Magnetic resonance image reconstruction method and device, imaging system and storage medium

Abstract

The invention provides a magnetic resonance image reconstruction method and device, an imaging system and a storage medium, and the method comprises the steps: inputting to-be-reconstructed undersampling K space data into an image reconstruction model, and obtaining a reconstructed image; and the image reconstruction model is trained by taking model training loss value minimization as a target: performing inverse Fourier transform and down-sampling on the full-sampling K-space data to obtain a full-sampling image and under-sampling K-space data, and performing randomization, interpolation and inverse Fourier transform on the under-sampling K-space data in sequence to obtain a frequency domain training image; and fusing an undersampling training image obtained by inverse Fourier transform of the undersampling K space data and an undersampling image obtained by sequentially carrying out multiple times of second downsampling and interpolation on the undersampling training image to obtain an image domain training image, and calculating a loss value of the training image obtained by fusing the frequency domain and the image domain and the full-sampling image. According to the invention, through adaptive filling and reconstruction of the image domain and the frequency domain, efficient, rapid and high-quality magnetic resonance image recovery is realized.

Description

Technical Field

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

[0002] Magnetic Resonance Imaging (MRI) is a radiation-free medical imaging technology with high soft tissue resolution, which is widely used in neurology, cardiology, osteoarthritis and other fields. However, since MRI scanning usually relies on data acquisition in K space, the scanning process is slow and the patient needs to remain still in the instrument for a long time. Long scanning time is not only prone to motion artifacts, but also increases costs and reduces the real-time performance of imaging. 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 methods are 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 such as compressed sensing (CS) and low-rank matrix recovery mainly use the sparsity and redundancy of the signal to achieve image reconstruction by optimizing the solution of the undersampling problem. However, the mathematical optimization calculation cost is high, and it is difficult to accurately model the complex image structure and signal noise, and it is difficult to ensure the quality and efficiency of reconstruction at the same time. At the same time, the reconstruction based on machine learning models is mainly divided into two strategies: image domain reconstruction that directly processes the undersampled image in the image space to remove artifacts and restore details, and frequency domain reconstruction that directly fills or interpolates the undersampled data in the K space and then reconstructs the image. Although the image domain reconstruction strategy has a strong feature extraction capability in the image space, it is difficult to fully utilize the frequency information distribution law, resulting in the loss of some high-frequency details. In addition, although the frequency domain reconstruction strategy can better utilize the structural information of the K space, it is difficult to effectively handle the non-locality, data sparsity and complex phase relationship of the K space data, which limits its application. Summary of the invention

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

[0005] In a first aspect, an embodiment of the present invention provides a method for reconstructing a magnetic resonance image, the method comprising the following steps:

[0006] Acquiring undersampled K-space data to be reconstructed;

[0007] Obtaining an image reconstruction model;

[0008] The undersampled K-space data to be reconstructed is input into the image reconstruction model to obtain a magnetic resonance reconstructed image; wherein the image reconstruction model is obtained through the following iterative training process based on minimizing the model training loss value:

[0009] Acquire full-sampled K-space data;

[0010] Obtaining a fully sampled image according to an inverse Fourier transform performed on the fully sampled K-space data;

[0011] Obtaining under-sampled K-space data corresponding to the fully sampled K-space data according to a first down-sampling performed on the fully sampled K-space data;

[0012] Obtaining randomized under-sampled K-space data according to randomization of the under-sampled K-space data corresponding to the fully-sampled K-space data;

[0013] Obtaining interpolated K-space data according to interpolation performed on the randomized under-sampled K-space data;

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

[0015] Obtaining a first under-sampled training image according to an inverse Fourier transform performed on the under-sampled K-space data corresponding to the fully-sampled K-space data;

[0016] Obtaining a second under-sampled training image according to a plurality of second down-samplings performed on the first under-sampled training image;

[0017] Obtaining an interpolated second training under-sampled image according to interpolation performed on the second under-sampled training image;

[0018] Obtaining an image domain training image by performing image fusion on the interpolated second training under-sampled image and the first under-sampled training image;

[0019] Obtaining a training image by weighted fusion of the frequency domain training image and the image domain training image;

[0020] A model training loss value is obtained based on the loss calculation performed on the training image and the fully sampled image.

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

[0022] A data acquisition module, used for acquiring under-sampled K-space data to be reconstructed;

[0023] A model acquisition module, used for acquiring an image reconstruction model;

[0024] An image reconstruction module is used to input the undersampled K-space data to be reconstructed into the image reconstruction model to obtain a magnetic resonance reconstructed image; wherein the image reconstruction model is obtained through the following iterative training process based on minimizing the model training loss value:

[0025] Acquire full-sampled K-space data;

[0026] Obtaining a fully sampled image according to an inverse Fourier transform performed on the fully sampled K-space data;

[0027] Obtaining under-sampled K-space data corresponding to the fully sampled K-space data according to a first down-sampling performed on the fully sampled K-space data;

[0028] Obtaining randomized under-sampled K-space data according to randomization of the under-sampled K-space data corresponding to the fully-sampled K-space data;

[0029] Obtaining interpolated K-space data according to interpolation performed on the randomized under-sampled K-space data;

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

[0031] Obtaining a first under-sampled training image according to an inverse Fourier transform performed on the under-sampled K-space data corresponding to the fully-sampled K-space data;

[0032] Obtaining a second under-sampled training image according to a plurality of second down-samplings performed on the first under-sampled training image;

[0033] Obtaining an interpolated second training under-sampled image according to the interpolation performed on the second under-sampled training image;

[0034] Obtaining an image domain training image by performing image fusion on the interpolated second under-sampled training image and the first under-sampled training image;

[0035] Obtaining a training image by weighted fusion of the frequency domain training image and the image domain training image;

[0036] A model training loss value is obtained based on the loss calculation performed on the training image and the fully sampled image.

[0037] In a third aspect, an embodiment of the present invention further provides a magnetic resonance imaging system, comprising 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, the processor is used to execute the computer programs or instructions in the memory, so that the imaging controller performs the steps of the above-mentioned magnetic resonance image reconstruction method.

[0038] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the above-mentioned magnetic resonance image reconstruction method when executed by a processor.

[0039] Compared with the prior art, the magnetic resonance image reconstruction method provided by the present invention disrupts the specific sampling mode of K-space data, destroys the original sampling structure of K-space points, breaks the structural characteristics of artifacts and increases data diversity by performing random operations on under-sampled K-space data. It can not only reduce the risk of overfitting of the model, improve the generalization ability of the model and the ability to capture global feature consistency, enhance the robustness of the model, but also more effectively identify and suppress artifacts to ensure the frequency domain reconstruction effect; at the same time, by performing multiple repeated downsampling operations on image domain image features, it is ensured that the spatial characteristics of the under-sampled image are effectively captured, and important anatomical structure information is provided for the model. The obtained frequency domain features and image domain features are weightedly 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 of image domain data and frequency domain data, efficient, fast and high-quality magnetic resonance image reconstruction is achieved, thereby effectively improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a schematic flow chart of a magnetic resonance image reconstruction method according to an embodiment of the present invention;

[0041] Figure 2 Schematic diagram of the iterative training process of the image reconstruction model in an embodiment of the present invention;

[0042] Figure 3 is a schematic diagram of the first downsampling effect in an embodiment of the present invention;

[0043] Figure 4 is a schematic diagram of randomization results of under-sampled K-space data in an embodiment of the present invention;

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

[0045] Figure 6is a schematic diagram of the second downsampling process in an embodiment of the present invention;

[0046] Figure 7 Schematic diagram of SiLU activation function in an embodiment of the present invention;

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

[0048] Fig. 9 Schematic diagram of the structure of a magnetic resonance image reconstruction system in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical scheme and beneficial effects of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Obviously, the embodiments described below are part of the embodiments of the present invention and are only used to illustrate the present invention, but are not used to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

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

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

[0052] S1. Acquire under-sampled K-space data to be reconstructed; wherein, the under-sampled K-space data to be reconstructed can be understood as K-space data obtained by reducing the sampling points of K-space in order to shorten the acquisition time and need to be directly used for magnetic resonance image reconstruction. The specific acquisition method can be obtained through existing acquisition methods and will not be described in detail here.

[0053] S2. Input the under-sampled K-space data to be reconstructed into the image reconstruction model to obtain a magnetic resonance reconstructed image; wherein the image reconstruction model can be understood as a network model that can directly process the input under-sampled K-space data to be reconstructed to obtain a full-space magnetic resonance reconstructed image.

[0054] In order to achieve efficient and rapid reconstruction of magnetic resonance images based on undersampled K-space data while effectively ensuring the reconstruction quality of magnetic resonance images, the image reconstruction model is designed and improved by constructing adaptive filling reconstruction in both the image domain and the frequency domain; in some embodiments, such as Figure 2 As shown in the figure, the image reconstruction model is based on minimizing the model training loss value and is obtained through the following iterative training process:

[0055] Acquire full-sampled K-space data; the full-sampled K-space data can be understood as the full K-space data with frequency domain information obtained by using a nuclear magnetic resonance device to collect magnetic resonance signals comprehensively and without omission, and converting these signals into digital signals and completely filling them in the K-space, which is recorded as

[0056] According to the inverse Fourier transform of the fully sampled K-space data, a fully sampled image is obtained; wherein the inverse Fourier transform (IFFT) corresponds to the Fourier transform, and the process of transforming the frequency domain data back to the time domain data can be expressed as F -1 Represents the inverse Fourier transform. The fully sampled image X obtained by inverse Fourier transform is the same as the fully sampled K space data The image is composed of pixels of the same size as , and the pixel value range of each pixel is 0 to 255. The fully sampled image X obtained here is used as a sample label for calculating the model training loss in subsequent model training.

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

[0058] In some embodiments, Figure 3 As shown, the step of obtaining the under-sampled K-space data corresponding to the fully sampled K-space data according to the first down-sampling of the fully sampled K-space data comprises:

[0059] According to the preset downsampling multiple and the number of K-space lines in the fully sampled K-space data, the number of sampled and retained K-space lines is obtained; wherein the preset downsampling multiple can be determined according to actual application requirements; in order to fit the actual undersampling effect as much as possible, in some embodiments, the preset downsampling multiple is set to 8 times. The number of K-space lines in the fully sampled K-space data is determined according to the actual sampled data size. Assuming that the size of the fully sampled K-space data is 256*256, the number of K-space lines is 256.

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

[0061] According to the number of sampled and retained K-space lines, the K-space lines outside the preset center reserved area of ​​the fully sampled K-space data are randomly deleted to obtain the undersampled K-space data; wherein the preset center reserved area can be understood as an area that is pre-set according to application requirements and is located in the center of the fully sampled K-space data and the data remains unchanged during the downsampling process. The specific number of K-space lines included in the area is not specifically limited. Corresponding undersampled K-space data The acquisition process can be understood as first calculating the difference between the number of sampled and retained K-space lines and the K-space lines in the preset central reserved area, and then randomly deleting the K-space lines corresponding to the difference in the area other than the preset central reserved area in the full sampled K-space data, and finally obtaining an undersampled K-space data with the number of sampled and retained K-space lines.

[0062] It should be noted that, in practical applications, the corresponding under-sampled K-space data obtained by down-sampling the fully-sampled K-space data can also be obtained by adjusting the sampling acceleration multiple of the nuclear magnetic resonance device.

[0063] According to the randomization of the under-sampled K-space data corresponding to the fully sampled K-space data, randomized under-sampled K-space data is obtained; wherein the under-sampled K-space data is randomized as follows: Figure 4 The figure can be understood as rearranging the already filled K-space data, that is, expanding the sampled data into a one-dimensional array and randomly shuffling the order, and then rearranging it into the original sampling shape, so that the shuffled sampling points are applied to the K-space data, so that the original under-sampled K-space data is randomly distributed on the K-space map, and the K-space image size remains unchanged before and after the randomization. For the frequency domain, this randomization processing method is similar to redesigning the filling method of the K-space, that is, selectively selecting which positions on the K-space need to be filled and which positions are not necessary to be filled, so as to simulate as much as possible the effect of selectively filling the K-space data sampling points, and this random effect is difficult to obtain by manipulating the magnetic resonance device.

[0064] By using the randomized under-sampled K-space data obtained based on the above randomization processing as the basic data for frequency domain reconstruction, not only can the specific sampling pattern of the K-space data be disrupted to avoid overfitting of the model to a certain fixed sampling pattern during model training, thereby improving the generalization ability of the model and enhancing the robustness of the model, but also the randomization operation is equivalent to "data enhancement" processing of the data. For the same under-sampled K-space data, different randomization schemes will produce different training samples. By increasing the data diversity, the risk of model overfitting can be further reduced. By destroying the original sampling structure of the K-space point, the information of the surrounding points must be comprehensively considered during interpolation and 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 characteristics, thereby improving the model's ability to capture global consistency while breaking the structural characteristics of the artifact, so that the artifact can be more effectively identified and suppressed during the model training process, ensuring the generation of clearer reconstructed images.

[0065] According to the interpolation performed on the randomized under-sampled K-space data, the interpolated K-space data is obtained; wherein the interpolation process can be understood as estimating and filling the unknown sampling point data based on the existing sampling point data in the under-sampled K-space data; in order to ensure the data filling effect, in some embodiments, the interpolation method used for the randomized under-sampled K-space data is a bilinear interpolation method, for each pixel point of the target image, the four nearest neighboring points on the original image are found, and its pixel value is obtained by interpolation calculation based on the neighboring points, and the pixel value is filled to the corresponding position in the K-space, so that all places without data points in the K-space are filled, and the following is obtained. Figure 5 All points in the K space shown have numerical interpolated K space data, where the blue points are original sampling points in the K space and the red points are filled points after bilinear interpolation.

[0066] By performing bilinear interpolation on the randomized undersampled K-space data, it is possible to reduce information loss while also reducing the abrupt boundary effect between the undersampled area and the sampled area, so that the interpolated K-space data has good smoothness, which can provide a better initial solution for subsequent model training, making the gradient smoother and the convergence more stable during model training.

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

[0068] A first under-sampled training image is obtained by performing an inverse Fourier transform on the under-sampled K-space data corresponding to the fully-sampled K-space data.

[0069] A second under-sampled training image is obtained based on several times of second down-sampling performed on the first under-sampled training image; wherein, the specific number of second down-sampling performed on the first under-sampled training image can be set according to the amount of data in actual applications, different sampling methods or other factors so that the subsequent model training achieves the best effect; in some embodiments, the number of second down-sampling is set to 3 times, that is, the second down-sampling is performed on the first under-sampled training image continuously for 3 times.

[0070] In some embodiments, Figure 6 As shown, the second downsampling process includes convolution, batch normalization, activation function and pooling processes in sequence; correspondingly, the step of obtaining the second under-sampled training image according to the second downsampling performed several times on the first under-sampled training image includes:

[0071] According to the convolution operation performed on the first under-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 the actual situation, and in some embodiments, a convolution process with a convolution kernel of 3*3 and a step size of 1 is selected; the corresponding first feature map is consistent with the size of the first under-sampled training image. It should be noted that the convolution operation here can adopt the convolution operation in existing deep learning.

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

[0073] According to the nonlinear transformation of 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 complex data distribution and features and effectively improve the expression ability, in some embodiments, the activation function in the nonlinear transformation is set to the SiLU activation function. Figure 7 As shown, the SiLU activation function has smoothness and nonlinear characteristics, and is more effective than other activation functions in dealing with gradient disappearance and explosion problems. It is defined as follows:

[0074]

[0075] Where 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 process performed on the third feature map, a fourth feature map is obtained; wherein, the pooling process can be selected from the existing maximum pooling, average pooling, global pooling, random pooling and mixed pooling methods according to actual needs, and is not specifically limited here; the corresponding fourth feature map obtained by the pooling process can be understood as a two-dimensional array, which contains various features extracted from the input third feature map. The role of the pooling process is to reduce the height and width of the feature map, reduce the complexity and amount of calculation 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, improve the calculation efficiency, and reduce the risk of overfitting of the model; the corresponding fourth feature map is consistent with the number of channels of the first under-sampled training image, but the height and width are 1 / 2 of the height and width of the first under-sampled training image.

[0077] The second under-sampled training image is obtained by downsampling the fourth feature map for a preset number of times as input data for the next downsampling process; wherein the preset number of downsampling can be understood as the remaining consecutive downsampling times except for the first downsampling; the processing process of the second downsampling performed each time is the same, and the above description can be referred to. Assuming that the second downsampling process is performed 3 times in total, the size of the second under-sampled training image finally obtained is 1 / 8 of the first under-sampled training image. If the size of the original first under-sampled training image is 256*256, the size of the second under-sampled training image is 32*32.

[0078] According to the interpolation of the second under-sampled training image, an interpolated second under-sampled training image is obtained; wherein the interpolated second under-sampled training image can be understood as the height and width of the second under-sampled training image being proportionally enlarged to the size before downsampling, so that the height, width and number of channels of the image are consistent with those of the first under-sampled training image. In order to ensure the interpolation effect, in some embodiments, the interpolation method in the interpolation of the second under-sampled training image is also set to bilinear interpolation to ensure the smoothness of the interpolated image.

[0079] According to the image fusion of the interpolated second training under-sampled image and the first under-sampled training image, an image domain training image is obtained; wherein image fusion can be understood as fusing pixel values ​​at the same pixel position in the interpolated second training under-sampled image and the first under-sampled training image. In some embodiments, the step of obtaining the image domain training image according to the image fusion of the interpolated second training under-sampled image and the first under-sampled training image includes:

[0080] The image domain training image is obtained by averaging the pixel values ​​at the same position of the interpolated second training under-sampled image and the first under-sampled training image; that is, the pixel values ​​at each corresponding position of the interpolated second training under-sampled image and the first under-sampled training image are added and averaged to obtain the required image domain training image.

[0081] The above-mentioned image domain training image acquisition method can effectively capture the spatial features of the undersampled image by performing multiple repeated downsampling operations, providing important anatomical structure information for the model, and further providing reliable support for the reconstruction effect of the model.

[0082] The training image is obtained by weighted fusion of the frequency domain training image and the image domain training image; wherein weighted fusion can be understood as weighted fusion processing of pixel values ​​at 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 by weighted fusion of the frequency domain training image and the image domain training image comprises:

[0083] The training image is obtained by weighted averaging the pixel values ​​at the same position of the frequency domain training image and the image domain training image; that is, the pixel values ​​at each corresponding position of the frequency domain training image and the image domain training image are multiplied by the weight, and then added and averaged to obtain the required training image, that is, based on the undersampled K-space data The reconstructed image Z is obtained.

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

[0085] According to the loss calculation performed on the training image and the fully sampled image, a model training loss value is obtained; wherein the model training loss value can be understood as a function value for evaluating the error between the reconstructed image obtained after model processing and the real image; to ensure the simplicity and efficiency of the model training loss value calculation, in some embodiments, the loss function in the loss calculation performed on the training image and the fully sampled image is a mean square error loss function, which is expressed as:

[0086]

[0087] In the formula, Represents the reconstruction loss value of the model; N represents the number of samples; X i and Z i They respectively represent the fully sampled image (true target value) and training image (model reconstructed image) corresponding to the i-th undersampled K-space data.

[0088] It should be noted that the above model training process only gives the processing process of a training sample. In practical applications, it is necessary to collect multiple sets of fully sampled K-space data for training based on the above method. In this way, an image reconstruction model based on under-sampled K-space data can be obtained to achieve efficient, fast and high-quality magnetic resonance image restoration, and it can ensure that the image reconstruction model finally trained has reliable reconstruction effect and generalization ability.

[0089] In addition, in order to ensure that the weighted fusion of frequency domain training images and image domain training images in the image reconstruction model processing can adaptively meet the high-quality reconstruction requirements of different under-sampled data, in some embodiments, the parameters updated 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 images and the image domain training images.

[0090] After the image reconstruction model is obtained through the above method steps, the undersampled K-space data actually collected by the 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 invention, in some embodiments, the full-sampled K-space data is set to be downsampled 8 times to obtain undersampled K-space data, and the undersampled K-space data is subjected to inverse Fourier transform to obtain the first undersampled training image, and the second downsampling process is performed three times in a row. The image reconstruction model is constructed based on this condition, and the following is obtained: Figure 8 The reconstructed image shown; Figure 8 It can be seen from the comparison effect that the magnetic resonance reconstructed image obtained by the method of the present invention is almost the same as the real full-sampling image, which is sufficient to prove that the method of the present invention can not only ensure the high efficiency of image reconstruction, but also ensure that the reconstructed image meets the high quality requirements.

[0091] A magnetic resonance image reconstruction method provided by an embodiment of the present invention disrupts the specific sampling pattern of K-space data, destroys the original sampling structure of K-space points, breaks the structural characteristics of artifacts and increases data diversity by performing random operations on under-sampled K-space data. This method can not only reduce the risk of overfitting of the model, improve the generalization ability of the model and the ability to capture global feature consistency, and enhance the robustness of the model, but also more effectively identify and suppress artifacts to ensure the frequency domain reconstruction effect. At the same time, by performing multiple repeated downsampling operations on image domain image features to ensure that the spatial characteristics of the under-sampled image are effectively captured and important anatomical structure information is provided to the model, the obtained frequency domain features and image domain features are weightedly 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 of image domain data and frequency domain data, efficient, fast and high-quality magnetic resonance image reconstruction is achieved, thereby effectively improving the user experience.

[0092] It should be noted that although the steps in the above flowchart are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders.

[0093] like Fig. 9 As shown, the present invention provides a magnetic resonance image reconstruction device, the device comprising:

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

[0095] Model acquisition module 2, used to acquire the image reconstruction model;

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

[0097] Acquire full-sampled K-space data;

[0098] Obtaining a fully sampled image according to an inverse Fourier transform performed on the fully sampled K-space data;

[0099] Obtaining under-sampled K-space data corresponding to the fully sampled K-space data according to a first down-sampling performed on the fully sampled K-space data;

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

[0101] Obtaining interpolated K-space data according to interpolation performed on the randomized under-sampled K-space data;

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

[0103] Obtaining a first under-sampled training image according to an inverse Fourier transform performed on the under-sampled K-space data corresponding to the fully-sampled K-space data;

[0104] Obtaining a second under-sampled training image according to a plurality of second down-samplings performed on the first under-sampled training image;

[0105] Obtaining an interpolated second training under-sampled image according to interpolation performed on the second under-sampled training image;

[0106] Obtaining an image domain training image by performing image fusion on the interpolated second training under-sampled image and the first under-sampled training image;

[0107] Obtaining a training image by weighted fusion of the frequency domain training image and the image domain training image;

[0108] A model training loss value is obtained based on the loss calculation performed on the training image and the fully sampled image.

[0109] For the specific definition of the magnetic resonance image reconstruction system, please refer to the definition of the magnetic resonance image reconstruction method above, and the corresponding technical effects can also be obtained equivalently, which will not be repeated here. Each module in the above-mentioned magnetic resonance image reconstruction system can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0110] In some embodiments, the present invention provides a magnetic resonance imaging system, characterized in that the magnetic resonance imaging system includes an imaging controller; the imaging controller includes 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 program or instructions in the memory, so that the imaging controller performs the steps of the above-mentioned magnetic resonance image reconstruction method.

[0111] In some embodiments, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the above-mentioned magnetic resonance image reconstruction method are implemented.

[0112] In summary, the magnetic resonance image reconstruction method, device, imaging system and storage medium provided by the embodiments of the present invention disrupt 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 data diversity by performing random operations on under-sampled K-space data. This can not only reduce the risk of overfitting of the model, improve the generalization ability of the model and the ability to capture global feature consistency, enhance the robustness of the model, but also more effectively identify and suppress artifacts to ensure the frequency domain reconstruction effect; at the same time, by performing multiple repeated downsampling operations on image domain image features to ensure that the spatial characteristics of the under-sampled image are effectively captured and important anatomical structure information is provided to the model, the frequency domain features and image domain features are weightedly 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 of image domain data and frequency domain data, efficient, fast and high-quality magnetic resonance image reconstruction is achieved, thereby effectively improving the user experience.

[0113] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be directly referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0114] The above-mentioned embodiments only express several preferred implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principle of the present invention, and these improvements and substitutions should also be regarded as the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the protection scope of the claims.

Claims

1. A magnetic resonance image reconstruction method, characterized in that: The method comprises the following steps: Acquiring undersampled K-space data to be reconstructed; Obtaining an image reconstruction model; The undersampled K-space data to be reconstructed is input into the image reconstruction model to obtain a magnetic resonance reconstructed image; wherein the image reconstruction model is obtained through the following iterative training process based on minimizing the model training loss value: Acquire full-sampled K-space data; Obtaining a fully sampled image according to an inverse Fourier transform performed on the fully sampled K-space data; Obtaining under-sampled K-space data corresponding to the fully sampled K-space data according to a first down-sampling performed on the fully sampled K-space data; Obtaining randomized under-sampled K-space data according to randomization of the under-sampled K-space data corresponding to the fully-sampled K-space data; Obtaining interpolated K-space data according to interpolation performed on the randomized under-sampled K-space data; Obtaining a frequency domain training image according to an inverse Fourier transform performed on the interpolated K-space data; Obtaining a first under-sampled training image according to an inverse Fourier transform performed on the under-sampled K-space data corresponding to the fully-sampled K-space data; Obtaining a second under-sampled training image according to a plurality of second down-samplings performed on the first under-sampled training image; Obtaining an interpolated second training under-sampled image according to interpolation performed on the second under-sampled training image; Obtaining an image domain training image by performing image fusion on the interpolated second training under-sampled image and the first under-sampled training image; Obtaining a training image by weighted fusion of the frequency domain training image and the image domain training image; A model training loss value is obtained based on the loss calculation performed on the training image and the fully sampled image.

2. The magnetic resonance image reconstruction method according to claim 1, characterized in that: The parameters updated in the iterative training of the image reconstruction model include fusion weights in the weighted fusion of the frequency domain training image and the image domain training image.

3. The magnetic resonance image reconstruction method according to claim 1, characterized in that: The step of obtaining the under-sampled K-space data corresponding to the fully sampled K-space data according to the first down-sampling of the fully sampled K-space data comprises: Obtaining the number of sampled and retained K-space lines according to a preset downsampling multiple and the number of K-space lines in the fully sampled K-space data; The K-space lines outside a preset central reserved area are randomly deleted from the fully sampled K-space data according to the number of sampled reserved K-space lines to obtain the under-sampled K-space data.

4. The magnetic resonance image reconstruction method according to claim 1, characterized in that: The interpolation method in the interpolation performed on the randomized under-sampled K-space data and the interpolation performed on the second under-sampled training image is a bilinear interpolation method.

5. The magnetic resonance image reconstruction method according to claim 1, characterized in that: The step of obtaining a second under-sampled training image according to a plurality of second down-samplings performed on the first under-sampled training image comprises: Obtaining a first feature map according to a convolution operation performed on the first under-sampled training image; Obtaining a second feature map according to batch normalization processing performed on the first feature map; Obtaining a third characteristic map according to a nonlinear transformation performed on the second characteristic map; Obtaining a fourth feature map according to the pooling process performed on the third feature map; The second under-sampled training image is obtained by downsampling a preset number of times using the fourth feature map as input data for the next downsampling process.

6. The magnetic resonance image reconstruction method according to claim 5, characterized in that: The activation function in the nonlinear transformation performed on the second feature map is a SiLU activation function.

7. The magnetic resonance image reconstruction method according to claim 1, characterized in that: The step of obtaining an image domain training image by performing image fusion on the interpolated second training under-sampled image and the first under-sampled training image comprises: The image domain training image is obtained by averaging the pixel values ​​at the same position of the interpolated second under-sampled training image and the first under-sampled training image.

8. The magnetic resonance image reconstruction method according to claim 1, characterized in that: The step of obtaining the training image by weighted fusion of the frequency domain training image and the image domain training image comprises: The training image is obtained by calculating the weighted average of the pixel values ​​at the same position of the frequency domain training image and the image domain training image.

9. The magnetic resonance image reconstruction method according to claim 1, characterized in that: The loss function in the loss calculation based on the training image and the fully sampled image is a mean square error loss function.

10. A magnetic resonance image reconstruction device, characterized in that: The device comprises: A data acquisition module, used for acquiring under-sampled K-space data to be reconstructed; A model acquisition module, used for acquiring an image reconstruction model; An image reconstruction module is used to input the undersampled K-space data to be reconstructed into the image reconstruction model to obtain a magnetic resonance reconstructed image; wherein the image reconstruction model is obtained through the following iterative training process based on minimizing the model training loss value: Acquire full-sampled K-space data; Obtaining a fully sampled image according to an inverse Fourier transform performed on the fully sampled K-space data; Obtaining under-sampled K-space data corresponding to the fully sampled K-space data according to a first down-sampling performed on the fully sampled K-space data; Obtaining randomized under-sampled K-space data according to randomization of the under-sampled K-space data corresponding to the fully-sampled K-space data; Obtaining interpolated K-space data according to interpolation performed on the randomized under-sampled K-space data; Obtaining a frequency domain training image according to an inverse Fourier transform performed on the interpolated K-space data; Obtaining a first under-sampled training image according to an inverse Fourier transform performed on the under-sampled K-space data corresponding to the fully-sampled K-space data; Obtaining a second under-sampled training image according to a plurality of second down-samplings performed on the first under-sampled training image; Obtaining an interpolated second training under-sampled image according to interpolation performed on the second under-sampled training image; Obtaining an image domain training image by performing image fusion on the interpolated second training under-sampled image and the first under-sampled training image; Obtaining a training image by weighted fusion of the frequency domain training image and the image domain training image; A model training loss value is obtained based on the loss calculation performed on the training image and the fully sampled image.

11. A magnetic resonance imaging system, characterized in that: The magnetic resonance imaging system includes an imaging controller; the imaging controller includes 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 program or instructions in the memory, so that the imaging controller performs the steps of the magnetic resonance image reconstruction method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the magnetic resonance image reconstruction method according to any one of claims 1 to 9 are implemented.

Images