Magnetic resonance image processing method and device, and image processing method based on EPI

Abstract

The invention relates to a magnetic resonance image processing method and device, and an EPI-based image processing method. The method comprises the steps of obtaining a first sampling image of a target object according to sensitivity information of a receiving coil corresponding to the target object and K space data; correcting artifacts in the first sampling image to obtain a second sampling image of the target object; and iterating data of a preset first magnetic resonance image by taking the second sampling image as a constraint condition and taking the K space data as a fidelity term in an iterative reconstruction process to obtain a magnetic resonance image of the target object. By adopting the method, the problem of aliasing artifacts existing in magnetic resonance imaging images can be solved.

Description

Technical Field

[0001] This application relates to the field of magnetic resonance imaging, and in particular to magnetic resonance image processing methods, apparatus, and EPI-based image processing methods. Background Technology

[0002] K-space data recovery technology based on high-magnification uniform downsampling of separated acquisition has wide applications in scenarios such as parallel imaging acceleration, parity-even echo calibration of high-field EPI (Echo Planar Imaging) with magnetic field strength greater than or equal to 5T (Tesla), and multi-shot diffusion-weighted imaging reconstruction. However, due to the missing K-space data obtained from downsampling, and the influence of natural physiological movements such as respiration, gastrointestinal motility, and heartbeat during the parallel imaging process, artifacts exist in the image domain under downsampling scenarios, leading to a decrease in magnetic resonance imaging quality.

[0003] The algorithms used in related technologies, such as SENSE_type, GRAPPA_type, Low-Rank based, and Residual mapping network, cannot achieve a recovery capability of more than 5 times in uniform downsampling scenarios, and cannot effectively eliminate noise and artifacts in the image domain under downsampling scenarios.

[0004] There is currently no effective solution to the problems of severe artifacts and low image quality in magnetic resonance imaging (MRI) in related technologies. Summary of the Invention

[0005] Therefore, it is necessary to provide a magnetic resonance image processing method, apparatus, and EPI-based image processing method that can solve the problems of severe magnetic resonance artifacts and low image quality, in order to address the above-mentioned technical issues.

[0006] In a first aspect, this embodiment provides a magnetic resonance image processing method, the method comprising:

[0007] The first sampled image of the target object is obtained based on the sensitivity information of the receiving coil corresponding to the target object and the K-space data;

[0008] The artifacts in the first sampled image are corrected to obtain the second sampled image of the target object;

[0009] Using the second sampled image as a constraint and the K-space data as a fidelity term in the iterative reconstruction process, the data of the preset first magnetic resonance image are iterated to obtain the magnetic resonance image of the target object.

[0010] In some embodiments, the second sampled image is used as a constraint, and the K-space data is used as a fidelity term in the iterative reconstruction process to iterate the preset magnetic resonance data to obtain the magnetic resonance image of the target object, including:

[0011] A regularization term is obtained based on the difference between the first magnetic resonance image and the second sampled image;

[0012] Iterate over the first magnetic resonance image, and stop iterating when the difference between the first magnetic resonance image and the first sampled image, and the value of the regularization term, converge to the first convergence condition, and obtain the magnetic resonance image of the target object.

[0013] In some embodiments, iterating over data from a preset first magnetic resonance image further includes:

[0014] Based on the similarity of amplitude and the difference in phase between the odd-numbered and even-numbered echo data in the K-space data, a low-rank matrix is ​​constructed.

[0015] The low-rank matrix is ​​expanded in the scanning channel of the target object, and the expanded matrix is ​​used as a low-rank constraint in the iterative process of the first magnetic resonance image.

[0016] In some embodiments, taking a first sampled image of the target object includes:

[0017] The second magnetic resonance image is obtained based on the sensitivity information of the receiving coil and the K-space data;

[0018] Iterate over the second magnetic resonance image;

[0019] If the difference between the data of the second magnetic resonance image and the K-space data satisfies the second convergence condition, then the iteration stops and the second magnetic resonance image is used as the first sampled image.

[0020] In some embodiments, artifacts in the first sampled image are corrected to obtain a second sampled image of the target object, including:

[0021] Obtain a mapping function, which is used to map the data distribution of the first sampled image to the data distribution of the second sampled image;

[0022] Based on the mapping function, a second sampled image corresponding to the first sampled image is obtained.

[0023] In some embodiments, obtaining the mapping function includes:

[0024] The adversarial network is trained based on a training dataset, which includes sampled images of the training object with and without artifacts.

[0025] The parameters of the adversarial network are adjusted, and the mapping function is obtained based on the adversarial network when the cycle consistency loss and discriminator loss of the adversarial network converge.

[0026] Secondly, this embodiment provides a magnetic resonance image processing apparatus, the apparatus comprising:

[0027] The acquisition module is used to obtain a first sampled image of the target object based on the sensitivity information of the receiving coil corresponding to the target object and the K-space data;

[0028] The correction module is used to correct artifacts in the first sampled image to obtain a second sampled image of the target object;

[0029] An iterative module is used to iterate the data of a preset first magnetic resonance image using the second sampled image as a constraint and the K-space data as a fidelity term in the iterative reconstruction process to obtain the magnetic resonance image of the target object.

[0030] Thirdly, this embodiment provides an EPI-based image processing method, including:

[0031] Acquire EPI data of the target object, wherein the EPI data is obtained through undersampling;

[0032] The EPI data is divided into odd echo datasets and even echo datasets;

[0033] Initialization and reconstruction based on machine learning network constraints are performed on the odd echo dataset and even echo dataset respectively to obtain the odd echo image and even echo image of the target object;

[0034] The odd-echo and even-echo images are regularized and reconstructed to obtain the magnetic resonance image of the target object.

[0035] In some embodiments, the initial reconstruction of the odd-echo and even-echo datasets based on machine learning network constraints includes:

[0036] The odd echo dataset is input into the machine learning network for reconstruction and restoration, and the intermediate dataset obtained by the SENSE reconstruction algorithm during the reconstruction and restoration process is used as a constraint term for the reconstruction and restoration of the machine learning network.

[0037] In some embodiments, the regularization reconstruction process of the odd-echo and even-echo images includes:

[0038] Based on the odd-echo and even-echo images, a low-rank matrix is ​​determined, which reflects the phase difference between the odd-echo and even-echo images.

[0039] The low-rank constraint is obtained from the low-rank matrix, and the low-rank constraint is used as a constraint term in the regularization reconstruction process of the odd echo image and the even echo image.

[0040] The aforementioned magnetic resonance image processing method, apparatus, and EPI-based image processing method, after acquiring a first sampled image of the target object, remove artifacts from the first sampled image to obtain a second sampled image. By iterating over a preset first magnetic resonance image, the first magnetic resonance image is made to move closer to the K-space data during the iteration process. Furthermore, using the second sampled image as a constraint, the magnetic resonance image of the target object obtained after iteration possesses the characteristics of high signal-to-noise ratio and low artifacts found in the second sampled image. Attached Figure Description

[0041] Figure 1 This is an application environment diagram of a magnetic resonance image processing method in one embodiment;

[0042] Figure 2 This is a flowchart illustrating a magnetic resonance image processing method in one embodiment;

[0043] Figure 3 This is a schematic diagram of the generator structure in one embodiment;

[0044] Figure 4 This is a schematic diagram of the discriminator in one embodiment;

[0045] Figure 5 This is a schematic diagram illustrating the training of an adversarial network in one embodiment;

[0046] Figure 6 This is a schematic diagram illustrating the generation of the Hank matrix in one embodiment;

[0047] Figure 7 This is a flowchart illustrating an EPI-based image processing method in one embodiment;

[0048] Figure 8 This is a structural block diagram of a magnetic resonance image processing device in one embodiment;

[0049] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0051] The magnetic resonance image processing method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed in the cloud or on other network servers. A first sampled image and a second sampled image are generated based on the terminal or the interaction between the terminal and the server. A preset first magnetic resonance image is iterated based on the first and second sampled images. The magnetic resonance image of the target object is obtained based on the iterated first magnetic resonance image. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, etc. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0052] In one embodiment, such as Figure 2 As shown, a magnetic resonance image processing method is provided, which can be applied to... Figure 1 Taking terminal 102 as an example, the explanation includes the following steps:

[0053] Step 202: Obtain the first sampled image of the target object based on the sensitivity information of the receiving coil corresponding to the target object and the K-space data.

[0054] The sensitivity information of the receiving coil when sampling the target object is used to estimate the relative signal response of each coil at different positions when sampling the target object. This information can include the sensitivity of each coil when sampling the target object, the coil sensitivity matrix, etc. The sensitivity information of the receiving coil is used to calibrate the differences in data obtained from sampling by different coils. Optionally, if a parallel acquisition method is used to obtain K-space data during data acquisition, the missing K-space data is estimated based on the sensitivity information of the multi-channel coils and the existing K-space data to obtain the filled K-space data. After processing the K-space data based on the sensitivity information of the receiving coil, the frequency information in the K-space data is converted into spatial domain image information to obtain the first sampled image.

[0055] Step 204: Correct the artifacts in the first sampled image to obtain the second sampled image of the target object.

[0056] Optionally, artifacts in the first sampled image can be corrected based on the trained convolutional neural network. For example, an adversarial network can be trained using downsampled images and their corresponding full-sampled images as training data. The first sampled image is then input into the trained adversarial network, which outputs a second sampled image. The convolutional neural network can also be a network, a ResNet network, etc., and is not limited here.

[0057] Alternatively, artifacts in the first sampled image can be removed based on existing filters.

[0058] Step 206: Using the second sampled image as a constraint and K-space data as the fidelity term in the iterative reconstruction process, the data of the preset first magnetic resonance image is iterated to obtain the magnetic resonance image of the target object.

[0059] The first magnetic resonance image can be any magnetic resonance image. Using the second sampled image as a constraint condition limits the divergence of the first magnetic resonance image during the iteration process, thus reducing artifacts in the iterated magnetic resonance image. Using K-space data as the fidelity term in the iterative reconstruction process allows the first magnetic resonance image to continuously approach the K-space data during the iteration process.

[0060] Optionally, the iteration of the first magnetic resonance image can be terminated when the difference between the data of the first magnetic resonance image and the K-space data converges to a minimum, thereby improving the magnetic resonance image of the target object. Alternatively, the iteration of the first magnetic resonance image can be terminated after a preset number of iterations, ensuring the quality of the magnetic resonance image while improving the processing efficiency of the magnetic resonance image.

[0061] In the aforementioned magnetic resonance image processing method, considering the differences in intensity and phase characteristics of the signal response of each coil at different positions, a first sampled image with relatively high signal-to-noise ratio (SNR) is obtained based on the sensitivity information of the receiving coil and K-space data, after removing some aliasing artifacts. Removing artifacts from the first sampled image further improves the SNR, resulting in a second sampled image. While using K-space data as a fidelity-preserving term in the iterative reconstruction process, the second sampled image obtained after artifact removal is used as an iterative constraint for the magnetic resonance image. This ensures that the iteratively obtained magnetic resonance image closely approximates the actual sampled result while possessing a high SNR, thus solving the problem of aliasing artifacts in magnetic resonance imaging images.

[0062] The iterative method for the preset first magnetic resonance image is described below. In some embodiments, the preset magnetic resonance data is iterated using a second sampled image as a constraint and K-space data as a fidelity term in the iterative reconstruction process to obtain the magnetic resonance image of the target object. This includes: obtaining a regularization term based on the difference between the first magnetic resonance image and the second sampled image; iterating the first magnetic resonance image until the difference between the first magnetic resonance image and the first sampled image, and the value of the regularization term, converge to a first convergence condition, at which point the iteration is stopped and the magnetic resonance image of the target object is obtained.

[0063] The first convergence condition indicates that the first magnetic resonance image satisfies the regularization constraint, and the difference between the first magnetic resonance image and the first sampled image converges. Specifically, when the difference between the first magnetic resonance image and the first sampled image converges to its minimum value, it means that the data of the iterated first magnetic resonance image after transformation to the K-space region is close to the actual sampled result, i.e., K-space data. When the first magnetic resonance image satisfies the regularization constraint, it means that the iterated first magnetic resonance image possesses a high signal-to-noise ratio.

[0064] In this embodiment, an iterative method is used to obtain a magnetic resonance image that closely approximates the actual sampled result, effectively reducing image aliasing artifacts. The iterative method for the first magnetic resonance image can be applied to a single-polarity K-space to fill the K-space data obtained based on downsampling, reducing aliasing artifacts in the first sampled image while significantly improving the image signal-to-noise ratio.

[0065] In some embodiments, acquiring a first sampled image of the target object includes: obtaining a second magnetic resonance image based on the sensitivity information of the receiving coil and K-space data; iterating over the second magnetic resonance image; and stopping the iteration if the difference between the data of the second magnetic resonance image and the K-space data satisfies a second convergence condition, and using the second magnetic resonance image as the first sampled image. The second convergence condition indicates that the difference between the data of the second magnetic resonance image and the K-space data converges to a specified value.

[0066] Optionally, the second convergence condition is used to indicate that the difference between the data of the second magnetic resonance image and the K-space data converges to a minimum. The method for obtaining the first sampled image is as follows: the K-space data is reconstructed based on the coil sensitivity matrix in the sensitivity information of the receiving coil to obtain the second magnetic resonance image. The second magnetic resonance image is converted to K-space through Fourier transform. The second magnetic resonance image is iterated, so that the second magnetic resonance image is adjusted accordingly. When it is determined that the difference between the data of the second magnetic resonance image and the K-space data converges to a minimum, the iteration stops. If Fourier transform is performed directly, the resulting image will have multiple aliasing layers and obvious aliasing artifacts. The number of aliasing layers depends on the undersampling factor, and aliasing artifacts exist in the image. However, the reconstruction and iteration method in this embodiment introduces the acquisition information of each channel in the multi-channel coil through the coil sensitivity matrix, which can reduce the artifacts caused by spatial coding information in the multi-coil channels and obtain an initialization result that conforms to the actual acquisition and has fewer artifacts.

[0067] To further improve the reconstruction quality of the first sampled image, wavelet transform can optionally be used as a regularization term during the iterative process of acquiring the first sampled image. The iterative method utilizes the multi-scale analysis characteristics of wavelet transform to obtain the first sampled image. The regularization term of wavelet transform effectively removes noise from the first sampled image while preserving edge details and preventing the first sampled image from becoming too smooth.

[0068] Related technologies exist that improve K-space recovery capabilities based on compressed sensing algorithms. However, compressed sensing requires the original downsampled K-space to be sparsity, meaning random downsampling is necessary during acquisition. A sparse K-space exhibits a low signal-to-noise ratio but almost no artifacts in the image domain. This also means that related technologies can only handle similar scenes from compressed sensing and cannot be used to recover uniformly downsampled K-spaces with significant aliasing artifacts. In contrast to these technologies, this embodiment iteratively recovers K-space data with significant aliasing artifacts by comparing the difference between the data from the second magnetic resonance image and the K-space data. This method can be applied to both uniformly downsampled and randomly downsampled K-spaces.

[0069] However, the distribution of the first sampled image in K-space may still contain erroneous phase encoding lines, leading to artifacts in the first sampled image. In some embodiments, artifacts in the first sampled image are corrected to obtain a second sampled image of the target object, including: obtaining a mapping function to map the data distribution of the first sampled image to the data distribution of the second sampled image; and obtaining the second sampled image corresponding to the first sampled image based on the mapping function.

[0070] In this design, the second sampled image is the image after artifact removal, while the first sampled image contains artifacts. A mapping function is used to indicate the correlation between erroneous coding lines and corrected coding lines in the sampled image. Based on the mapping function, the data distribution of the first sampled image can be adjusted to correct erroneous coding lines and remove artifacts from the first sampled image.

[0071] Optionally, the first sampled image input to the mapping function can be an initial downsampling result obtained by directly filling the downsampled K-space data based on the sensitivity information of the receiving coil. Alternatively, the downsampled K-space data can be split into even and odd echoes, and the split even and odd echoes can be filled based on the sensitivity information of the receiving coil. The two downsampling results are then input to the mapping function, and the two results output by the mapping function are merged to obtain the second sampled image with artifacts removed.

[0072] The mapping function can be an existing function or a trained neural network. Optionally, the mapping function is an adversarial network. Obtaining the mapping function includes: training the adversarial network based on a training dataset, which includes sampled images of the training object with and without artifacts; adjusting the parameters of the adversarial network; and obtaining the mapping function based on the adversarial network after the recurrent consistency loss and discriminator loss of the adversarial network converge. The adversarial network includes a discriminator and a generator.

[0073] Optionally, the downsampling and full-sampling results of the training objects are used as the sampled image with artifacts and the sampled image without artifacts, respectively, to train the adversarial network. Specifically, the downsampling result is obtained by reconstructing the image using the K-space undersampling result; the full-sampling image is obtained by sampling in the full K-space.

[0074] In this embodiment, by training an adversarial network, aliasing artifacts in the first sampled image are modeled as missing phase coding lines in Fourier space. The adversarial network then recovers the K-space information in the first sampled image, effectively reducing image-domain aliasing artifacts caused by insufficient K-space information. Compared to related techniques that use image-domain residual networks to recover K-space data, the adversarial network in this embodiment is a generative network that incorporates K-space information. This makes the training method of the adversarial network more closely aligned with the physical properties of MR (Magnetic Resonance). The trained adversarial network has a stronger ability to correct the tissue structure in the image than the image-domain residual network, resulting in better K-space data recovery.

[0075] In one embodiment, a magnetic resonance image processing method for a uniform downsampling scenario is provided.

[0076] First, the K-space data is initialized using the regularized SENSE technique based on wavelet transform constraints to obtain the first sampled image:

[0077]

[0078] Where d represents the actual downsampling K-space data, Ω is a logic matrix used to mark the two-dimensional downsampling trajectory, C is the coil sensitivity matrix, F refers to the Fourier transform, ρ is the preset first magnetic resonance image, and N s (numberof shots) represents the total number of excitations, and t represents the current number of excitations. λ1 is the weight coefficient used to constrain the regularization term, and Ψ(ρ) is the regularization term obtained based on wavelet transform.

[0079] The convergence to the minimum on the left side of the formula signifies that the image result is close to the actual sampled result within the K-space region of the downsampling trajectory. The wavelet transform on the right side, used as a regularization term, leverages the multi-scale analysis characteristics of wavelet transforms. This means that wavelet transform-based regularization can effectively remove noise from the image while preserving image edge details and preventing the image from becoming overly smooth. In addition to the wavelet transform-based regularization term, other denoising constraints can be used during the iteration process, such as constraints based on low-rank principles or total variation constraints, etc., which will not be elaborated upon here.

[0080] Obtaining the first sampled image through initialization can accelerate the convergence process of the adversarial network and reduce the network's dependence on the amount of data stored in the dataset.

[0081] Subsequently, an adversarial network is trained, and based on the adversarial network, the K-space information of the first sampled image is recovered to obtain the second sampled image. The adversarial network consists of a generator and a discriminator. Figure 3 This is a schematic diagram of the generator structure in this embodiment. Figure 4 This is a schematic diagram of the discriminator in this embodiment. Figure 3 and Figure 4 The arrows in the diagram indicate the steps each layer of the network takes in processing the input data, including: sliding convolutional kernels across the image with a stride to extract features, performing instance normalization on the input data, and using the Lealy ReLU activation function to help the network learn complex data distributions.

[0082] Figure 5 This is a schematic diagram of the training of the adversarial network in this embodiment, as shown below. Figure 5 As shown, Gθ is the generator, Dθ is the discriminator, and L... cycle The cycle consistency loss is the L1 loss, L patchGAN The discriminator loss is the MSE loss, S is channel combining, F is the two-dimensional Fourier transform, and F... -1For a two-dimensional inverse Fourier transform, D represents downsampling, where the downsampling mask is used to indicate the downsampling position. Figure 5 The formula for channel merging shown is:

[0083] sqrt(sum(image real 2 +image imag 2 ))

[0084] Among them, image real This refers to the input image of the generator. imag This refers to the output image of the generator.

[0085] The initialized downsampling result, i.e., the first sampled image, is input into the trained network to obtain the final artifact-free full-sampled image, i.e., the second sampled image. By introducing a generative adversarial network to restore the K-space information of the first sampled image, the image domain aliasing artifacts caused by insufficient high-magnification downsampling information during initialization are effectively reduced.

[0086] This embodiment constructs an algorithmic framework based on iterative methods and adversarial networks for recovering the K-space of high-magnification downsampling, K-space recovery of multi-shot EPI sequences, and image domain signal-to-noise ratio (SNR) enhancement. Specifically, an iterative SENSE algorithm based on wavelet transform constraints is first used to initialize the downsampling K-space, reducing aliasing artifacts while significantly improving the image SNR. Subsequently, utilizing the physical laws of magnetic resonance imaging (MRI) acquisition, the aliasing artifacts remaining in the initialized image domain are modeled as erroneous phase coding lines in the K-space. These coding lines are then corrected using an adversarial network, thereby recovering a full-sampled K-space with low aliasing artifacts and high SNR.

[0087] In related technologies, EPI, due to its unique sampling method, is relatively susceptible to Nyquist ghosting artifacts caused by phase inconsistencies between odd and even echoes in K-space acquisition. In clinical use, such artifacts severely degrade image quality. This situation is exacerbated by higher-order eddies in high-field environments. To address artifacts caused by phase inconsistencies between odd and even echoes, some embodiments iterate the data of a preset first magnetic resonance image, further including: constructing a low-rank matrix based on the similarity of amplitude and the difference in phase between odd and even echo rows in the K-space data; expanding the low-rank matrix in the scanning channel of the target object; and using the expanded matrix as a low-rank constraint in the iteration process of the first magnetic resonance image.

[0088] In this model, the odd and even echo image domains exhibit similar amplitudes but differ in phase, and the differences between the odd and even echo image domains are sparse. The K-space data of the target object is split into odd-row echo data and even-row echo data. Based on the information redundancy between the odd-row and even-row echo data in the K-space data, a low-rank constraint is constructed. The low-rank matrix can be a Hankel matrix or other matrices; no restriction is placed here.

[0089] Optionally, Figure 6 A schematic diagram for generating the Hank matrix is ​​provided. For example... Figure 6 As shown, the acquired EPIPPA2 sequence is split to obtain an odd-echo dataset containing odd-numbered echo data and an even-echo dataset containing even-numbered echo data. R_SENSE (regularization SENSE) is then applied to both datasets. Specifically, after obtaining the odd-echo and even-echo datasets, odd-echo images are obtained based on the receiver coil's sensitivity information and the odd-echo dataset; even-echo images are obtained based on the receiver coil's sensitivity information and the even-echo dataset. The R_SENSE method is used to perform regularization reconstruction processing on both the odd-echo and even-echo images, and a low-rank matrix is ​​constructed based on the difference between the odd-echo and even-echo images.

[0090] In some of these embodiments, Figure 7 A flowchart of an EPI-based image processing method is provided, including the following steps:

[0091] Step 702: Obtain the EPI data of the target object. The EPI data is obtained through undersampling.

[0092] In the undersampling process, because the sampling rate is lower than the sampling rate required by the Nyquist sampling theorem, there may be missing data information, which may lead to aliasing artifacts in the image. However, sampling the target object by undersampling can reduce scanning time and accelerate imaging speed.

[0093] Step 704: Divide the EPI data into odd echo datasets and even echo datasets. The odd echo dataset is the set of odd-numbered echo data in the EPI data, and the even echo dataset is the set of even-numbered echo data in the EPI data.

[0094] Step 706: Initialize and reconstruct the odd echo dataset and even echo dataset based on machine learning network constraints to obtain the odd echo image and even echo image of the target object.

[0095] In this context, machine learning network constraints refer to inputting odd-echo and even-echo datasets into a machine learning network and using the network's output as constraints during the initial reconstruction process. Optionally, the machine learning network recovers insufficient data information introduced by undersampling by inputting odd-echo and even-echo datasets. Initial reconstruction refers to transforming the undersampled EPI data to the spatial domain using inverse Fourier transform. Optionally, the initial reconstruction of the odd-echo and even-echo datasets can be achieved iteratively. Initial reconstruction based on machine learning network constraints ensures that the reconstructed odd-echo and even-echo images are constrained by the machine learning network's output, effectively reducing image domain aliasing artifacts caused by insufficient sampling information during initialization.

[0096] Step 708: Regularize the odd echo image and even echo image to obtain the magnetic resonance image of the target object.

[0097] Regularized reconstruction refers to the process of reconstructing magnetic resonance images based on odd-echo and even-echo images by restricting the solution space through regularization terms, so that the reconstructed magnetic resonance images have the characteristics of low artifacts and low noise.

[0098] In this embodiment, odd-echo and even-echo datasets are obtained by dividing the EPI data, and odd-echo and even-echo images with better image quality and partial aliasing artifacts are reconstructed based on machine learning network constraints. Regularized reconstruction processing is then applied to the odd-echo and even-echo images to further suppress noise in the magnetic resonance images, resulting in magnetic resonance images with low aliasing artifacts and high imaging quality.

[0099] In some embodiments, the initial reconstruction of odd echo datasets and even echo datasets based on machine learning network constraints includes: inputting the odd echo dataset into the machine learning network for reconstruction and recovery, and using the intermediate dataset obtained by the SENSE reconstruction algorithm during the reconstruction and recovery process as a constraint term for the reconstruction and recovery of the machine learning network.

[0100] The SENSE reconstruction algorithm is a parallel imaging technique that utilizes multiple receiving coils to acquire data simultaneously. It estimates missing K-space data based on the odd-echo dataset, achieving low-noise reconstruction results with shorter acquisition and reconstruction times. Optionally, before reconstruction using the SENSE algorithm, data received by different coils can be registered based on the sensitive area matrix of the receiving coils to calibrate the odd-echo dataset. Alternatively, during reconstruction using the SENSE algorithm, regularization can be used to further remove noise from the image. In this embodiment, the intermediate dataset reconstructed by the SENSE algorithm is used as a constraint term to reduce aliasing artifacts in both odd-echo and even-echo images while significantly improving the image signal-to-noise ratio.

[0101] Based on the same principle, the initial reconstruction of odd-echo and even-echo datasets using machine learning network constraints involves: inputting the even-echo dataset into the machine learning network for reconstruction, and using the intermediate dataset obtained by the SENSE reconstruction algorithm as a constraint term in the machine learning network reconstruction process. Specific limitations can be found in the method for initial reconstruction based on odd-echo datasets using machine learning network constraints, which will not be elaborated further.

[0102] In some embodiments, the regularization reconstruction of odd-echo and even-echo images includes: determining a low-rank matrix based on the odd-echo and even-echo images, the low-rank matrix reflecting the phase difference between the odd-echo and even-echo images; obtaining a low-rank constraint based on the low-rank matrix, the low-rank constraint serving as a constraint term in the regularization reconstruction of odd-echo and even-echo images.

[0103] For example, the phase difference between odd-echo and even-echo images in the image domain is obtained, and a low-rank matrix is ​​constructed based on the phase difference. The low-rank matrix is ​​then expanded in the scanning channel of the target object to obtain low-rank constraints. In this embodiment, the low-rank constraints are used as constraints in the regularized reconstruction processing of odd-echo and even-echo images. This allows for further regularization constraints on the initialization results of the odd-even echo separation, thereby obtaining a reconstruction result with a high signal-to-noise ratio.

[0104] In one embodiment, a magnetic resonance image processing method for high-field EPI is provided. First, the K-space data obtained by high-field EPI sampling is split into even and odd echoes, and the split even and odd echo data is initialized to obtain downsampling results.

[0105] This approach transforms the solution to the phase difference mismatch between odd and even echoes into an accelerated high-magnification downsampling problem by splitting the odd and even echoes. Therefore, considering the doubling of the downsampling factor in a single-polarity K-space, a parallel imaging technique like SENSE can be used for image initialization. For example, multi-channel images with aliasing artifacts are acquired from an MRI machine. The coil sensitivity map is estimated using the separately acquired odd and even echo data. Regularized SENSE techniques are then used to initialize the multi-channel image based on the coil sensitivity matrix, resulting in an initialized downsampling result. It is understood that other algorithms for high-magnification downsampling, including advanced parallel imaging algorithms such as the ESPIRiT algorithm and the RAKI algorithm, can also be used to initialize the split odd and even echo data; these will not be elaborated upon here.

[0106] Secondly, the initialized downsampling results are input into the trained network to obtain the final artifact-free full-sampled image. The artifact-free full-sampled image includes the full-sampled image corresponding to odd-echo data and the full-sampled image corresponding to even-echo data.

[0107] Optionally, the trained network route generator and discriminator are constructed based on downsampled and full-sampled images. Cycle consistency loss is used as the L1 loss, and discriminator loss is used as the MSE (mean-square error) loss. Training terminates when the loss converges. It is understood that other network structures, such as UNet and ResNet, can also be used, along with other existing loss functions, to generate the network and obtain high signal-to-noise ratio, smooth sampled images.

[0108] After obtaining the full-sampled images corresponding to odd-echo data and even-echo data, an annihilation filter that meets the low-rank condition can be constructed by utilizing the information redundancy of the similar amplitudes of the odd and even echo images.

[0109] The expression for the EPI K-space is as follows:

[0110]

[0111] Based on the odd and even echoes, the EPI K-space is equivalently separated into two sets of virtual K-spaces.

[0112] The K-space expression for EPI odd echoes is as follows:

[0113]

[0114] The K-space expression for EPI even echoes is as follows:

[0115]

[0116] Where, A(x, y) = a(x, y)e ι2π[Δf(x，y)((TE+(n-N / 2)ESP))] ;

[0117] k x To read the direction data, k y TE represents the data in the direction of the encoded frequency; ESP represents the echo time; Δf(x,y) represents the offset from the resonant frequency at the spatial location (x,y); and a(x,y) represents the magnetization intensity of the protons after transverse magnetization in the direction perpendicular to the main magnetic field, which is the source of the MRI signal.

[0118] In the above formula and The difference between them is the main source of eddy current artifacts. Calculation and The difference between them:

[0119]

[0120]

[0121]

[0122] in, It has sparsity, therefore it corresponds to the Hankel matrix constructed using this difference in K-space.

[0123]

[0124]

[0125] By introducing the Hankel matrix into the scan channel (r) and expanding it, we can obtain a low-rank matrix:

[0126]

[0127] Using this low-rank matrix, we can further regularize the initialization results of the odd-even echo separation to obtain a reconstruction result with a high signal-to-noise ratio.

[0128]

[0129] Where d represents the actual downsampling K-space data, Ω is a logical matrix used to mark the two-dimensional downsampling trajectory, C is the coil sensitivity matrix, F refers to the Fourier transform, and ρ is the preset first magnetic resonance image. λ1 is the weight coefficient used to constrain the AI ​​term, and Ψ(ρ) is the regularization term obtained based on wavelet transform. netThis is the output of the convolutional neural network. λ2 is the weight coefficient used to constrain the low-rank matrix. It is a regularization constraint derived from a low-rank matrix.

[0130] The convergence of the left side of the formula to the minimum means that the image result is close to the actual sampled result in the K-space region within the downsampling trajectory. The network output on the right side, as a regularization term, aims to constrain the iteration with the high signal-to-noise ratio result learned by the network. The low-rank matrix on the right side, as a regularization term, is used to reduce artifacts caused by the inconsistency between the phases of odd and even echoes.

[0131] This embodiment combines a method that uses prior information obtained from convolutional neural networks and K-space filtering based on low-rank class constraints. It leverages the robustness of convolutional neural network constraints to significantly optimize the image initialization quality under different whole-body scenarios, and optimizes the image quality degradation caused by natural physiological movements such as breathing, gastrointestinal peristalsis, and heartbeat during the traditional parallel imaging solution process. Furthermore, it reduces artifacts caused by phase inconsistency by constructing a low-rank annihilation filter through information redundancy between odd and even echo data, thereby significantly improving the signal-to-noise ratio of the final image and obtaining a high-field EPI scanning result with low artifacts.

[0132] Based on the same inventive concept, this application also provides a magnetic resonance image processing apparatus for implementing the magnetic resonance image processing method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more embodiments of the magnetic resonance image processing apparatus provided below can be found in the limitations of the magnetic resonance image processing method described above, and will not be repeated here.

[0133] In one embodiment, such as Figure 8 As shown, a magnetic resonance image processing device is provided, comprising: an acquisition module, a correction module, and an iteration module, wherein:

[0134] The acquisition module is used to obtain the first sampled image of the target object based on the sensitivity information of the receiving coil corresponding to the target object and the K-space data;

[0135] The correction module is used to correct artifacts in the first sampled image to obtain a second sampled image of the target object;

[0136] The iterative module is used to iterate the data of the preset first magnetic resonance image with the second sampled image as a constraint and the K-space data as the fidelity term in the iterative reconstruction process to obtain the magnetic resonance image of the target object.

[0137] In some embodiments, the iterative module uses the second sampled image as a constraint and K-space data as a fidelity term in the iterative reconstruction process to iterate the preset magnetic resonance data to obtain the magnetic resonance image of the target object, including: obtaining a regularization term based on the difference between the first magnetic resonance image and the second sampled image; iterating the first magnetic resonance image, and stopping the iteration and obtaining the magnetic resonance image of the target object when the difference between the first magnetic resonance image and the first sampled image and the value of the regularization term converge to a first convergence condition.

[0138] In some embodiments, the iteration module iterates over the data of a preset first magnetic resonance image, and further includes: constructing a low-rank matrix based on the similarity of amplitude and the difference in phase between odd-numbered and even-numbered echo data in the K-space data; expanding the low-rank matrix in the scanning channel of the target object, and using the expanded matrix as a low-rank constraint in the iteration process of the first magnetic resonance image.

[0139] In some embodiments, the acquisition module acquires a first sampled image of the target object, including: obtaining a second magnetic resonance image based on the sensitivity information of the receiving coil and K-space data; iterating on the second magnetic resonance image; if the difference between the data of the second magnetic resonance image and the K-space data satisfies a second convergence condition, then stopping the iteration and using the second magnetic resonance image as the first sampled image.

[0140] In some embodiments, the correction module corrects artifacts in the first sampled image to obtain a second sampled image of the target object, including: obtaining a mapping function, which maps the data distribution of the first sampled image to the data distribution of the second sampled image; and obtaining the second sampled image corresponding to the first sampled image according to the mapping function.

[0141] Optionally, obtaining the mapping function includes: training an adversarial network based on a training dataset, the training dataset including sampled images of the training object with and without artifacts; adjusting the parameters of the adversarial network, and obtaining the mapping function based on the adversarial network after the cycle consistency loss and discriminator loss of the adversarial network converge.

[0142] Each module in the aforementioned magnetic resonance imaging processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0143] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data from the magnetic resonance image processing and acquisition process. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a magnetic resonance image processing method.

[0144] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0145] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0146] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0147] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0148] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0149] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0150] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A magnetic resonance image processing method, characterized in that, The method includes: The first sampled image of the target object is obtained based on the sensitivity information of the receiving coil corresponding to the target object and the K-space data; The artifacts in the first sampled image are corrected to obtain the second sampled image of the target object; Using the second sampled image as a constraint and the K-space data as a fidelity term in the iterative reconstruction process, the data of the preset first magnetic resonance image are iterated to obtain the magnetic resonance image of the target object.

2. The method according to claim 1, characterized in that, Using the second sampled image as a constraint and the K-space data as a fidelity term in the iterative reconstruction process, the preset magnetic resonance data is iterated to obtain the magnetic resonance image of the target object, including: A regularization term is obtained based on the difference between the first magnetic resonance image and the second sampled image; Iterate over the first magnetic resonance image, and stop iterating when the difference between the first magnetic resonance image and the first sampled image, and the value of the regularization term, converge to the first convergence condition, and obtain the magnetic resonance image of the target object.

3. The method according to claim 1, characterized in that, The method further includes iterating over the data of a preset first magnetic resonance image. Based on the similarity of amplitude and the difference in phase between the odd-numbered and even-numbered echo data in the K-space data, a low-rank matrix is ​​constructed. The low-rank matrix is ​​expanded in the scanning channel of the target object, and the expanded matrix is ​​used as a low-rank constraint in the iterative process of the first magnetic resonance image.

4. The method according to claim 1, characterized in that, Obtain the first sampled image of the target object, including: The second magnetic resonance image is obtained based on the sensitivity information of the receiving coil and the K-space data; Iterate over the second magnetic resonance image; If the difference between the data of the second magnetic resonance image and the K-space data satisfies the second convergence condition, then the iteration stops and the second magnetic resonance image is used as the first sampled image.

5. The method according to claim 1, characterized in that, Correcting artifacts in the first sampled image to obtain a second sampled image of the target object includes: Obtain a mapping function, which is used to map the data distribution of the first sampled image to the data distribution of the second sampled image; Based on the mapping function, a second sampled image corresponding to the first sampled image is obtained.

6. The method according to claim 5, characterized in that, Obtaining the mapping function includes: The adversarial network is trained based on a training dataset, which includes sampled images of the training object with and without artifacts. The parameters of the adversarial network are adjusted, and the mapping function is obtained based on the adversarial network when the cycle consistency loss and discriminator loss of the adversarial network converge.

7. A magnetic resonance image processing device, characterized in that, The device includes: The acquisition module is used to obtain a first sampled image of the target object based on the sensitivity information of the receiving coil corresponding to the target object and the K-space data; The correction module is used to correct artifacts in the first sampled image to obtain a second sampled image of the target object; An iterative module is used to iterate the data of a preset first magnetic resonance image using the second sampled image as a constraint and the K-space data as a fidelity term in the iterative reconstruction process to obtain the magnetic resonance image of the target object.

8. An image processing method based on EPI, characterized in that, The method includes: Acquire EPI data of the target object, wherein the EPI data is obtained through undersampling; The EPI data is divided into odd echo datasets and even echo datasets; Initialization and reconstruction based on machine learning network constraints are performed on the odd echo dataset and even echo dataset respectively to obtain the odd echo image and even echo image of the target object; The odd echo and even echo images are regularized and reconstructed to obtain the magnetic resonance image of the target object.

9. The method according to claim 8, characterized in that, The initial reconstruction of the odd-echo and even-echo datasets based on machine learning network constraints includes: The odd echo dataset is input into the machine learning network for reconstruction and restoration, and the intermediate dataset obtained by the SENSE reconstruction algorithm during the reconstruction and restoration process is used as a constraint term for the reconstruction and restoration of the machine learning network.

10. The method according to claim 8, characterized in that, The regularized reconstruction process for the odd-echo and even-echo images includes: Based on the odd-echo and even-echo images, a low-rank matrix is ​​determined, which reflects the phase difference between the odd-echo and even-echo images. The low-rank constraint is obtained from the low-rank matrix, and the low-rank constraint is used as a constraint term in the regularization reconstruction process of the odd echo image and the even echo image.

Images