solving operators, and process the second sample data, initial parameters, and initial images in the sample set based on the undetermined parameters and preset structural relationships to obtain the first Two input information; iterative training is performed on the initial imaging model based on the second input information, the pending parameters and the network parameters of the initial imaging model are determined, and a second magnetic resonance imaging model for magnetic resonance imaging is generated. The initial imaging model includes a preset number of iterative sub-network models, the iterative sub-network models are connected in sequence and used to iteratively process the second input information, and the network parameters of the initial imaging model are used to replace the preset Solve operator. Train the initial imaging model through the second sample data, learn undetermined parameters and undetermined solving operators, and generate a second magnetic resonance imaging model. The second magnetic resonance imaging model can be used to perform step S130, the under-sampled K-space data to be processed Process to obtain magnetic resonance images.

[0040] It should be noted that the structure of the first magnetic resonance imaging model may be adjusted, based on a preset number of iterative sub-network models to replace the preset number of iterative modules in the first magnetic resonance imaging model, and the above-mentioned second sample data pair The adjusted first magnetic resonance imaging model is trained to obtain a second magnetic resonance imaging model. Among them, the training may be performed on the basis of the undetermined parameters obtained by the training of the first magnetic resonance imaging model, which is beneficial to fast training of the adjusted first magnetic resonance imaging model and improves training efficiency.

[0041] Optionally, after generating the second magnetic resonance imaging model for magnetic resonance imaging, the method further includes: using the third sample data, initial parameters, and initial images in the unprocessed sample set as the third input information; The third input information trains the second magnetic resonance imaging model, learns the pending structural relationship and the pending solving operator, and generates a third magnetic resonance imaging model for magnetic resonance imaging. On the basis of the second magnetic resonance imaging model, the preset structural relationship in the second magnetic resonance imaging model is replaced based on the pending structural relationship, the second magnetic resonance imaging model is updated, and the updated second magnetic resonance imaging model is updated based on the third sample data. The resonance imaging model is trained, and the undetermined structural relationship and the undetermined solving operator are learned during the training process (the solving operator in the second magnetic resonance imaging model is updated), and the third magnetic resonance imaging model is obtained. Wherein, the second sample data and the third sample data may be the same or different. In this embodiment, on the basis of the second magnetic resonance imaging model obtained through training, the model structure of the second magnetic resonance imaging model is adjusted, new undetermined factors are introduced, and the third magnetic resonance imaging model is obtained after further training. The progressive training method sequentially learns the pending factors in the iterative algorithm, which can reduce the number of samples used in the training process. At the same time, training is performed on the basis of the network parameters in the second MRI model, which improves the training of the third MRI model. effectiveness.

[0042] It should be noted that, on the basis of the first magnetic resonance imaging model, a preset number of iterative sub-network models may be used to replace the preset number of iterative modules in the first magnetic resonance imaging model, where the adjusted first magnetic The resonance imaging model includes undetermined solving factors and undetermined structural relationships. Training the adjusted first magnetic resonance imaging model using the third sample data to obtain a third magnetic resonance imaging model.

[0043] On the basis of the above-mentioned embodiment, the original model of magnetic resonance imaging further includes a sparse transformation algorithm. Accordingly, the method includes: establishing a sub-network model for executing the sparse transformation algorithm, and the sub-network model and the The output terminal of the magnetic resonance imaging model is connected; the sub-network model is trained based on the fourth sample data in the sample set, and the fourth magnetic resonance imaging model is generated based on the trained sub-network model. The magnetic resonance imaging model connected to the sub-network model may be any one of the first magnetic resonance imaging model, the second magnetic resonance imaging model, or the third magnetic resonance imaging model.

[0044] In some embodiments, the imaging model may be: Among them, λ is a regular parameter, and Ψ represents sparse transformation. In traditional magnetic resonance imaging, the fixed-base sparse transformation is used, and the fixed-base sparse transformation cannot completely sparsely represent all the information of the image. In this embodiment, a sub-network model for executing the sparse transformation algorithm is established and connected to the output terminal of the initial imaging model to form a new magnetic resonance imaging model. The sub-network model may be a neuromagnetic resonance imaging model, such as a CNN model, including a convolutional layer and an activation layer of preset layers. The convolution kernel of the convolution layer may be 3×3. The sub-network model may be connected to the output terminal of the first magnetic resonance imaging model, the second magnetic resonance imaging model, or the third magnetic resonance imaging model to form a new magnetic resonance imaging model. The sub-network model in the new magnetic resonance imaging model is trained based on the fourth sample data. Specifically, the fourth sample data is input into the new magnetic resonance imaging model to obtain the output magnetic resonance image, which is based on the output magnetic resonance image and The standard magnetic resonance image generated by the fully sampled K-space data corresponding to the fourth sample data determines the loss function, adjusts the network parameters of the sub-network model, and when the sub-network model training is completed, generates the fourth magnetic resonance imaging Resonance imaging model, where the fourth magnetic resonance imaging model may include the trained sub-network model and the first magnetic resonance imaging model, may include the trained sub-network model and the second magnetic resonance imaging model, or include the completed training The sub-network model and the third MRI model.

[0045] In the foregoing embodiment, in the training process of any magnetic resonance imaging model or sub-network model, a stochastic gradient descent algorithm (stochastic gradient descent, SGD) may be used to adjust the network parameters of the model.

[0046] The technical solution provided in this embodiment replaces the traditional iterative algorithm with a magnetic resonance imaging model, passes the degrees of freedom of the magnetic resonance imaging model, and uses undetermined factors to replace fixed factors one by one, adjusts the magnetic resonance imaging model, and adjusts the magnetic resonance imaging model separately based on sample data After the MRI model is trained, the MRI model for MRI is obtained. The sample data in each training process may overlap, which reduces the requirement for the number of sample data, and the results of the previous training Training on the basis of MRI reduces the difficulty of model training, improves the efficiency of model training and the imaging accuracy of the MRI model.

[0047] Take the basic dual algorithm as an example to introduce the construction method of the initial imaging model: For example, the original model of MRI can be the minimization formula (1):

[0048]

[0049] Among them, x is the magnetic resonance image that needs to be reconstructed, y is the under-sampled K-space data obtained by magnetic resonance scanning, and F u It is the under-picked Fourier transform operator, R(x) is the constraint term.

[0050] The original model of MRI can be solved iteratively through the basic dual algorithm. In traditional MRI, it can be A=F u , Using the following iterative algorithm to solve:

[0051]

[0052] Among them, in the basic dual algorithm formula (2), F * Represents the adjoint function of the function F, x is the magnetic resonance image to be reconstructed, d is the dual parameter, prox is the approximate mapping function, σ, τ and θ are free parameters, n is a positive integer greater than or equal to 0, p n It is the connection factor of dual iteration and basic iteration. However, the structural relationship between the solving operator and the parameter in the above iterative algorithm is artificially set, and its accuracy cannot be guaranteed.

[0053] In this embodiment, the above-mentioned basic dual algorithm is improved, and one or more of the solving operators, free parameters, and structural relationships in the iterative algorithm are learned in the manner of the magnetic resonance imaging model to improve the freedom of the magnetic resonance imaging model To improve the imaging accuracy of the magnetic resonance imaging model and further improve the quality of the magnetic resonance imaging. Optionally, based on undetermined solving operators, free parameters, and structural relations, replace the fixed solving operators, free parameters, and structural relations in formula (2), and learn the aforementioned undetermined factors through a magnetic resonance imaging model. Exemplarily, see basic dual algorithm formula (3) and basic dual algorithm formula (4):

[0054]

[0055]

[0056] Among them, in the above-mentioned basic dual algorithm formula (3) or basic dual algorithm formula (4), Γ is the undetermined dual iterative function, and Λ is the undetermined basic iterative function, where Γ in formula (3) or formula (4) The specific function corresponding to Λ can be different, A=F u , A * Is the adjoint function of function A, σ, τ and θ are undetermined parameters, p n It is the connection factor of dual iteration and basic iteration. It should be noted that the above-mentioned basic dual algorithm formula (3) or basic dual algorithm formula (4) is only an achievable way. The undetermined factors in the algorithm can be set according to user needs to form the corresponding iterative algorithm. limited.

[0057] Optionally, the iterative algorithm is selected according to the number of samples and the accuracy requirements of the magnetic resonance imaging model. Among them, since the basic dual algorithm (3) includes undetermined solving operators and undetermined parameters, the basic dual algorithm (4) includes undetermined solving operators , Undetermined parameters and undetermined structural relationship, so the number of samples required in the training process of the magnetic resonance imaging model established based on the basic dual algorithm (3) is relative to the samples required in the training process of the magnetic resonance imaging model established based on the basic dual algorithm (4) The number is small, and accordingly, the accuracy of the MRI model established and trained based on the basic dual algorithm (3) is lower than that of the MRI model established and trained based on the basic dual algorithm (4).

[0058] Exemplarily, performing an initial imaging model based on the basic duality algorithm (3) may be establishing at least one first sub-model for performing the dual iteration; establishing at least one second sub-model for performing the basic iteration; The association relationship between the dual iteration and the basic iteration determines the associated module of the first sub-magnetic model and the second sub-model; the second iteration is determined according to the iterative relationship between the dual iteration and the basic iteration A connection relationship between a sub-model, the second sub-model, and the association module; the at least one first sub-model, the at least one second sub-model, and the association module are performed according to the connection relationship Connecting to generate the initial imaging model. Among them, the initial imaging model established based on the basic dual algorithm (3) includes a first sub-model and a second sub-model of a preset level. The first sub-model and the second sub-model may both be network models, including multiple network layers, See image 3 , image 3 It is a schematic diagram of an initial imaging model provided by Embodiment 1 of the present invention. in image 3 In, the first sub-magnetic resonance imaging model receives the output information of the first sub-model in the upper level and the output information of the second sub-model processed by the associated module in the upper level, and performs processing on the received information based on the first pending parameter. After processing, it is input to the network layer of the first sub-model. The second sub-model is used to receive the output results of the first sub-model in the same level and the output results of the second sub-model in the previous level, and the received information is based on the first sub-model. After the pending parameters are processed, they are input to the network layer of the second sub-model, and the correlation module is used to process the output information of the second sub-network module and the output information of the previous sub-network module based on the third pending parameters, and process The result is sent to the first sub-model of the next level. Wherein, the first sub-model and the second sub-model of each level perform iterative reconstruction on the input information to generate a magnetic resonance image. The first sub-model and the second sub-model in the first level are used to receive initial input information of the magnetic resonance imaging model.

[0059] Exemplarily, performing an initial imaging model based on the basic dual algorithm (4) may be establishing at least one first sub-model for executing the dual iterative algorithm; establishing at least one second sub-model for executing the basic iterative algorithm Determine the connection relationship between the first sub-model and the second sub-model according to the iterative relationship between the dual iterative algorithm and the basic iterative algorithm; determine the connection relationship between the at least one first sub-model according to the connection relationship The sub-model and the at least one second sub-model are connected to generate the initial imaging model. The initial imaging model established based on the basic dual algorithm (4) includes a first sub-model and a second sub-model of a preset level, and the output end of the first sub-model is respectively connected to the input end and the input end of the second sub-model in the same level. The input end of the first sub-model of the next level is connected, and the output layer of the second network sub-model is respectively connected to the input ends of the first sub-model and the second sub-model in the next level. See Figure 4 , Figure 4 It is a schematic diagram of another initial imaging model provided in the first embodiment of the present invention.

[0060] Among them, the preset level of the initial imaging model may be determined according to the number of iterations. Illustratively, the preset level may be but not limited to 10 or 15 layers, which can be determined according to the accuracy requirements of the magnetic resonance image. The more accurate the magnetic resonance image is Higher, the more levels of the initial imaging model. It should be noted, image 3 with Figure 4 It is only a schematic diagram of the initial imaging model. In other embodiments, the level of the initial imaging model can be set according to user requirements. Optionally, before the initial neural network is established, the number of levels input by the user may be received; or the accuracy requirement of the magnetic resonance image input by the user is received, and the number of levels is determined according to the accuracy requirement of the magnetic resonance image.

[0061] In this embodiment, the basic dual algorithms (3) and (4) both include dual iteration and basic iteration. Accordingly, the initial imaging model both include the first sub-model and the second sub-model. The initial imaging model can be a network model. Correspondingly, the first sub-model and the second sub-model are both sub-network models, which can be composed of at least one convolutional layer, an activation function layer, a pooling layer, etc., for example, the first The sub-model and the second sub-model may include a preset number of convolutional layers for convolution processing on input information.

[0062] Optionally, the first sub-model is a first residual network. In a neural network, the depth of the network is an important factor that affects the effect. The greater the depth of the network, the higher the level of the extracted feature information, which is more conducive to improving the accuracy of the output result. However, as the depth of the neural network increases, gradient dispersion/explosion problems tend to occur, which causes the neural network to fail to converge. The residual network is used to increase the depth of the neural network without changing the expressive power and complexity of the network, and to improve the output accuracy of the neural network. Optionally, the front end and the tail end of the first residual network are jump-connected to directly connect the input and output of each residual block, which is conducive to the extraction and preservation of image details and features, and improves the training process of the neural network. convergence speed. Among them, the first residual network includes a convolutional layer and an activation layer, where there can be three convolutional layers, the activation layer is arranged between two adjacent convolutional base layers, and the convolution kernel of the convolutional layer is 3×3. Exemplarily, in one embodiment, the number of output channels of the three convolutional layers in the first residual network is 32, 32, and 2 in order. Optionally, the first sub-model further includes a first pre-processing layer, connected to the first residual network, and configured to pre-process different types of input parameters received by the first sub-model according to preset rules , Sending the generated first multi-dimensional matrix data to the first residual network. Among them, due to the need to learn the structural relationship between multiple input data, the first preprocessing layer is used to preprocess multiple received data. Optionally, the preprocessing may include stacking multiple received data. Specifically, the received multiple data may be generated according to a preset rule to generate a first multi-dimensional matrix, and the generated first multi-dimensional matrix data may be sent to the convolutional layer in the first residual network for convolution processing. Optionally, the number of channels of the first preprocessing layer is determined according to the input information of the first sub-network module. Specifically, the number of channels of the first preprocessing layer can be twice the number of types of input information, and they are used to process the input respectively. The real and imaginary data of the information. For example, in this embodiment, the number of channels in the first pretreatment layer may be 4.

[0063] Optionally, the second sub-model includes a second residual network and a second pre-processing layer, and the second pre-processing layer is connected to the second residual network and is used for receiving different types of the second sub-model The input parameters of is preprocessed according to preset rules, and the generated second multi-dimensional matrix data is sent to the second residual network. The front end and the tail end of the second sub-model are jump-connected, including a convolutional layer and an activation layer. The convolution kernel of the convolutional layer is 3×3. Exemplarily, in one embodiment, the number of output channels of the three convolutional layers in the first residual network is 32, 32, and 2, respectively, and the second preprocessing layer preprocesses the received input information, such as preprocessing The data can be stacked to generate the second multi-dimensional matrix data. In this embodiment, the number of channels of the second preprocessing layer is determined according to the input information of the second sub-network module. For example, the number of channels of the second preprocessing layer can be It is 2, so I won't repeat it here.

[0064] Exemplary, see Figure 5-Figure 10 ,among them, Figure 5 Is a schematic diagram of the first sub-model in the initial imaging model established based on the basic dual algorithm (3) provided by an embodiment of the present invention, Image 6 It is a schematic diagram of the second sub-model and the associated module in the initial imaging model established based on the basic dual algorithm (3) provided by an embodiment of the present invention; Figure 7 Is a schematic diagram of the first sub-model in the initial imaging model established based on the basic dual algorithm (4) provided by an embodiment of the present invention; Picture 8 It is a schematic diagram of the second sub-model in the initial imaging model established based on the basic dual algorithm (4) provided by an embodiment of the present invention.

[0065] Optionally, each convolutional layer in the first residual network and the second residual network includes a real channel and an imaginary channel, and the real channel is used to input the convolutional layer The real part data of the information is subjected to convolution processing, and the imaginary part channel is used to perform convolution processing on the imaginary part data of the input information of the convolution layer. Since the magnetic resonance signal is a complex signal, the neural network model cannot directly process the complex data. Therefore, the magnetic resonance signal is represented as real part data and imaginary part data. Based on the characteristics of the magnetic resonance signal, the first residual network and the second residual Each convolutional layer in the difference network includes a real channel and an imaginary channel. Optionally, the first preprocessing layer and the second preprocessing layer may also extract the real part data and the imaginary part data in the magnetic resonance signal in the input information after receiving the input information, respectively Partial data is preprocessed. Correspondingly, based on the combination of the output result of the real channel and the output result of the imaginary channel, the output magnetic resonance image of the magnetic resonance imaging model can be determined.

[0066] Example two

[0067] Picture 9 It is a schematic structural diagram of a magnetic resonance imaging apparatus provided in the second embodiment of the present invention, and the magnetic resonance imaging apparatus includes:

[0068] The initial imaging model establishment module 210 is used to obtain the original model of magnetic resonance imaging, and establish the initial imaging model according to the iterative algorithm used to solve the original model. The iterative algorithm includes undetermined parameters, undetermined solving operators, and undetermined structural relationships At least one of

[0069] The model training module 220 is configured to train the initial imaging model based on sample data to generate a magnetic resonance imaging model, wherein the training of the initial imaging model is used to learn the undetermined parameters and undetermined solving operators in the iterative algorithm At least one of the relationship with the pending structure;

[0070] The magnetic resonance imaging module 230 is configured to obtain under-sampled K-space data to be processed, and input the under-sampled K-space data into the magnetic resonance imaging model to generate a magnetic resonance image.

[0071] Optionally, the model training module 220 includes:

[0072] The first input information determining unit is configured to process the first sample data, the initial dual parameters and the initial image in the sample set based on the undetermined parameters and the preset structural relationship to obtain the first input information;

[0073] The first magnetic resonance imaging model generating unit is configured to iteratively train the initial imaging model based on the first input information, learn the pending parameters, and generate a first magnetic resonance imaging model for magnetic resonance imaging, wherein: The initial imaging model includes a preset number of iterative modules, which are sequentially connected to perform iterative processing on the first input information, and the iterative module includes a preset solution operator.

[0074] Optionally, the model training module 220 includes:

[0075] The second input information determining unit is configured to process the second sample data, the initial dual parameters and the initial image in the sample set based on the undetermined parameters and the preset structural relationship to obtain the second input information;

[0076] The second magnetic resonance imaging model generating unit is configured to perform iterative training on the initial imaging model based on the second input information, learn the pending parameters and the network parameters of the initial imaging model, and generate a magnetic resonance imaging model The second magnetic resonance imaging model, wherein the initial imaging model includes a preset number of iterative sub-network models, and the iterative sub-network models are sequentially connected for iterative processing on the second input information.

[0077] Optionally, the model training module 220 further includes:

[0078] The third input information determining unit is configured to use the third sample data, the initial dual parameter and the initial image in the sample set as the third input information after the second magnetic resonance imaging model for magnetic resonance imaging is generated;

[0079] The third magnetic resonance imaging model generating unit is configured to train the second magnetic resonance imaging model based on the third input information, and learn the pending structural relationship and the network parameters in the second magnetic resonance imaging model, Generate a third magnetic resonance imaging model for magnetic resonance imaging.

[0080] Optionally, the original model of magnetic resonance imaging further includes a sparse transformation algorithm;

[0081] The device includes:

[0082] A third sub-magnetic resonance imaging model establishment module, configured to establish a third sub-magnetic resonance imaging model for executing the sparse transformation algorithm, the third sub-magnetic resonance imaging model being connected to the output end of the initial imaging model;

[0083] The fourth magnetic resonance imaging model generation module is configured to train the third sub-magnetic resonance imaging model based on the third sample data in the sample set, and generate the third sub-magnetic resonance imaging model based on the trained third sub-magnetic resonance imaging model. The fourth magnetic resonance imaging model for imaging.

[0084] Optionally, the model training module 220:

[0085] The loss function determining unit is configured to input the sample data into the initial imaging model to obtain an output magnetic resonance image of the magnetic resonance imaging model, and according to the output magnetic resonance image and the fully sampled K-space corresponding to the sample data The standard magnetic resonance image generated from the data determines the loss function, where the loss function loss is determined according to the following formula:

[0086] Among them, the Is the output magnetic resonance image of the magnetic resonance imaging model, the x ref A standard magnetic resonance image generated for the fully sampled K-space data of the sample.

[0087] The magnetic resonance imaging model adjustment unit is configured to adjust the network parameters in the initial imaging model according to the loss function to generate a magnetic resonance imaging model.

[0088] Optionally, the iterative algorithm includes a basic dual algorithm, an alternating direction multiplier algorithm, and an iterative threshold shrinkage algorithm.

[0089] The magnetic resonance imaging apparatus provided in the embodiment of the present application can execute the magnetic resonance imaging method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects for executing the magnetic resonance imaging method.

[0090] Example three

[0091] Picture 10 Is a schematic structural diagram of a magnetic resonance system provided in the third embodiment of the present invention, Picture 10 Showing a block diagram of an exemplary medical imaging system suitable for implementing embodiments of the present invention, Picture 10 The medical imaging system shown is only an example, and should not bring any limitation to the function and application scope of the embodiment of the present invention.

[0092] The magnetic resonance system includes a magnetic resonance device 300 and a computer 400.

[0093] The computer 400 may be used to implement specific methods and apparatuses disclosed in some embodiments of the present invention. The specific device in this embodiment uses a functional block diagram to show a hardware platform including a display module. In some embodiments, the computer 400 may implement specific implementations of some embodiments of the present invention through its hardware devices, software programs, firmware, and combinations thereof. In some embodiments, the computer 400 may be a general purpose computer or a special purpose computer.

[0094] Such as Picture 10 As shown, the computer 400 may include an internal communication bus 401, a processor 402, a read only memory (ROM) 403, a random access memory (RAM) 404, a communication port 405, an input/output component 406, a hard disk 407, and User interface 408. The internal communication bus 401 can implement data communication between the components of the computer 400. The processor 402 can make a judgment and issue a prompt. In some embodiments, the processor 402 may consist of one or more processors. The communication port 405 can implement data communication between the computer 400 and other components (not shown in the figure), such as external devices, image acquisition devices, databases, external storage, and image processing workstations. In some embodiments, the computer 400 can send and receive information and data from the network through the communication port 405. The input/output component 406 supports the input/output data flow between the computer 400 and other components. The user interface 408 can implement interaction and information exchange between the computer 400 and the user. The computer 400 may also include different forms of program storage units and data storage units, such as a hard disk 407, a read only memory (ROM) 403, and a random access memory (RAM) 404, which can store various data for computer processing and/or communication. Files, and possible program instructions executed by the processor 402.

[0095] The processor may be used to execute a magnetic resonance imaging method when executing a program, and the method includes:

[0096] Acquiring an original model of magnetic resonance imaging, and establishing an initial imaging model according to an iterative algorithm for solving the original model, the iterative algorithm including at least one of undetermined parameters, undetermined solving operators, and undetermined structural relationships;

[0097] The initial imaging model is trained based on the sample data to generate a magnetic resonance imaging model, wherein the training of the initial imaging model is used to learn at least one of undetermined parameters, undetermined solving operators, and undetermined structural relationships in the iterative algorithm ;

[0098] Obtain under-sampled K-space data to be processed, and input the under-sampled K-space data into the magnetic resonance imaging model to generate a magnetic resonance image.

[0099] Although the present invention has been disclosed as above in preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can use the methods and technical content disclosed above to improve the present invention without departing from the spirit and scope of the present invention. The technical solution makes possible changes and modifications. Therefore, all simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention belong to the technical solution of the present invention. protected range.

[0100] At the same time, this application uses specific words to describe the embodiments of the application. For example, "one embodiment", "an embodiment", and/or "some embodiments" mean a certain feature, structure, or characteristic related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that "an embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more in different positions in this specification does not necessarily refer to the same embodiment. . In addition, some features, structures, or characteristics in one or more embodiments of the present application can be appropriately combined.

[0101] In addition, those skilled in the art can understand that various aspects of this application can be illustrated and described through a number of patentable categories or situations, including any new and useful process, machine, product, or combination of substances, or for them Any new and useful improvements. Correspondingly, various aspects of the present application can be completely executed by hardware, can be completely executed by software (including firmware, resident software, microcode, etc.), or can be executed by a combination of hardware and software. The above hardware or software can be called "data block", "module", "submodule", "engine", "unit", "subunit", "component" or "system". In addition, various aspects of this application may be embodied as a computer product located in one or more computer-readable media, and the product includes computer-readable program codes.