A magnetic resonance image reconstruction method and system based on deep learning

Abstract

The present invention provides a deep learning-based magnetic resonance image reconstruction method and system. The method acquires magnetic resonance image data consisting of multiple layers of continuous undersampled slice images, uses the slice image to be reconstructed as the target image, and inputs the magnetic resonance image data into a pretrained neural network model to obtain a reconstructed magnetic resonance image. The neural network model includes a slice fusion module, a slice translation module, a channel splicing module, and an image reconstruction network module, which are connected in sequence. Leveraging the strong structural and phase correlations between consecutive slices, the neural network model is used to reconstruct the magnetic resonance image, reducing artifacts caused by undersampling in the magnetic resonance data. This eliminates the need for additional calibration data, reduces magnetic resonance scanning time, and improves the flexibility of spatial encoding.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a magnetic resonance image reconstruction method and system based on deep learning. Background Art

[0002] In the field of medical imaging, magnetic resonance imaging (MRI) is a non-invasive imaging technique that enables high-resolution, high-contrast imaging of the human body and is now widely used in clinical diagnosis. However, high-quality MR images often require extensive data acquisition and long scanning times. K-space undersampling, particularly uniform k-space undersampling, is often used to accelerate scanning. However, traditional MRI techniques based on uniform undersampling, such as parallel imaging, require additional calibration data (e.g., CSM or ACSs) for reconstruction, reducing scanning efficiency and spatial encoding flexibility. Summary of the Invention

[0003] Based on the above problems, the present invention provides a magnetic resonance image reconstruction method and system based on deep learning, aiming to solve technical problems in the prior art such as low magnetic resonance scanning efficiency and inflexible spatial encoding.

[0004] A magnetic resonance image reconstruction method based on deep learning, comprising:

[0005] Step A1, acquiring magnetic resonance image data consisting of multiple layers of continuous undersampled slice images, and using the slice image to be reconstructed in the magnetic resonance image data as a target image;

[0006] Step A2: inputting the magnetic resonance image data into a pre-trained neural network model, wherein the neural network model includes a slice fusion module, a slice translation module, a channel stitching module, and an image reconstruction network module connected in sequence, wherein the channel stitching module is further connected to the slice fusion module;

[0007] Step A3: The slice fusion module fuses all slice images in the magnetic resonance image data and outputs a first image;

[0008] Step A4: the slice translation module performs a translation operation on the first image to output a second image;

[0009] Step A5: The channel stitching module stitches the target image, the first image, and the second image to obtain a third image;

[0010] Step A6: The image reconstruction network module processes the third image and outputs a magnetic resonance reconstructed image corresponding to the target image.

[0011] Furthermore, in step A2, the pre-training process of the neural network model includes:

[0012] Step B1, obtaining unreconstructed multi-layer continuous fully sampled slice K-space data as fully sampled sample data;

[0013] Step B2, performing undersampling processing on the fully sampled sample data to obtain undersampling sample data;

[0014] Step B3, performing an inverse Fourier transform on the fully sampled sample data to obtain a fully sampled sample image, and performing an inverse Fourier transform on the undersampled sample data to obtain an undersampled sample image;

[0015] In step B4, the under-sampled sample images are used as training data and the fully sampled sample images are used as labels to train the pre-established neural network to obtain a neural network model.

[0016] Furthermore, step A4 includes:

[0017] Step A41, determining an offset for the first image data;

[0018] Step A42: cyclically shift the first image data in a predetermined direction according to the offset to obtain second image data.

[0019] Furthermore, in step A41, the offset is determined according to the undersampling factor and the field of view of the magnetic resonance scan;

[0020] The offset is calculated as follows:

[0021] D = FOV*(Rn) / R;

[0022] in,

[0023] D represents the offset;

[0024] FOV represents the field of view of the MRI scan;

[0025] R is the downsampling factor of the slice image;

[0026] n is a positive integer from 1 to R-1.

[0027] Furthermore, in step A6, the image reconstruction network module is a multi-layer residual network module;

[0028] The multi-layer residual network module includes: a first convolutional layer, several convolutional modules, a data consistency module, a data conversion module, and an output module connected in sequence;

[0029] The processing of the third image in step A6 includes:

[0030] Step A61: taking the third image as an input image, sequentially passing the input image through a first convolutional layer and a plurality of convolutional modules, and outputting a first result image;

[0031] Step A62: The data consistency module converts the first result image into corresponding first K-space data, and fills the real sampled K-space data corresponding to the target image into the corresponding position of the first K-space data to obtain second K-space data.

[0032] Step A63: The data conversion module performs inverse Fourier transform on the second K-space data to obtain a second result image;

[0033] Step A64: output the second result image as a magnetic resonance reconstructed image.

[0034] Furthermore, the multi-layer residual network module further includes a judgment module, which is respectively connected to the first convolutional layer, the data conversion module, and the output module;

[0035] After step A63 and before step A64, the method further includes:

[0036] Step A63a: Determine whether the second result image satisfies any one of the first iteration stop condition and the second iteration stop condition:

[0037] If yes, stop the iteration and execute step A64;

[0038] If not, continue with step A61 using the second result image as the input image of the first convolutional layer;

[0039] The first iteration stop condition is: the number of iterations is greater than the preset number;

[0040] The second iteration stopping condition is: the image error between the second result image generated by the current iteration and the corresponding input image is less than a preset value.

[0041] Furthermore, in step A1, the undersampling factor of each slice image is the same, and in continuous slice images, the real sampled K-space phase encoding lines corresponding to two adjacent slice images are different.

[0042] Furthermore, the number of input images of the neural network model is a first predetermined number of layers, and the output of the neural network model is a single-slice magnetic resonance reconstructed image;

[0043] In step A1 , a first predetermined number of continuous under-sampled slice images are acquired to form magnetic resonance image data, and the number of target images is 1.

[0044] Furthermore, the number of input images of the neural network model is a second predetermined number of layers, and the output of the neural network model is a magnetic resonance reconstructed image of a third predetermined number of layers;

[0045] In step A1, a second predetermined number of continuous under-sampled slice images are acquired to form magnetic resonance image data, and the number of target images is a third predetermined number of slices.

[0046] A deep learning-based magnetic resonance image reconstruction system, used in the aforementioned deep learning-based magnetic resonance image reconstruction method, comprising:

[0047] a data acquisition module, configured to acquire magnetic resonance image data consisting of multiple layers of continuous undersampled slice images, and to use the slice image to be reconstructed in the magnetic resonance image data as a target image;

[0048] An image input module, connected to the data acquisition module, is used to input magnetic resonance image data into a pre-trained neural network model;

[0049] An image reconstruction module, connected to the image input module, is used to reconstruct the target image through a neural network model and output a magnetic resonance reconstructed image corresponding to the target image;

[0050] The neural network model includes a slice fusion module, a slice translation module, a channel stitching module and an image reconstruction network module connected in sequence, and the channel stitching module is also connected to the slice fusion module;

[0051] The slice fusion module is used to fuse all slice images in the magnetic resonance image data and output a first image;

[0052] The slice translation module is used to perform a translation operation on the first image and output a second image;

[0053] The channel stitching module is used to stitch the target image, the first image and the second image to obtain a third image;

[0054] The image reconstruction network module is used to process the third image to obtain a magnetic resonance reconstructed image and output it.

[0055] The beneficial technical effects of the present invention are: the present invention reconstructs magnetic resonance images by using a neural network model, reduces artifacts caused by undersampling of magnetic resonance data, reconstructs high-quality magnetic resonance images from undersampled magnetic resonance data without the need for additional calibration data, reduces magnetic resonance scanning time, and improves the flexibility of spatial encoding. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 、 Figure 3-4 and Figure 7 This is a flowchart of the steps of a magnetic resonance image reconstruction method based on deep learning of the present invention;

[0057] Figure 2 、 Figure 5 、 Figure 6 and Figure 9 This is a module schematic diagram of a deep learning-based magnetic resonance image reconstruction system of the present invention;

[0058] Figure 8 This is a flow chart of the neural network model training steps for a deep learning-based magnetic resonance image reconstruction system of the present invention. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0060] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0061] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0062] See also Figure 1 The present invention provides a magnetic resonance image reconstruction method based on deep learning, comprising:

[0063] Step A1, acquiring magnetic resonance image data consisting of multiple layers of continuous undersampled slice images, and using the slice image to be reconstructed in the magnetic resonance image data as a target image;

[0064] Step A2: inputting the magnetic resonance image data into a pre-trained neural network model, wherein the neural network model includes a slice fusion module, a slice translation module, a channel stitching module, and an image reconstruction network module connected in sequence, wherein the channel stitching module is further connected to the slice fusion module;

[0065] Step A3: The slice fusion module fuses all slice images in the magnetic resonance image data and outputs a first image;

[0066] Step A4: the slice translation module performs a translation operation on the first image to output a second image;

[0067] Step A5: The channel stitching module stitches the target image, the first image, and the second image to obtain a third image;

[0068] Step A6: The image reconstruction network module processes the third image and outputs a magnetic resonance reconstructed image corresponding to the target image.

[0069] The present invention utilizes the similarity between adjacent slices and uses a neural network model to reconstruct magnetic resonance images, thereby reducing artifacts caused by undersampling of magnetic resonance data. High-quality magnetic resonance images can be reconstructed from undersampled magnetic resonance data without the need for additional calibration data, reducing magnetic resonance scanning time and improving the flexibility of spatial encoding.

[0070] In step A1, the slice images obtained are continuous slices, and each slice image is undersampled. Furthermore, in step A1, the undersampling factor of each slice image is the same, and in continuous slice images, the real sampled K-space phase encoding lines corresponding to two adjacent slice images are different.

[0071] By using the same downsampling factor, the undersampling of the slice images is uniform. In addition, the undersampling of consecutive slice images is complementary, that is, the summary of the actual scanned k-space phase encoding lines involved in all slice images involves all k-space phase encoding lines of the magnetic resonance scan field of view FOV. Figure 2 As shown, three consecutive undersampled slice images were acquired: slice images of layer i, layer i-1, and layer i+1. The downsampling factor R for the three-layer undersampling was 2. The image reconstruction network module was a multi-layer network residual module. The k-space phase encoding lines actually acquired for layer i were odd-numbered lines 1, 3, and 5. For slice images of layers i-1 and i+1 adjacent to layer i, the k-space phase encoding lines actually acquired were even-numbered lines 2, 4, and 6. Through uniform undersampling and the complement of uniform undersampling, all phase encoding lines in the magnetic resonance scanning field of view (FOV) were acquired. For example, if the uniform undersampling factor is 3, then every other phase encoding line is acquired. If the phase encoding lines actually acquired for layer i are 1, 4, 7, and 10, then for the slice image of layer i-1, the phase encoding lines actually acquired are 2, 5, and 8, and for the slice image of layer i+1, the phase encoding lines actually acquired are 3, 6, and 9. In this way, the phase encoding lines acquired for the slice images of layer i, layer i+1, and layer i+2 encompass the phase encoding lines of all MRI scan fields. This uniform and complementary undersampling method further enhances the accuracy of target image reconstruction.

[0072] In step A1, the slice image to be reconstructed is used as the target image, for example Figure 2The under-sampled slice image of the i-th layer is used as the reconstruction target. When performing image stitching in step A5, the target image is also used as one of the stitching channels. The image data obtained by the slice fusion module and the slice translation module are stitched with the target image. After stitching, the image data is input into the image reconstruction network module to reconstruct the target image, eliminate the artifacts caused by undersampling, and obtain a high-quality magnetic resonance reconstructed image.

[0073] In step A3, the slice fusion module fuses the slice images in the magnetic resonance image data according to the weights and outputs a first image, such as Figure 2 As shown in the figure, the undersampled K-space data of the i, i+1, and i-1 consecutive layers are converted from K-space to image domain through inverse fast Fourier transform IFFT. After obtaining the slice images of three consecutive layers, the i-1, i and i+1 slice images are converted according to the weight w. i-1 、w i and w i+1 The fused image is then joined with the target image at the channel junction to assist in the reconstruction of the target image.

[0074] See also Figure 3 , further, step A4 includes:

[0075] Step A41, determining an offset for the first image data;

[0076] Step A42: cyclically shift the first image data in a predetermined direction according to the offset to obtain second image data.

[0077] Furthermore, in step A41, the offset is determined according to the undersampling factor and the field of view of the magnetic resonance scan;

[0078] The offset is calculated as follows:

[0079] D = FOV*(Rn) / R;

[0080] in,

[0081] D represents the offset;

[0082] FOV represents the field of view of the MRI scan;

[0083] R is the downsampling factor of the slice image;

[0084] n is a positive integer from 1 to R-1.

[0085] The first image obtained after the magnetic resonance image data is fused by the slice fusion module is translated by the slice translation module to form a second image. The second image is further spliced with the target image to assist in the reconstruction of the target image and help remove artifacts during the reconstruction of the target image.

[0086] Different downsampling factors and n values can result in multiple second images. If R is 2, n can only be 1, resulting in a single second image. If the downsampling factor R is greater than 2, for example, 3, and n is 1, 2, or R-1, multiple offsets can be used, resulting in multiple second images with different offsets.

[0087] The offset of the slice translation module is related to the downsampling factor of the slice image and the magnetic resonance scanning field of view. For example, if the downsampling factor of the slice image is 2, the offset is 1 / 2 of the magnetic resonance scanning field of view. After translation, a second image is obtained. Figure 2 In the example, after circular shifting, the lower half of the first image is in the upper half of the second image, and vice versa. If the downsampling factor is 3, there are two offsets: 1 / 3 FOV and 2 / 3 FOV. The first image is circularly shifted by 1 / 3 in the predetermined direction. Each row of data in the first image is sequentially shifted by 1 / 3 in the predetermined direction. The first 1 / 3 rows of data in the first image exceed the front boundary of the first image, the middle 1 / 3 rows of data sequentially occupy the first 1 / 3 rows of the original first image, and the last 1 / 3 rows of data occupy the middle 1 / 3 rows of the original first image. The first 1 / 3 rows of data that exceed the boundary are sequentially attached to the back of the last 1 / 3 rows of data, occupying the last 1 / 3 rows of the original first image, thus forming a second image corresponding to an offset of 1 / 3 FOV. For an offset of 2 / 3 FOV, the 2 / 3 rows of data that exceed the front boundary circularly fill the last 2 / 3 rows of the original first image, thus forming a second image corresponding to an offset of 2 / 3 FOV. The second image at different offsets is spliced with the target image and the first image to assist in the reconstruction of the target image. This fully utilizes the strong similarity in structure and phase between adjacent slice images, suppresses aliasing artifacts introduced by uniform undersampling, and accurately reconstructs the size and phase of the target image without the need for additional calibration information, thereby improving the efficiency and flexibility of magnetic resonance data acquisition.

[0088] See also Figure 4 ,Further, in step A6, the image reconstruction network module is a multi-layer residual network module;

[0089] The multi-layer residual network module includes: a first convolutional layer, several convolutional modules, a data consistency module, a data conversion module, and an output module connected in sequence;

[0090] The processing of the third image in step A6 includes:

[0091] Step A61: taking the third image as an input image, sequentially passing the input image through a first convolutional layer and a plurality of convolutional modules, and outputting a first result image;

[0092] Step A62: The data consistency module converts the first result image into corresponding first K-space data, and fills the real sampled K-space data corresponding to the target image into the corresponding position of the first K-space data to obtain second K-space data.

[0093] Step A63: The data conversion module performs inverse Fourier transform on the second K-space data to obtain a second result image;

[0094] Step A64: output the second result image as a magnetic resonance reconstructed image.

[0095] The spliced third image is optimized and reconstructed through a multi-layer residual network module, and finally a high-quality magnetic resonance reconstructed image without artifacts is output.

[0096] See also Figure 5 The convolution module includes a second convolutional layer, an activation layer, and a third convolutional layer, connected in sequence. The activation function of the activation layer is Reinforced Lu (ReLU), and the input of the second convolutional layer and the output of the third convolutional layer have a jump connection. The data consistency module replaces the corresponding position of the first result image with the actual collected data, ensuring that the reconstructed image and the original target image are consistent with the actual collected data.

[0097] See also Figure 6-7 ,Furthermore, the multi-layer residual network module also includes a judgment module, ,which is respectively connected to the first convolutional layer, the data ,conversion module, and the output module;

[0098] After step A63 and before step A64, the method further includes:

[0099] Step A63a: Determine whether the second result image satisfies any one of the first iteration stop condition and the second iteration stop condition:

[0100] If yes, stop the iteration and execute step A64;

[0101] If not, continue with step A61 using the second result image as the input image of the first convolutional layer;

[0102] The first iteration stop condition is: the number of iterations is greater than the preset number;

[0103] The second iteration stopping condition is: the image error between the second result image generated by the current iteration and the corresponding input image is less than a preset value.

[0104] Specifically, the image error between the second result image generated by the previous iteration and the corresponding input image refers to the average value of the pixel difference between the two images. If it is the first iteration, the input image is the third image; if it is not the first iteration, the input image is the second result image of the previous iteration. Preferably, the image error is 2×10-6 .

[0105] Specifically, the preset number of times is 3.

[0106] join Figure 8 ,Further, in step A2, the pre-training process of the neural network model includes:

[0107] Step B1, obtaining unreconstructed multi-layer continuous fully sampled slice K-space data as fully sampled sample data;

[0108] Step B2, performing undersampling processing on the fully sampled sample data to obtain undersampling sample data;

[0109] Step B3, performing an inverse Fourier transform on the fully sampled sample data to obtain a fully sampled sample image, and performing an inverse Fourier transform on the undersampled sample data to obtain an undersampled sample image;

[0110] In step B4, the under-sampled sample images are used as training data and the fully sampled sample images are used as labels to train the pre-established neural network to obtain a neural network model.

[0111] Fully sampled MRI slice K-space data, serving as fully sampled sample data, can be obtained from the public dataset Calgary-Campinas Public Brain MR Database. The fully sampled slice K-space data in each fully sampled sample data set is a continuous slice. In step B2, each fully sampled sample data set is undersampled using a complementary uniform undersampling method, similar to the uniform undersampling and complementation of the magnetic resonance image data to be reconstructed described above. Uniform undersampling is performed using the same downsampling factor, with phase encoding lines acquired from continuous slices of different layers being different, so that the sum of the phase k-space phase encoding lines involved in the undersampling is ensured to include every phase encoding line in the FOV.

[0112] In step B4, during training, the optimizer uses the Adam optimizer, and the loss function uses the L1 norm, L2 norm, or the weighted sum of the L1 norm and L2 norm.

[0113] Furthermore, the number of input images of the neural network model is a first predetermined number of layers, and the output of the neural network model is a single-slice magnetic resonance reconstructed image;

[0114] In step A1 , a first predetermined number of continuous under-sampled slice images are acquired to form magnetic resonance image data, and the number of target images is 1.

[0115] The neural network model of the present invention can reconstruct one of the target images based on multiple layers of continuous slice images and output a single-layer magnetic resonance reconstruction image.

[0116] Preferably, the first predetermined number of layers is greater than 3.

[0117] Furthermore, the number of input images of the neural network model is a second predetermined number of layers, and the output of the neural network model is a magnetic resonance reconstructed image of a third predetermined number of layers;

[0118] In step A1, a second predetermined number of continuous under-sampled slice images are acquired to form magnetic resonance image data, and the number of target images is a third predetermined number of slices.

[0119] The neural network model of the present invention can also reconstruct multiple target images based on multiple layers of continuous slice images, outputting magnetic resonance reconstructed images corresponding to each target image, that is, outputting multiple magnetic resonance reconstructed images. In this case, the channel splicing module in the neural network model splices each target image together with the first image and the second image to form a third image, and the multi-layer residual network module reconstructs the spliced third image to obtain multiple magnetic resonance reconstructed images.

[0120] Preferably, the second predetermined number of layers is greater than 3, and the third predetermined number of layers is greater than 2.

[0121] Different numbers of input and output layers result in different parameters after training for each neural network model. When using it, you can select a neural network model with the corresponding number of input and output layers for reconstruction as needed. For example, if there are five consecutive slice images and two target images to be reconstructed, you can select a neural network model with five inputs and two outputs to reconstruct the target images.

[0122] See also Figure 9 The present invention further provides a deep learning-based magnetic resonance image reconstruction system for executing the aforementioned deep learning-based magnetic resonance image reconstruction method, comprising:

[0123] A data acquisition module (1) is used to acquire magnetic resonance image data consisting of multiple layers of continuous undersampled slice images, and to use the slice image to be reconstructed in the magnetic resonance image data as a target image;

[0124] An image input module (2), connected to the data acquisition module (1), is used to input the magnetic resonance image data into a pre-trained neural network model;

[0125] An image reconstruction module (3) is connected to the image input module (2) and is used to reconstruct the target image through a neural network model and output a magnetic resonance reconstructed image corresponding to the target image;

[0126] The neural network model includes a slice fusion module (31), a slice translation module (32), a channel splicing module (33) and an image reconstruction network module (34) which are connected in sequence, and the channel splicing module (33) is also connected to the slice fusion module (34);

[0127] The slice fusion module (31) is used to fuse all slice images in the magnetic resonance image data and output a first image;

[0128] The slice translation module (32) is used to perform a translation operation on the first image and output a second image;

[0129] The channel splicing module (33) is used to splice the target image, the first image and the second image to obtain a third image;

[0130] The image reconstruction network module (34) is used to process the third image to obtain a magnetic resonance reconstructed image and output it.

[0131] The above are only preferred embodiments of the present invention and do not limit the implementation mode and protection scope of the present invention. For those skilled in the art, it should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A magnetic resonance image reconstruction method based on deep learning, characterized in that: include: Step A1, acquiring magnetic resonance image data consisting of multiple layers of continuous undersampled slice images, and using a slice image to be reconstructed in the magnetic resonance image data as a target image; Step A2, inputting the magnetic resonance image data into a pre-trained neural network model, wherein the neural network model includes a slice fusion module, a slice translation module, a channel stitching module, and an image reconstruction network module connected in sequence, wherein the channel stitching module is also connected to the slice fusion module; Step A3, the slice fusion module fuses all the slice images in the magnetic resonance image data and outputs a first image; Step A4: the slice translation module performs a translation operation on the first image to output a second image; Step A5: the channel stitching module stitches the target image, the first image, and the second image to obtain a third image; Step A6: the image reconstruction network module processes the third image and outputs a magnetic resonance reconstructed image corresponding to the target image.

2. The magnetic resonance image reconstruction method based on deep learning according to claim 1, characterized in that: In step A2, the pre-training process of the neural network model includes: Step B1, obtaining unreconstructed multi-layer continuous fully sampled slice K-space data as fully sampled sample data; Step B2, performing undersampling processing on the fully sampled sample data to obtain undersampling sample data; Step B3, performing an inverse Fourier transform on the fully sampled sample data to obtain a fully sampled sample image, and performing an inverse Fourier transform on the undersampled sample data to obtain an undersampled sample image; Step B4: Using the under-sampled sample images as training data and the fully sampled sample images as labels, the pre-established neural network is trained to obtain the neural network model.

3. The deep learning-based magnetic resonance image reconstruction method according to claim 1, wherein: The step A4 comprises: Step A41, determining an offset for the first image data; Step A42: cyclically shift the first image data in a predetermined direction according to the offset to obtain the second image data.

4. The magnetic resonance image reconstruction method based on deep learning according to claim 3, characterized in that: In the step A41, the offset is determined according to the undersampling factor and the field of view of the magnetic resonance scan; The calculation formula of the offset is as follows: D = FOV*(Rn) / R; in, D represents the offset; FOV represents the field of view of the MRI scan; R is the downsampling factor of the slice image; n is a positive integer from 1 to R-1.

5. The deep learning-based magnetic resonance image reconstruction method according to claim 1, wherein: In step A6, the image reconstruction network module is a multi-layer residual network module; The multi-layer residual network module includes: a first convolution layer, a plurality of convolution modules, a data consistency module, a data conversion module, and an output module connected in sequence; The processing of the third image in step A6 includes: Step A61: taking the third image as an input image, and sequentially passing the input image through the first convolution layer and a plurality of convolution modules to output a first result image; Step A62: the data consistency module converts the first result image into corresponding first K-space data, and fills the real sampling data of the K-space corresponding to the target image into the corresponding position of the first K-space data to obtain second K-space data; Step A63: the data conversion module performs inverse Fourier transform on the second K-space data to obtain a second result image; Step A64: output the second result image as the magnetic resonance reconstructed image.

6. The deep learning-based magnetic resonance image reconstruction method according to claim 5, wherein: The multi-layer residual network module further includes a judgment module, which is respectively connected to the first convolutional layer, the data conversion module, and the output module; After step A63 and before step A64, the method further includes: Step A63a: Determine whether the second result image satisfies any one of the first iteration stop condition and the second iteration stop condition: If yes, stop the iteration and execute step A64; If not, continue with step A61 using the second result image as the input image of the first convolutional layer; The first iteration stopping condition is: the number of iterations is greater than a preset number; The second iteration stopping condition is: an image error between the second result image generated by the current iteration and the corresponding input image is less than a preset value.

7. The deep learning-based magnetic resonance image reconstruction method according to claim 1, wherein: In step A1, the undersampling factor of each slice image is the same, and in the continuous slice images, the real sampled K-space phase encoding lines corresponding to two adjacent slice images are different.

8. The deep learning-based magnetic resonance image reconstruction method according to claim 1, wherein: The number of input images of the neural network model is a first predetermined number of layers, and the output of the neural network model is a single-slice magnetic resonance reconstructed image; In the step A1, a number of continuous under-sampled slice images equal to the first predetermined number of layers is acquired to form the magnetic resonance image data, and the number of the target images is one.

9. The deep learning-based magnetic resonance image reconstruction method according to claim 1, wherein: The number of input images of the neural network model is a second predetermined number of layers, and the output of the neural network model is a magnetic resonance reconstructed image of a third predetermined number of layers; In the step A1, a number of continuous under-sampled slice images equal to the second predetermined number of layers is acquired to form the magnetic resonance image data, and the number of the target images is equal to the third predetermined number of layers.

10. A magnetic resonance image reconstruction system based on deep learning, characterized in that: A method for reconstructing a magnetic resonance image based on deep learning according to any one of claims 1 to 9, comprising: a data acquisition module, configured to acquire magnetic resonance image data consisting of multiple layers of continuous undersampled slice images, and use the slice image to be reconstructed in the magnetic resonance image data as a target image; an image input module, connected to the data acquisition module, for inputting the magnetic resonance image data into a pre-trained neural network model; an image reconstruction module, connected to the image input module, configured to reconstruct a target image using the neural network model and output a magnetic resonance reconstructed image corresponding to the target image; The neural network model includes a slice fusion module, a slice translation module, a channel splicing module and an image reconstruction network module connected in sequence, and the channel splicing module is also connected to the slice fusion module; The slice fusion module is used to fuse all the slice images in the magnetic resonance image data and output a first image; The slice translation module is used to perform a translation operation on the first image to output a second image; The channel stitching module is used to stitch the target image, the first image and the second image to obtain a third image; The image reconstruction network module is used to process the third image to obtain the magnetic resonance reconstructed image and output it.

Images