A magnetic resonance image reconstruction method and an image reconstruction device

Abstract

The present invention relates to the technical field of magnetic resonance imaging, and specifically relates to a magnetic resonance image reconstruction method and an image reconstruction device. First, the present invention uses a deep neural network for the image structure information of a depth image to obtain a bottom layer image, a background phase, the coil sensitivity of a target object to a magnetic field coil, and the conjugate sensitivity of the coil sensitivity. Then, the signal received from the target object by a magnetic resonance device is sampled to generate a sampled template signal. After that, according to each magnetic field information applied in the sampling environment, a magnetic field phase difference formed by each magnetic field is obtained. Finally, according to the bottom layer image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampled template signal, and the magnetic field phase difference, a magnetic resonance image of the target object is reconstructed. In the process of calculating the coil sensitivity in the present invention, the low-frequency ACS signal is not involved, so the quality of the reconstructed image of the present invention is improved. Also, the speed of reconstructing the image of the present invention is increased.

Description

Technical Field

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

[0002] Magnetic resonance imaging (MRI) has a slow scanning speed. The excessively long scanning time not only causes discomfort to patients, but also easily introduces motion artifacts into the images, thereby affecting the image quality. Magnetic resonance parallel imaging methods are a class of methods for accelerating the MRI scanning speed, such as sensitivity encoding (SENSE) and generalized autocalibrating partially parallel acquisitions (GRAPPA). Such methods reduce the amount of data acquired and use the redundant information contained in multi-channel coils to reconstruct the undersampled data, so as to achieve the purpose of fast scanning.

[0003] Magnetic resonance imaging includes wave gradient field encoding parallel imaging technology, Wave encoding imaging technology, and virtual conjugate coil (VCC) imaging technology.

[0004] Among them, wave gradient field encoding parallel imaging technology (Wave encoding) is a parallel imaging technology for accelerating the magnetic resonance scanning speed. It utilizes the more efficient spatial encoding characteristics of multi-channel coils. The virtual conjugate coil technology (Virtual Conjugate Coil, VCC) is a parallel imaging technology with a similar effect to Wave encoding and can provide more prior information on spatial encoding of multiple channels. The combination of Wave-VCC can give play to the characteristics of both and provide a higher multiple of acceleration technology. However, in the conventional Wave-VCC reconstruction method, only the low-frequency ACS signal in the middle of k-space is used to estimate the background phase. Due to the lack of surrounding high-frequency information, the estimated background phase is difficult to represent the image with high-frequency phase changes.

[0005] Wave encoding technology is a parallel imaging technology for accelerating the three-dimensional magnetic resonance scanning speed. During the acquisition of MRI signals (while applying the readout gradient field), it applies phase differences in the slice selection and phase directions respectively using the MRI gradient coils, and applies a phase difference of a sine gradient field, and uses a fast parallel imaging technique with controllable aliasing (2D CAIPIRINHA, two-dimension Controlled Aliasing In Parallel Imaging Results In Higher Acceleration) to undersample the data, so that the aliasing artifacts caused by undersampling are dispersed along the readout, slice selection, and phase directions, reducing the degree of image aliasing artifacts in each pixel, thereby greatly reducing the signal-to-noise ratio loss of the geometric factor (g-factor, geometry factor) in parallel imaging reconstruction and achieving the purpose of high-fold acceleration.

[0006] The virtual conjugate coil (VCC) is another technique to improve the conditions of the encoding matrix system in parallel imaging. The idea is to incorporate the object background and coil phase into the reconstruction process, and provide additional encoding capabilities by adding virtual coils, which are generated from the conjugate symmetric k-space signals of the actual physical coils.

[0007] The prior art needs to acquire low-frequency auto-calibration signals (ACS), and then calculate the high-frequency coil sensitivities based on the low-frequency ACS. However, during the acquisition of the low-frequency ACS, motion errors are introduced, resulting in large errors in the calculated coil sensitivities, and further leading to poor image quality reconstructed based on the coil sensitivities.

[0008] In summary, the quality of the reconstructed images calculated by the prior art is poor.

[0009] Therefore, the prior art still needs to be improved. Summary of the Invention

[0010] To solve the above technical problems, the present invention provides a magnetic resonance image reconstruction method and an image reconstruction device, which solve the problem of poor quality of the reconstructed images calculated by the prior art.

[0011] To achieve the above object, the present invention adopts the following technical solutions:

[0012] In the first aspect, the present invention provides a magnetic resonance image reconstruction method, which includes:

[0013] Applying a network structure composed of several neural networks to the image structure information of the depth image to obtain the underlying image, background phase, coil sensitivity of the target object to the magnetic field coil, and the conjugate sensitivity of the coil sensitivity output by the network structure. The depth image is used to characterize the depth information of the target object relative to the magnetic resonance device, and the magnetic field coil is the coil inside the magnetic resonance device;

[0014] Sample the signals received by the magnetic resonance device from the target object to generate a sampled template signal;

[0015] Obtain the magnetic field phase differences formed by each magnetic field based on the magnetic field information applied in the sampling environment;

[0016] Reconstruct the magnetic resonance image of the target object based on the underlying image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampled template signal, and the magnetic field phase differences.

[0017] In one implementation, applying a network structure composed of several neural networks to the image structure information of the depth image to obtain the underlying image, the background phase, the coil sensitivity of the target object to the magnetic field coil, and the conjugate sensitivity of the coil sensitivity output by the network structure. The depth image is used to represent the depth information of the target object relative to the magnetic resonance device, and the magnetic field coil is a coil inside the magnetic resonance device, including:

[0018] Apply a first deep convolutional neural network with a decoding structure to the image structure information of the depth image to obtain the underlying image output by the first deep convolutional neural network;

[0019] Apply a second deep convolutional neural network to the image structure information of the depth image to obtain the background phase output by the second deep convolutional neural network;

[0020] Apply a third deep convolutional neural network to the image structure information of the depth image to obtain the coil sensitivity output by the third deep convolutional neural network;

[0021] Apply a fourth deep convolutional neural network to the image structure information of the depth image to obtain the conjugate sensitivity output by the fourth deep convolutional neural network.

[0022] In one implementation, the sampling of the signals received by the magnetic resonance device from the target object to generate a sampled template signal includes:

[0023] Sample the signals received by the magnetic resonance device from the target object using a three-dimensional MRI sequence to obtain the sampling signals in the slice selection direction and the sampling signals in the phase direction;

[0024] Generate a sampled template signal based on the sampling signals in the slice selection direction and the sampling signals in the phase direction.

[0025] In one implementation, while sampling the signals received by the magnetic resonance device from the target object using a three-dimensional MRI sequence, a sinusoidal gradient field is applied in the slice selection direction, and a truncated sinusoidal gradient field is applied in the phase direction; alternatively, a truncated sinusoidal gradient field is applied in the slice selection direction, and a sinusoidal gradient field is applied in the phase direction.

[0026] In one implementation, the zeroth moment of the truncated sinusoidal gradient field is zero.

[0027] In one implementation, obtaining the magnetic field phase difference formed by each magnetic field according to each magnetic field information applied in the sampling environment includes:

[0028] Calculating the magnetic field phase difference between the sinusoidal gradient field and the truncated sinusoidal gradient field according to the magnetic field phase of the sinusoidal gradient field and the magnetic field phase of the truncated sinusoidal gradient field.

[0029] In one implementation, reconstructing the magnetic resonance image of the target object according to the underlying image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampling template signal, and the magnetic field phase difference includes:

[0030] Multiplying the underlying image by the coil sensitivity to obtain a first result;

[0031] Applying a Fourier transform to the first result to obtain a second result;

[0032] Multiplying the second result by the magnetic field phase difference to obtain a third result;

[0033] Applying an inverse Fourier transform to the third result to obtain a fourth result;

[0034] Multiplying the fourth result by the sampling template signal to obtain a target signal;

[0035] Multiplying the background phase, the conjugate sensitivity, and the underlying image to obtain a fifth result;

[0036] Applying a Fourier transform to the fifth result to obtain a sixth result;

[0037] Multiplying the sixth result by the magnetic field phase difference to obtain a seventh result;

[0038] Applying an inverse Fourier transform to the seventh result to obtain an eighth result;

[0039] Multiplying the eighth result by the sampling template signal to obtain the conjugate symmetric signal of the target signal;

[0040] Reconstruct the magnetic resonance image of the target object based on the target signal and the conjugate symmetric signal.

[0041] In a second aspect, an embodiment of the present invention further provides a magnetic resonance image reconstruction device, where the device includes the following components:

[0042] An information analysis module, configured to apply a network structure composed of several neural networks to the image structure information of a depth image to obtain a bottom layer image output by the network structure, a background phase, a coil sensitivity of the target object to a magnetic field coil, and a conjugate sensitivity of the coil sensitivity, where the depth image is used to characterize the depth information of the target object relative to a magnetic resonance device, and the magnetic field coil is a coil inside the magnetic resonance device;

[0043] A signal sampling module, configured to sample the signal received by the magnetic resonance device from the target object to generate a sampling template signal;

[0044] A phase difference calculation module, configured to obtain a magnetic field phase difference formed by each magnetic field according to each magnetic field information applied in a sampling environment;

[0045] An image reconstruction module, configured to reconstruct the magnetic resonance image of the target object according to the bottom layer image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampling template signal, and the magnetic field phase difference.

[0046] In a third aspect, an embodiment of the present invention further provides a terminal device, where the terminal device includes a memory, a processor, and a magnetic resonance image reconstruction program stored in the memory and executable on the processor. When the processor executes the magnetic resonance image reconstruction program, the steps of the magnetic resonance image reconstruction method described above are implemented.

[0047] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, on which a magnetic resonance image reconstruction program is stored. When the magnetic resonance image reconstruction program is executed by a processor, the steps of the magnetic resonance image reconstruction method described above are implemented.

[0048] Advantageous effects: The bottom layer image, the background phase, the coil sensitivity, and the conjugate sensitivity required for reconstructing the image in the present invention all come from neural networks. Since the present invention does not involve low-frequency ACS signals in the process of calculating the coil sensitivity, the quality of the reconstructed image in the present invention is improved. In addition, since the bottom layer image of the present invention is also calculated through a neural network, the speed of reconstructing the image in the present invention is improved. Description of the Drawings

[0049] Figure 1 is the overall flowchart of the present invention;

[0050] Figure 2 It is the structural diagram of the deep convolutional neural network in the embodiment of the present invention;

[0051] Figure 3 It is the schematic diagram of the data undersampling strategy in the embodiment of the present invention;

[0052] Figure 4 It is the schematic diagram of the truncated sine gradient field in the embodiment of the present invention;

[0053] Figure 5 It is the schematic diagram of the sine gradient field in the embodiment of the present invention;

[0054] Figure 6 It is the schematic diagram of the truncated gradient field for the bSSFP sequence in the embodiment of the present invention;

[0055] Figure 7 It is the internal structure principle block diagram of the terminal device provided by the embodiment of the present invention. Specific embodiments

[0056] The following combines the embodiments and the drawings of the specification to clearly and completely describe the technical solutions in the present invention. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the protection scope of the present invention.

[0057] It has been found through research that the prior art needs to collect low-frequency ACS signals and then calculate the high-frequency coil sensitivity based on the low-frequency ACS signals. However, during the process of collecting low-frequency ACS signals, motion errors are introduced, resulting in large errors in the calculated coil sensitivity, and further leading to poor quality of the reconstructed images based on the coil sensitivity.

[0058] To solve the above technical problems, the present invention provides a magnetic resonance image reconstruction method and an image reconstruction device, which solve the problem of poor quality of the reconstructed images calculated by the prior art. Specifically, during implementation, first, the image structure information of the depth image is applied to a network structure composed of several neural networks to obtain the underlying image output by the network structure, the background phase, the coil sensitivity of the target object to the magnetic field coil, and the conjugate sensitivity of the coil sensitivity; sample the signals received by the magnetic resonance device from the target object to generate a sampling template signal; then, based on the magnetic field information applied in the sampling environment, obtain the magnetic field phase difference formed by each magnetic field; finally, based on the underlying image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampling template signal, and the magnetic field phase difference, reconstruct the magnetic resonance image of the target object. The present invention improves the quality of the reconstructed images.

[0059] For example, magnetic resonance imaging technology is required to collect images of the lesion of a patient. The magnetic resonance device is fixed at a position, and depth images of the patient's lesion are collected by the magnetic resonance device (used to characterize the distance between each point of the lesion and the magnetic resonance device). The image information of the depth image is respectively input into four neural networks to obtain a bottom layer image, a background phase, a coil sensitivity (the sensitivity of the lesion to the internal coil of the magnetic resonance device, that is, how much the change of the coil will cause the change of the image information collected at the lesion), and a conjugate sensitivity. In addition, while sampling the signal (the signal is the signal formed after the magnetic resonance device sends the original signal to the lesion and the lesion acts on the original signal) to obtain a sampling template signal, various magnetic fields are applied to the environment where the signal is located. There will be a magnetic field phase difference between various magnetic fields. Finally, by combining the bottom layer image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampling template signal, and the magnetic field phase difference, the magnetic resonance image of the lesion (target object) can be reconstructed.

[0060] Exemplary method

[0061] The magnetic resonance image reconstruction method of this embodiment can be applied to a terminal device. The terminal device can be a terminal product with an image acquisition function, such as a magnetic resonance device, etc. In this embodiment, as Figure 1 shown, the magnetic resonance image reconstruction method specifically includes the following steps:

[0062] S100, applying a network structure composed of several neural networks to the image structure information of the depth image to obtain the bottom layer image, the background phase, the coil sensitivity of the target object to the magnetic field coil, and the conjugate sensitivity of the coil sensitivity output by the network structure. The depth image is used to characterize the depth information of the target object relative to the magnetic resonance device, and the magnetic field coil is the internal coil of the magnetic resonance device.

[0063] In this embodiment, the image structure information is utilized. This kind of image structure information generally represents the sparse property of the image itself in the MRI image, and this property can be characterized by the optimized network structure.

[0064] This embodiment includes four neural networks as Figure 2 shown, Figure 2 from top to bottom in the figure are the first deep convolutional neural network CNN k 、the second deep convolutional neural network CNN, the third deep convolutional neural network CNN C 、the fourth deep convolutional neural network CNN C*。The four deep convolutional neural network structures in this embodiment are similar, except for the different number of intermediate channels. In this embodiment, the Adam optimizer is used to iteratively optimize the parameters of the network. After the parameters of the network are optimized, a small noise can be randomly input to generate a complete image.

[0065] The optimized parameters are based on the following formula:

[0066]

[0067] In the formula, is the second norm, C is the input of the deep convolutional neural network, is the output of the deep convolutional neural network, is the loss function of the deep convolutional neural network, are the parameters of the above four deep convolutional neural networks respectively. The values of the network parameters corresponding to the minimum value of the second norm of the loss function are calculated through the above formula to complete the optimization of the network parameters.

[0068] After optimizing the network, the image structure information of the depth image is input into the first deep convolutional neural network CNN k of the decoding structure, k and the underlying image m output by the first deep convolutional neural network CNN The image structure information of the depth image is input into the second deep convolutional neural network and the background phase output by the second deep convolutional neural network C The image structure information of the depth image is input into the third deep convolutional neural network CNN C and the coil sensitivity CSM output by the third deep convolutional neural network CNN C* The image structure information of the depth image is input into the fourth deep convolutional neural network CNN C* and the conjugate sensitivity CSM* output by the fourth deep convolutional neural network CNN k where the input of CNN CNN c and CNN φ The input of is

[0069] In this embodiment, instead of directly using the ACS signal that only contains low-frequency phase information to estimate the background phase, the image and its phase information are characterized according to the image degradation process and the finally acquired signal. Therefore, the image and the included background phase information can be generated more accurately.

[0070] S200, sample the signals received by the magnetic resonance device from the target object to generate a sampled template signal.

[0071] For example, the magnetic resonance device sends an original signal to a human body lesion (target object), and the human body lesion interacts with the original signal, causing the original signal to become a new signal. Sample this new signal to generate a Figure 3 sampled template signal as shown.

[0072] Before sampling in step S200, a sinusoidal gradient field and a truncated sinusoidal gradient field need to be applied.

[0073] In one embodiment, before signal sampling, a sinusoidal gradient field is applied in the slice selection direction using an MRI gradient field coil, and a Figure 4 truncated sinusoidal gradient field as shown is applied in the phase direction using an MRI gradient field coil. Alternatively, a Figure 4 truncated sinusoidal gradient field as shown is applied in the slice selection direction using an MRI gradient field coil, and a Figure 5 sinusoidal gradient field as shown is applied in the phase direction using an MRI gradient field coil. And the zero-order moment of the truncated sinusoidal gradient field is zero.

[0074] In this embodiment, the zero-order moment (in the phase direction and the slice selection direction) of the truncated sinusoidal gradient field is zero, combined with the direction of the truncated sinusoidal gradient field in the phase direction as Figure 4 shown, which can not only effectively disperse the aliasing artifacts caused by undersampling, reduce the signal-to-noise ratio loss caused by the g-factor to achieve high-fold acceleration, but also avoid the imaging artifacts caused by the gradient field with non-zero zero-order moment.

[0075] In one embodiment, the magnetic field phase difference Psf between the sinusoidal gradient field and the truncated sinusoidal gradient field is

[0076] In one embodiment, the expression of the sinusoidal gradient field is as follows:

[0077]

[0078]

[0079] where t is time, and the time point t = 0 has been marked in Figure 5 ; and are the gradient fields applied in the phase and slice selection directions by the wave-CAIPI technique respectively; A is the amplitude of the sinusoidal gradient field; D C is the duration of one period of the sinusoidal gradient field (as marked in Figure 5 ); D RFor the platform duration of the readout gradient field (as Figure 5 marked).

[0080] After applying the above-mentioned sine gradient field and truncated sine gradient field, step S200 starts sampling the signal. Step S200 includes steps S201 and S202 as follows:

[0081] S201, sampling the signal received by the magnetic resonance device from the target object using a three-dimensional MRI sequence to obtain a sampling signal in the slice selection direction and a sampling signal in the phase direction.

[0082] In this embodiment, the three-dimensional MRI sequence includes a GRE sequence, an SE sequence, a bSSFP sequence, etc.

[0083] S202, generating a sampling template signal based on the sampling signal in the slice selection direction and the sampling signal in the phase direction.

[0084] In this embodiment, a three-dimensional MRI sequence is used to undersample the signal to achieve the purpose of accelerating the scan. The truncated gradient field is applied to the bSSFP sequence, and the signal acquisition strategy of the 2D CAIPIRINHA technique is used to reduce the amount of data acquisition. Through the combination of the truncated gradient field and the 2D CAIPIRINHA sampling strategy, the aliasing caused by undersampling can be simultaneously dispersed into the readout, phase, and slice selection directions, making more effective use of the background region in the FOV, increasing the sensitivity difference between different pixel points, and thus further reducing the loss of the g-factor signal-to-noise ratio.

[0085] The truncated gradient field is applied to the bSSFP sequence in the manner as Figure 6 shown, and at the same time, the 2D CAIPIRINHA data acquisition strategy is as Figure 3 shown. Figure 3 In Figure 3 , the direction perpendicular to both the phase direction and the slice selection direction is the readout direction. The intersection of the dotted lines is the readout line required for full sampling, and the readout lines required for the undersampling strategy adopted by the present invention are represented by solid dots. Figure 3 Shown in Figure 3 is 3×3 times undersampling (3 times undersampling in the phase direction and 3 times undersampling in the slice selection direction). The total acceleration factor is 9, and the required acquisition time is the repetition time (TR, repetition time) × the number of phase encoding lines (Np) × the number of slice encoding lines (Ns) / 9.

[0086] S300, obtaining the magnetic field phase difference formed by each magnetic field based on the information of each magnetic field applied in the sampling environment.

[0087] In this embodiment, the information of each magnetic field is the magnetic field phase of the sine gradient field and the magnetic field phase of the truncated sine gradient field in step S200. In the two-dimensional case, the point spread function PsfY For any point (k x , y) in it, the point spread function value is Psf Y (k x , y) = waveP y (k x , y) / P y (k x , y). If in another frequency encoding direction, represented by the z direction, the above formula will become Psf z (k x , y) = waveP z (k x , z) / P z (k x , z). In the three - dimensional case, sinusoidal gradient magnetic fields and truncated sinusoidal gradient magnetic fields are added to both phase encoding directions, then the three - dimensional point spread function is Psf yz (k x , y, z) = Psf z (k x , z) · Psf y (k x , y), that is, the value of the three - dimensional point spread function Psf yz at any point (k x , y, z) is Psf yz (k x , y, z).

[0088] S400. Reconstruct the magnetic resonance image of the target object according to the underlying image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampling template signal, and the magnetic field phase difference.

[0089] Step S400 includes the following steps S401 to S4011:

[0090] S401. Multiply the underlying image by the coil sensitivity to obtain a first result.

[0091] S402. Apply a Fourier transform to the first result to obtain a second result.

[0092] S403. Multiply the second result by the magnetic field phase difference Psf to obtain a third result.

[0093] S404. Apply an inverse Fourier transform to the third result to obtain a fourth result.

[0094] S405. Multiply the fourth result by the sampling template signal to obtain the target signal b:

[0095]

[0096] The underlying image m corresponding to the output of the first deep convolutional neural network corresponding to the third deep convolutional neural network CNN C the coil sensitivity CSM output corresponding to the first result is the second result obtained by applying the Fourier transform to the first result is the third result Apply the inverse Fourier transform to the third result, where M is Figure 3 the sampling template signal in

[0097] S406, multiply the background phase, the conjugate sensitivity, and the underlying image to obtain a fifth result

[0098] S407, apply the Fourier transform to the fifth result to obtain a sixth result

[0099] S408, multiply the sixth result by the magnetic field phase difference Psf to obtain a seventh result

[0100] S409, apply the inverse Fourier transform to the seventh result to obtain an eighth result

[0101] S4010, multiply the eighth result by the sampling template signal to obtain the conjugate symmetric signal b of the target signal H :

[0102]

[0103] The underlying image m corresponding to the output of the first deep convolutional neural network the conjugate sensitivity CSM* corresponding to the output of the fourth deep convolutional neural network the background phase corresponding to the output of the second deep convolutional neural network is the fifth result is the sixth result is the seventh result is the inverse Fourier transform

[0104] In one embodiment, use b i to replace b, and use to replace b H :

[0105] b i = MF y,z Psf(k x ,y,z)F x C i m

[0106]

[0107] b i and is the forward model of wave encoding, where M is the CAIPI sampling template. As Figure 3 shown, in the actual magnetic resonance imaging process, is the background phase, and D i is the coil sensitivity inherent in the receiving coil. In general models, the background phase is not considered separately but is included in C i for subsequent reconstruction work. The data extended through VCC is

[0108]

[0109] is the conjugate symmetry of, and Psf * is the conjugate symmetry of Psf. The received signal is the conjugate symmetry of the original signal b i . By such virtual conjugate symmetry, it is equivalent to providing another set of phase information, that is, providing an additional prior encoding information of the receiving coil in the wave encoding framework, further providing an acceleration multiple.

[0110] S4011, based on the target signal b and the conjugate symmetry signal b H , reconstruct the magnetic resonance image of the target object.

[0111] According to the target signal b and the conjugate symmetry signal b H the true magnetic resonance image of the target object can be reconstructed. Obtaining the magnetic resonance image through b and b H is the prior art.

[0112] In summary, all the underlying image, background phase, coil sensitivity, and conjugate sensitivity required for reconstructing the image in the present invention come from the neural network. Since the present invention does not involve the low-frequency ACS signal in the process of calculating the coil sensitivity, the quality of the reconstructed image of the present invention is improved. In addition, since the underlying image of the present invention is also calculated through the neural network, the speed of reconstructing the image of the present invention is improved.

[0113] In addition, the present invention uses a depth convolutional neural network (Decoder) that does not require training to represent the underlying image, CSM, and the background phase of high-frequency changes that cannot be estimated by only the low-frequency part of the ACS after Wave-VCC encoding and expansion. The convolutional neural network is used to indirectly generate the above three variables first. In the present invention, a decoding convolutional neural network that does not require training is used. Compared with traditional supervised neural networks or unsupervised neural networks, there is no need to collect training data, and the update of network parameters is optimized through an optimization algorithm, which is more in line with the characteristics of difficult-to-collect fully sampled data in clinical magnetic resonance imaging.

[0114] Exemplary device

[0115] This embodiment also provides a magnetic resonance image reconstruction device, which includes the following components:

[0116] An information analysis module, which is used to apply a network structure composed of several neural networks to the image structure information of a depth image to obtain the underlying image, background phase, coil sensitivity of the target object to the magnetic field coil, and the conjugate sensitivity of the coil sensitivity output by the network structure. The depth image is used to characterize the depth information of the target object relative to the magnetic resonance device, and the magnetic field coil is the coil inside the magnetic resonance device;

[0117] A signal sampling module, which is used to sample the signal received by the magnetic resonance device from the target object to generate a sampling template signal;

[0118] A phase difference calculation module, which is used to obtain the magnetic field phase difference formed by each magnetic field according to each magnetic field information applied in the sampling environment;

[0119] An image reconstruction module, which is used to reconstruct the magnetic resonance image of the target object according to the underlying image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampling template signal, and the magnetic field phase difference.

[0120] Based on the above embodiments, the present invention also provides a terminal device, and its principle block diagram can be as Figure 7As shown in the figure. The terminal device includes a processor, a memory, a network interface, a display screen, and a temperature sensor connected by a system bus. Among them, the processor of the terminal device is used to provide computing and control capabilities. The memory of the terminal device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the terminal device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a magnetic resonance image reconstruction method. The display screen of the terminal device can be a liquid crystal display screen or an electronic ink display screen. The temperature sensor of the terminal device is pre-set inside the terminal device and is used to detect the operating temperature of the internal device.

[0121] Those skilled in the art can understand that Figure 7 the block diagram of the principle shown in the figure is only the block diagram of some structures related to the solution of the present invention, and does not constitute a limitation on the terminal device to which the solution of the present invention is applied. The specific terminal device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0122] In one embodiment, a terminal device is provided. The terminal device includes a memory, a processor, and a magnetic resonance image reconstruction program stored in the memory and executable on the processor. When the processor executes the magnetic resonance image reconstruction program, the following operation instructions are implemented:

[0123] Apply a network structure composed of several neural networks to the image structure information of the depth image to obtain the underlying image, background phase, coil sensitivity of the target object to the magnetic field coil, and conjugate sensitivity of the coil sensitivity output by the network structure. The depth image is used to characterize the depth information of the target object relative to the magnetic resonance device, and the magnetic field coil is the coil inside the magnetic resonance device;

[0124] Sample the signal received by the magnetic resonance device from the target object to generate a sampling template signal;

[0125] According to the magnetic field information applied in the sampling environment, obtain the magnetic field phase difference formed by each magnetic field;

[0126] Reconstruct the magnetic resonance image of the target object according to the underlying image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampling template signal, and the magnetic field phase difference.

[0127] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or equivalently replace some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A magnetic resonance image reconstruction method, characterized in that, Including: Applying a network structure composed of a number of neural networks to the image structure information of a depth image to obtain a bottom layer image, a background phase, a coil sensitivity of a target object to a magnetic field coil, and a conjugate sensitivity of the coil sensitivity, where the depth image is used to characterize depth information of the target object relative to a magnetic resonance device, and the magnetic field coil is a coil inside the magnetic resonance device; Sampling signals received by the magnetic resonance device from the target object to generate a sampling template signal; Obtaining a magnetic field phase difference formed by each magnetic field according to each magnetic field information applied in a sampling environment; Reconstructing a magnetic resonance image of the target object according to the bottom layer image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampling template signal, and the magnetic field phase difference.

2. The magnetic resonance image reconstruction method according to claim 1, wherein The applying a network structure composed of a number of neural networks to the image structure information of a depth image to obtain a bottom layer image, a background phase, a coil sensitivity of a target object to a magnetic field coil, and a conjugate sensitivity of the coil sensitivity, where the depth image is used to characterize depth information of the target object relative to a magnetic resonance device, and the magnetic field coil is a coil inside the magnetic resonance device, includes: Applying a first deep convolutional neural network with a decoding structure to the image structure information of the depth image to obtain the bottom layer image output by the first deep convolutional neural network; Applying a second deep convolutional neural network to the image structure information of the depth image to obtain the background phase output by the second deep convolutional neural network; Applying a third deep convolutional neural network to the image structure information of the depth image to obtain the coil sensitivity output by the third deep convolutional neural network; Applying a fourth deep convolutional neural network to the image structure information of the depth image to obtain the conjugate sensitivity output by the fourth deep convolutional neural network.

3. The magnetic resonance image reconstruction method according to claim 1, wherein The sampling signals received by the magnetic resonance device from the target object to generate a sampling template signal, includes: Sampling the signals received by the magnetic resonance device from the target object using a three-dimensional MRI sequence to obtain a sampling signal in a slice selection direction and a sampling signal in a phase direction; Generating a sampling template signal according to the sampling signal in the slice selection direction and the sampling signal in the phase direction.

4. The magnetic resonance image reconstruction method according to claim 3, wherein, When sampling the signals received by the magnetic resonance device from the target object using a three-dimensional MRI sequence, applying a sinusoidal gradient field in the slice selection direction and a truncated sinusoidal gradient field in the phase direction; or, applying a truncated sinusoidal gradient field in the slice selection direction and a sinusoidal gradient field in the phase direction.

5. The magnetic resonance image reconstruction method according to claim 4, characterized in that, The zero-order moment of the truncated sinusoidal gradient field is zero.

6. The magnetic resonance image reconstruction method according to claim 4, wherein The obtaining a magnetic field phase difference formed by each magnetic field according to each magnetic field information applied in a sampling environment, includes: Calculating a magnetic field phase difference between the sinusoidal gradient field and the truncated sinusoidal gradient field according to the magnetic field phase of the sinusoidal gradient field and the magnetic field phase of the truncated sinusoidal gradient field.

7. The magnetic resonance image reconstruction method according to claim 2, wherein Reconstructing the magnetic resonance image of the target object based on the underlying image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampling template signal, and the magnetic field phase difference includes: Multiplying the underlying image by the coil sensitivity to obtain a first result; Applying a Fourier transform to the first result to obtain a second result; Multiplying the second result by the magnetic field phase difference to obtain a third result; Applying an inverse Fourier transform to the third result to obtain a fourth result; Multiplying the fourth result by the sampling template signal to obtain a target signal; Multiplying the background phase, the conjugate sensitivity, and the underlying image to obtain a fifth result; Applying a Fourier transform to the fifth result to obtain a sixth result; Multiplying the sixth result by the magnetic field phase difference to obtain a seventh result; Applying an inverse Fourier transform to the seventh result to obtain an eighth result; Multiplying the eighth result by the sampling template signal to obtain the conjugate symmetric signal of the target signal; Reconstructing the magnetic resonance image of the target object based on the target signal and the conjugate symmetric signal.

8. A magnetic resonance image reconstruction device, characterized in that, The device includes the following components: An information analysis module for applying a network structure composed of several neural networks to the image structure information of the depth image to obtain the underlying image, the background phase, the coil sensitivity of the target object to the magnetic field coil, and the conjugate sensitivity of the coil sensitivity output by the network structure. The depth image is used to characterize the depth information of the target object relative to the magnetic resonance device, and the magnetic field coil is the coil inside the magnetic resonance device; A signal sampling module for sampling the signal received by the magnetic resonance device from the target object to generate a sampling template signal; A phase difference calculation module for obtaining the magnetic field phase difference formed by each magnetic field based on the magnetic field information applied in the sampling environment; An image reconstruction module for reconstructing the magnetic resonance image of the target object based on the underlying image, the coil sensitivity, the conjugate sensitivity, the background phase, the sampling template signal, and the magnetic field phase difference.

9. A terminal device, characterized in that, The terminal device includes a memory, a processor, and a magnetic resonance image reconstruction program stored in the memory and executable on the processor. When the processor executes the magnetic resonance image reconstruction program, the steps of the magnetic resonance image reconstruction method according to any one of claims 1-7 are implemented.

10. A computer-readable storage medium, characterized in that, A magnetic resonance image reconstruction program is stored on the computer-readable storage medium. When the magnetic resonance image reconstruction program is executed by a processor, the steps of the magnetic resonance image reconstruction method according to any one of claims 1-7 are implemented.

Images