Magnetic resonance image reconstruction method and magnetic resonance image reconstruction apparatus

Abstract

To improve stability of image reconstruction.SOLUTION: A magnetic resonance image reconstruction method according to an embodiment includes an image space regularization step, a deviation correction step, a data consistency processing step, and an output step. The image space regularization step generates second image data by performing regularization processing in an image space using a first neural network on first image data generated based on the undersampled K-space data. In the deviation correction step, third image data is generated by correcting the second image data so that a pixel value statistical feature of the second image data approaches the pixel value statistical feature of the first image data. In the data consistency processing step, data consistency processing is performed on the third image data so that the K-space data corresponding to the third image data approaches the undersampled K-space data to generate fourth image data.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The embodiments disclosed in this specification and the drawings relate to a magnetic resonance image reconstruction method and a magnetic resonance image reconstruction apparatus. [Background technology]

[0002] Magnetic resonance imaging (MRI) is a non-invasive medical imaging technique that utilizes the magnetic resonance phenomenon, in which hydrogen nuclei placed in a static magnetic field resonate with a radio-frequency magnetic field of a specific frequency. MRI has the advantages of high resolution, no trauma, and no radiation, and can examine various solid organs in the human body, making it widely used in the diagnosis of clinical diseases.

[0003] In magnetic resonance imaging, a subject (patient) is scanned to acquire K-space data for generating a magnetic resonance image, and the magnetic resonance image is reconstructed based on the K-space data. However, acquiring all of the K-space data requires a long scan time, which is likely to cause discomfort to the patient and increase the likelihood of motion artifacts. Therefore, in magnetic resonance imaging, a portion of the K-space data is typically acquired by undersampling, and the magnetic resonance image is reconstructed based on the undersampled K-space data to shorten the scan time.

[0004] When reconstructing a magnetic resonance image based on undersampled K-space data, there is insufficient information for reconstruction, making it impossible to reconstruct the ground truth of the magnetic resonance image. Therefore, an estimated result of the ground truth of the magnetic resonance image is reconstructed. In other words, reconstructing a magnetic resonance image based on undersampled K-space data is an ill-posed problem, and an accurate magnetic resonance image cannot be uniquely reconstructed.

[0005] A conventional magnetic resonance image reconstruction method is known in which a magnetic resonance image is reconstructed based on undersampled K-space data using a compressed sensing algorithm. In this magnetic resonance image reconstruction method, a constraint is imposed on the reconstruction to uniquely determine an appropriate solution.

[0006] In recent years, with the development of machine learning technology, magnetic resonance image reconstruction methods using neural networks have been proposed. Neural networks can be used to impose appropriate constraints on magnetic resonance image reconstruction by performing regularization processing in image space or K-space. One proposed magnetic resonance image reconstruction technique using neural networks is an end-to-end unrolled reconstruction network that reconstructs magnetic resonance images based on undersampled K-space data. The advantages of this technique are high quality reconstructed magnetic resonance images and easy training. Here, the reconstruction network is a large network consisting of multiple serially connected reconstruction modules, each of which includes a neural network.

[0007] For example, one known example proposes a magnetic resonance image reconstruction method using an end-to-end unrolled reconstruction network in which a reconstruction module included in the reconstruction network is provided with an image space regularization block that performs image space regularization processing using a neural network and a data consistency block that performs data consistency processing. In this known example, the input data or output data of the image space regularization block is not processed. Another known example also proposes a magnetic resonance image reconstruction method using an end-to-end unrolled reconstruction network in which a reconstruction module included in the reconstruction network is provided with an image space regularization block and a data consistency block. In this known example, normalization processing is performed on input data of the image space regularization block, and un-normalization processing is performed on output data of the image space regularization block.

[0008] In magnetic resonance image reconstruction using the above-mentioned end-to-end unrolled reconstruction network, when the neural networks included in the reconstruction network do not share parameters, the degree of freedom of the neural network parameters is higher than when the parameters are shared, and the accuracy of the reconstructed magnetic resonance image is often higher. On the other hand, when the neural networks included in the reconstruction network do not share parameters, the neural networks are more likely to output unintended results, which may result in the reconstruction of a magnetic resonance image that is completely deviated from the ground truth, resulting in a problem of low stability of the magnetic resonance image reconstruction.

[0009] Furthermore, the reconstruction networks of the above-mentioned known examples all have low image reconstruction stability, and there is a problem in that a magnetic resonance image that is completely deviated from the ground truth may be reconstructed. [Prior art documents] [Non-patent literature]

[0010] [Non-Patent Document 1] MoDL: Model Based Deep Learning Architecture for Inverse Problems, Aggarwal HK, Mani MP, Jacob M. IEEE Trans Med Imaging. 2019;38(2):394-405. doi:10.1109 / TMI.2018.2865356 [Non-patent document 2] End-to-End Variational Networks for Accelerated MRI ReconStruction, Anuroop Sriram, Jure Zbontar, Tullie Murrell, Aaron Defazio, C. Lawrence Zitnick, Nafissa Yakubova, Florian Knoll, & Patricia Johnson. (2020). arXiv:2004.06688 Summary of the Invention [Problem to be solved by the invention]

[0011] One of the problems to be solved by the embodiments disclosed in this specification and the drawings is to improve the stability of image reconstruction. However, the problems to be solved by the embodiments disclosed in this specification and the drawings are not limited to the above problem. Problems corresponding to the effects of each configuration shown in the embodiments described below can also be positioned as other problems. [Means for solving the problem]

[0012] A magnetic resonance image reconstruction method according to an embodiment is a magnetic resonance image reconstruction method for reconstructing magnetic resonance image data based on undersampled K-space data, and includes an image space regularization step, a deviation correction step, a data consistency processing step, and an output step. The image space regularization step performs regularization processing in image space using a first neural network on first image data generated based on the undersampled K-space data to generate second image data. The deviation correction step corrects the second image data so that pixel value statistical features of the second image data approach the pixel value statistical features of the first image data to generate third image data. The data consistency processing step performs data consistency processing on the third image data so that K-space data corresponding to the third image data approach the undersampled K-space data to generate fourth image data. The output step outputs image data based on the fourth image data as the reconstructed magnetic resonance image data. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a diagram showing an example of the configuration of a magnetic resonance image reconstruction apparatus according to the first embodiment. [Figure 2] FIG. 2 is a flowchart showing the flow of the magnetic resonance image reconstruction method according to the first embodiment. [Figure 3] FIG. 3 is a data flow diagram for explaining the processing of step S102 of the magnetic resonance image reconstruction method according to the first embodiment. [Figure 4] FIG. 4 is a data flow diagram for explaining the processing of steps S104 to S106 of the magnetic resonance image reconstruction method according to the first embodiment. [Figure 5] FIG. 5 is a diagram showing an example of the configuration of a magnetic resonance image reconstruction apparatus according to the second embodiment. [Figure 6] FIG. 6 is a flowchart showing the flow of the magnetic resonance image reconstruction method according to the second embodiment. [Figure 7]FIG. 7 is a diagram illustrating an example of the configuration of a magnetic resonance image reconstruction apparatus according to a comparative example. [Figure 8] FIG. 8 is a flowchart showing the flow of a magnetic resonance image reconstruction method according to a comparative example. [Figure 9] FIG. 9 is a diagram for explaining examples of regularized image data and corrected image data generated in each correction process of the magnetic resonance image reconstruction method according to the comparative example. [Figure 10] FIG. 10 is a diagram for explaining an example of regularized image data and corrected image data generated in each correction process of the magnetic resonance image reconstruction method according to the embodiment. [Figure 11] FIG. 11 is a diagram comparing a magnetic resonance image reconstructed by the magnetic resonance image reconstruction method according to the embodiment with a magnetic resonance image reconstructed by a comparative example. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, a magnetic resonance image reconstruction method and a magnetic resonance image reconstruction apparatus according to an embodiment will be described with reference to the drawings.

[0015] (First embodiment) A magnetic resonance image reconstruction device according to the first embodiment performs image reconstruction of image space image data based on undersampled K-space data obtained by scanning a subject. The purpose is to estimate ground truth image data corresponding to fully sampled K-space data and reconstruct magnetic resonance image data similar to the ground truth image data. The K-space data is obtained by a magnetic resonance imaging device transmitting pulse signals to a subject in a frequency-encoded and phase-encoded magnetic field and receiving echo signals due to specific nuclear magnetic resonance with multiple receiving coils.

[0016] FIG. 1 is a diagram showing an example of the configuration of a magnetic resonance image reconstruction apparatus 1 according to the first embodiment. The configuration of the magnetic resonance image reconstruction apparatus 1 according to the first embodiment will be described with reference to FIG. 1. The magnetic resonance image reconstruction apparatus 1 according to the first embodiment includes an input / output interface 10, a display interface 20, a communication interface 30, a storage unit 40, a pre-processing unit 50, an image space regularization unit 60, a deviation correction unit 70, a data consistency processing unit 80, and an output unit 90. The input / output interface 10, the display interface 20, the communication interface 30, the storage unit 40, the pre-processing unit 50, the image space regularization unit 60, the deviation correction unit 70, the data consistency processing unit 80, and the output unit 90 are connected to each other so that they can communicate with each other.

[0017] The input / output interface 10 is an interface for connecting the magnetic resonance image reconstruction apparatus 1 to an input device (not shown), receives a user's input operation from the input device, and transmits a signal based on the received input operation to the magnetic resonance image reconstruction apparatus 1. The input / output interface 10 is, for example, a serial bus interface such as a USB (Universal Serial Bus). The input device includes a mouse, keyboard, trackball, switch, button, joystick, touch screen, microphone, etc. The input / output interface 10 may also be connected to a storage device to read and write various data to and from the storage device. The storage device may be, for example, a hard disk drive (HDD), a solid state drive (SSD), etc.

[0018] The display interface 20 is an interface for connecting the magnetic resonance image reconstruction apparatus 1 to a display device (not shown), and transmits data to the display device to display an image. The display interface 20 is, for example, a video output interface such as a DVI (Digital Visual Interface) or an HDMI (registered trademark) (High-Definition Multimedia Interface). The display device includes an LCD (Liquid Crystal Display) or an organic EL (Electroluminescence) display. The display device displays a user interface for accepting input operations from a user and magnetic resonance image data output from the magnetic resonance image reconstruction apparatus 1, and the user interface is, for example, a GUI (Graphical User Interface).

[0019] The communication interface 30 is an interface for connecting the magnetic resonance image reconstruction apparatus 1 to a server (not shown), and is capable of transmitting and receiving various data to and from the server. The communication interface 30 is, for example, a network card such as a wireless network card or a wired network card.

[0020] The storage unit 40 stores user data such as image data and K-space data used for image reconstruction. The storage unit 40 also stores parameters, such as neural network parameters, used by the magnetic resonance image reconstruction device 1 when performing image reconstruction. The storage unit 40 also stores training data for training each neural network and other learnable parameters used by the magnetic resonance image reconstruction device 1. The storage unit 40 is implemented by a storage device such as a ROM, flash memory, random access memory (RAM), hard disk drive (HDD), solid state drive (SSD), or register. Flash memory, HDD, SSD, etc. are non-volatile storage media. These non-volatile storage media may also be implemented by other storage devices connected via a network, such as a network-attached storage (NAS) or an external storage server device. Here, the network includes, for example, the Internet, a wide area network (WAN), a local area network (LAN), a carrier terminal, a wireless communication network, a wireless base station, and a dedicated line.

[0021] The pre-processing unit 50 generates initial image data by pre-processing the undersampled K-space data, which is input data to the magnetic resonance image reconstruction device 1. The pre-processing unit 50 has an inverse Fourier transform function 51 and a channel integrating function 52. The inverse Fourier transform function 51 performs an inverse Fourier transform on the data using an algorithm such as an inverse fast Fourier transform. The channel integrating function 52 integrates multi-channel data corresponding to each receive coil of the magnetic resonance imaging device into single-channel data.

[0022] The image space regularization unit 60 performs regularization processing on input image data in image space using a first neural network 61 to generate regularized image data, which is image data that has been regularized. The image space regularization unit 60 includes the first neural network 61. The first neural network 61 is, for example, a feedforward neural network, a convolutional neural network, or a Transformer. The first neural network 61 is preferably a convolutional neural network. More preferably, the first neural network 61 is a U-Net. In this embodiment, the first neural network 61 is a convolutional neural network that includes an input layer, an output layer, a convolutional layer, an excitation layer, a pooling layer, a batch normalization layer, and a fully connected layer, and the input layer and output layer are equal in size. The first neural network 61 realizes the function of image space regularization processing by loading neural network parameters dedicated to the first neural network 61 stored in the storage unit 40. By loading different neural network parameters, the first neural network 61 can be made to perform different regularization processes.

[0023] The deviation correction unit 70 corrects the regularized image data generated by the image space regularization unit 60 so that pixel value statistical features of the regularized image data approach pixel value statistical features of the initial image data corresponding to the undersampled K-space data, thereby generating deviation-corrected image data. The deviation correction unit 70 has a first statistical value calculation function 71, a second statistical value calculation function 72, a deviation calculation function 73, and a correction function 74. The first statistical value calculation function 71 calculates a first statistical value for the initial image data based on pixel values ​​of the initial image data. The second statistical value calculation function 72 calculates a second statistical value for the regularized image data based on pixel values ​​of the regularized image data. The deviation calculation function 73 calculates a deviation value between the initial image data and the regularized image data based on the first statistical value and the second statistical value. The correction function 74 corrects the regularized image data using the calculated deviation value to generate deviation-corrected image data.

[0024] The data consistency processor 80 performs data consistency processing on the deviation-corrected image data so that the K-space data corresponding to the deviation-corrected image data approaches the undersampled K-space data, thereby generating corrected image data.

[0025] The output unit 90 outputs image data based on the corrected image data as reconstructed magnetic resonance image data.

[0026] 2 is a flowchart showing the flow of the magnetic resonance image reconstruction method according to the first embodiment. Hereinafter, the flow of the magnetic resonance image reconstruction method according to the first embodiment will be described with reference to FIG.

[0027] First, proceed to step S101.

[0028] In step S101, the user selects undersampled K-space data K0 and mask data M stored in the memory unit 40 or input from an external storage device using the input device in accordance with the user interface displayed on the display device, and inputs them to the magnetic resonance image reconstruction device 1.

[0029] In the following description, K-space data K0 is described as three-dimensional tensor data with a width W × height H × number of channels (number of receiving coils) C, where the width direction is the frequency encoding direction and the height direction is the phase encoding direction. Magnetic resonance scans are undersampling, omitting specific frequency encoding or phase encoding, to reduce scan time. Therefore, magnetic resonance scans are performed by skipping coordinates corresponding to specific frequency encoding or phase encoding. As a result, data does not exist at some coordinates in the K-space data K0, and zero padding is performed on the data at these coordinates. Because data near the center of the K-space data has a significant impact on the contrast of the reconstructed image data, undersampling typically involves concentrating sampling on data near the center in the frequency encoding and phase encoding directions and skipping data at some positions far from the center.

[0030] The mask data M indicates which coordinates are sampled and which coordinates are omitted in the K-space data K0 during a magnetic resonance scan. The mask data M is, for example, a matrix of width W x height H, where the width direction is the frequency encoding direction and the height direction is the phase encoding direction. In the mask data M, the value of the coordinates where the sampled frequency encoding and phase encoding are located is set to 1, and the value of the coordinates where the unsampled frequency encoding and phase encoding are located is set to 0.

[0031] When the process of step S101 is completed, the process proceeds to step S102.

[0032] In step S102, the pre-processing unit 50 generates initial image data X0 based on the undersampled K-space data K0. The processing of step S102 will be described below with reference to Fig. 3. Fig. 3 is a data flow diagram for explaining the processing of step S102 of the magnetic resonance image reconstruction method according to the first embodiment. In Fig. 3, the flow of data is indicated by solid arrows.

[0033] First, the preprocessing unit 50 reads the K-space data K0 from the storage unit 40 and performs an inverse Fourier transform on the K-space data K0 using the inverse Fourier transform function 51 to generate multi-channel image space data I0. The multi-channel image space data I0 is image space data that has the same width, height, and number of channels as the K-space data K0. The data of each channel of the multi-channel image space data I0 is image space data converted from the K-space data collected by each receive coil.

[0034] Next, the pre-processing unit 50 uses a channel integration function 52 to integrate the data from multiple channels of the multi-channel image space data I0 into one channel of data according to the sensitivity of each receiving coil, thereby generating initial image data X0. Here, the initial image data X0 is two-dimensional image data with width W and height H that is generated directly based on the undersampled K-space data K0, and therefore suffers from problems such as many artifacts and noise, a lack of detail, and blurred images.

[0035] When the process of step S102 is completed, the process proceeds to step S103.

[0036] In steps S103 to S108, the image data is subjected to a predetermined number of correction processes based on the initial image data X0. Here, the predetermined number of times is preferably 8 to 10 times. Hereinafter, the image data that has been subjected to the correction process based on the initial image data X0 t times (t is an integer equal to or greater than 0) is referred to as corrected image data X0. t The corrected image data X0 at t=0 is the same image as the initial image data X0.

[0037] In step S103, the magnetic resonance image reconstruction apparatus 1 sets the current number of corrections (number of iterations) to 0. When the process of step S103 is completed, the process proceeds to step S104.

[0038] The processing of steps S104 to S106 will be described below with reference to Fig. 4. Fig. 4 is a data flow diagram for explaining the processing of steps S104 to S106 of the magnetic resonance image reconstruction method according to the first embodiment. In Fig. 4, the flow of data is indicated by solid arrows.

[0039] In step S104, the image space regularization unit 60 regularizes the corrected image data X t is subjected to regularization processing in the image space using the first neural network 61 to obtain the regularized image data Z t where t is the current revision number.

[0040] First, the image space regularization unit 60 reads out the neural network parameters of the first neural network 61 corresponding to the current number of corrections from the storage unit 40, and loads the parameters into the first neural network 61. Then, the image space regularization unit 60 reads out the corrected image data X t and reads out the corrected image data X t is input to the first neural network 61 whose parameters have already been loaded. Then, the image space regularization unit 60 inputs the corrected image data X t , and the regularized image data Z t Here, the regularized image data Z t is the corrected image data X t It is 2D image data of the same size as

[0041] The processing by the first neural network 61 is to obtain the corrected image data X t Therefore, the regularized image data Z t is the corrected image data X that has undergone noise removal, artifact removal, and anti-aliasing processing. tIt may be considered that the first neural network 61 uses different neural network parameters for the correction process with different correction times. By using different parameters for the first neural network 61 at different correction stages, the corrected image data X t Appropriate processing can be performed for the above.

[0042] When the process of step S104 is completed, the process proceeds to step S105.

[0043] In step S105, the deviation correction unit 70 calculates the regularized image data Z t The regularized image data Z t is corrected to obtain the deviation corrected image data D t Generate.

[0044] First, the deviation correction unit 70 calculates a first statistical value S0 for the initial image data X0 based on the pixel values ​​of the initial image data X0 using the first statistical value calculation function 71, and calculates the regularized image data Z t Based on the pixel values ​​of t The second statistical value S t Here, the first statistical value S0 and the second statistical value S t are the same type of statistics. For example, the first statistical value S0 and the second statistical value S t As the pixel values ​​of the image data, the average value, standard deviation, median value, minimum value, maximum value, etc. are applied.

[0045] Next, the deviation correction unit 70 calculates the second statistical value S t The deviation value V is calculated by dividing the above by the first statistical value S0.

[0046] Finally, the deviation correction unit 70 corrects the regularized image data Z t The deviation corrected image data D is the value obtained by dividing by the deviation value V. t Here, the deviation correction image data Dt is the regularized image data Z t The pixel value statistical characteristics are closer to the initial image data X0 than

[0047] When the process of step S105 is completed, the process proceeds to step S106.

[0048] In step S106, the data consistency processing unit 80 converts the deviation-corrected image data D t The deviation correction image data D is calculated so that the K-space data corresponding to t Data consistency processing is performed on the corrected image data X t+1 Generate.

[0049] The data consistency processing unit 80 calculates the corrected image data X t+1 Calculate.

number

[0050] In equation (1), λ is a data consistency coefficient, and A is a forward operator. λ may be a preset fixed value or a trainable value. The calculation A(X) by the forward operator A on arbitrary image data X means that the image data X is first converted into multi-channel image data and then subjected to a Fourier transform to obtain K-space data corresponding to the image data X, and then the Hadamard product of the obtained K-space data and mask data M is calculated.

[0051] Equation (1) can be solved by optimization algorithms such as gradient descent and proximal mapping, where proximal mapping can be further solved using conjugate gradient methods.

[0052] Here, the corrected image data X generated by the data consistency processing t+1The pixel value of deviation corrected image data D t The pixel value is close to that of the corrected image data X t+1 The corresponding K-space data is the deviation-corrected image data D t is closer to the undersampled K-space data K0 than the K-space data corresponding to

[0053] When the process of step S106 is completed, the process proceeds to step S107.

[0054] In step S107, 1 is added to the current number of corrections.

[0055] In the magnetic resonance image reconstruction method according to this embodiment, the corrected image data X t Correction processing is performed once on the corrected image data X t+1 Here, the corrected image data X t+1 is the corrected image data X t This image data is closer to the ground truth image data than the original image data.

[0056] When the process of step S107 is completed, the process proceeds to step S108.

[0057] In step S108, it is determined whether the current number of corrections has reached a predetermined number, and if it is determined that the predetermined number has been reached, the process proceeds to step S109, and if it is determined that the predetermined number has not been reached, the process proceeds to step S104.

[0058] In step S109, the corrected image data that has been corrected a predetermined number of times is output as an estimate for the ground truth image data.

[0059] When the process of step S109 is completed, the process of the magnetic resonance image reconstruction method ends.

[0060] Here, the processing in steps S104 to S106 is the correction of the corrected image data X by the correction module including the correction processing of the image space regularization block, the deviation correction block, and the data consistency block. t Among these, the processes in steps S104 to S106 correspond to the processes in the image space regularization block, the deviation correction block, and the data consistency block, respectively.

[0061] Furthermore, the processing of steps S103 to S109 can be considered as processing the initial image data X0 using a reconstruction network including multiple serially connected correction modules, i.e., an end-to-end unrolled reconstruction network, to generate an estimate for the ground truth image data.

[0062] (Second embodiment) A magnetic resonance image reconstruction method and a magnetic resonance image reconstruction device according to the second embodiment will be described below. In the second embodiment, differences from the first embodiment will be mainly described, and explanations of commonalities with the first embodiment will be omitted. In the description of the second embodiment, parts that are the same as those in the first embodiment will be described with the same reference numerals.

[0063] Fig. 5 is a diagram showing an example of the configuration of a magnetic resonance image reconstruction device 1A according to the second embodiment. The configuration of the magnetic resonance image reconstruction device 1A according to the second embodiment will be described with reference to Fig. 5. Compared to the magnetic resonance image reconstruction device 1 according to the first embodiment, the magnetic resonance image reconstruction device 1A according to the second embodiment has a deviation correction unit 70A instead of the deviation correction unit 70.

[0064] The deviation correction unit 70A corrects the regularized image data generated by the image space regularization unit 60 so that pixel value statistical features of the regularized image data approach pixel value statistical features of the initial image data corresponding to the undersampled K-space data, thereby generating deviation-corrected image data. The deviation correction unit 70A includes a second neural network 71A. The second neural network 71A is, for example, a feedforward neural network, a convolutional neural network, or a Transformer. The second neural network 71A is preferably a convolutional neural network. More preferably, the second neural network 71A is a U-Net. In this embodiment, the second neural network 71A is a convolutional neural network including an input layer, an output layer, a convolutional layer, an excitation layer, a pooling layer, a batch normalization layer, and a fully connected layer, and the input layer and output layer are of equal size. The second neural network 71A realizes the deviation correction processing function by loading neural network parameters dedicated to the second neural network 71A stored in the storage unit 40.

[0065] Fig. 6 is a flowchart showing the flow of a magnetic resonance image reconstruction method according to the second embodiment. The flow of the magnetic resonance image reconstruction method according to the second embodiment will be described with reference to Fig. 6. Compared to the magnetic resonance image reconstruction method according to the first embodiment, the magnetic resonance image reconstruction method according to the second embodiment has step S105A instead of step S105.

[0066] In the second embodiment, when the process of step S104 is completed, the process proceeds to step S105A.

[0067] In step S105A, the deviation correction unit 70A calculates the regularized image data Z t The regularized image data Z t is corrected to obtain the deviation corrected image data D t Generate.

[0068] First, the deviation correction unit 70A reads out the neural network parameters of the second neural network 71A from the storage unit 40 and loads the parameters into the second neural network 71A. Then, the deviation correction unit 70A converts the initial image data X0 into regularized image data Z t The deviation correction unit 70A then inputs the initial image data X0 and the regularized image data Z to the second neural network 71A. t The forward propagation based on the above is performed to obtain the deviation-corrected image data D t Calculate.

[0069] When the process of step S105A is completed, the process proceeds to step S106.

[0070] (Comparative Example) A magnetic resonance image reconstruction method and a magnetic resonance image reconstruction device according to a comparative example of the prior art will be described below, but in the comparative example, differences from the first embodiment will be mainly described, and explanations of commonalities with the first embodiment will be omitted. In the description of the comparative example, parts that are the same as those in the first embodiment will be described using the same reference numerals.

[0071] Fig. 7 is a diagram showing an example of the configuration of a magnetic resonance image reconstruction device 1B according to a comparative example. The configuration of the magnetic resonance image reconstruction device 1B according to the comparative example will be described with reference to Fig. 7. Compared to the magnetic resonance image reconstruction device 1 according to the first embodiment, the magnetic resonance image reconstruction device 1B according to the comparative example does not have a deviation correction unit 70, and has a data consistency processing unit 80B instead of the data consistency processing unit 80.

[0072] The data consistency processing unit 80B performs data consistency processing on the regularized image data so that the K-space data corresponding to the regularized image data approaches the undersampled K-space data, thereby generating corrected image data.

[0073] Fig. 8 is a flowchart showing the flow of a magnetic resonance image reconstruction method according to a comparative example. The flow of the magnetic resonance image reconstruction method according to the comparative example will be described with reference to Fig. 8. Compared to the magnetic resonance image reconstruction method according to the first embodiment, the magnetic resonance image reconstruction method according to the comparative example does not include step S105, and includes step S106B instead of step S106.

[0074] In the comparative example, once the processing in step S104 is completed, the process proceeds to step S106B.

[0075] In step S106B, the data consistency processing unit 80B converts the regularized image data Z t The regularized image data Z is calculated so that the K-space data corresponding to t Data consistency processing is performed on the corrected image data X t+1 Generate.

[0076] The data consistency processing unit 80B calculates the corrected image data X t+1 Calculate.

number

[0077] In equation (2), λ is the data consistency coefficient and A is the forward operator. Equation (2) can be solved by optimization algorithms such as gradient descent and proximal mapping. Here, proximal mapping can be further solved using conjugate gradient.

[0078] When the process of step S106B is completed, the process proceeds to step S107.

[0079] (Technical effect) The effects of the magnetic resonance image reconstruction method and the magnetic resonance image reconstruction apparatus according to the embodiment will be described below.

[0080] In magnetic resonance image reconstruction technology based on an end-to-end unrolled reconstruction network, initial image data is subjected to multiple corrections to generate estimates for ground truth image data. However, when regularization processing is performed on image data using a neural network, the neural network may not process the image data correctly, and the regularization processing may add artifacts or noise to the image or blur the image, resulting in the regularized image data generated by the regularization processing deviating from the original distribution.

[0081] In the comparative example, regularized image data is generated using a neural network, and then data consistency processing is performed using the regularized image data itself without processing the regularized image data to generate corrected image data. Therefore, if regularized image data that deviates from the original distribution is generated in a certain correction process, the corrected image data generated based on that regularized image data will also deviate from the original distribution. Furthermore, because the neural network cannot correctly process image data that deviates from the original distribution, subsequent correction processes will not be able to correctly correct the image data. As a result, artifacts and noise become increasingly severe as the correction process progresses, significantly reducing the accuracy of the reconstructed magnetic resonance image.

[0082] 9 is a diagram for explaining examples of regularized image data and corrected image data generated in each correction process of the magnetic resonance image reconstruction method according to the comparative example. Problems with the magnetic resonance image reconstruction method according to the comparative example will be explained with reference to FIG.

[0083] FIG. 9 shows an example of reconstructing a magnetic resonance image of an axial plane near the prostate using a magnetic resonance image reconstruction method according to a comparative example. As shown in FIG. 9, the magnetic resonance image reconstruction method according to the comparative example performs a total of eight correction processes. The regularized image data generated in the second correction process contains artifacts and noise caused by the neural network. Subsequently, in the third to eighth correction processes, the neural network is affected by these artifacts and noise, making it impossible to perform accurate regularization processing, and the generated regularized image data increasingly deviates from the ground truth image data. Finally, in the eighth correction process, the corrected image data is generated based on regularized image data that deviates significantly from the ground truth image data, so artifacts and noise similarly occur in the generated corrected image data, significantly reducing the accuracy of the generated corrected image data.

[0084] On the other hand, in an embodiment, regularized image data is generated using a neural network, and then deviation correction is performed on the regularized image data based on pixel value statistical features of the initial image data to generate a deviation-corrected image. Data consistency processing is then performed using the deviation-corrected image to generate corrected image data. As a result, even if regularized image data that deviates from the original distribution is generated in a certain correction process, deviation of the corrected image data from the original distribution can be suppressed. According to the embodiment, a significant decrease in the accuracy of the reconstructed magnetic resonance image can be suppressed, thereby improving the stability of the reconstruction of the magnetic resonance image. Furthermore, according to the embodiment, the convergence speed of the corrected image data can be increased, thereby reducing the number of correction processes required.

[0085] Fig. 10 is a diagram for explaining examples of regularized image data and corrected image data generated by each correction process of the magnetic resonance image reconstruction method according to the embodiment. Fig. 10 is a diagram showing an example of reconstructing an axial plane near the prostate shown in Fig. 9 by the magnetic resonance image reconstruction method according to the embodiment.

[0086] As shown in Fig. 10, in the magnetic resonance image reconstruction method according to the embodiment, a total of eight correction processes are performed. Of these, artifacts and noise caused by the neural network appear in the regularized image data generated in the first and second correction processes. However, because deviation correction is performed on the regularized image data in which artifacts and noise have occurred, the subsequent correction processes are not affected by the artifacts and noise generated in the first and second correction processes, and the corrected image data converges by the fifth correction process.

[0087] FIG. 11 is a diagram comparing a magnetic resonance image reconstructed by the magnetic resonance image reconstruction method according to the embodiment with a magnetic resonance image reconstructed by a comparative example.

[0088] (a) of Figure 11 shows a local ground truth image of the knee, (b) of Figure 11 shows a local image of the knee reconstructed by the magnetic resonance image reconstruction method of the comparative example, and (c) of Figure 11 shows a local image of the knee reconstructed by the magnetic resonance image reconstruction method of the embodiment.

[0089] 11, in the local image of the knee reconstructed by the magnetic resonance image reconstruction method according to the comparative example, noticeable artifacts appear on the upper and lower sides of the image, and the fine structure of the area indicated by the arrow (the position indicated by the arrow in the figure) was not reconstructed. On the other hand, in the local image of the knee reconstructed by the magnetic resonance image reconstruction method according to the embodiment, there are no noticeable artifacts or noise in the image, and the fine structure of the area indicated by the arrow was reconstructed.

[0090] In the case of 4X undersampling, the stability of reconstruction by the magnetic resonance image reconstruction method according to the embodiment is relatively high, and in tests on various parts of the body such as the head, chest, spine, and waist, the magnetic resonance image reconstruction method according to the embodiment has a higher average structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) than the magnetic resonance image reconstruction method according to the comparative example.

[0091] (Methods for training and evaluating neural networks) In the above description, the magnetic resonance image reconstruction method and magnetic resonance image reconstruction device according to the embodiment use the first neural network 61, the second neural network 71A, and the data consistency coefficient λ, but these neural networks and parameters must be trained in advance to operate properly. Below, we will explain how to train the above neural networks and parameters.

[0092] First, multiple sets of training data stored in advance are read out from the storage unit 40. Each set of training data includes undersampled K-space data K0 and corresponding mask data M as input data, and ground truth image data as output data.

[0093] Next, the multiple sets of teacher data are divided into a training set, a test set, and a cross-validation set. Examples of the ratios of the training set, the test set, and the cross-validation set include 80%, 10%, 10%, or 90%, 5%, and 5%. For example, if the total number of sets of teacher data is 10,000, the teacher data of data #1 to #10,000 are divided so that data #1 to #8000 are the training set, data #8001 to #9000 are the test set, and data #9001 to #10,000 are the cross-validation set. In this case, input data of each set of teacher data in the training set is input to the magnetic resonance image reconstruction device 1, and the magnetic resonance image reconstruction method according to the embodiment is executed to calculate an estimated value of ground truth image data. Then, a difference value between the estimated value of the ground truth image data and the ground truth image data is calculated, and backpropagation is performed based on this difference value. This changes the parameters of each neural network and other machine-learnable parameters so as to reduce the difference between the estimated value of the ground truth image data output by the magnetic resonance image reconstruction device 1 and the ground truth image data. The above procedure is repeated until the difference between the estimated value of the ground truth image data output by the magnetic resonance image reconstruction device 1 and the ground truth of the ground truth image data becomes smaller than a preset threshold for the majority of data in the test set. At this point, it is determined that training of each neural network and parameter is complete.

[0094] Next, the input data of the cross-validation data (data #9001 to #10000) is input to the trained magnetic resonance image reconstruction device 1, and the peak signal-to-noise ratio of the estimated value for the ground truth image data output by the magnetic resonance image reconstruction device 1 and the structural similarity index between the estimated value and the ground truth image data are calculated as evaluation data.

[0095] Note that in the above-described embodiment, an example has been described in which the image space regularizer, deviation correction unit, data consistency processor, and output unit in this specification are realized by the image space regularizer 60, deviation correction unit 70 or 70A, data consistency processor 80, and output unit 90 described in the above-described embodiment, respectively, but the embodiment is not limited to this. For example, the image space regularizer, deviation correction unit, data consistency processor, and output unit in this specification may be realized by the image space regularizer 60, deviation correction unit 70 or 70A, data consistency processor 80, and output unit 90 described in the above-described embodiment, or processing units having the same functions may be realized by hardware only, software only, or a combination of hardware and software.

[0096] In the above-described embodiment, each of the processing units, i.e., the image space regularizer 60, the deviation correctors 70 and 70A, the data consistency processor 80, and the output unit 90, is implemented by a processing circuit such as a processor. In this case, the processing functions of the processing circuit are stored in the storage unit 40, for example, in the form of a computer-executable program. The processing circuit then reads and executes each program from the storage unit 40 to implement the processing function corresponding to each program. In other words, each processing unit has the configuration shown in FIGS. 1 and 5 when the corresponding processing circuit reads the program. Note that, although the programs corresponding to the processing functions of the processing circuit are stored in a single storage unit here, the embodiment is not limited thereto. For example, the programs corresponding to the processing functions may be stored in multiple storage units in a distributed manner, and the processing circuit may read and execute each program from each storage unit.

[0097] In the above description, each of the image space regularization unit 60, the deviation correction units 70 and 70A, the data consistency processing unit 80, and the output unit 90 is implemented by a single processing circuit, but the embodiment is not limited to this. For example, each processing unit may be configured by combining multiple independent processing circuits, and each processing circuit may execute a program to implement each processing function. Furthermore, the processing functions of each circuit and each unit may be implemented by being appropriately distributed or integrated into a single or multiple processing circuits. Furthermore, the processing functions of each circuit and each unit may be implemented by a combination of hardware and software, such as circuits.

[0098] Furthermore, the term "processor" used in the above description refers to a circuit such as a central processing unit (CPU), a graphics processing unit (GPU), an application specific integrated circuit (ASIC), or a programmable logic device (e.g., a simple programmable logic device (SPLD), a complex programmable logic device (CPLD), or a field programmable gate array (FPGA)). If the processor is a CPU, for example, the processor realizes its function by reading and executing a program stored in a memory unit. On the other hand, if the processor is an ASIC, for example, instead of storing the program in a memory unit, the function is directly incorporated into the processor circuit as a logic circuit. Note that each processor in this embodiment is not limited to being configured as a single circuit for each processor, but may be configured as a single processor by combining multiple independent circuits to realize its function. Furthermore, multiple components in FIG. 1 may be integrated into a single processor to realize its function.

[0099] Here, the program executed by the processor is provided in advance in a read-only memory (ROM) or storage unit. The program may be provided by being recorded on a computer-readable storage medium such as a compact disk (CD)-ROM, a flexible disk (FD), a recordable CD-R (CD-R), or a digital versatile disk (DVD) in a format that can be installed on these devices or in an executable format. The program may also be provided or distributed by being stored on a computer connected to a network such as the Internet and downloaded via the network. For example, the program may be composed of modules including the above-mentioned functional units. In actual hardware, a CPU reads and executes the program from a storage medium such as a ROM, whereby each module is loaded into a main memory device and generated on the main memory device.

[0100] In the above-described embodiments, the components of each device shown in the drawings are conceptual functional units and do not necessarily have to be physically configured as shown. In other words, the specific form of distribution or integration of each device is not limited to that shown in the drawings, and all or part of the devices can be functionally or physically distributed or integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions performed by each device can be realized by a CPU and a program analyzed and executed by the CPU, or can be realized as hardware using wired logic.

[0101] Furthermore, among the processes described in the above embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using a known method.In addition, the information including the processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings can be changed as desired unless otherwise specified.

[0102] According to at least one of the embodiments described above, it is possible to improve the stability of image reconstruction.

[0103] Although several embodiments have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, modifications, and combinations of embodiments can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, as well as within the scope of the invention and its equivalents as defined in the claims. [Explanation of symbols]

[0104] 1,1A Magnetic Resonance Image Reconstruction Device 60 Image Space Regularization Unit 70,70A deviation correction unit 80 Data integrity processing section 90 Output section

Claims

1. 1. A magnetic resonance image reconstruction method for reconstructing magnetic resonance image data based on undersampled K-space data, comprising: an image space regularization step of performing a regularization process in image space on the first image data generated based on the undersampled K-space data using a first neural network to generate second image data; a deviation correction step of correcting the second image data so that pixel value statistical characteristics of the second image data approach the pixel value statistical characteristics of the first image data to generate third image data; a data consistency processing step of performing data consistency processing on the third image data to generate fourth image data so that K-space data corresponding to the third image data approaches the undersampled K-space data; an output step of outputting image data based on the fourth image data as the reconstructed magnetic resonance image data; A magnetic resonance image reconstruction method comprising:

2. the deviation correction step calculates a first parameter related to the pixel value statistical feature of the first image data and a second parameter related to the pixel value statistical feature of the second image data, and corrects the second image data based on the first parameter and the second parameter to generate third image data.

2. The magnetic resonance image reconstruction method according to claim 1.

3. the deviation correction step calculates deviation data indicating a deviation between the second image data and the first image data based on the first parameter and the second parameter, and corrects the second image data based on the deviation data to generate the third image data.

3. The magnetic resonance image reconstruction method according to claim 2.

4. the deviation correction step calculates a value obtained by dividing the second parameter by the first parameter as the deviation data, and divides the second image data by the deviation data to generate the third image data.

4. The magnetic resonance image reconstruction method according to claim 3.

5. the deviation correction step generates the third image data by a second neural network; 3. The magnetic resonance image reconstruction method according to claim 2.

6. the first neural network is a convolutional neural network; 6. The magnetic resonance image reconstruction method according to claim 1.

7. 1. A magnetic resonance image reconstruction apparatus for reconstructing magnetic resonance image data based on undersampled K-space data, comprising: an image space regularization unit that performs a regularization process in an image space using a first neural network on the first image data generated based on the undersampled K-space data to generate second image data; a deviation correction unit that corrects the second image data so that pixel value statistical features of the second image data approach the pixel value statistical features of the first image data, thereby generating third image data; a data consistency processing unit that performs data consistency processing on the third image data so that K-space data corresponding to the third image data approaches the undersampled K-space data, thereby generating fourth image data; an output unit that outputs image data based on the fourth image data as the reconstructed magnetic resonance image data; A magnetic resonance image reconstruction apparatus comprising:

Images