A magnetic resonance image reconstruction method, device, imaging system and storage medium

Abstract

The application provides a magnetic resonance image reconstruction method and device, an imaging system and a storage medium. The method is to input the to-be-reconstructed undersampled K-space data into an image reconstruction model to obtain a reconstructed image, and the image reconstruction model is trained to minimize a model training loss value, and the training includes: performing inverse Fourier transform and downsampling on full-sampling K-space data to obtain a full-sampling image and the undersampled K-space data, sequentially performing randomization, interpolation and inverse Fourier transform on the undersampled K-space data to obtain a frequency-domain training image, fusing the undersampled training image obtained by inversely Fourier transforming the undersampled K-space data and the undersampled image obtained by performing a plurality of second downsampling and interpolation on the undersampled K-space data to obtain an image-domain training image, and calculating the loss value of the training image fused in the frequency domain and the image domain and the full-sampling image. Through adaptive filling and reconstruction in the image domain and the frequency domain, the application realizes efficient, fast and high-quality magnetic resonance image recovery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of magnetic resonance image reconstruction, in particular to a magnetic resonance image reconstruction method and device, an imaging system and a storage medium. BACKGROUND

[0002] Magnetic Resonance Imaging (MRI) is a kind of medical imaging technology without radiation and with high soft tissue resolution, which is widely used in the fields of neurology, cardiology, and orthopedics. However, since MRI scanning usually relies on data acquisition in K-space, the scanning process is slow, and the patient needs to remain motionless in the instrument for a long time. Longer scanning time not only easily causes motion artifacts, but also increases the cost and reduces the imaging real-time performance. In order to shorten the imaging time of MRI, accelerating MRI reconstruction has become a key research direction.

[0003] The existing magnetic resonance image reconstruction method is to shorten the acquisition time by reducing the sampling points in K-space, and then use mathematical methods or machine learning models to restore the image. Among them, mathematical methods based on compressed sensing (CS) and low-rank matrix restoration mainly utilize the sparsity, redundancy and other characteristics of signals to solve the under-sampling problem through optimization, so as to realize image reconstruction. However, the mathematical optimization calculation cost is high, it is difficult to accurately model the complex image structure and signal noise, and it is difficult to ensure the quality and efficiency of the reconstruction at the same time. At the same time, the reconstruction based on machine learning model mainly includes two strategies: image domain reconstruction which directly processes the under-sampled image in the image space to remove artifacts and restore details, and frequency domain reconstruction which directly fills or interpolates the under-sampled data in K-space and then reconstructs the image. Although the image domain reconstruction strategy has strong feature extraction ability in the image space, it is difficult to fully utilize the frequency information distribution law, resulting in the loss of some high-frequency details. And the frequency domain reconstruction strategy can better utilize the structural information of K-space, but it is difficult to effectively handle the non-locality, data sparsity and complex phase relationship of K-space data, which limits its application. SUMMARY

[0004] In order to solve the problems in the prior art, it is necessary to provide a magnetic resonance image reconstruction method, device, imaging system and storage medium in view of the above technical problems.

[0005] In a first aspect, an embodiment of the present application provides a magnetic resonance image reconstruction method, which comprises the following steps:

[0006] Obtaining under-sampled K-space data to be reconstructed;

[0007] Obtaining an image reconstruction model;

[0008] inputting the to-be-reconstructed undersampling K-space data into the image reconstruction model to obtain a magnetic resonance reconstruction image; wherein the image reconstruction model is obtained through an iterative training process as follows, with a model training loss value minimization as an objective:

[0009] acquiring full-sampling K-space data;

[0010] obtaining a full-sampling image according to inverse Fourier transform on the full-sampling K-space data;

[0011] obtaining undersampling K-space data corresponding to the full-sampling K-space data according to first down-sampling on the full-sampling K-space data;

[0012] obtaining randomized undersampling K-space data according to randomization on the undersampling K-space data corresponding to the full-sampling K-space data;

[0013] obtaining interpolated K-space data according to interpolation on the randomized undersampling K-space data;

[0014] obtaining a frequency-domain training image according to inverse Fourier transform on the interpolated K-space data;

[0015] obtaining a first undersampling training image according to inverse Fourier transform on the undersampling K-space data corresponding to the full-sampling K-space data;

[0016] obtaining a second undersampling training image according to a plurality of times of second down-sampling on the first undersampling training image;

[0017] obtaining an interpolated second training undersampling image according to interpolation on the second undersampling training image;

[0018] obtaining an image-domain training image according to image fusion on the interpolated second training undersampling image and the first undersampling training image;

[0019] obtaining a training image according to weighted fusion on the frequency-domain training image and the image-domain training image;

[0020] obtaining a model training loss value according to loss calculation on the training image and the full-sampling image.

[0021] In a second aspect, an embodiment of the present application provides a magnetic resonance image reconstruction device, the device comprising:

[0022] a data acquisition module configured to acquire to-be-reconstructed undersampling K-space data;

[0023] a model acquisition module configured to acquire an image reconstruction model;

[0024] an image reconstruction module, configured to input the to-be-reconstructed under-sampling K-space data into the image reconstruction model to obtain a magnetic resonance reconstruction image; wherein the image reconstruction model is obtained through an iterative training process as follows, with a model training loss value minimization as an objective:

[0025] obtain full-sampling K-space data;

[0026] obtain a full-sampling image according to inverse Fourier transform on the full-sampling K-space data;

[0027] obtain under-sampling K-space data corresponding to the full-sampling K-space data according to first down-sampling on the full-sampling K-space data;

[0028] obtain randomized under-sampling K-space data according to randomization on the under-sampling K-space data corresponding to the full-sampling K-space data;

[0029] obtain interpolated K-space data according to interpolation on the randomized under-sampling K-space data;

[0030] obtain a frequency domain training image according to inverse Fourier transform on the interpolated K-space data;

[0031] obtain a first under-sampling training image according to inverse Fourier transform on the under-sampling K-space data corresponding to the full-sampling K-space data;

[0032] obtain a second under-sampling training image according to a plurality of times of second down-sampling on the first under-sampling training image;

[0033] obtain an interpolated second training under-sampling image according to interpolation on the second under-sampling training image;

[0034] obtain an image domain training image according to image fusion on the interpolated second training under-sampling image and the first under-sampling training image;

[0035] obtain a training image according to weighted fusion on the frequency domain training image and the image domain training image;

[0036] obtain a model training loss value according to loss calculation on the training image and the full-sampling image.

[0037] In a third aspect, the embodiments of the present application also provide a magnetic resonance imaging system, which comprises an imaging controller; the imaging controller comprises a processor and a memory, the processor and the memory are coupled, the memory is used to store a computer program or instructions, and the processor is used to execute the computer program or instructions in the memory, so that the imaging controller executes the steps of the above-mentioned magnetic resonance image reconstruction method.

[0038] In a fourth aspect, the embodiments of the present application also provide a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the above-mentioned magnetic resonance image reconstruction method.

[0039] Compared with the prior art, the magnetic resonance image reconstruction method provided by the present application can not only reduce the overfitting risk of the model, improve the generalization ability and the ability to capture the consistency of global features of the model, and enhance the robustness of the model, but also can more effectively identify and suppress artifacts, and ensure the frequency domain reconstruction effect; at the same time, by repeatedly downsampling the image domain image features, the spatial features of the undersampled image can be effectively captured, and important anatomical structure information can be provided for the model, and the frequency domain features and the image domain features are weighted and fused, which can effectively improve the comprehensiveness of the model learning features, and further improve the generalization ability and reconstruction effect of the model; that is, by simultaneously performing reliable adaptive filling and reconstruction on the image domain data and the frequency domain data, efficient, fast and high-quality magnetic resonance image reconstruction is realized, and the user experience is effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 is a flowchart of the magnetic resonance image reconstruction method in the embodiments of the present application;

[0041] Figure 2 is a schematic diagram of the iterative training process of the image reconstruction model in the embodiments of the present application;

[0042] Figure 3 is a first downsampling effect schematic diagram in the embodiments of the present application;

[0043] Figure 4 is a randomization result schematic diagram of the undersampled K-space data in the embodiments of the present application;

[0044] Figure 5 is a bilinear interpolation effect schematic diagram of the randomized undersampled K-space data in the embodiments of the present application;

[0045] Figure 6is a schematic diagram of a second down-sampling process in an embodiment of the present application;

[0046] Figure 7 is a schematic diagram of a SiLU activation function in an embodiment of the present application;

[0047] Figure 8 is a schematic diagram of a magnetic resonance image reconstruction effect of an image reconstruction model in an embodiment of the present application;

[0048] Figure 9 is a schematic diagram of a structure of a magnetic resonance image reconstruction system in an embodiment of the present application. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and beneficial effects of the present application clearer, further detailed description will be made to the present application in combination with the accompanying drawings and embodiments. Obviously, the following described embodiments are only a part of the embodiments of the present application, and are only used to illustrate the present application, but not to limit the scope of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0050] In the description of the present application, it should be noted that, unless otherwise defined, all technical and scientific terms used in the present application are the same as the meanings commonly understood by those skilled in the art. The terms used in the specification of the present application are only for the purpose of describing the embodiments, and are not intended to limit the present application. For those skilled in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0051] As shown in Figure 1 The present application provides a magnetic resonance image reconstruction method, comprising the following steps:

[0052] S1, acquiring the to-be-reconstructed undersampled K-space data; wherein the to-be-reconstructed undersampled K-space data can be understood as the K-space data obtained by reducing the sampling points of the K-space in order to shorten the acquisition time, which is directly used for magnetic resonance image reconstruction. The specific acquisition method can be obtained by the existing acquisition method, which is not described in detail here.

[0053] S2, inputting the to-be-reconstructed undersampled K-space data into the image reconstruction model to obtain a magnetic resonance reconstruction image; wherein the image reconstruction model can be understood as a network model capable of directly processing the input to-be-reconstructed sampling K-space data to obtain a full-space magnetic resonance reconstruction image.

[0054] To achieve efficient and rapid reconstruction of magnetic resonance images based on undersampled K-space data while effectively ensuring the reconstruction quality, an improved image reconstruction model is designed by simultaneously performing adaptive padding reconstruction in both the image and frequency domains. In some embodiments, such as... Figure 2 As shown, the image reconstruction model is based on minimizing the model training loss value and is obtained through the following iterative training process:

[0055] Acquire fully sampled K-space data; where fully sampled K-space data can be understood as the complete and comprehensive acquisition of magnetic resonance signals using nuclear magnetic resonance equipment, converting these signals into digital signals, and then filling the K-space with the resulting full K-space data containing frequency domain information, denoted as .

[0056] The fully sampled image is obtained by performing an inverse Fourier transform on the fully sampled K-space data; where the inverse Fourier transform (IFFT) corresponds to the Fourier transform and is the process of transforming frequency domain data back to time domain data, which can be represented as follows: F -1 This represents the inverse Fourier transform. The fully sampled image X obtained through the inverse Fourier transform is the K-space data of the fully sampled image. The images are of the same size, composed of pixels, with each pixel having a value ranging from 0 to 255. The fully sampled image X obtained here is used as the sample label for calculating the model training loss during subsequent model training.

[0057] Based on the first downsampling of the fully sampled K-space data, the undersampled K-space data corresponding to the fully sampled K-space data is obtained; wherein, downsampling can be understood as the process of reducing the amount of data while keeping the data size unchanged.

[0058] In some embodiments, such as Figure 3 As shown, the step of obtaining the undersampled K-space data corresponding to the full-sampled K-space data based on the first downsampling of the full-sampled K-space data includes:

[0059] The number of K-space lines retained for sampling is obtained based on a preset downsampling factor and the number of K-space lines in the fully sampled K-space data. The preset downsampling factor can be determined according to actual application requirements. To closely approximate the actual undersampling effect, in some embodiments, the preset downsampling factor is set to 8 times. The number of K-space lines in the fully sampled K-space data depends on the actual sampled data size. Assuming the size of the fully sampled K-space data is 256*256, then the number of K-space lines is 256.

[0060] The sampling reserved K-space line number can be understood as the K-space line number that needs to be reserved in the original full sampling K-space data after the downsampling processing. The K-space line number is divided by the preset downsampling multiple to obtain the required sampling reserved K-space line number. For example, if the K-space line number is 256 and the preset downsampling multiple is 8, the sampling reserved K-space line number is 32 (256 / 8). If the actual K-space line number cannot be divided by the preset downsampling multiple, the sampling reserved K-space line number can be obtained by rounding up or rounding down.

[0061] According to the random deletion of the K-space lines outside the preset central reserved region of the full sampling K-space data according to the sampling reserved K-space line number, the undersampling K-space data is obtained. The preset central reserved region can be understood as a region located in the center of the full sampling K-space data and the data remains unchanged in the downsampling processing, which is set in advance according to application requirements. The number of K-space lines in this region is not limited. The corresponding undersampling K-space data The acquisition process of the corresponding undersampling K-space data can be understood as follows: first, calculate the difference between the sampling reserved K-space line number and the K-space line number in the preset central reserved region; then, randomly delete the K-space lines corresponding to the difference number in the region of the full sampling K-space data except the preset central reserved region; finally, obtain an undersampling K-space data with a sampling reserved K-space line number.

[0062] It should be noted that the corresponding undersampling K-space data obtained by the above-mentioned downsampling processing of the full sampling K-space data can also be obtained by adjusting the sampling acceleration multiple of the nuclear magnetic resonance device in actual application.

[0063] According to the randomization of the undersampling K-space data corresponding to the full sampling K-space data, the randomized undersampling K-space data is obtained. As shown in Figure 4 The randomization of the undersampling K-space data can be understood as rearranging the K-space data that has been filled, i.e., expanding the sampling data into a one-dimensional array and randomly rearranging the order, and then rearranging it into the original sampling shape, so as to apply the rearranged sampling points to the K-space data, so that the original undersampling K-space data is randomly distributed on the K-space graph, and the K-space image size remains unchanged before and after randomization. This randomization processing method is similar to redesigning the filling method of K-space in the frequency domain, i.e., selectively choosing which positions in K-space need to be filled and which positions do not need to be filled, so as to simulate the effect of selectively filling the sampling points of K-space data as much as possible, and this random effect is difficult to obtain by operating the magnetic resonance device.

[0064] By using the randomized undersampled K-space data obtained based on the above-described randomization process as the basic data for frequency domain reconstruction, not only can the model training avoid overfitting of the model to a certain fixed sampling mode, improve the generalization ability of the model, and enhance the robustness of the model by disrupting the specific sampling mode of the K-space data, but also the randomization operation is equivalent to a "data augmentation" process. For the same undersampled K-space data, different randomization schemes will generate different training samples, which can further reduce the risk of model overfitting by increasing data diversity. In addition, by destroying the original sampling structure of the K-space points, the model must consider the information of the surrounding points during interpolation completion, and the model training based on the interpolated data will learn more about the global characteristics of the K-space rather than just the local features, thereby improving the model's ability to capture global consistency while breaking the structural characteristics of artifacts, allowing the model to more effectively identify and suppress artifacts during the training process, and ensuring the generation of clearer reconstructed images.

[0065] According to the interpolation of the randomized undersampled K-space data, interpolated K-space data is obtained; wherein the interpolation process can be understood as estimating and filling in the unknown sampling point data based on the existing sampling point data in the undersampled K-space data; in order to ensure the data filling effect, in some embodiments, the interpolation method used for the randomized undersampled K-space data is the bilinear interpolation method. For each pixel point of the target image, find its four nearest neighboring points on the original image, perform interpolation calculation based on the neighboring points to obtain its pixel value, and fill in the pixel value to the corresponding position on the K-space, so that the places without data points on the K-space are all filled up, and the interpolated K-space data with all points having values on the K-space as shown in Figure 5 is obtained.

[0066] By performing bilinear interpolation on the randomized undersampled K-space data, the information loss can be reduced while reducing the abrupt boundary effect between the undersampled region and the already sampled region, so that the interpolated K-space data obtained has good smoothness, can provide a better initial solution for subsequent model training, and make the gradient smoother and the convergence more stable during model training.

[0067] According to the inverse Fourier transform of the interpolated K-space data, a frequency domain training image is obtained; wherein the inverse Fourier transform to obtain the frequency domain training image is described above and will not be repeated here.

[0068] According to the inverse Fourier transform of the undersampled K-space data corresponding to the fully sampled K-space data, a first undersampled training image is obtained.

[0069] According to a plurality of second down-sampling operations performed on the first down-sampled training image, a second down-sampled training image is obtained; wherein the number of times of the second down-sampling operations performed on the first down-sampled training image can be set according to the data size, the sampling mode or other factors in actual application so that the subsequent model training achieves the best effect; in some embodiments, the number of times of the second down-sampling operations is set to 3, that is, the first down-sampled training image is subjected to the second down-sampling operation for 3 times in succession.

[0070] In some embodiments, as shown in Figure 6 the second down-sampling processing sequentially includes convolution, batch normalization, activation function and pooling processing; correspondingly, the step of obtaining the second down-sampled training image according to a plurality of second down-sampling operations performed on the first down-sampled training image includes:

[0071] According to a convolution operation performed on the first down-sampled training image, a first feature map is obtained; wherein the size of the convolution kernel used in the convolution operation can be adjusted according to actual conditions, and in some embodiments, a convolution processing with a convolution kernel of 3*3 and a step of 1 is selected; correspondingly, the size of the obtained first feature map is consistent with that of the first down-sampled training image. It should be noted that the convolution operation herein can adopt the convolution operation in the existing deep learning.

[0072] According to a batch normalization processing performed on the first feature map, a second feature map is obtained; wherein the batch normalization processing can be understood as normalizing each first feature map in a small batch to make its mean close to 0 and its standard deviation close to 1, which helps to accelerate the training process of the model and improve the generalization ability of the model.

[0073] According to a nonlinear transformation performed on the second feature map, a third feature map is obtained; wherein the nonlinear transformation can be understood as introducing an activation function to enable the network model to learn and represent more complex nonlinear relationships. In order to ensure that the model can better learn the complex data distribution and features and effectively improve the expression ability, in some embodiments, the activation function in the nonlinear transformation is set to SiLU activation function. As shown in Figure 7 the SiLU activation function has smoothness and nonlinearity, and is more effective than other activation functions in dealing with the problem of gradient vanishing and explosion, and is defined as follows:

[0074]

[0075] In the formula, x is the second feature map obtained after batch normalization; SiLU(x) represents the third feature map corresponding to the second feature map.

[0076] According to the pooling processing on the third feature map, a fourth feature map is obtained; wherein the pooling processing can be selected from the existing maximum pooling, average pooling, global pooling, random pooling and hybrid pooling according to actual needs, which is not limited here; the fourth feature map obtained by the pooling processing can be understood as a two-dimensional array containing various features extracted from the input third feature map. The role of the pooling processing is to reduce the height and width of the feature map, reduce the complexity and calculation amount of the data while maintaining the main features, which helps to reduce the number of parameters and the calculation burden in the subsequent network layer, improves the calculation efficiency, and reduces the risk of model overfitting; the fourth feature map obtained corresponds to the channel number of the first undersampling training image, but the height and width are 1 / 2 of the height and width of the first undersampling training image.

[0077] According to the preset number of downsampling with the fourth feature map as the input data of the next downsampling processing, the second undersampling training image is obtained; wherein the preset number of downsampling can be understood as the number of continuous downsampling except the first downsampling; the processing process of each second downsampling is the same, which can be referred to the description above. Assuming that a total of 3 second downsampling processes are performed, the size of the finally obtained second undersampling training image is 1 / 8 of the first undersampling training image, and if the size of the original first undersampling training image is 256*256, the size of the second undersampling training image is 32*32.

[0078] According to the interpolation on the second undersampling training image, an interpolated second training undersampling image is obtained; wherein the interpolated second training undersampling image can be understood as the high and wide of the second undersampling training image are enlarged to the size before downsampling, so that the height, width and channel number of the image are consistent with the first undersampling training image. In order to ensure the interpolation effect, in some embodiments, the interpolation method in the interpolation on the second undersampling training image is also set to the bilinear interpolation method to ensure the smoothness of the image after interpolation.

[0079] According to the image fusion on the interpolated second training undersampling image and the first undersampling training image, an image domain training image is obtained; wherein the image fusion can be understood as the fusion processing of the pixel values at the same pixel position in the interpolated second training undersampling image and the first undersampling training image. In some embodiments, the step of obtaining the image domain training image according to the image fusion on the interpolated second training undersampling image and the first undersampling training image comprises:

[0080] According to the same position pixel value average calculation on the interpolated second training undersampling image and the first undersampling training image, the image domain training image is obtained; that is, the pixel values of each corresponding position of the interpolated second training undersampling image and the first undersampling training image are added and averaged to obtain the required image domain training image.

[0081] The image domain training image acquisition method can effectively capture the spatial features of the undersampling image by repeatedly performing the downsampling operation, thereby providing important anatomical structure information for the model, and further providing reliable support for the reconstruction effect of the model.

[0082] According to the weighted fusion of the frequency domain training image and the image domain training image, the training image is obtained; wherein the weighted fusion can be understood as the weighted fusion processing of the pixel values of the same pixel position in the frequency domain training image and the image domain training image. In some embodiments, the step of obtaining the training image according to the weighted fusion of the frequency domain training image and the image domain training image comprises:

[0083] According to the weighted fusion of the frequency domain training image and the image domain training image, the training image is obtained; wherein the weighted fusion can be understood as the weighted fusion processing of the pixel values of the same pixel position in the frequency domain training image and the image domain training image. In some embodiments, the step of obtaining the training image according to the weighted fusion of the frequency domain training image and the image domain training image comprises: The reconstructed image Z is obtained.

[0084] By adding the corresponding weights of the features learned in the frequency domain and the image domain (spatial domain), the model can learn more comprehensive features, has stronger generalization ability, and has better reconstruction effect on the basis of learning the K-space features.

[0085] According to the loss calculation on the training image and the full sampling image, the model training loss value is obtained; wherein the model training loss value can be understood as the function value of evaluating the error between the reconstructed image obtained by the model processing and the real image; in order to ensure the simplicity and efficiency of the model training loss value calculation, in some embodiments, the loss function in the loss calculation on the training image and the full sampling image is a mean square error loss function, which is represented as:

[0086]

[0087] In the formula, indicates the reconstruction loss value of the model; N indicates the number of samples; X i and Z i respectively indicate the full sampling image (real target value) and the training image (model reconstructed image) corresponding to the i-th undersampling K-space data.

[0088] It should be noted that the above model training process only gives the processing process of one training sample. In actual application, a plurality of sets of full sampling K-space data are collected for training based on the above method, so as to obtain an image reconstruction model for realizing efficient, fast and high-quality magnetic resonance image recovery based on undersampling K-space data, and to ensure that the finally trained image reconstruction model has reliable reconstruction effect and generalization ability.

[0089] In addition, in order to ensure that the weighted fusion of the frequency domain training image and the image domain training image in the image reconstruction model processing can adaptively meet the high-quality reconstruction requirements of different undersampling data, in some embodiments, the updated parameters in the iterative training of the image reconstruction model include not only the hyperparameters in the image reconstruction model, but also the fusion weight λ in the weighted fusion of the frequency domain training image and the image domain training image.

[0090] After obtaining the image reconstruction model through the above method steps, undersampling K-space data actually collected by a magnetic resonance device can be directly processed to obtain the required magnetic resonance image. In order to verify the reconstruction quality of the image reconstruction model provided by the present application, in some embodiments, full sampling K-space data is set to be 8 times down-sampled to obtain undersampling K-space data, and the inverse Fourier transform is performed on the undersampling K-space data to obtain a first undersampling training image, which is processed by second down-sampling three times in succession, and the image reconstruction model is constructed based on this condition, and the reconstructed image as shown in Figure 8 Figure 8 It can be seen from the contrast effect of

[0091] ​The magnetic resonance image reconstruction method provided by the embodiment of the application can not only reduce the overfitting risk of the model, improve the generalization ability and the ability to capture the global feature consistency of the model, and enhance the robustness of the model, but also can more effectively identify and suppress artifacts, and ensure the frequency domain reconstruction effect; meanwhile, through the repeated downsampling operation on the image domain image features, the spatial features of the undersampled image can be effectively captured, and the frequency domain features and the image domain features are weighted and fused on the basis of providing important anatomical structure information for the model, so that the comprehensiveness of the model learning features can be effectively improved, and the generalization ability and the reconstruction effect of the model are further improved; that is, through reliable adaptive filling and reconstruction of the image domain data and the frequency domain data, efficient, fast and high-quality magnetic resonance image reconstruction is realized, and the user experience is effectively improved.

[0092] It should be noted that although each step in the above flowchart is displayed in sequence according to the arrow indication, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, the execution of these steps has no strict order limitation, and these steps can be executed in other orders.

[0093] As shown in Figure 9 The application provides a magnetic resonance image reconstruction device, the device comprises:

[0094] A data acquisition module 1 is configured to acquire undersampled K-space data to be reconstructed;

[0095] A model acquisition module 2 is configured to acquire an image reconstruction model;

[0096] An image reconstruction module 3 is configured to input the undersampled K-space data to be reconstructed into the image reconstruction model to obtain a magnetic resonance reconstruction image; wherein the image reconstruction model is obtained through an iterative training process as follows based on minimizing the model training loss value as an objective:

[0097] Acquire fully sampled K-space data;

[0098] Obtain a fully sampled image according to an inverse Fourier transform of the fully sampled K-space data;

[0099] Obtain undersampled K-space data corresponding to the fully sampled K-space data according to a first downsampling of the fully sampled K-space data;

[0100] Obtain randomized undersampled K-space data according to randomization of the undersampled K-space data corresponding to the fully sampled K-space data;

[0101] obtaining interpolated K-space data according to interpolation on the randomized undersampled K-space data;

[0102] obtaining a frequency domain training image according to inverse Fourier transform on the interpolated K-space data;

[0103] obtaining a first undersampled training image according to inverse Fourier transform on the undersampled K-space data corresponding to the fully sampled K-space data;

[0104] obtaining a second undersampled training image according to a plurality of second down-sampling on the first undersampled training image;

[0105] obtaining an interpolated second training undersampled image according to interpolation on the second undersampled training image;

[0106] obtaining an image domain training image according to image fusion on the interpolated second training undersampled image and the first undersampled training image;

[0107] obtaining a training image according to weighted fusion on the frequency domain training image and the image domain training image;

[0108] obtaining a model training loss value according to loss calculation on the training image and the fully sampled image.

[0109] The specific limitations of the magnetic resonance image reconstruction system can refer to the limitations of the magnetic resonance image reconstruction method in the above, and the corresponding technical effects can also be obtained equally, which will not be repeated here. Each module in the above magnetic resonance image reconstruction system can be realized by software, hardware and their combination in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in the form of software, so as to call and execute the operations corresponding to the above modules by the processor.

[0110] In some embodiments, the present application provides a magnetic resonance imaging system, characterized in that the magnetic resonance imaging system comprises an imaging controller; the imaging controller comprises a processor and a memory, the processor and the memory are coupled, the memory is used to store computer programs or instructions, and the processor is used to execute the computer programs or instructions in the memory, so that the imaging controller executes the steps of the above magnetic resonance image reconstruction method.

[0111] In some embodiments, the present application provides a computer readable storage medium having computer programs stored thereon, and the computer programs are executed by a processor to realize the steps of the above magnetic resonance image reconstruction method.

[0112] In summary, the magnetic resonance image reconstruction method, device, imaging system and storage medium provided by the embodiments of the present application can break the specific sampling mode of K-space data, destroy the original sampling structure of K-space points, break the structural characteristics of artifacts and increase the diversity of data by randomizing the undersampled K-space data, not only can reduce the overfitting risk of the model, improve the generalization ability and the ability to capture global feature consistency of the model, and enhance the robustness of the model, but also can more effectively identify and suppress artifacts to ensure the frequency domain reconstruction effect; at the same time, by repeatedly downsampling the image domain image features, the spatial features of the undersampled image can be effectively captured, and important anatomical structure information can be provided to the model, and the frequency domain features and the image domain features are weighted and fused, which can effectively improve the comprehensiveness of the model learning features and further improve the generalization ability and reconstruction effect of the model; that is, by simultaneously performing reliable adaptive filling and reconstruction on the image domain data and the frequency domain data, efficient, fast and high-quality magnetic resonance image reconstruction is realized, and the user experience is effectively improved.

[0113] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts of each embodiment can be referred to each other, and each embodiment mainly describes the difference from other embodiments. Especially, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the related parts can be referred to the part of the method embodiment. It should be noted that, the technical features of the above embodiments can be combined arbitrarily, in order to make the description simple, all possible combinations of the technical features of the above embodiments are not described, however, as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present application.

[0114] The above-described embodiments only express several preferred embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the patent. It should be noted that, for ordinary skilled in the art, without departing from the technical principles of the present application, some improvements and replacements can be made, and these improvements and replacements should be considered as the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the protection scope of the claims.

Claims

1. A method of magnetic resonance image reconstruction, characterized by, The method comprises the following steps: Obtain the to-be-reconstructed undersampled K-space data; Obtain an image reconstruction model; Input the to-be-reconstructed undersampled K-space data into the image reconstruction model to obtain a magnetic resonance reconstruction image; wherein the image reconstruction model is obtained through an iterative training process as follows: Obtain fully-sampled K-space data; Obtain a fully-sampled image according to inverse Fourier transform of the fully-sampled K-space data; Obtain undersampled K-space data corresponding to the fully-sampled K-space data according to first down-sampling of the fully-sampled K-space data; Obtain randomized undersampled K-space data according to randomization of the undersampled K-space data corresponding to the fully-sampled K-space data; Obtain interpolated K-space data according to interpolation of the randomized undersampled K-space data; Obtain a frequency-domain training image according to inverse Fourier transform of the interpolated K-space data; Obtain a first undersampled training image according to inverse Fourier transform of the undersampled K-space data corresponding to the fully-sampled K-space data; Obtain a second undersampled training image according to a plurality of times of second down-sampling of the first undersampled training image; Obtain an interpolated second training undersampled image according to interpolation of the second undersampled training image; Obtain an image-domain training image according to image fusion of the interpolated second training undersampled image and the first undersampled training image; Obtain a training image according to weighted fusion of the frequency-domain training image and the image-domain training image; Obtain a model training loss value according to loss calculation of the training image and the fully-sampled image.

2. The magnetic resonance image reconstruction method of claim 1, wherein, The updated parameters in the iterative training of the image reconstruction model comprise a fusion weight in the weighted fusion of the frequency-domain training image and the image-domain training image.

3. The magnetic resonance image reconstruction method of claim 1, wherein, The step of obtaining the undersampled K-space data corresponding to the fully-sampled K-space data according to first down-sampling of the fully-sampled K-space data comprises: Obtain a sampling-retained K-space line number according to a preset down-sampling multiple and a K-space line number in the fully-sampled K-space data; Obtain the undersampled K-space data by randomly deleting K-space lines outside a preset center-retained region of the fully-sampled K-space data according to the sampling-retained K-space line number.

4. The magnetic resonance image reconstruction method of claim 1, wherein, The interpolation method in the interpolation of the randomized undersampled K-space data and the interpolation of the second undersampled training image is a bilinear interpolation method.

5. The magnetic resonance image reconstruction method of claim 1, wherein, The step of obtaining the second undersampled training image according to a plurality of times of second down-sampling of the first undersampled training image comprises: Obtain a first feature map according to convolution operation of the first undersampled training image; Obtain a second feature map according to batch normalization processing of the first feature map; Obtain a third feature map according to nonlinear transformation of the second feature map; Obtain a fourth feature map according to pooling processing of the third feature map; According to a preset number of downsampling performed on the fourth feature map as input data of a next down-sampling processing, the second under-sampling training image is obtained.

6. The magnetic resonance image reconstruction method of claim 5, wherein, According to an activation function in the nonlinear transformation performed on the second feature map, the activation function is a SiLU activation function.

7. The magnetic resonance image reconstruction method of claim 1, wherein, The step of obtaining the image domain training image according to the image fusion performed on the interpolated second training under-sampling image and the first under-sampling training image comprises: The image domain training image is obtained according to the same-position pixel value average calculation performed on the interpolated second training under-sampling image and the first under-sampling training image.

8. The magnetic resonance image reconstruction method of claim 1, wherein, The step of obtaining the training image according to the weighted fusion performed on the frequency domain training image and the image domain training image comprises: The training image is obtained according to the same-position pixel value weighted average calculation performed on the frequency domain training image and the image domain training image.

9. The magnetic resonance image reconstruction method of claim 1, wherein, The loss function in the loss calculation performed on the training image and the full-sampling image is a mean square error loss function.

10. A magnetic resonance image reconstruction apparatus, characterized by The device comprises: a data acquisition module configured to acquire under-sampling K-space data to be reconstructed; a model acquisition module configured to acquire an image reconstruction model; an image reconstruction module configured to input the under-sampling K-space data to be reconstructed into the image reconstruction model to obtain a magnetic resonance reconstruction image, wherein the image reconstruction model is obtained through an iterative training process as follows, with the objective of minimizing a model training loss value: acquire full-sampling K-space data; obtain a full-sampling image according to inverse Fourier transform performed on the full-sampling K-space data; obtain under-sampling K-space data corresponding to the full-sampling K-space data according to first down-sampling performed on the full-sampling K-space data; obtain randomized under-sampling K-space data according to randomization performed on the under-sampling K-space data corresponding to the full-sampling K-space data; obtain interpolated K-space data according to interpolation performed on the randomized under-sampling K-space data; obtain a frequency domain training image according to inverse Fourier transform performed on the interpolated K-space data; obtain a first under-sampling training image according to inverse Fourier transform performed on the under-sampling K-space data corresponding to the full-sampling K-space data; obtain a second under-sampling training image according to a plurality of times of second down-sampling performed on the first under-sampling training image; obtain an interpolated second training under-sampling image according to interpolation performed on the second under-sampling training image; obtain an image domain training image according to image fusion performed on the interpolated second training under-sampling image and the first under-sampling training image; obtain a training image according to weighted fusion performed on the frequency domain training image and the image domain training image; obtain a model training loss value according to loss calculation performed on the training image and the full-sampling image.

11. A magnetic resonance imaging system, characterized by The magnetic resonance imaging system comprises an imaging controller; the imaging controller comprises a processor and a memory, the processor and the memory are coupled, the memory is used to store a computer program or instructions, the processor is used to execute the computer program or instructions in the memory, so that the imaging controller executes the steps of the magnetic resonance image reconstruction method as claimed in any one of claims 1-9.

12. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the magnetic resonance image reconstruction method as claimed in any one of claims 1-9.

Images