Motion artifact correction method, motion artifact correction model training method, and apparatus

Abstract

The application discloses a motion artifact correction method and a motion artifact correction model training method and device, and relates to the technical fields of deep learning and image processing. The motion artifact correction method comprises the following steps: acquiring an undersampling factor, wherein the undersampling factor comprises a plurality of regional undersampling factors, and each regional undersampling factor is related to the number of phase encoding lines of the motion artifact existing in the region; performing undersampling processing on to-be-corrected frequency domain data based on the undersampling factor to obtain undersampled frequency domain data; inputting the undersampled frequency domain data into a frequency domain repair model to obtain repaired frequency domain data; and acquiring a corrected image based on the repaired frequency domain data.

Description

Technical Field

[0001] This application relates to the technical fields of deep learning and image processing, and in particular to a motion artifact correction method, a motion artifact correction model training method, and an apparatus. Background Technology

[0002] Motion artifacts are the most common image artifacts in clinical magnetic resonance imaging (MRI), and mitigating them remains a significant challenge in the development of MRI. Motion artifacts in MRI are caused by the patient's (voluntary or involuntary) movements (random) or the pulsatile flow of blood vessels (periodic). The occurrence of motion artifacts is the result of a complex interaction between image structure, type of motion, the specificity of the MR pulse sequence, and k-space acquisition strategies. Typical motion effects in MRI images include: blurred image boundaries, ghosting caused by moving structures, signal loss, and unwanted strong signals. How to better eliminate motion artifacts in images has always been a pressing problem in this field. Summary of the Invention

[0003] The embodiments of this application aim to at least partially solve one of the technical problems in the related art. Therefore, the first objective of the embodiments of this application is to provide a motion artifact correction method, a motion artifact correction model training method, apparatus, device, medium, and program product.

[0004] This application provides a motion artifact correction method, which includes: acquiring undersampling factors, wherein the undersampling factors include multiple regional undersampling factors, and the regional undersampling factors are related to the number of phase coding lines of motion artifacts existing in the region; based on the undersampling factors, performing downsampling processing on the frequency domain data to be corrected to obtain downsampling frequency domain data; inputting the downsampling frequency domain data into a frequency domain restoration model to obtain restored frequency domain data; and acquiring a corrected image based on the restored frequency domain data.

[0005] In some embodiments, obtaining undersampling factors includes: performing residual processing on the motion artifact dataset and the reference dataset to obtain a residual dataset; performing threshold processing on the residual dataset to determine a thresholded artifact dataset; and performing data region division and regional undersampling factor calculation on the thresholded artifact dataset to obtain multiple regional undersampling factors.

[0006] In some embodiments, the thresholded artifact dataset includes Z thresholded artifact data, Z≥1, each thresholded artifact data being a two-dimensional data with X rows and Y columns, where the row direction represents the frequency domain coding direction of image acquisition and the column direction represents the phase coding direction of image acquisition, X>1, Y>1; the thresholded artifact dataset is divided into data regions and regional undersampling factors are calculated to obtain multiple regional undersampling factors, including: dividing the z-th thresholded artifact data into data regions to determine a low-frequency data region and N high-frequency data regions, 1≤z≤Z, N≥1, where the number of rows in the low-frequency data region and each high-frequency data region is less than X and the number of columns is Y; in each high-frequency data region, determining whether the artifact degree represented by each row of data is greater than a preset artifact degree, if so, marking the row as a phase coding line of motion artifacts; determining the number of phase coding lines in each high-frequency data region; and determining N regional undersampling factors based on the number of N phase coding lines corresponding to each of the N high-frequency data regions.

[0007] In some embodiments, the regional undersampling factor is positively correlated with the number of phase coding lines of motion artifacts present in the corresponding high-frequency data region.

[0008] In some embodiments, determining N region undersampling factors based on the number of N phase coding lines corresponding to each of the N high-frequency data regions includes: determining the proportion of the number of each phase coding line in the corresponding high-frequency data region; obtaining N sets of row number proportion values ​​based on the row number proportion, each set of row number proportion values ​​including Z row number proportion values; averaging the Z row number proportion values ​​in each set to obtain N average proportion values; and determining the N region undersampling factors based on the N average proportion values.

[0009] In some embodiments, residual processing is performed on the motion artifact dataset and the reference dataset to obtain a residual dataset, including: calculating the image difference between each motion artifact image in the motion artifact dataset and the corresponding parameter image in the reference dataset; and transforming the image difference into K space to obtain the residual dataset.

[0010] In some embodiments, the frequency domain data to be corrected is downsampled based on an undersampling factor to obtain downsampled frequency domain data, including: obtaining artifact mask data based on the undersampling factor; and downsampling the frequency domain data to be corrected based on the artifact mask data to obtain downsampled frequency domain data.

[0011] In some embodiments, obtaining artifact mask data based on the undersampling factor includes: determining the number of artifact rows corresponding to the nth high-frequency data region based on the undersampling factor of the nth region and the total number of uplink data in the nth high-frequency data region, thereby obtaining N artifact row numbers corresponding to the N high-frequency data regions, where 1≤n≤N; determining the first row of data randomly corresponding to the N artifact row numbers in the corresponding high-frequency data region as a first value, and determining the second row of data other than the first row of data in the corresponding high-frequency data region as a second value; and generating two-dimensional data of X rows and Y columns as artifact mask data based on the first row of data, the second row of data, and the low-frequency data region.

[0012] In some embodiments, obtaining a corrected image based on frequency domain inpainting data includes: converting the frequency domain inpainting data into image domain data; inputting the image domain data into an image domain inpainting model to obtain the corrected image.

[0013] This application provides a method for training a motion artifact correction model, which includes a frequency domain restoration model. The method includes: acquiring undersampling factors, wherein the undersampling factors include multiple regional undersampling factors, and the regional undersampling factors are related to the number of phase coding lines of motion artifacts existing in the region; acquiring frequency domain data samples to be corrected and reference frequency domain data samples corresponding to the frequency domain data samples to be corrected; performing downsampling processing on the frequency domain data samples to be corrected based on the undersampling factors to obtain downsampling frequency domain data samples; inputting the downsampling frequency domain data samples into the frequency domain restoration model to obtain restored frequency domain data samples; determining a frequency domain loss value based on the restored frequency domain data samples and the reference frequency domain data samples; and adjusting the model parameters of the frequency domain restoration model based at least on the frequency domain loss value to obtain a trained motion artifact correction model.

[0014] In some embodiments, the motion artifact correction model further includes an image domain inpainting model; the method further includes: acquiring a reference image sample corresponding to the frequency domain data sample to be corrected; converting the frequency domain inpainting data sample into an image domain inpainting data sample, and inputting the image domain inpainting data sample into the image domain inpainting model to obtain a corrected sample image; determining an image domain loss value based on the corrected image sample and the reference image sample; and adjusting the model parameters of the image domain inpainting model based at least on the image domain loss value to obtain a trained motion artifact correction model.

[0015] In some embodiments, the frequency domain inpainting model and the image domain inpainting model are jointly trained.

[0016] This application provides a motion artifact correction device, comprising: an acquisition module for acquiring undersampling factors, wherein the undersampling factors include multiple regional undersampling factors, and the regional undersampling factors are related to the number of phase coding lines of motion artifacts existing in the region; a downsampling module for downsampling the frequency domain data to be corrected based on the undersampling factors to obtain downsampling frequency domain data; a repair module for inputting the downsampling frequency domain data into a frequency domain repair model to obtain repaired frequency domain data; and a correction module for acquiring a corrected image based on the repaired frequency domain data.

[0017] This application provides a motion artifact correction model training device, the motion artifact correction model including a frequency domain repair model; the device includes: a factor acquisition module, used to acquire undersampling factors, wherein the undersampling factors include multiple regional undersampling factors, and the regional undersampling factors are related to the number of phase coding lines of motion artifacts existing in the region; a sample acquisition module, used to acquire frequency domain data samples to be corrected and reference frequency domain data samples corresponding to the frequency domain data samples to be corrected; a sample downsampling module, used to downsampling the frequency domain data samples to be corrected based on the undersampling factors to obtain downsampling frequency domain data samples; a sample repair module, used to input the downsampling frequency domain data samples into the frequency domain repair model to obtain repaired frequency domain data samples; a determination module, used to determine the frequency domain loss value based on the repaired frequency domain data samples and the reference frequency domain data samples; and an adjustment module, used to adjust the model parameters of the frequency domain repair model based at least on the frequency domain loss value to obtain a trained motion artifact correction model.

[0018] In some embodiments, the motion artifact correction model further includes an image domain inpainting model; the method further includes: acquiring a reference image sample corresponding to the frequency domain data sample to be corrected; converting the frequency domain inpainting data sample into an image domain inpainting data sample, and inputting the image domain inpainting data sample into the image domain inpainting model to obtain a corrected sample image; determining an image domain loss value based on the corrected image sample and the reference image sample; and adjusting the model parameters of the image domain inpainting model based at least on the image domain loss value to obtain a trained motion artifact correction model.

[0019] In some embodiments, the frequency domain inpainting model and the image domain inpainting model are jointly trained.

[0020] This application provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the above embodiments.

[0021] This application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method of any of the above embodiments.

[0022] This application provides a computer program product that includes instructions that, when executed by a processor of a computer device, enable the computer device to perform the steps of the method described in any of the above embodiments. Attached Figure Description

[0023] Figure 1 A schematic diagram of a retrospective correction based on the image domain is shown.

[0024] Figure 2 A schematic diagram of a k-space-based retrospective correction is shown.

[0025] Figure 3 A flowchart of the motion artifact correction method provided in the embodiments of this application.

[0026] Figure 4 This is a schematic diagram of a residual image acquisition method provided in an embodiment of this application.

[0027] Figure 5 A schematic diagram illustrating the method for acquiring artifact mask data provided in this application.

[0028] Figure 6 This is a schematic diagram illustrating the region division of thresholded artifact data provided in an embodiment of this application.

[0029] Figure 7 A flowchart of the motion artifact correction model training method provided in the embodiments of this application.

[0030] Figure 8 A schematic diagram of the motion artifact correction model training method provided in the embodiments of this application.

[0031] Figure 9 A schematic diagram of the motion artifact correction model provided in the embodiments of this application.

[0032] Figure 10 This is a schematic diagram of motion artifact correction results provided in an embodiment of this application.

[0033] Figure 11 A schematic diagram of a motion artifact correction device provided in an embodiment of this application.

[0034] Figure 12 A schematic diagram of a motion artifact correction model training device provided in an embodiment of this application.

[0035] Figure 13 A block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0036] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0037] Motion of objects in MRI imaging affects the magnetic resonance signal in k-space. On one hand, changes in position can impair the ability to encode spatial information in the acquired signal; on the other hand, the physical properties of the MR signal are negatively affected by second-order motion effects, such as magnetic field inhomogeneities or spin history effects caused by motion. Regardless of the type of motion, motion artifacts ultimately corrupt the k-space data, and this corruption is generally only reflected in the phase-encoded direction. 1. Motion along any magnetic field gradient direction will cause the accumulation of abnormal phases, which will lead to incorrect plotting of the signal in the phase encoding direction.

[0038] 2. The sampling time (milliseconds) for a signal in the frequency encoding direction is significantly shorter than the time for a single phase encoding (seconds). In clinical MRI applications, most motion is significantly slower than the rapid sampling process along the frequency encoding axis. This inconsistency between the frequency and phase encoding periods causes motion artifacts to mainly appear along the phase encoding axis (of course, motion artifacts may also appear in the frequency encoding direction, but they are usually insignificant and at most only cause very minor blurring).

[0039] In magnetic resonance imaging (MRI), complex artifacts may appear after reconstructing motion-destroyed data from the frequency domain to the image domain. The different types of motion associated with MRI in vivo imaging mainly include: involuntary rigid motion, elastic motion caused by heartbeats, respiration, and blood flow. These motions can be broadly classified into three types: periodic motion, random motion, and flow.

[0040] Periodic motion is caused by the pulsation or periodic movement of blood vessels, the heart, or cerebrospinal fluid. Common pulsation artifacts are equidistant replicas of pulsating structures, such as the aorta, in the phase-encoding direction, and generally decrease in intensity as the distance from the original structure increases. Clinical countermeasures: 1. Use spatial pre-saturation pulses to saturate the incoming protons and reduce artifacts. 2. Increase TR, Ny, or NEX (equivalent to increasing scan time), thereby increasing the interval between artifacts. 3. Exchange the phase and frequency encoding directions to change the direction of artifacts, thus differentiating lesions from artifacts. 4. Use ECG gating. 5. Use flow compensation.

[0041] Random motion is caused by the patient's voluntary or involuntary movements (e.g., breathing, changing position, swallowing, tremors, coughing, etc.). It can cause image blurring or result in many parallel bands in the phase-encoding direction. Clinical countermeasures: 1. The most effective approach is to instruct the patient not to move. 2. Respiratory compensation (RC). 3. Use glucagon in the abdomen to reduce artifacts caused by intestinal peristalsis. 4. Sedation and analgesia. 5. Faster scanning (FSE, GRE, EPI, etc.).

[0042] Proton dephase caused by cerebrospinal fluid flow can sometimes resemble a pathological condition. Flow compensation techniques can reduce this effect.

[0043] The aforementioned periodic motion, random motion, and cerebrospinal fluid flow are currently corrected primarily by interventions performed before MRI scans (such as adding saturation bands, adjusting TR, changing phase encoding direction, administering medication, etc.) or during MRI acquisition (such as ECG gating, rapid FSE scanning, etc.), rather than by techniques to correct data images that already exhibit motion artifacts.

[0044] For motion correction, the following two methods are generally included: 1. Look-ahead calibration is an intuitive method in which the relative spatial position and orientation between the scan coordinates and the object of interest remain constant. For example, in rigid head movements, if the position and orientation of the moving object can be measured in real time, the magnetic field gradient, radio frequency pulse, and receiving frequency and phase can be adjusted accordingly (if the moving object rotates, the encoding ladder needs to be rotated; if the moving object translates, the transmission and reception frequencies and phases need to be changed). In short, look-ahead calibration maintains the relative stillness of the scan coordinates and the object of interest by adjusting the magnetic field gradient, radio frequency pulse, and other related measures in real time, thereby ensuring data quality during acquisition.

[0045] 2. Retrospective correction improves data consistency after acquisition by modifying the acquired data or reconstruction process, thereby correcting motion artifacts. The basic idea of ​​retrospective correction is to eliminate motion-related changes in magnetic resonance data. For rigid body motion, these changes are described by Fourier's theorem: translation of the object causes a phase ramp in the acquired k-space, while rotation corresponds to rotation in k-space. Translation is relatively easy to correct by applying phase changes to the acquired data, while rotation correction requires the use of non-Cartesian reconstruction methods.

[0046] Figure 1 A schematic diagram of a retrospective correction based on the image domain is shown.

[0047] like Figure 1As shown, image-based motion correction methods take an image affected by motion as input to generate a corrected image, similar to denoising and deblurring tasks in the image domain. It mainly has two directions: classical training, which uses an image affected by motion to predict the predicted image. x Based on predicted images x and real images x GT Adjust the model parameters based on the loss between them to make the predicted image x As close as possible to the real image x GT Adversarial training involves predicting images based on the effects of motion. x Additional training of a discriminant network is added to predict the image. x With real images x GT To facilitate competition and differentiation, the loss includes generative loss and adversarial loss. Image domain-based correction typically uses an encoder-decoder structure with convolutional operations, mainly consisting of a downsampling part for feature encoding and an upsampling part for feature extraction during decoding. The differences between image domain correction methods lie primarily in network architecture, the use of prior information, and training methods.

[0048] Figure 2 A schematic diagram of a k-space-based retrospective correction is shown.

[0049] like Figure 2 As shown, motion artifact correction based on k-space can utilize additional information from the original data during reconstruction. The main directions include: 1. Using deep learning modules to replace different components in the reconstruction process, targeting image initialization, loss functions, or motion parameters, representing optimization of parameters. θ To minimize the loss function L To obtain the reconstructed image 2. Or combine classical motion detection or estimation with deep learning reconstruction; 3. Or combine the reconstruction process with deep learning's detection of corrupted k-space data (data consistency module), etc.

[0050] Since it is impossible to obtain perfectly consistent pairs of data with and without motion artifacts simultaneously in reality, data simulation is necessary for both the image domain and k-space. However, even the best data simulation currently available cannot perfectly match real data; data with and without motion artifacts need to be strictly paired, requiring high consistency between pairs; incomplete removal of motion artifacts results in poor image quality and often introduces blur artifacts; and the supported scenarios are limited (only supporting specific parts of the whole body or partial sequences).

[0051] In view of this, embodiments of this application propose a retrospective motion correction method. Specifically, embodiments of this application propose a motion artifact correction method and a motion artifact correction model training method. The data required by the embodiments of this application can be either collected paired data with / without motion artifacts or simulated data with / without motion artifacts; and the requirements for data consistency are low, the motion artifact removal effect is good, and it supports 2D or 3D sequences of various parts of the body.

[0052] Figure 3 A flowchart of the motion artifact correction method provided in the embodiments of this application.

[0053] like Figure 3 As shown, the motion artifact correction method 300 provided in this application includes steps S310-S340.

[0054] Step S310: Obtain the undersampling factor, wherein the undersampling factor includes multiple regional undersampling factors, and the regional undersampling factor is related to the number of phase coding lines of motion artifacts existing in the region.

[0055] Step S320: Based on the undersampling factor, the frequency domain data to be corrected is downsampled to obtain downsampled frequency domain data.

[0056] The undersampling factor can be obtained by performing residual processing on the image. For example, residual processing can be performed on a set of motion artifact images and a corresponding set of reference images to obtain the undersampling factor. The set of motion artifact images may include at least one motion artifact image, and the set of reference images may include at least one corresponding reference image. The reference image is, for example, an image without motion artifacts. A motion artifact image and a corresponding reference image constitute an image pair. The undersampling factor is obtained by performing residual processing on each image pair. The undersampling factor represents the artifact information present in the motion artifact image.

[0057] Image data can be divided into multiple regions, each corresponding to a region undersampling factor. For any given region, there are multiple phase-coded lines. Some phase-coded lines contain motion artifacts, while others do not. The region undersampling factor is related to the number of phase-coded lines with motion artifacts in that region. For example, the region undersampling factor is positively correlated with the number of phase-coded lines with motion artifacts in the corresponding region; the larger the number of phase-coded lines with motion artifacts in the region, the larger the region undersampling factor.

[0058] The frequency domain data to be corrected can be obtained directly through scanning or acquisition, or it can be obtained based on the motion artifact image to be corrected. For example, the motion artifact image to be corrected can be acquired and then converted into the frequency domain data to be corrected. The motion artifact image to be corrected can be one or more motion artifact images from a set of motion artifact images, or it can be other images. For example, based on the Fourier transform method, a Fourier transform can be performed on the motion artifact image to be corrected, transforming the motion artifact image to be corrected into the frequency domain, thus obtaining the frequency domain data to be corrected.

[0059] After obtaining the undersampling factor, downsampling can be performed based on the undersampling factor. Specifically, artifact mask data can be obtained based on the undersampling factor, and then downsampling can be performed on the frequency domain data to be corrected based on the artifact mask data to obtain downsampling frequency domain data.

[0060] Since the undersampling factor is obtained by processing the historical motion artifact image set and the corresponding reference image set, it is a position estimate or position statistics of the phase coding line of the motion artifact in the K space. It reflects a probability distribution of the phase coding line of the motion artifact in the K space. The purpose is to roughly eliminate the phase coding line of the motion artifact in the frequency domain data to be processed. Therefore, the consistency requirement for pairing the historical motion artifact image set and the corresponding reference image set is relatively low.

[0061] Step S330: Input the downsampled frequency domain data into the frequency domain repair model to obtain the repaired frequency domain data.

[0062] Step S340: Obtain the corrected image based on the frequency domain repair data.

[0063] The downsampled frequency domain data undergoes preliminary motion artifact correction using an undersampling factor. After this initial correction, the downsampled frequency domain data may contain little or no motion artifact information (the artifact severity has been reduced). Next, the downsampled frequency domain data can be restored using a frequency domain inpainting model to obtain fully sampled or equivalent repaired frequency domain data. If the repaired frequency domain data does not contain artifact information, it can be directly inverse Fourier transformed to the image domain to obtain a corrected image. Alternatively, the image transformed to the image domain can be filtered or processed using a deep learning model to obtain a corrected image. In other words, this operation can include further artifact correction to generate a corrected image in the image domain, but it is not mandatory.

[0064] In the embodiments of this application, an undersampling factor is obtained based on the number of phase coding lines of motion artifacts in each region. Then, a downsampling process is performed based on the undersampling factor to achieve preliminary motion artifact correction. After preliminary artifact correction using the undersampling factor, a frequency domain repair model is further used to repair the downsampled frequency domain data to achieve artifact correction. This reduces the consistency requirements for image pairs, reduces the difficulty of data acquisition or data simulation, and improves the effect of artifact correction.

[0065] Figure 4 This is a schematic diagram of a residual image acquisition method provided in an embodiment of this application.

[0066] like Figure 4 As shown, the motion artifact image set and the reference image set can be three-dimensional data, i.e., including Z two-dimensional motion artifact images and Z reference images, where Z represents the number of scan slices or the number of image channels, or the product of the number of scan slices and the number of image channels, Z≥1. The image difference between each motion artifact image in the motion artifact dataset and the corresponding parameter image in the reference dataset is calculated, and the image difference is transformed into K-space to obtain the residual dataset.

[0067] Figure 4 This demonstrates how a motion artifact image (Motion_Image) and its corresponding reference image (GT_Image, the gold standard data) are normalized, followed by residual processing (which includes image subtraction) to obtain a residual image (Residue_Image). The normalization process ensures that the image values ​​are within the same range. Finally, Z residual images (Residue_Image) are obtained, forming a residual image set.

[0068] Figure 5 A schematic diagram illustrating the method for acquiring artifact mask data provided in this application.

[0069] like Figure 5 As shown, residual processing is performed on the motion artifact image Motion_Image and the corresponding reference image GT_Image (gold standard data) to obtain the residual image Residue_Image. A frequency domain transformation is then performed on the residual image set, that is, each image difference is transformed into k-space to obtain the residual dataset. The residual dataset can be a residual frequency domain dataset, which also includes Z residual frequency domain data Residue_k. For example, a Fourier transform is performed on the residual image to transform it into the k-space frequency domain, resulting in the residual frequency domain data Residue_k.

[0070] Then, the residual frequency domain data Residue_k is subjected to thresholding and expansion processing to obtain undersampling factors. Based on multiple undersampling factors, artifact mask data Mask_Resi is obtained. This Mask_Resi is then used to downsample the frequency domain data to be corrected, resulting in downsampled frequency domain data. For example, thresholding is performed on the residual dataset (residual frequency domain dataset) to determine the thresholded artifact dataset. The thresholded artifact dataset includes Z thresholded artifact datasets Residue_th_k. Specifically, for each residual frequency domain dataset Residue_k, data exceeding the artifact severity threshold are selected, retaining the more obvious artifact information to obtain the thresholded artifact dataset Residue_th_k. After obtaining the thresholded artifact dataset Residue_th_k, it can be expanded, divided into regions, and its regional undersampling factors calculated to obtain undersampling factors. The specific process is detailed below. Figure 6 .

[0071] Figure 6 This is a schematic diagram illustrating the region division of thresholded artifact data provided in an embodiment of this application.

[0072] like Figure 6 As shown, the thresholded artifact dataset after thresholding includes Z thresholded artifact data points, Residue_th_k (…). Figure 6 The diagram illustrates a thresholded artifact data set (Residue_th_k), where Z≥1. Each thresholded artifact data set (Residue_th_k) is a two-dimensional dataset with X rows and Y columns. The row direction represents the frequency domain coding direction of the image acquisition, and the column direction represents the phase coding direction of the image acquisition, where X>1 and Y>1. The PE (Phase Encoding Line) is a single row of data in the K-space, corresponding to a complete signal acquired under a specific phase coding.

[0073] The thresholded artifact dataset is divided into data regions and the regional undersampling factor is calculated to obtain multiple regional undersampling factors, including the following:

[0074] For the z-th thresholded artifact data Residue_th_k, divide the data into low-frequency data regions and N high-frequency data regions, where 1≤z≤Z, N≥1. The number of rows in both the low-frequency and high-frequency data regions is less than X, and the number of columns is Y. For example... Figure 6As shown, taking a thresholded artifact data Residue_th_k as an example, there are N high-frequency data regions U1, U2, U3, U4, D1, D2, D3, D4, and a low-frequency data region ACS_line. The low-frequency data region ACS_line is located in the middle, and the high-frequency data regions U1, U2, U3, U4 and D1, D2, D3, D4 are located at the two ends respectively.

[0075] For example, the size of the thresholded artifact dataset is [X, Y, Z], and the size of the z-th thresholded artifact data Residue_th_k (i.e., each thresholded artifact data Residue_th_k) is [X, Y]. We divide p*Y into 2M equal parts, where p is a number less than 1, and p can be 0.9, 0.85, 0.8, etc., and 2M = N, where M is a positive integer. Figure 6 The diagram shows that M=4, i.e. N=8, and the partitioning result is divided into two parts, upper and lower M parts, which are U1+U2+…+UM+DM+…+D2+D1. The remaining part is denoted as the low-frequency data area ACS_line and all its data is set to 1 (that is, it represents important low-frequency data that needs to be retained).

[0076] In each high-frequency data region, it is determined whether the artifact level represented by each row of data is greater than a preset artifact level. If so, the phase coding line of the motion artifact present in that row is marked. For example, if more than 10% of the data points in a row indicate the presence of artifact information, the phase coding line of the motion artifact present in that row is marked. Then, the number of phase coding lines in each high-frequency data region is determined. Based on the number of N phase coding lines corresponding to each of the N high-frequency data regions, N region undersampling factors are determined. The region undersampling factor is positively correlated with the number of phase coding lines of motion artifacts present in the corresponding high-frequency data region. In some embodiments, the preset artifact level threshold may not be set; as long as data points with artifact information are present, the phase coding line of the motion artifact present in that row is marked, although this will result in the loss of some original information.

[0077] For example, determine the proportion of the number of rows in the corresponding high-frequency data region for each phase coding line; based on the proportion of the number of rows, obtain N sets of proportion values, each set of proportion values ​​including Z proportion values; calculate the average of the Z proportion values ​​in each set to obtain N average proportion values; based on the N average proportion values, determine N regional undersampling factors.

[0078] Specifically, for a thresholded artifact data Residue_th_k, the proportion of the number of marked rows in each high-frequency data region is determined. For example, for a high-frequency data region U1, if region U1 includes 100 rows (i.e., 100 PE lines), and the number of marked rows in region U1 is 20, then the proportion of the number of rows in this region is 0.2. The proportion of the number of rows corresponding to each high-frequency data region is calculated in this way.

[0079] Based on the row quantity ratio, N sets of row quantity ratio values ​​are obtained for N regions. For Z thresholded artifact data Residue_th_k, each set of row quantity ratio values ​​includes Z row quantity ratio values. The average of the Z row quantity ratio values ​​in each set is calculated to obtain N average ratio values. Based on the N average ratio values, the under-collection factor for N regions is determined.

[0080] For example, for Z thresholded artifact data Residue_th_k, there are Z high-frequency data regions U1. The percentage of rows in each of the Z high-frequency data regions U1 is calculated, yielding the Z row percentages. The average of these Z row percentages is the region undersampling factor corresponding to the high-frequency data region U1, denoted as DSF. u1 As shown in formula (1), the number of rows (PE lines) marked in the U1 region for the Z thresholded artifact data Residue_th_k are S u11 ,S u12 ...,S u1Z .

[0081] (1) Where Z represents the number of scan slices or the number of image channels, or the product of the number of scan slices and the number of image channels.

[0082] Similarly, the average of the row count proportions of the Z high-frequency data regions U2 is denoted as DSF. u2 ...; The average percentage of the number of rows in the Z high-frequency data regions D1 is denoted as DSF. d1 Therefore, the under-collection factors for N regions include DSF. u1 DSF u2 ... DSF uM DSF dM ... DSF d2 DSF d1 .

[0083] In another example of this application, N regional undersampling factors correspond to N high-frequency data regions. After obtaining the N regional undersampling factors, the artifact mask data Mask_Resi can be obtained based on the N regional undersampling factors.

[0084] For example, for any one of the N regional undersampling factors (such as the nth regional undersampling factor), based on the nth regional undersampling factor and the total number of uplink data in the nth high-frequency data region, the number of artifact lines corresponding to the nth high-frequency data region is determined. This yields the N artifact line counts corresponding to the N high-frequency data regions, where 1 ≤ n ≤ N. For instance, if the nth regional undersampling factor is 0.3 and the total number of uplink data in the nth high-frequency data region is 100, multiplying the nth regional undersampling factor and the total number of uplink data in the nth high-frequency data region yields the number of artifact lines corresponding to the nth high-frequency data region (e.g., 30). This results in the number of artifact lines corresponding to each high-frequency data region.

[0085] It can be seen that the total number of N artifact lines Mnum As shown in formula (2).

[0086] Mnum = (p*Y / 2M)*(DSF) u1 +……+DSF uM + DSF d1 +……+DSF dM (2) Then, based on the number of N artifact rows, artifact mask data Mask_Resi is generated. For example, in the corresponding high-frequency data region, the first row of data corresponding to the N random number of artifact rows is determined as the first value (the first value is, for example, 0), and the second row of data other than the first row of data in the corresponding high-frequency data region is determined as the second value (the first value is, for example, 1). Based on the first row of data, the second row of data, and the low-frequency data region, two-dimensional data of X rows and Y columns is generated as the artifact mask data Mask_Resi.

[0087] For example, after obtaining the number of N artifact lines, multiple rows of data are randomly selected from each of the N high-frequency data regions. The number of rows of randomly selected data is the number of artifact lines corresponding to each region. The row positions of the randomly selected data in the corresponding region are random. The N high-frequency data regions and the low-frequency data region ACS_line are combined in sequence (the values ​​in the low-frequency data region are the original values, for example, 1) to obtain the final artifact mask data Mask_Resi. The artifact mask data Mask_Resi is a two-dimensional data with X rows and Y columns. A value of 1 in the artifact mask data Mask_Resi indicates that the corresponding information in the frequency domain data to be rectified needs to be retained when rectifying the frequency domain data, and a value of 0 indicates that the corresponding information in the frequency domain data to be rectified needs to be removed when rectifying the frequency domain data.

[0088] After obtaining the artifact mask data Mask_Resi, the frequency domain data to be corrected is multiplied by the artifact mask data Mask_Resi to achieve motion artifact correction processing by downsampling the frequency domain data to be corrected, and the downsampled frequency domain data is obtained.

[0089] In another embodiment of this application, after obtaining the downsampled frequency domain data, frequency domain feature processing can be performed based on the downsampled frequency domain data to generate repaired frequency domain data. For example, a corrected image can be generated using a trained motion artifact correction model, which includes a trained frequency domain repair model.

[0090] For example, the downsampled frequency domain data is input into a trained frequency domain restoration model for feature extraction and inference prediction processing to generate restored frequency domain data. The restored frequency domain data is generated by further artifact correction of the frequency domain data to be corrected in the frequency domain to produce inference prediction results in the frequency domain, or in other words, the undersampled part of the data is supplemented or restored.

[0091] Then, the repaired frequency domain data is converted into image domain data, for example, by converting the repaired frequency domain data into the image domain through inverse Fourier transform, to obtain image domain data.

[0092] Next, image domain feature processing can be performed on the image domain data to generate a corrected image. For example, the trained motion artifact correction model also includes a trained frequency domain inpainting model. Image domain data is input into the trained image domain inpainting model for feature extraction and inference prediction processing to generate a corrected image. The corrected image is the inference prediction result generated in the image domain after further artifact correction of the image domain data.

[0093] Figure 7 A flowchart of the motion artifact correction model training method provided in the embodiments of this application.

[0094] like Figure 7 As shown, the motion artifact correction model training method 700 provided in this application includes steps S710-S760. The motion artifact correction model includes a frequency domain restoration model.

[0095] Step S710: Obtain the undersampling factor, wherein the undersampling factor includes multiple regional undersampling factors, and the regional undersampling factor is related to the number of phase coding lines of motion artifacts existing in the region.

[0096] Step S720: Obtain the frequency domain data sample to be calibrated and the reference frequency domain data sample corresponding to the frequency domain data sample to be calibrated.

[0097] Step S730: Based on the undersampling factor, the frequency domain data sample to be corrected is downsampled to obtain the downsampled frequency domain data sample.

[0098] Step S740: Input the downsampled frequency domain data sample into the frequency domain repair model to obtain the repaired frequency domain data sample.

[0099] Step S750: Determine the frequency domain loss value based on the repaired frequency domain data sample and the reference frequency domain data sample.

[0100] Step S760: Based at least on the frequency domain loss value, adjust the model parameters of the frequency domain inpainting model to obtain a trained motion artifact correction model.

[0101] The undersampling factor can be obtained by performing residual processing on the image. For example, residual processing can be performed on a set of motion artifact images and a corresponding set of reference images to obtain the undersampling factor. The set of motion artifact images may include at least one motion artifact image, and the set of reference images may include at least one corresponding reference image. The reference image is, for example, an image without motion artifacts. A motion artifact image and a corresponding reference image constitute an image pair. The undersampling factor is obtained by performing residual processing on each image pair. The undersampling factor represents the artifact information present in the motion artifact image.

[0102] Image data can be divided into multiple regions, each corresponding to a region undersampling factor. For any given region, there are multiple phase-coded lines. Some phase-coded lines contain motion artifacts, while others do not. The region undersampling factor is related to the number of phase-coded lines with motion artifacts in that region. For example, the region undersampling factor is positively correlated with the number of phase-coded lines with motion artifacts in the corresponding region; the larger the number of phase-coded lines with motion artifacts in the region, the larger the region undersampling factor.

[0103] The frequency domain data samples to be corrected can be obtained directly through scanning or acquisition, or they can be obtained based on the motion artifact sample images to be corrected. For example, the motion artifact sample images to be corrected can be acquired and then converted into frequency domain data samples to be corrected. The motion artifact sample images to be corrected can be one or more motion artifact images from a set of motion artifact images, or they can be other images. For example, based on the Fourier transform method, the motion artifact sample images to be corrected can be subjected to a Fourier transform to transform them into the frequency domain, thus obtaining the frequency domain data samples to be corrected.

[0104] After obtaining the undersampling factor, downsampling can be performed based on the undersampling factor. Specifically, artifact mask data can be obtained based on the undersampling factor, and then downsampling can be performed on the frequency domain data sample to be corrected based on the artifact mask data to obtain the downsampling frequency domain data sample.

[0105] The downsampled frequency domain data samples underwent preliminary motion artifact correction using an undersampling factor. After this initial correction, the downsampled frequency domain data samples contained less motion artifact information (the artifact severity was reduced). Next, based on the frequency domain inpainting model, feature extraction, prediction inference, and inverse Fourier transform to the image domain were performed on the downsampled frequency domain data samples. This process included further artifact correction, generating inpainted frequency domain data samples. A reference frequency domain data sample was used as a label to determine the frequency domain loss value between the inpainted and reference frequency domain data samples. Based on this loss value, the model parameters of the frequency domain inpainting model were adjusted in reverse. This process was repeated iteratively until the iteration conditions were met, at which point training ceased, resulting in a well-trained motion artifact correction model.

[0106] It is understood that the relevant content of the embodiments of this application can be referred to the embodiments above, and will not be repeated here.

[0107] In the embodiments of this application, an undersampling factor is obtained based on the number of phase coding lines of motion artifacts in each region. Then, a downsampling process is performed based on the undersampling factor to achieve preliminary motion artifact correction. After preliminary artifact correction using the undersampling factor, a frequency domain insulation model is further used to further insulate the downsampled frequency domain data to achieve artifact correction. This reduces the consistency requirements for image pairs, reduces the difficulty of data acquisition or data simulation, and improves the training and prediction effects of the dual-domain correction model.

[0108] In another example, the motion artifact correction model also includes an image domain inpainting model. The model training method further includes: acquiring a reference image sample corresponding to the frequency domain data sample to be corrected; converting the frequency domain inpainting data sample into an image domain inpainting data sample, and inputting the image domain inpainting data sample into the image domain inpainting model to obtain a corrected sample image; determining an image domain loss value based on the corrected image sample and the reference image sample; and adjusting the model parameters of the image domain inpainting model, at least based on the image domain loss value, to obtain a trained motion artifact correction model.

[0109] In another example, the frequency domain inpainting model and the image domain inpainting model are jointly trained. These two models constitute a dual-domain joint model, allowing for simultaneous adjustment of their parameters based on the frequency domain loss value, simultaneous adjustment based on the image domain loss value, and joint adjustment based on both the frequency domain and image domain loss values.

[0110] Figure 8 A schematic diagram of the motion artifact correction model training method provided in the embodiments of this application.

[0111] like Figure 8 As shown, residual processing is performed on the motion artifact image set and the corresponding reference image set to obtain an undersampling factor, and the artifact mask data Mask_Resi is obtained based on the undersampling factor. Then, the motion artifact image sample Motion_i to be corrected and the reference image sample (gold standard data) GT_i are obtained. The motion artifact image sample Motion_i to be corrected can be one or more images in the motion artifact image set, and the reference image sample GT_i can be one or more images in the reference image set.

[0112] The motion artifact image sample Motion_i to be corrected is converted into a frequency domain data sample Motion_k to be corrected through Fourier transform. The artifact mask data Mask_Resi and the frequency domain data sample Motion_k to be corrected are then multiplied and downsampled to perform preliminary motion artifact correction processing, resulting in downsampled frequency domain data samples.

[0113] Converting the reference image sample GT_i to the reference frequency domain data sample GT_k can be achieved by using Fourier transform to convert the image domain to the frequency domain.

[0114] Based on the frequency domain restoration model NN k Frequency domain feature processing is performed on the downsampled frequency domain data samples to obtain the repaired frequency domain data sample Temp_k. For example, the downsampled frequency domain data sample is input into the frequency domain repair model NN. k In the middle, through the frequency domain repair model NN k Feature extraction and inference prediction are performed on the downsampled frequency domain data samples to obtain the repaired frequency domain data sample Temp_k. Based on the repaired frequency domain data sample Temp_k and the reference frequency domain data sample GT_k, the frequency domain loss value loss1 is determined. Through training, the frequency domain loss value loss1 is gradually reduced to achieve wireless approximation between the two.

[0115] Then, the frequency domain data sample Temp_k is converted into the image domain data sample Temp_i by inverse Fourier transform, and the image domain data sample Temp_i is input into the image domain inpainting model NN. i In the middle, based on the image domain inpainting model NN i Image domain feature processing is performed on the image domain data sample Temp_i, such as feature extraction and inference prediction, to generate a corrected image sample Generate_i. The final corrected image sample Generate_i is the data with motion artifacts removed. Based on the corrected image sample Generate_i and the reference image sample GT_i, the image domain loss value loss2 is determined. Through training, the image domain loss value loss2 is gradually reduced to achieve infinite approximation between the two.

[0116] Finally, the model parameters of the motion artifact correction model are adjusted based on the frequency domain loss value (loss1) and the image domain loss value (loss2) to obtain a trained motion artifact correction model. The model parameters include at least the frequency domain inpainting model NN. k and image domain inpainting model NN i The model parameters in the model can also include parameters of other network modules, as the motion artifact correction model can also include parameters of other network modules.

[0117] After model training is complete, the trained model can be used for model prediction and inference. For example, the motion artifact image Motion_i to be corrected can be Fourier transformed to obtain its k-space representation (Motion_k), and then multiplied by the artifact mask data Mask_Resi before being input into the frequency domain inpainting model NN. k We obtain the restored frequency domain data Temp_k with reduced motion artifacts. We then perform an inverse Fourier transform on the restored frequency domain data Temp_k back to the image domain to obtain the image domain data Temp_i. Finally, we input the image domain data Temp_i into the image domain restoration model NN. i Image domain inpainting model NN i The output corrected image Generate_i is the final image with motion artifacts removed.

[0118] Figure 9 A schematic diagram of the motion artifact correction model provided in the embodiments of this application.

[0119] like Figure 9 As shown, the motion artifact correction model is a dual-domain network structure, which is a Unet network or a variant of Unet. This network can also be implemented by other generative networks.

[0120] Figure 10 This is a schematic diagram of motion artifact correction results provided in an embodiment of this application.

[0121] like Figure 10 As shown, the left image is the motion artifact data to be corrected, and the right image is the corrected image. It can be seen that the motion artifact correction effect of this application is better.

[0122] The motion artifact correction method of this application does not require strict adherence to data consistency in data collection or acquisition. The under-collection factor is obtained statistically from a large dataset rather than being manually designed, which improves reliability and accuracy. Furthermore, by combining the image domain and k-space (frequency domain), it can better preserve data fidelity.

[0123] Figure 11 A schematic diagram of a motion artifact correction device provided in an embodiment of this application.

[0124] This application provides a motion artifact correction device 1100, including: an acquisition module 1110, a downsampling module 1120, a repair module 1130, and a correction module 1140.

[0125] The acquisition module 1110 is used to acquire undersampling factors, wherein the undersampling factors include multiple regional undersampling factors, and the regional undersampling factors are related to the number of phase coding lines of motion artifacts present in the region.

[0126] The downsampling module 1120 is used to downsample the frequency domain data to be corrected based on the undersampling factor, so as to obtain downsampled frequency domain data.

[0127] The repair module 1130 is used to input the downsampled frequency domain data into the frequency domain repair model to obtain the repaired frequency domain data.

[0128] The correction module 1140 is used to obtain the corrected image based on frequency domain repair data.

[0129] In some embodiments, the acquisition module 1110 is further configured to: perform residual processing on the motion artifact dataset and the reference dataset to obtain a residual dataset; perform threshold processing on the residual dataset to determine a thresholded artifact dataset; and perform data region division and region undersampling factor calculation on the thresholded artifact dataset to obtain multiple region undersampling factors.

[0130] In some embodiments, the thresholded artifact dataset includes Z thresholded artifact data, Z≥1, each thresholded artifact data being a two-dimensional data with X rows and Y columns, where the row direction represents the frequency domain coding direction of image acquisition and the column direction represents the phase coding direction of image acquisition, X>1, Y>1; the thresholded artifact dataset is divided into data regions and regional undersampling factors are calculated to obtain multiple regional undersampling factors, including: dividing the z-th thresholded artifact data into data regions to determine a low-frequency data region and N high-frequency data regions, 1≤z≤Z, N≥1, where the number of rows in the low-frequency data region and each high-frequency data region is less than X and the number of columns is Y; in each high-frequency data region, determining whether the artifact degree represented by each row of data is greater than a preset artifact degree, if so, marking the row as a phase coding line of motion artifacts; determining the number of phase coding lines in each high-frequency data region; and determining N regional undersampling factors based on the number of N phase coding lines corresponding to each of the N high-frequency data regions.

[0131] In some embodiments, the regional undersampling factor is positively correlated with the number of phase coding lines of motion artifacts present in the corresponding high-frequency data region.

[0132] In some embodiments, determining N region undersampling factors based on the number of N phase coding lines corresponding to each of the N high-frequency data regions includes: determining the proportion of the number of each phase coding line in the corresponding high-frequency data region; obtaining N sets of row number proportion values ​​based on the row number proportion, each set of row number proportion values ​​including Z row number proportion values; averaging the Z row number proportion values ​​in each set to obtain N average proportion values; and determining the N region undersampling factors based on the N average proportion values.

[0133] In some embodiments, residual processing is performed on the motion artifact dataset and the reference dataset to obtain a residual dataset, including: calculating the image difference between each motion artifact image in the motion artifact dataset and the corresponding parameter image in the reference dataset; and transforming the image difference into K space to obtain the residual dataset.

[0134] In some embodiments, the downsampling module 1120 is further configured to: obtain artifact mask data based on the undersampling factor; and perform downsampling processing on the frequency domain data to be corrected based on the artifact mask data to obtain downsampling frequency domain data.

[0135] In some embodiments, obtaining artifact mask data based on the undersampling factor includes: determining the number of artifact rows corresponding to the nth high-frequency data region based on the undersampling factor of the nth region and the total number of uplink data in the nth high-frequency data region, thereby obtaining N artifact row numbers corresponding to the N high-frequency data regions, where 1≤n≤N; determining the first row of data randomly corresponding to the N artifact row numbers in the corresponding high-frequency data region as a first value, and determining the second row of data other than the first row of data in the corresponding high-frequency data region as a second value; and generating two-dimensional data of X rows and Y columns as artifact mask data based on the first row of data, the second row of data, and the low-frequency data region.

[0136] In some embodiments, the correction module 1140 is further configured to: convert frequency domain repair data into image domain data; input the image domain data into an image domain repair model to obtain a corrected image.

[0137] It is understood that a detailed description of the motion artifact correction device 1100 can be found in the description of the motion artifact correction method above, and will not be repeated here.

[0138] Figure 12 A schematic diagram of a motion artifact correction model training device provided in an embodiment of this application.

[0139] This application provides a motion artifact correction model training device 1200, including: a factor acquisition module 1210, a sample acquisition module 1220, a sample downsampling module 1230, a sample repair module 1240, a determination module 1250, and an adjustment module 1240.

[0140] The factor acquisition module 1210 is used to acquire undersampling factors, wherein the undersampling factors include multiple regional undersampling factors, and the regional undersampling factors are related to the number of phase coding lines of motion artifacts present in the region.

[0141] The sample acquisition module 1220 is used to acquire the frequency domain data sample to be corrected and the reference frequency domain data sample corresponding to the frequency domain data sample to be corrected.

[0142] The sample downsampling module 1230 is used to downsample the frequency domain data samples to be corrected based on the undersampling factor, so as to obtain downsampled frequency domain data samples.

[0143] The sample repair module 1240 is used to input the downsampled frequency domain data samples into the frequency domain repair model to obtain repaired frequency domain data samples.

[0144] The determination module 1250 is used to determine the frequency domain loss value based on the repaired frequency domain data sample and the reference frequency domain data sample.

[0145] The adjustment module 1260 is used to adjust the model parameters of the frequency domain inpainting model based at least on the frequency domain loss value, so as to obtain a trained motion artifact correction model.

[0146] In some embodiments, the motion artifact correction model further includes an image domain inpainting model; the motion artifact correction model training device further includes: an image sample acquisition module for acquiring a reference image sample corresponding to the frequency domain data sample to be corrected; an image domain sample inpainting module for converting the frequency domain inpainting data sample into an image domain inpainting data sample and inputting the image domain inpainting data sample into the image domain inpainting model to obtain a corrected sample image; an image domain loss determination module for determining an image domain loss value based on the corrected image sample and the reference image sample; and an image domain adjustment module for adjusting the model parameters of the image domain inpainting model based at least on the image domain loss value to obtain a trained motion artifact correction model.

[0147] In some embodiments, the frequency domain inpainting model and the image domain inpainting model are jointly trained.

[0148] It is understood that for a detailed description of the motion artifact correction model training device 1200, please refer to the description of the motion artifact correction model training method above, and will not be repeated here.

[0149] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method in any of the above embodiments.

[0150] One embodiment of this application provides a computer program product including instructions that, when executed by a processor of a computer device, enable the computer device to perform the steps of the method described in any of the above embodiments.

[0151] Figure 13 A block diagram of an electronic device provided in an embodiment of this application.

[0152] This application provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method in any of the above embodiments.

[0153] like Figure 13 As shown, for ease of understanding, an embodiment of this application illustrates a specific electronic device 1300.

[0154] Electronic device 1300 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic device 1300 may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0155] like Figure 13 As shown, device 1300 includes a computing unit 1301, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1302 or a computer program loaded from storage unit 1308 into random access memory (RAM) 1303. The RAM 1303 may also store various programs and data required for the operation of electronic device 1300. The computing unit 1301, ROM 1302, and RAM 1303 are interconnected via bus 1304. Input / output (I / O) interface 1305 is also connected to bus 1304.

[0156] Multiple components in electronic device 1300 are connected to I / O interface 1305. These components include: input unit 1306, such as a keyboard or mouse; output unit 1307, such as various types of displays or speakers; storage unit 1308, such as a hard disk or optical disk; and communication unit 1309, such as a network interface card (NIC), modem, or wireless transceiver. Communication unit 1309 allows electronic device 1300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0157] The computing unit 1301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1301 performs the various methods described above. For example, in some embodiments, any one or more of the methods described above can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1308. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 1300 via ROM 1302 and / or communication unit 1309. When the computer program is loaded into RAM 1303 and executed by the computing unit 1301, one or more steps of any one or more of the methods described above can be performed. Alternatively, in other embodiments, the computing unit 1301 can be configured to perform any one or more of the methods described above by any other suitable means (e.g., by means of firmware).

[0158] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this application, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0159] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0160] In the description of this application, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this application, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0161] In the description of this application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicating the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0162] Furthermore, the terms "first," "second," etc., used in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying relative importance, or implicitly specifying the number of technical features indicated in this embodiment. Therefore, features defined with terms such as "first" and "second" in the embodiments of this application can explicitly or implicitly indicate that the embodiment includes at least one of those features. In the description of this application, the word "multiple" means at least two or more, such as two, three, four, etc., unless otherwise explicitly and specifically defined in the embodiments.

[0163] In this application, unless otherwise explicitly specified or limited in the embodiments, the terms "installation," "connection," "joining," and "fixing" appearing in the embodiments should be interpreted broadly. For example, a connection can be a fixed connection, a detachable connection, or an integral part; it can also be a mechanical connection, an electrical connection, etc. Of course, it can also be a direct connection, or an indirect connection through an intermediate medium, or it can be the internal communication between two components, or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific implementation.

[0164] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

Claims

1. A method for correcting motion artifacts, characterized in that, The method includes: Obtain the undersampling factor, wherein the undersampling factor includes multiple regional undersampling factors, and the regional undersampling factor is related to the number of phase coding lines of motion artifacts present in the region; Based on the undersampling factor, the frequency domain data to be corrected is downsampled to obtain downsampled frequency domain data. Input the downsampled frequency domain data into the frequency domain repair model to obtain repaired frequency domain data; Based on the frequency domain repair data, the corrected image is obtained.

2. The method according to claim 1, characterized in that, The acquisition of the under-sampling factor includes: The motion artifact dataset and the reference dataset are subjected to residual processing to obtain the residual dataset; The residual dataset is thresholded to determine the thresholded artifact dataset; The thresholded artifact dataset is divided into data regions and the regional undersampling factor is calculated to obtain the multiple regional undersampling factors.

3. The method according to claim 2, characterized in that, The thresholded artifact dataset includes Z thresholded artifact data points, Z≥1. Each thresholded artifact data point is a two-dimensional data point with X rows and Y columns, where the row direction represents the frequency domain coding direction of image acquisition and the column direction represents the phase coding direction of image acquisition, X>1, Y>1. The process of dividing the thresholded artifact dataset into data regions and calculating regional undersampling factors to obtain the multiple regional undersampling factors includes: The data region is divided for the z-th thresholded artifact data to determine a low-frequency data region and N high-frequency data regions, where 1≤z≤Z and N≥1. The number of rows in the low-frequency data region and each of the high-frequency data regions is less than X and the number of columns is Y. In each of the high-frequency data regions, it is determined whether the degree of artifact represented by each row of data is greater than the preset artifact degree. If so, the phase encoding line of the row containing motion artifacts is marked. Determine the number of phase encoding lines in each of the high-frequency data regions; Based on the number of N phase encoding lines corresponding to each of the N high-frequency data regions, N undersampling factors for each region are determined.

4. The method according to claim 3, characterized in that, The undersampling factor in the region is positively correlated with the number of phase coding lines of motion artifacts present in the corresponding high-frequency data region.

5. The method according to claim 3 or 4, characterized in that, The determination of N undersampling factors for each of the N high-frequency data regions based on the number of N phase coding lines corresponding to each of the N high-frequency data regions includes: Determine the proportion of the number of each phase encoding line in the corresponding high-frequency data region; Based on the row quantity ratio, N sets of row quantity ratio values ​​are obtained, and each set of row quantity ratio values ​​includes Z row quantity ratio values. Calculate the average percentage of each of the Z rows in each group to obtain N average percentage values; Based on the average of the N percentages, N under-collection factors for the region are determined.

6. The method according to claim 2, characterized in that, The residual dataset is obtained by performing residual processing on the motion artifact dataset and the reference dataset, including: Calculate the image difference between each motion artifact image in the motion artifact dataset and the corresponding parameter image in the reference dataset; The image difference is transformed into K space to obtain the residual dataset.

7. The method according to claim 3, characterized in that, The step of downsampling the frequency domain data to be corrected based on the undersampling factor to obtain downsampling frequency domain data includes: Based on the undersampling factor, artifact mask data is obtained; Based on the artifact mask data, the frequency domain data to be corrected is downsampled to obtain downsampled frequency domain data.

8. The method according to claim 7, characterized in that, The process of obtaining artifact mask data based on the undersampling factor includes: Based on the under-collection factor of the nth region and the total number of uplink data in the nth high-frequency data region, the number of artifact lines corresponding to the nth high-frequency data region is determined, and the number of N artifact lines corresponding to the N high-frequency data regions is obtained, 1≤n≤N; In the corresponding high-frequency data region, the first row of data corresponding to the random number of the N artifact rows is determined as the first value, and the second row of data other than the first row of data is determined as the second value in the corresponding high-frequency data region. Based on the first row of data, the second row of data, and the low-frequency data region, two-dimensional data of X rows and Y columns is generated as the artifact mask data.

9. The method according to claim 1, characterized in that, The process of obtaining the corrected image based on the frequency domain restoration data includes: The frequency domain repair data is converted into image domain data; The image domain data is input into the image domain restoration model to obtain the corrected image.

10. A method for training a motion artifact correction model, characterized in that, The motion artifact correction model includes a frequency domain restoration model; the method includes: Obtain the undersampling factor, wherein the undersampling factor includes multiple regional undersampling factors, and the regional undersampling factor is related to the number of phase coding lines of motion artifacts present in the region; Obtain the frequency domain data sample to be corrected and the reference frequency domain data sample corresponding to the frequency domain data sample to be corrected; Based on the undersampling factor, the frequency domain data sample to be corrected is downsampled to obtain a downsampled frequency domain data sample. Input the downsampled frequency domain data sample into the frequency domain repair model to obtain the repaired frequency domain data sample; Based on the repaired frequency domain data sample and the reference frequency domain data sample, the frequency domain loss value is determined; Based at least on the frequency domain loss value, the model parameters of the frequency domain repair model are adjusted to obtain a trained motion artifact correction model.

11. The method according to claim 10, characterized in that, The motion artifact correction model further includes an image domain inpainting model; the method further includes: Obtain a reference image sample corresponding to the frequency domain data sample to be corrected; The frequency domain repair data samples are converted into image domain repair data samples, and the image domain repair data samples are input into the image domain repair model to obtain the corrected sample image; Based on the corrected image samples and the reference image samples, determine the image domain loss value; Based at least on the image domain loss value, the model parameters of the image domain inpainting model are adjusted to obtain a trained motion artifact correction model.

12. The method according to claim 11, characterized in that, The frequency domain restoration model and the image domain restoration model are trained together.

13. A motion artifact correction device, characterized in that, The device includes: An acquisition module is used to acquire undersampling factors, wherein the undersampling factors include multiple regional undersampling factors, and the regional undersampling factors are related to the number of phase coding lines of motion artifacts present in the region; The downsampling module is used to downsample the frequency domain data to be corrected based on the undersampling factor, so as to obtain downsampled frequency domain data. The repair module is used to input the downsampled frequency domain data into the frequency domain repair model to obtain repaired frequency domain data; The correction module is used to obtain the corrected image based on the frequency domain repair data.

14. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-12.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-12.

Images