Simultaneous reconstruction method of multi-layer imaging signals, storage medium and computer device

Abstract

The present invention discloses a method for reconstructing simultaneous multi-layer imaging signals. The reconstruction method comprises: generating virtual automatic calibration signals of multiple layers according to the acquired automatic calibration signals of multiple layers, wherein the automatic calibration signals of multiple layers and the virtual automatic calibration signals are both complex data of k-space; generating a training data set according to the automatic calibration signals of multiple layers and the virtual automatic calibration signals; training a neural network model using the training data set; inputting the acquired original aliased data of multiple layers into the trained neural network model to obtain corresponding network output data; obtaining complete reconstruction data of k-space according to the fusion of the original aliased data and the network output data; and generating a reconstructed image according to the complete reconstruction data. The phase information of the virtual automatic calibration signal is fully utilized to more fully train the neural network model, and the virtual signal is used as another training data set to obtain a higher quality reconstructed image without increasing the extra scanning time.

Description

Technical Field

[0001] The present invention belongs to the technical field of image reconstruction of magnetic resonance imaging signals, and in particular, relates to a method for reconstructing simultaneous multi-layer imaging signals, a computer-readable storage medium, and a computer device. Background Art

[0002] Simultaneous Multi-Slice (SMS) is a fast magnetic resonance (MR) scanning imaging method that can simultaneously excite and acquire aliased signals of multiple slices within the acquisition time of a traditional slice image, and reconstruct each slice using the redundant information contained in the multi-channel acquisition coil, thereby achieving accelerated acquisition and reconstruction of magnetic resonance. SMS reconstruction requires the use of redundant information contained in the multi-channel coil to separate multiple slice signals from the aliased signal. Therefore, the higher the utilization rate of redundant information by the SMS reconstruction algorithm, the higher the acceleration multiple that can be obtained. For example, there is a technology called Controlled Aliasing in Parallel Imaging Results in Higher Acceleration (CAIPIRINHA) that is used to improve the acceleration performance of SMS. This technology modulates the phase of a multi-band RF pulse or applies an additional gradient field to add different linear phases to the multiple slices that are excited simultaneously. This allows the multiple slices that are collected simultaneously to form different spatial offsets in the aliased image, thereby making better use of the information differences of the multi-channel coils within the imaging field of view and achieving a higher multiple of SMS acceleration.

[0003] Among the SMS reconstruction methods using CAIPIRINHA technology, Slice-GRAPPA is the most commonly used traditional reconstruction method, which uses the acquired low-resolution Auto Calibration Signal (ACS) data to estimate and reconstruct the linear kernel, and has a wide range of application convenience. This type of SMS reconstruction method is based on the linear kernel method, which can only linearly weight the selected area, cannot perform adaptive changes, and has a low utilization rate of the redundant information contained in the multi-channel receiving coil. Summary of the invention

[0004] (I) Technical Problems to be Solved by the Present Invention

[0005] The technical problem solved by the present invention is: how to better utilize the phase information of magnetic resonance data to improve the image reconstruction quality.

[0006] (II) Technical solution adopted by the present invention

[0007] A method for reconstructing simultaneous multi-layer imaging signals, the reconstruction method comprising:

[0008] Generate virtual automatic calibration signals of multiple slices according to the acquired automatic calibration signals of multiple slices, wherein the automatic calibration signals of multiple slices and the virtual automatic calibration signals are both complex data in k-space;

[0009] Generate a training data set according to the automatic calibration signals of multiple slices and the virtual automatic calibration signals;

[0010] Train the neural network model using the training dataset;

[0011] Input the obtained original aliased data of multiple layers into the trained neural network model to obtain corresponding network output data;

[0012] The complete reconstruction data of k-space is obtained by fusing the original aliased data and the network output data;

[0013] A reconstructed image is generated based on the complete reconstruction data.

[0014] Preferably, the specific method of generating a training data set according to the automatic calibration signals of multiple slices and the virtual automatic calibration signals includes:

[0015] A first training sample is generated according to the automatic calibration signal of multiple slices, and a second training sample is generated according to the virtual automatic calibration signal of multiple slices. The first training sample and the second training sample constitute two training data sets.

[0016] Preferably, the method of training the neural network model using the training data set is:

[0017] The neural network model is trained using the first training sample and the second training sample, wherein the first training sample includes first input training data and first output training data, the first input training data is used to train the input part of the neural network model, and the first output training data is used to train the output part of the neural network model; the second training sample includes second input training data and second output training data, the second input training data is used to train the input part of the neural network model, and the second output training data is used to train the output part of the neural network model.

[0018] Preferably, the specific method of generating the first training sample according to the automatic calibration signal of multiple slices includes:

[0019] Performing inverse Fourier transformation on each layer of the automatic calibration signal in the k-space to generate each layer of low-resolution image in the image space;

[0020] Connecting each layer of low-resolution images in the image space in the readout direction to generate a multi-layer connected image;

[0021] Performing Fourier transformation on the multi-layer connection image to generate k-space data corresponding to the multi-layer connection image;

[0022] Normalizing the k-space data corresponding to the multi-layer connected image to obtain normalized k-space data corresponding to the multi-layer connected image;

[0023] The k-space data corresponding to the normalized multi-layer connected image is downsampled by an acceleration factor MB to obtain first input training data, and the remaining data of the k-space data corresponding to the multi-layer connected image after the downsampling is used as the first output training data, where MB is equal to the number of slices of the multi-slice.

[0024] Preferably, the specific method of generating the second training sample according to the virtual automatic calibration signal of multiple slices includes:

[0025] Performing inverse Fourier transformation on virtual automatic calibration signals of each layer in k-space respectively to generate virtual low-resolution images of each layer in image space;

[0026] Connecting each layer of virtual low-resolution images in the image space in a readout direction to generate a multi-layer connected virtual image;

[0027] Performing Fourier transformation on the multi-layer connected virtual image to generate k-space data corresponding to the multi-layer connected virtual image;

[0028] Normalizing the k-space data corresponding to the multi-layer connected virtual image to obtain normalized k-space data corresponding to the multi-layer connected virtual image;

[0029] The k-space data corresponding to the multi-layer connected virtual image is downsampled by an acceleration factor MB to obtain second input training data, and the remaining data of the k-space data corresponding to the multi-layer connected virtual image after the downsampling processing is used as the second output training data, where MB is equal to the number of slices of the multi-slice.

[0030] Preferably, the specific method of fusing the original aliased data with the corresponding network output data to obtain the complete reconstructed data includes:

[0031] Construct a zero-filled data space where multiple slices are connected in the read direction;

[0032] Upsampling the original aliased data in the k-space by a factor of MB in the readout direction and interpolating the data into the zero-filled data space;

[0033] The network output data corresponding to the original aliased data are interpolated to the remaining positions of the zero-filled data space to obtain complete reconstructed data connected in the readout direction.

[0034] Preferably, the specific method of generating a reconstructed image according to the complete reconstruction data includes:

[0035] performing an inverse Fourier transform on the complete reconstruction data connected in the readout direction to generate a complete multi-slice image connected in the readout direction;

[0036] Cutting the complete multi-slice image connected in the readout direction into MB slices to obtain MB reconstructed images;

[0037] The coil fusion square sum method is used to fuse the multi-coil data of each image to obtain the MB layer image data.

[0038] The present invention also discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a reconstruction program for simultaneous multi-layer imaging signals, and when the reconstruction program for simultaneous multi-layer imaging signals is executed by a processor, the above-mentioned reconstruction method for simultaneous multi-layer imaging signals is implemented.

[0039] The present invention also discloses a computer device, which includes a computer-readable storage medium, a processor, and a reconstruction program for simultaneous multi-layer imaging signals stored in the computer-readable storage medium. When the reconstruction program for simultaneous multi-layer imaging signals is executed by the processor, the above-mentioned reconstruction method for simultaneous multi-layer imaging signals is implemented.

[0040] (III) Beneficial effects

[0041] The present invention discloses a method for reconstructing simultaneous multi-layer imaging signals, which has the following technical effects compared with traditional calculation methods:

[0042] By expanding the actual acquired auto-calibration signal to generate twice as many virtual auto-calibration signals, the data characteristics and information of the SMS aliasing data on the virtual conjugate coil are fully utilized, so that the neural network model can be trained more fully, and higher quality reconstructed images can be obtained without increasing the extra scanning time, thereby reducing the artifacts and noise of the reconstructed images. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flowchart of a method for reconstructing simultaneous multi-layer imaging signals according to an embodiment of the present invention;

[0044] Figure 2 A flowchart of constructing a first training sample according to an embodiment of the present invention;

[0045] Figure 3 A flowchart of constructing a second training sample according to an embodiment of the present invention;

[0046] Figure 4 A schematic diagram of a process of constructing training data according to an embodiment of the present invention;

[0047] Figure 5 A flowchart of fusing original aliased data and network output data according to an embodiment of the present invention;

[0048] Figure 6 A schematic diagram of a process of fusing original aliased data and network output data according to an embodiment of the present invention;

[0049] Figure 7 It is a fully sampled reference image of an embodiment of the present invention, and an image obtained by a traditional method and a reconstruction method of the present application when the acceleration factor is MB=2.

[0050] Figure 8 is the error image ratio obtained by the conventional method and the reconstruction method of the present application when the acceleration factor is MB=2 according to an embodiment of the present invention;

[0051] Fig. 9 It is a fully sampled reference image of an embodiment of the present invention, and an image obtained by a traditional method and a reconstruction method of the present application when the acceleration factor is MB=3.

[0052] Fig.10 is the error image ratio obtained by the conventional method and the reconstruction method of the present application when the acceleration factor is MB=3 according to an embodiment of the present invention;

[0053] Fig.11 It is a fully sampled reference image of an embodiment of the present invention, and an image obtained by a traditional method and a reconstruction method of the present application when the acceleration factor is MB=4.

[0054] Fig.12 is the error image ratio obtained by the conventional method and the reconstruction method of the present application when the acceleration factor is MB=4 according to an embodiment of the present invention;

[0055] Fig.13 A functional block diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0057] Before describing the various embodiments of the present application in detail, the inventive concept of the present application is first briefly described: in order to solve the technical problem that the existing SMS reconstruction method cannot effectively utilize the redundant information contained in the multi-channel receiving coil, the present application generates twice as much virtual data by utilizing the complex conjugate symmetry characteristics of the real data collected by the magnetic resonance multi-channel coil, and uses the real data and the corresponding generated virtual data as training samples of the convolutional neural network, thereby making full use of the channel redundant information of the magnetic resonance receiving coil and obtaining a higher quality reconstructed image.

[0058] Specifically, Figure 1 As shown, the method for reconstructing multi-layer imaging signals simultaneously in the present application includes the following steps:

[0059] Step S10: acquiring fully sampled multi-slice auto-calibration signals, and generating multi-slice virtual auto-calibration signals according to the multi-slice auto-calibration signals, wherein the multi-slice auto-calibration signals and the virtual auto-calibration signals are both complex data in k-space.

[0060] Specifically, the data that needs to be acquired in this application include the original aliased data and the automatic calibration signal collected by the simultaneous multi-plane imaging technology. The automatic calibration signal is a low-frequency signal, which is equivalent to the prior information and is used to train the neural network. The original aliased data includes low-frequency signals and high-frequency signals, which are used to obtain missing data through the neural network, so as to finally reconstruct the image. First of all, the SMS technology of this application is preferably SMSiES technology, which is specifically achieved by combining multi-echo offset technology and accelerated gradient echo (Gradient Echo, GRE) sequence to simultaneously excite imaging in multiple planes. This technology excites multiple slices simultaneously within an effective repetition time, applies an interlaced slice acquisition scheme, and combines echo offset technology to avoid signal loss without affecting the acquisition speed. At the same time, CAIPIRINHA technology is also used to increase the difference in coil sensitivity, so as to facilitate more accurate separation of each slice signal.

[0061] In the field of magnetic resonance, the data directly collected by the above technology is the data in the Fourier domain, which is called k-space data. In magnetic resonance, a set of original aliased data containing multiple slices is collected at the same time. In order to use a neural network to find the mapping relationship between the original aliased data and the missing data, it is also necessary to collect the corresponding multiple slices of ACS data, i.e., low-resolution automatic calibration signals, slice by slice, whose acquisition parameters are the same as those of the original aliased data. The present invention collects 15 central phase encoding lines as automatic calibration signals.

[0062] For example, the present application collects aliased data of a subject's head with MB=5 slices excited simultaneously, where slices 1, 3, and 5 have no intra-slice offset, and slices 2 and 4 have a 1 / 2FOV field of view offset in the phase encoding direction, and the acquisition parameters of each piece of data are the same. The data matrix finally obtained is [N pe , N fe , N coil , N sample ] = [192, 192, 32, 3], where N pe 、N fe 、N coil 、N sample Respectively represent the number of phase encoding lines, the number of readout points, the number of coils, and the number of samples. The automatic calibration signal required for training the neural network is collected slice by slice under the same parameters as the original aliased data. Therefore, the final data matrix is ​​[N acs , N fe , N coil , N sl , N sample ]=[15,192,32,5,3],N sl Indicates the number of slices, N acs Indicates the number of phase encoding lines collected. At the same time, for comparison purposes, the full sampling reference data of the above slices is also obtained, and the matrix size is [N pe , N fe , N coil , N sl , N sample ]=[192,192,32,5,3].

[0063] The present invention uses the above-mentioned fully sampled reference data and constructs the original aliased data with an acceleration factor of 3 (MB=3) and the corresponding automatic calibration signal through retrospective downsampling. There is a cyclic equidistant displacement offset of 2π / 3 between slices, which is manifested in the image as an offset of 2 / 3FOV in the phase encoding direction. After the reconstruction is completed, it is only necessary to restore the offset in the image space.

[0064] Furthermore, the multi-channel virtual coil (Virtual Conjugate Coils, VCC) technology is used to expand the automatic calibration signal of the above-mentioned fully sampled multi-layer to generate a virtual automatic calibration signal with twice the number of layers. Specifically, the automatic calibration signal of each layer collected by magnetic resonance is expanded by using the complex conjugate symmetry relationship to generate a virtual automatic calibration signal. According to formula (1), represents the virtual autocalibration signal of coil j at position k in Fourier space, represented by S j The complex conjugate of (k) Obtained. Wherein, L represents the actual number of coils, j is the coil index, and * represents the complex conjugate operation. In the present application, L is equal to 32. After applying the VCC method, L remains unchanged, which is an important point different from the application of the VCC method in traditional reconstruction. The present invention uses the expanded automatic calibration signal as another sample for training, which is equivalent to increasing the amount of training data without having to modify the specific network parameters. The reason why this is feasible is that the artificial neural network method has integrated the phase information of the automatic calibration signal and the virtual automatic calibration signal during the model training process, thereby stabilizing the training effect.

[0065]

[0066] Step S20: generating a training data set according to the automatic calibration signals of multiple slices and the virtual automatic calibration signals.

[0067] Specifically, a first training sample is generated according to the automatic calibration signal of multiple slices, and a second training sample is generated according to the virtual automatic calibration signal of multiple slices. The first training sample and the second training sample constitute two training data sets.

[0068] As a preferred embodiment, the first training sample includes first input training data and first output training data, the first input training data is used to train the input part of the neural network model, and the first output training data is used to train the output part of the neural network model. Figure 2 As shown, the specific method of generating the first training sample according to the automatic calibration signal of multiple slices includes the following:

[0069] Step S21: performing inverse Fourier transform on each layer of the automatic calibration signal in the k-space to generate each layer of low-resolution image in the image space.

[0070] Specifically, taking three layers of data as an example, the three layers of automatic calibration signals in the acquired k-space are respectively subjected to two-dimensional inverse Fourier transform to obtain three layers of low-resolution images in the image space, wherein the two-dimensional inverse Fourier transform method is a prior art and will not be described in detail here.

[0071] Step S22: Connecting each layer of low-resolution images in the image space in the readout direction to generate a multi-layer connected image.

[0072] As a preferred embodiment, three layers of low-resolution images are connected in the readout direction (X direction) to form a multi-layer connected image. At this time, the multi-layer connected image is an image space image.

[0073] Step S23: Perform Fourier transform on the multi-layer connected image to generate k-space data corresponding to the multi-layer connected image.

[0074] Specifically, the multi-layer connected images in the image space are transformed into the k-space through Fourier transform to generate k-space data corresponding to the multi-layer connected images, wherein each data point in the k-space contains the information of the entire image.

[0075] Step S24: performing normalization processing on the k-space data corresponding to the multi-layer connected images to obtain normalized k-space data corresponding to the multi-layer connected images.

[0076] Step S25: downsample the k-space data corresponding to the normalized multi-layer connected image by an acceleration factor MB to obtain first input training data, and use the remaining data of the k-space data corresponding to the multi-layer connected image after the downsampling as the first output training data, where MB is equal to the number of slices of the multi-slice layer.

[0077] like Figure 4 As shown, the k-space data of the present application is arranged in Cartesian form, the y-direction is the phase line, the x-direction is the frequency-coded data, referred to herein as coding lines, each coding line is arranged in sequence, and each coding line is filled with data, the data of different coding lines represent data at different positions, and there is a nonlinear relationship between the data on the coding lines. The present application achieves the purpose of predicting the data through a part of the data by learning the nonlinear relationship between the data on different coding lines.

[0078] For the sake of vivid description, the present application takes the acceleration factor MB equal to 3 as an example, and performs downsampling processing of the k-space data corresponding to the normalized multi-layer connected image with an acceleration factor of 3 to obtain the first input training data. In the figure, it can be corresponded to that the data on the 1st, 4th, 7th, 10th... encoding lines are used as the first input training data; the remaining data of the k-space data corresponding to the multi-layer connected image after the downsampling processing is used as the first output training data. Figure 4 In the above construction method, the downsampled data can be used to train the neural network, thereby obtaining a nonlinear relationship between the downsampled data and the data at its adjacent positions.

[0079] As a preferred embodiment, the second training sample includes second input training data and second output training data, the second input training data is used to train the input part of the neural network model, and the second output training data is used to train the output part of the neural network model. The specific method of generating the second training sample according to the virtual automatic calibration signal of the multi-layer includes:

[0080] Step S21 ′: performing inverse Fourier transform on virtual automatic calibration signals of each layer in k-space respectively, so as to generate virtual low-resolution images of each layer in image space.

[0081] Correspondingly, the automatic calibration signal is a three-layer data, and the generated virtual automatic calibration signal is also a three-layer data. The three layers of virtual automatic calibration signals collected in the k-space are respectively subjected to two-dimensional inverse Fourier transform to obtain three layers of virtual low-resolution images in the image space, wherein the two-dimensional inverse Fourier transform method is a prior art and will not be described in detail here.

[0082] Step S22 ′: connecting each layer of virtual low-resolution images in the image space in the readout direction to generate a multi-layer connected virtual image.

[0083] As a preferred embodiment, three layers of virtual low-resolution images are connected in the readout direction (X direction) to form a multi-layer connected virtual image. At this time, the multi-layer connected virtual image is an image in the image space.

[0084] Step S23 ′: Performing Fourier transformation on the multi-layer connected virtual images to generate k-space data corresponding to the multi-layer connected virtual images.

[0085] Specifically, the multi-layer connected virtual image in the image space is transformed into the k-space through Fourier transform to generate k-space data corresponding to the multi-layer connected virtual image, wherein each data point in the k-space contains the information of the entire virtual image.

[0086] Step S24 ′: performing normalization processing on the k-space data corresponding to the multi-layer connected virtual image to obtain normalized k-space data corresponding to the multi-layer connected virtual image.

[0087] Step S25': downsample the k-space data corresponding to the multi-layer connected virtual image by an acceleration factor MB to obtain second input training data, and use the remaining data of the k-space data corresponding to the multi-layer connected virtual image after the downsampling as the second output training data, where MB is equal to the number of slices of the multi-slice.

[0088] like Figure 4 As shown, the k-space data of the present application is arranged in Cartesian form, that is, arranged in sequence according to each coding line, different coding lines represent data at different positions, and there is a nonlinear relationship between the data on the coding lines. The present application achieves the purpose of predicting another part of the data through a part of the data by learning the nonlinear relationship between the data on different coding lines.

[0089] For the sake of vivid description, this application takes the acceleration factor MB equal to 3 as an example, and performs downsampling processing of the k-space data corresponding to the normalized multi-layer connected virtual image with an acceleration factor of 3 to obtain the second input training data. Figure 4In the figure, the data on the 1st, 4th, 7th, 10th, ... coding lines can be used as the second input training data; the remaining data of the k-space data corresponding to the multi-layer connection image after downsampling processing can be used as the second output training data. In the figure, the data on the 2nd, 3rd, 5th, 6th, ... coding lines can be used as the second output training data. Through the above construction method, the downsampled data can be used to train the neural network, thereby obtaining a nonlinear relationship between the downsampled data and its adjacent position data.

[0090] Step S30: Train the neural network model using the training data set.

[0091] Specifically, the neural network model is trained using the first training sample and the second training sample constructed in step S20, that is, the input part of the neural network model is trained using the first input training data and the second input training data, and the output part of the neural network model is trained using the first output training data and the second output training data. It should be noted that in each round of training, the training data of the input part and the output part of the neural network model are all from the same training sample, that is, the first training sample or the second training sample.

[0092] As a preferred embodiment, the neural network model of the present application is a two-layer convolutional neural network, and the training parameters are set as follows: the input part is [N s , N pe , N fe , N c *2], the output part is [N s , N pe ,,N fe , N c *2*(MB-1)], the loss function is the mean absolute error (MAE), and the optimizer is Adam. The network weights are initialized with normal distribution, and the weights of the first layer are [5, 3, N c *2, N c *2*(MB-1)], the weight of the second layer network is [2, 2, N c *2*(MB-1), N c *2*(MB-1)], with an initial learning rate of 0.001 and an exponential decay of 0.98, for a total of 1500 rounds of training.

[0093] Furthermore, the fitting equation of the k-space signal is as follows:

[0094] S j,1..L (k x ,k y )≈f j ({S l,collapse (k x -b x Δkx ,k y -b y Δk y )} bx∈[-Bx,Bx],by∈[-By,By],l∈[1,L] )(2)

[0095] This equation indicates the nonlinear mapping relationship between the input signal and the output signal of the network, and the learned parameters are reflected in the variable b, i.e., the weight. Where j∈1..L represents the coil index; k x , k y They represent the position indexes in the readout direction and the phase encoding direction, respectively, z∈1..MB, l∈1..MB, and both represent the number of slices collected simultaneously; b x and b y They represent the index of the fitting kernel, and their sizes are 2·B respectively. x and 2.B y . S j,1..L represents the full k-space data of all slices of coil j, S l,collapse represents the aliased k-space signal of coil l, f j Represents the nonlinear mapping function from the aliased signal to the reconstructed signal. This formula expresses that from the aliased signal S l,collapse The nonlinear mapping f of the network weights is used in j Restore all layers of coil j in k x , k y Reconstruction signal of position.

[0096] Step S40: input the acquired original aliased data of multiple layers into the trained neural network model to obtain corresponding network output data.

[0097] Specifically, after training in step S30, the neural network model obtains the nonlinear relationship between the aliased data and the adjacent missing k-space data in the readout direction, that is, the mapping relationship between the input data and the output data. Then, the original aliased data of the multiple slices obtained in step S10 is input into the neural network model, and the neural network model outputs the corresponding network output data. Since the original aliased data can be regarded as undersampled data in the readout direction, the network output data obtained here is equivalent to another part of the missing data of the image to be reconstructed in the readout direction. The original aliased data and the network output data are used together to reconstruct the complete image.

[0098] Step S50: Obtain complete reconstructed data of k-space by fusing the original aliased data and the network output data.

[0099] The network output data predicted by the neural network model in step S40 cannot be directly used to reconstruct the image, and needs to be fused with the original aliased data to obtain complete reconstructed data. Specifically, Figure 5As shown, step S50 specifically includes the following:

[0100] Step S51: construct a zero-filled data space in which multiple slices are connected in the read direction.

[0101] Step S52: upsample the original aliased data in the k-space by a factor of MB along the readout direction and interpolate the data into the zero-filled data space.

[0102] like Figure 6 As shown, the zero-filled data space is filled in a Cartesian manner, and each encoding line is arranged in sequence along the readout direction. When no interpolation is performed, the value of each encoding line in the zero-filled data space is zero. Taking MB equal to 3 as an example, the original aliased data of each layer in the k-space is upsampled by 3 times and interpolated into the zero-filled data space, corresponding to Figure 6 In the figure, the original aliased data is filled into the 1st, 4th, 7th, 10th, ... encoding lines in sequence.

[0103] Step S53: interpolating the network output data corresponding to the original aliased data to the remaining positions of the zero-filled data space to obtain MB layer complete reconstruction data connected in the read direction.

[0104] Taking MB equal to 3 as an example, the network output data corresponding to the original aliased data is interpolated to the remaining positions of the multi-layer zero-filled data, corresponding to Figure 6 In the figure, the network output data is sequentially filled into the 2nd, 3rd, 5th, 6th, 8th, 9th ... encoding lines, so that the original aliased data and the network output data fill up the zero-filled data space together, thereby obtaining complete k-space reconstruction data with multi-layer connections.

[0105] Step S60: Generate a reconstructed image based on the multi-layer connected complete k-space reconstruction data.

[0106] Specifically, the complete reconstruction data connected in the readout direction is subjected to inverse Fourier transform to generate a reconstructed image of multiple slices connected in the readout direction, and the reconstructed image of multiple slices connected in the readout direction is further cut into MB slices to obtain reconstructed MB images, and then the multi-coil data of each image is fused using the coil fusion method to obtain the final MB layer image data, thereby obtaining the MB layer reconstructed image. In the present invention, the coil fusion method is a square sum method, that is, the images of all coils are squared and summed. Due to the use of the CAIPIRINHA acquisition method, in the reconstructed multi-layer image, each layer will have an image displacement in the phase encoding direction, and only needs to be moved and restored.

[0107] The present application discloses a method for reconstructing simultaneous multi-layer imaging signals, which generates twice as many virtual automatic calibration signals based on the actual acquired automatic calibration signals, and fully utilizes the data characteristics and information of the down-sampled data in the slice direction on the virtual conjugate coil, so as to more fully train the neural network model, obtain higher quality reconstructed images without increasing the extra scanning time, and reduce the artifacts and noise of the reconstructed images.

[0108] In order to demonstrate the advantages of the image reconstruction method of the present application, experiments were conducted to compare the fully sampled image with the reconstructed image obtained by the reconstruction method of the present application, and to compare the reconstructed image obtained by the reconstruction method of the present application with the reconstructed image based on the convolutional neural network method. Evaluation indicators included peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and root mean square error (RMSE), where the higher the PSNR and SSIM, the higher the image quality, and the smaller the RMSE, the higher the image quality.

[0109] like Figure 7 As shown, from left to right are the fully sampled reference image, the image reconstructed by the traditional method when the acceleration factor is MB=2, and the image obtained by the reconstruction method of the present application, as shown in FIG. Figure 8 As shown, by comparing the image reconstructed by the traditional method and the image error obtained by the reconstruction method of the present application, it can be seen that the numerical value in the lower left corner of each image represents the PSNR, SSIM, and RMSE values ​​respectively. It can be found that the PSNR and SSIM are improved, while the RMSE is reduced, indicating that the image quality is improved.

[0110] like Figure 7 As shown, from left to right are the fully sampled reference image, the image reconstructed by the traditional method when the acceleration factor is MB=2, and the image obtained by the reconstruction method of the present application, as shown in FIG. Figure 8 As shown, by comparing the image reconstructed by the traditional method and the image error obtained by the reconstruction method of the present application, it can be seen that the numerical value in the lower left corner of each image represents the PSNR, SSIM, and RMSE values ​​respectively. It can be found that the PSNR and SSIM are improved, while the RMSE is reduced, indicating that the image quality is improved.

[0111] like Fig. 9 As shown, from left to right are the full sampling reference image, the image reconstructed by the traditional method when the acceleration factor is MB=3, and the image obtained by the reconstruction method of the present application, as shown in FIG. Fig.10 As shown, by comparing the image reconstructed by the traditional method and the image error obtained by the reconstruction method of the present application, it can be seen that the numerical value in the lower left corner of each image represents the PSNR, SSIM, and RMSE values ​​respectively. It can be found that the PSNR and SSIM are improved, while the RMSE is reduced, indicating that the image quality is improved.

[0112] like Fig.11 As shown, from left to right are the fully sampled reference image, the image reconstructed by the traditional method when the acceleration factor is MB=4, and the image obtained by the reconstruction method of the present application, as shown in FIG. Fig.12 As shown, by comparing the image reconstructed by the traditional method and the image error obtained by the reconstruction method of the present application, it can be seen that the numerical value in the lower left corner of each image represents the PSNR, SSIM, and RMSE values ​​respectively. It can be found that the PSNR and SSIM are improved, while the RMSE is reduced, indicating that the image quality is improved.

[0113] The present application also discloses a computer-readable storage medium storing a reconstruction program of simultaneous multi-layer imaging signals. When the reconstruction program of simultaneous multi-layer imaging signals is executed by a processor, the above-mentioned reconstruction method of simultaneous multi-layer imaging signals is implemented.

[0114] The present application also discloses a computer device, at the hardware level, such as Fig.13 As shown, the computer device includes a processor 12, an internal bus 13, a network interface 14, and a computer-readable storage medium 11. The processor 12 reads the corresponding computer program from the computer-readable storage medium and then runs it, forming a request processing device at the logical level. Of course, in addition to the software implementation, one or more embodiments of this specification do not exclude other implementations, such as logic devices or a combination of software and hardware, etc., that is, the execution subject of the following processing flow is not limited to each logical unit, but can also be hardware or logic devices. The computer-readable storage medium 11 stores a reconstruction program for simultaneous multi-layer imaging signals, and when the reconstruction program for simultaneous multi-layer imaging signals is executed by the processor, the above-mentioned reconstruction method for simultaneous multi-layer imaging signals is implemented.

[0115] Computer-readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer-readable instructions, data structures, modules of programs or other data. Examples of computer-readable storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage, quantum memory, graphene-based storage media or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0116] The specific implementation methods of the present invention are described in detail above. Although some embodiments have been shown and described, those skilled in the art should understand that these embodiments can be modified and improved without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents. These modifications and improvements should also be within the scope of protection of the present invention.

Claims

1. A method for reconstructing simultaneous multi-layer imaging signals, characterized in that: The reconstruction method comprises: Generate virtual automatic calibration signals of the multiple slices according to the acquired automatic calibration signals of the multiple slices, wherein the automatic calibration signals of the multiple slices and the virtual automatic calibration signals are both complex data of k-space, wherein the automatic calibration signals of each slice are expanded to obtain virtual automatic calibration signals by using complex conjugate symmetry relationship; Generating a training data set according to the automatic calibration signals of the multiple slices and the virtual automatic calibration signals, including: generating a first training sample according to the automatic calibration signals of the multiple slices, generating a second training sample according to the virtual automatic calibration signals of the multiple slices, the first training sample and the second training sample forming two training data sets; Training a neural network model using a training data set includes: training the neural network model using a first training sample and a second training sample, wherein the first training sample includes first input training data and first output training data, the first input training data is used to train the input part of the neural network model, and the first output training data is used to train the output part of the neural network model; the second training sample includes second input training data and second output training data, the second input training data is used to train the input part of the neural network model, and the second output training data is used to train the output part of the neural network model; Input the obtained original aliased data of multiple layers into the trained neural network model to obtain corresponding network output data; The complete reconstruction data of k-space is obtained by fusing the original aliased data and the network output data; A reconstructed image is generated based on the complete reconstruction data.

2. The method for reconstructing simultaneous multi-layer imaging signals according to claim 1, characterized in that: The specific method of generating the first training sample according to the automatic calibration signal of multiple slices includes: Performing inverse Fourier transformation on each layer of the automatic calibration signal in the k-space to generate each layer of low-resolution image in the image space; Connecting each layer of low-resolution images in the image space in the readout direction to generate a multi-layer connected image; Performing Fourier transformation on the multi-layer connection image to generate k-space data corresponding to the multi-layer connection image; Normalizing the k-space data corresponding to the multi-layer connected image to obtain normalized k-space data corresponding to the multi-layer connected image; The k-space data corresponding to the normalized multi-layer connected image is downsampled by an acceleration factor MB to obtain first input training data, and the remaining data of the k-space data corresponding to the multi-layer connected image after the downsampling is used as the first output training data, where MB is equal to the number of slices of the multi-slice.

3. The method for reconstructing simultaneous multi-layer imaging signals according to claim 1, characterized in that: The specific method for generating the second training sample according to the virtual automatic calibration signal of multiple slices includes: Performing inverse Fourier transformation on virtual automatic calibration signals of each layer in k-space respectively to generate virtual low-resolution images of each layer in image space; Connecting each layer of virtual low-resolution images in the image space in a readout direction to generate a multi-layer connected virtual image; Performing Fourier transformation on the multi-layer connected virtual image to generate k-space data corresponding to the multi-layer connected virtual image; Normalizing the k-space data corresponding to the multi-layer connected virtual image to obtain normalized k-space data corresponding to the multi-layer connected virtual image; The k-space data corresponding to the multi-layer connected virtual image is downsampled by an acceleration factor MB to obtain second input training data, and the remaining data of the k-space data corresponding to the multi-layer connected virtual image after the downsampling processing is used as the second output training data, where MB is equal to the number of slices of the multi-slice.

4. The method for reconstructing simultaneous multi-layer imaging signals according to claim 1, characterized in that: The specific method of fusing the original aliased data with the corresponding network output data to obtain the complete reconstructed data includes: Construct a zero-filled data space where multiple slices are connected in the read direction; Upsampling the original aliased data in the k-space by a factor of MB in the readout direction and interpolating the data into the zero-filled data space; The network output data corresponding to the original aliased data are interpolated to the remaining positions of the zero-filled data space to obtain complete reconstructed data connected in the readout direction.

5. The method for reconstructing simultaneous multi-layer imaging signals according to claim 4, characterized in that: The specific method of generating a reconstructed image based on the complete reconstruction data includes: performing an inverse Fourier transform on the complete reconstruction data connected in the readout direction to generate a complete multi-slice image connected in the readout direction; Cutting the complete multi-slice image connected in the readout direction into MB slices to obtain MB reconstructed images; The coil fusion square sum method is used to fuse the multi-coil data of each image to obtain the MB layer image data.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a reconstruction program for simultaneous multi-layer imaging signals, and when the reconstruction program for simultaneous multi-layer imaging signals is executed by a processor, the reconstruction method for simultaneous multi-layer imaging signals according to any one of claims 1 to 5 is implemented.

7. A computer device, characterized in that: The computer device includes a computer-readable storage medium, a processor, and a reconstruction program for simultaneous multi-layer imaging signals stored in the computer-readable storage medium. When the reconstruction program for simultaneous multi-layer imaging signals is executed by the processor, the reconstruction method for simultaneous multi-layer imaging signals according to any one of claims 1 to 5 is implemented.

Images