Method and apparatus for motion artifact correction using artificial neural networks

Abstract

Neural network-based systems, methods, and apparatus can be used to remove motion artifacts from magnetic resonance (MR) images. Such neural network-based systems can be trained to perform motion artifact removal tasks without a reference (e.g., without using pairs of motion-contaminated and motion-free MR images). Various training techniques are described herein, including techniques that present pairs of MR images with varying levels of motion contamination to a neural network and force the neural network to learn to correct for the motion contamination by transforming the first image of the contaminated pair into the second image of the contaminated pair. Other neural network training techniques are also described that aim to reduce reliance on difficult-to-obtain training data.

Description

Technical Field

[0001] The present application relates to magnetic resonance imaging (MRI) procedures, and more particularly to procedures for dynamically scanning an object. Background Art

[0002] Motion artifacts (such as those caused by the overall motion of the patient) are commonly seen during magnetic resonance imaging (MRI) procedures, especially those involving dynamically scanning subjects. Using cardiac MRI as an example, the captured images are often contaminated by artifacts caused by respiratory motion, blood flow, and other types of motion of the patient. It is reported that approximately 20% of repeated MRI scans are attributable to motion artifacts, which places a heavy burden on hospitals and other medical institutions. In recent years, deep learning-based techniques have brought great progress to MRI image analysis and post-processing, but the removal of motion artifacts remains a challenging task. The main obstacle is the lack of training data. After the scan, motion-contaminated images are usually discarded, and it is even more difficult to collect data with controlled motion (such as image pairs consisting of clean and motion-contaminated images that can be used for supervised learning).

[0003] On the other hand, many MRI applications involve acquiring multiple MR images within a single scan. For example, in cardiac cine MRI, a time series of images is acquired to record the contractile motion of the heart; in quantitative T1 imaging, multiple images are captured to record the T1 relaxation of different tissues. These correlated images often contain valuable information indicating the source and extent of motion artifacts and can provide key insights into how to remove these artifacts. However, deep learning-based image processing systems that utilize multiple correlated images for motion artifact correction are still lacking. Summary of the Invention

[0004] Described herein are systems, methods, and apparatus associated with removing (e.g., correcting or reducing) motion artifacts from magnetic resonance (MR) images using an artificial neural network (ANN). The ANN can learn parameters associated with removing motion artifacts (e.g., an artifact removal model) through a training process. Learning can be performed using a training network and a training dataset comprising a plurality of paired MR images containing different motion artifacts, without reference to corresponding motion-free images. Each MR image pair of the training dataset can include a first MR image containing a first motion artifact and a second MR image containing a second motion artifact. The first motion artifact and the second motion artifact can be randomly generated, for example, based on computer simulation or patient motion.

[0005] During learning, the training network can generate an output image for each MR image pair based on the first MR image in the MR image pair to be similar to the second MR image in the MR image pair. The training network can determine the difference between the output image and the second MR image in the MR image pair and adjust one or more parameters of the training network with the goal of minimizing the difference. By performing the above operations on a large number of images included in the training dataset, the training network can exploit the randomness of motion artifacts included in each motion-contaminated MR image pair to learn parameters (e.g., a machine learning model) to remove (e.g., correct or reduce) motion artifacts from the motion-contaminated MR images. These parameters of the trained network can be stored upon completion of the learning process and used to implement an ANN to remove motion artifacts from actual MR images.

[0006] In an example, motion artifacts included in the training datasets described herein can be simulated based on k-space data (e.g., by manipulating (e.g., randomly) the order in which the k-space data is acquired). In an example, motion artifacts included in the training datasets described herein can be generated based on patient motion that occurs during an actual MRI procedure (e.g., deep breathing). In an example, the training dataset can include MR images of different scanned subjects (e.g., belonging to different patients), but the individual MR image pairs provided to the training network can be associated with the same subject (e.g., with different degrees or realizations of motion artifacts).

[0007] Other network training techniques associated with no-reference (e.g., unsupervised) learning are also described herein, including, for example, providing multiple MR images (e.g., included in a movie) as input to the network (e.g., different input channels), and forcing the network to learn associations between the multiple images, for example, by applying linear or fully connected operations to time frames associated with the multiple images. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] A more detailed understanding of the examples disclosed herein can be obtained from the following description, which is given by way of example in conjunction with the accompanying drawings.

[0009] Figure 1 is a block diagram illustrating an example neural network system for removing (eg, correcting) motion artifacts in magnetic resonance (MR) images.

[0010] Figure 2 is a block diagram illustrating an example of using motion-contaminated MR image pairs to train a neural network to remove motion artifacts.

[0011] Figure 3A 、 Figure 3B and Figure 3C Example results that may be produced by a motion artifact removal neural network as described herein are illustrated.

[0012] Figure 4 is a block diagram illustrating an example technique for training a motion artifact removal network based on domain transfer.

[0013] Figure 5 is a flow chart illustrating an example process for training a neural network system to remove motion artifacts.

[0014] Figure 6 is a block diagram illustrating example components of a motion artifact removal neural network system as described herein. DETAILED DESCRIPTION

[0015] The present disclosure is illustrated by way of example and not limitation in the figures of the accompanying drawings.

[0016] Figure 1 The present invention is a block diagram illustrating an example neural network system 100 (e.g., a motion artifact removal device) for removing (e.g., correcting or reducing) motion artifacts in magnetic resonance (MR) images. As shown, system 100 may take a motion-contaminated MR image 102 as input, process the image using a neural network (e.g., one or more neural networks) 104 trained for motion artifact removal, and generate an MR image 104 (e.g., having similar dimensions to input image 102) that is substantially free of (e.g., having significantly reduced) motion artifacts compared to MR image 102. MR image 102 may include a single image or a cine video (e.g., including multiple MR images) obtained from various sources, such as a magnetic resonance imaging (MRI) device, a database storing patient imaging records, and the like. Motion artifacts included in MR image 102 may have various causes, including patient motion, device errors, and the like. For example, motion artifacts included in MR image 102 may include artifacts caused by overall patient motion (such as respiratory motion), irregular heartbeats, signal acquisition errors (e.g., hardware glitches), and the like.

[0017] The neural network 104 may include one or more convolutional neural networks (CNNs) trained to correct (e.g., remove) motion artifacts from the MR images 102. The neural network 104 may include one or more convolutional layers, each of which includes a plurality of kernels or filters configured to identify artifact components from the MR images 102 via one or more convolution operations. Each kernel or filter may be associated with a corresponding set of weights that may be optimized through a training process for extracting specific features from the MR images 102. The convolution operations may be followed by batch normalization and / or linear or nonlinear activation (e.g., using one or more rectified linear units (ReLUs)) to separate image features (e.g., non-artifact features) from noise features (e.g., artifacts). For example, the neural network 104 may include a first convolutional layer having a plurality of (e.g., 64) filters (with a kernel size of 3×3) and a ReLU as an activation function. The neural network 104 may include multiple intermediate layers (e.g., additional convolutional layers after the first convolutional layer), and each of these intermediate layers may include a similar number of filters (e.g., 64 filters with a kernel size of 3×3), followed by batch normalization and ReLU. The neural network 104 may include additional convolutional layers (e.g., the last layer before the output) that utilize filters of similar kernel size (e.g., 3×3) to construct an output image that represents a clean version of the MR image 102 (e.g., substantially free of motion artifacts) or artifacts contained in the MR image 102. In the latter case, an artifact-free or artifact-reduced image (e.g., MR image 106) may be generated by subtracting the predicted output image from the MR image 102.

[0018] In an example, the neural network 104 may include an encoder configured to extract image features at multiple spatial resolution levels, e.g., via multiple convolution operations, and / or a decoder configured to decode the image features to generate the output image 106. In an example, the neural network 104 may include a recurrent neural network (RNN), e.g., when receiving multiple images at an input.

[0019] The motion artifact removal neural network described in this paper (e.g., Figure 1The neural network 104 in the present invention can be trained using techniques that do not require training images with controlled motion (e.g., clean and contaminated image pairs, the acquisition of which may require specially designed sequences, extended scan times, patient cooperation, technician experience, etc.). In contrast, the motion artifact removal neural network can be trained using motion-contaminated image pairs (e.g., different image pairs can be associated with different scan objects, but each image pair used during an iteration of training can include the same underlying object, but with motion artifacts of different types or severities). These images can be obtained from actual MRI procedures where motion contamination occurs (e.g., due to patient movement), or can be artificially created by computer simulation or other means. For example, motion artifacts can be randomly introduced during an MRI acquisition sequence, such as by manipulating (e.g., changing) the order in which k-space data is acquired (e.g., the order in which k-space is covered), and / or by simulating multiple acquisitions to generate randomly perturbed artifacts. As another example, motion artifacts can be introduced during an MRI scanning procedure by instructing the patient to perform certain movements (e.g., heavy breathing movements) at different stages of the procedure. As yet another example, motion artifacts can be simulated based on an underlying image that may or may not contain artifacts and by adding artificial noise to the underlying image (e.g., based on a parameterized model that simulates respiratory motion). In any of these techniques, artifact simulation can be performed using parameters that allow for random variation in the severity or type of artifact to be introduced, or the manner in which the artifact is implemented. These parameters can include, for example, the number of indices of phase-encoding lines that may be affected by motion, the direction and / or magnitude of the motion, etc. Thus, the motion artifacts thus generated may have a structured appearance, but have a random cause or implementation.

[0020] During training of a neural network, the motion-contaminated MR training images described herein may be arranged in pairs, e.g., each pair having the same scan object but different motion artifacts (e.g., randomly generated artifacts), and the MR training images provided to the neural network for processing. Figure 2An example operation of neural network 202 during such a training process is illustrated. As shown, during one or more training iterations, neural network 202 may receive an input MR image 204a, which may be part of a motion-contaminated image pair. The motion-contaminated image pair may be selected such that input MR image 204a and second MR image 204b in the pair are associated with the same scan object (e.g., the same anatomical structure of a patient), but contain different amounts, types, or severities of motion artifacts. Neural network 202 may process input MR image 204a (e.g., via one or more convolution, batch normalization, and / or activation operations) to generate an output image 206 that resembles the other image 204b of the motion-contaminated image pair. Neural network 202 may then determine the difference between output image 206 and image 204b, for example, based on a loss function 208 (e.g., various loss functions may be used for this purpose, including, for example, L1, L2, mean squared error (MSE), structural similarity index measure (SSIM), etc.). Based on the determined difference, neural network 202 may adjust one or more of its operating parameters (e.g., weights associated with the neural network) via a backpropagation process (e.g., using gradient descent associated with a loss function) with the goal of minimizing the difference between output image 206 and image 204 b.

[0021] Using the methods described herein, the variance of artifacts / noise in the estimates (e.g., output images 206) produced by neural network 202 can be inversely proportional to the number of images used to train neural network 202. Thus, for a sufficiently large number of training images, the error in motion removal or correction can approach zero. In other words, based on a large training dataset with randomly generated motion artifacts, neural network 202 can learn and maintain consistent information included in the training images (e.g., image information without random artifacts). For example, let A denote the underlying artifact-free object, let R(A,ε) i ) represents the real image of the object (where ε i can represent the random motion artifacts introduced into the image by the random function R), then the neural network described in this paper can be used to derive B = G{R(A,ε i ),θ}, where G can represent the prediction or mapping performed by the network (e.g., a denoising operation), can be a parameter of the network, and B can be the output of the network. The neural network can then be trained using a loss function to minimize the difference between A and B (e.g., which may contain one or more other artifacts). Such a loss function can be based on L2, L1, structural similarity index measure (SSIM), and / or one or more other principles. Thus, when using such a loss function and a large number of image pairs (R(A, ε i ),R(A,ε j)) During training, the mapping G may be able to retain only the i ),θ} information.

[0022] The motion artifact removal neural networks described herein (e.g., Figure 1 The neural network 104 in FIG. 104 is a block diagram of a method for learning a motion artifact correction algorithm. The images may be acquired within a single scan, and therefore there may be associations (e.g., correlations) between the images (e.g., along a time axis) that may be used to guide the removal or correction of motion artifacts. Techniques that treat the time dimension as an additional spatial dimension (e.g., using convolutional kernels) may not be suitable for these images because the motion artifacts may span multiple image frames and the receptive field of the convolutional neural network may be limited. The benefits of recurrent neural networks may also be limited in these cases, for example, because associations between image frames that are further apart in time may be difficult to learn. To improve the accuracy and effectiveness of motion artifact correction, the motion artifact removal neural network described herein may be trained to receive a video (e.g., a movie) at input and learn associations between different images of the video by performing one or more linear or fully connected operations along a timeline of the images. In an example, multiple images may be stacked, and the number of images may determine the number of channels included in the input matrix of the neural network. Through training (e.g., the training can be performed unsupervised (without reference) or supervised (with reference)), the motion artifact removal neural network can learn time-specific interactions for each individual image frame (e.g., the weights relative to other images may be different for each time point), and the performance of the motion artifact removal operation can therefore be improved, for example, at least for those motion artifacts associated with temporally more distant information, by exploiting spatial-temporal correlations in the time series of images.

[0023] Figure 3A 、 Figure 3B and Figure 3C Shown are example results that may be produced by a motion artifact removal neural network trained using the techniques described herein. Figure 3A The left image in the figure is the motion-contaminated image to be corrected, and the right image in the figure shows the motion correction result produced by the neural network trained using the motion-contaminated image pair, as described in this paper. Figure 3B The leftmost image in the figure is the motion-contaminated image to be corrected, the middle image in the figure shows the motion removal result produced by a neural network trained using a single image or frame as input, and the rightmost image in the figure shows the motion correction produced by a neural network trained using multiple interrelated images or frames as input, as described in the paper. Similarly, Figure 3CThe leftmost image in the figure is the motion-contaminated image to be corrected, the middle image in the figure shows the motion correction result produced by a neural network trained using conventional techniques (e.g., CNN), and the rightmost image in the figure shows the motion correction produced by a neural network trained using multiple interrelated images or frames at its input, as described in this article.

[0024] Figure 4 exemplifies a method for training a motion artifact removal (eg, correction) network 402 (eg, Figure 1 Another example technique for training a neural network 104 in a MRI machine learning environment is to use a plurality of image sets 404 for training. The image set 404 used for training may include conventional MR images (e.g., without special manipulation or arrangement) collected from a clinical setting. These clinical images may contain a variety of motion artifact types and / or severities, but may also include motion-free images. In preparation for training, the image set 404 may be divided into multiple groups, including, for example, a motion-free group and a motion-contaminated group. The motion-contaminated group may be further divided into subgroups, wherein each subgroup may include images with motion contamination of a certain severity. The grouping of images may be performed manually (e.g., by a medical expert) or automatically (e.g., by a computing program), or using a combination of manual and automatic techniques. Images in different groups may include different scan objects (e.g., different tissues) belonging to the same patient or different patients.

[0025] Training may be performed in an unsupervised manner, during which the motion artifact removal network 402 may learn a model for removing or reducing motion artifacts from MR images by transferring (e.g., translating) training images 404 from one domain (e.g., a motion-contaminated group) to another domain (e.g., a motion-free group). Figure 4 As shown, D a can represent a no-motion domain (e.g., a no-motion group), D b The image pair may be provided as input to a motion artifact removal network 402, which may be referred to herein as a first neural network. a ) can be obtained from D a In is selected (e.g., randomly), and another image (x b ) can be obtained from D b The motion artifact removal network 402 may be configured to decompose the input image into corresponding motion-free images (e.g., f a and f b ) and motion artifact images (e.g., m a and m b , which may be a feature map containing motion artifacts detected in the input image).

[0026] The second neural network 406 can be used to facilitate the training of the motion artifact removal neural network. The second neural network 406 can be pre-trained to synthesize a motion-contaminated image based on a motion-free image and an image (or feature map) containing motion artifacts. As can be seen, the second neural network 406 can also be referred to as a motion artifact synthesis network in this article. Figure 4 As shown, once the input image x a and x b is decomposed into corresponding motion-free images f a and f b and the motion artifact image m a and m b , then with the image x a and x b The associated f and m can be switched, and the motion artifact synthesis network 406 can be used to generate the corresponding motion-contaminated image x b-a and x a-b , these motion contaminated images include those from f a and f b The non-moving object information and the information from m b and m a sports information.

[0027] Synthesized image x b-a and x a-b The motion artifact removal network 402 can be used to further decompose the image, and the resulting motion-free image f′ a and f' b and the motion artifact image m' a and m' b It can be switched again, and then the motion artifact synthesis network 406 is used to generate the image including the image from f' a and f' b The non-moving object information and the information from m' b and m' a The motion information of the additional motion contaminated image x a -ba and x b -ab. Parameters of the motion artifact removal network 402 (e.g., weights associated with individual filters of the neural network) may then be adjusted based on one or more differences between the processed results. For example, the parameters of the motion artifact removal network 402 may be adjusted based on f a and f' a (and / or f b and f' b ) between the first consistency loss (L f cc ), m a and m' a (and / or m b and m'b ) between the second consistency loss (L m cc ) and / or input image x a With the output image x a-b-a (and / or input image x b With the output image x b-a-b ) between the third consistency loss (L cx cc ) to adjust. In an example, the parameters of the motion artifact removal network 402 can also be based on the adversarial loss (L ad ) (e.g., in the form of a binary cross entropy loss). For example, a discriminator can be used to distinguish between "real" input images and "fake" input images generated by an image generator under training (e.g., such as motion artifact synthesis network 406), and the image generator can be trained to output images that mimic the "real" input images (e.g., to fool the discriminator). Through this adversarial process, the generator can learn to generate images that are as realistic as the "real" input images (e.g., the images generated by the generator can have the same distribution as the "real" input images).

[0028] The consistency and / or adversarial losses described herein can be determined based on corresponding loss functions. These loss functions can be based on, for example, mean squared error (MSE), minimum absolute deviation (L1 loss), least squares error (L2 loss), cross entropy, SSIM, perceptual loss, etc. For example, the adversarial loss described herein can include a binary cross entropy loss, and the consistency loss described herein can include an SSIM loss or a VGG-based perceptual loss. The same loss function can be applied to calculate all losses described herein, or different loss functions can be applied to calculate different losses. Adjustment of the parameters of the motion artifact removal network 402 can be implemented by a backpropagation process (e.g., gradient descent based on the adopted loss function (e.g., learning gradient descent)). More details about the training of the motion artifact removal network 402 will be provided below. Once trained, the parameters of the motion artifact removal network 402 can be stored as coefficients of a motion artifact removal model, which can then be used to take a motion-contaminated MR image as input, separate object information and artifact information in the image, and produce an output image in which motion artifacts are substantially removed or reduced.

[0029] Figure 5is a flow chart illustrating an example process 500 that may be used to train a motion artifact removal network described herein. Process 500 may begin at 502, and at 504, initial parameters of a neural network (e.g., weights associated with various filters or kernels of the neural network) may be initialized. These parameters may be initialized, for example, based on samples collected from one or more probability distributions or parameter values ​​of another neural network having a similar architecture. At 506, the neural network may receive a motion-contaminated MR image, and the image may be processed by the neural network to remove motion artifacts from the image (e.g., predict a corresponding clean MR image without motion artifacts). At 508, the results of the processing may be compared to a reference image to determine adjustments that need to be made to the currently assigned neural network parameters. In an example, the reference image may be an image in a contaminated image pair (e.g., such as Figure 2 exemplified) or images generated using the domain transfer techniques described herein (e.g., Figure 4 In an example, the reference image may be another motion-contaminated image containing the same scanned object but with motion contamination of different severity. The adjustment of the network parameters may be determined based on a loss function (e.g., based on MSE, L1 loss, L2 loss, etc.) and a gradient descent associated with the loss function (e.g., stochastic gradient descent).

[0030] At 510, the neural network may apply adjustments to the currently assigned network parameters, for example, via a backpropagation process. At 512, the neural network may determine whether one or more training termination criteria have been met. For example, if the neural network has completed a predetermined number of training iterations, if the difference between the processed result and the reference result is below a predetermined threshold, or if the change in the value of the loss function between two training iterations is below a predetermined threshold, the neural network may determine that the training termination criteria have been met. If it is determined at 512 that the training termination criteria have not been met, the neural network may return to 506. If it is determined at 512 that the training termination criteria have been met, the neural network may end the training process 500 at 514.

[0031] For simplicity of illustration, the training steps are depicted and described herein in a particular order. However, it should be understood that the training operations can occur in various orders, simultaneously, and / or with other operations not presented or described herein. Furthermore, it should be noted that not all operations that may be included in the training process are depicted and described herein, and not all illustrated operations need to be performed.

[0032] The systems, methods, and / or devices described herein may be implemented using one or more processors, one or more storage devices, and / or other suitable auxiliary devices (such as display devices, communication devices, input / output devices, etc.). Figure 66 is a block diagram illustrating an example system (e.g., device) 600 that can be configured to perform one or more functions described herein. As shown, system 600 may include a processor (e.g., one or more processors) 602, which may be a central processing unit (CPU), a graphics processing unit (GPU), a microcontroller, a reduced instruction set computer (RISC) processor, an application-specific integrated circuit (ASIC), an application-specific instruction set processor (ASIP), a physical processing unit (PPU), a digital signal processor (DSP), a field programmable gate array (FPGA), or any other circuit or processor capable of performing the functions described herein. System 600 may also include communication circuitry 604, a memory 606, a mass storage device 608, an input device 610, and / or a communication link 612 (e.g., a communication bus) through which one or more components shown in the figure can exchange information.

[0033] The communication circuit 604 can be configured to send and receive information using one or more communication protocols (e.g., TCP / IP) and one or more communication networks, including a local area network (LAN), a wide area network (WAN), the Internet, a wireless data network (e.g., Wi-Fi, 3G, 4G / LTE, or 5G network). The memory 606 may include a storage medium (e.g., a non-transitory storage medium) configured to store machine-readable instructions, which, when executed, causes the processor 602 to perform one or more functions described herein. Examples of machine-readable media may include volatile or non-volatile memory, including but not limited to semiconductor memory (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM)), flash memory, etc. The mass storage device 608 may include one or more disks, such as one or more internal hard disks, one or more removable disks, one or more magneto-optical disks, one or more CD-ROM or DVD-ROM disks, etc., on which instructions and / or data may be stored to facilitate the operation of the processor 602. The input device 610 may include a keyboard, a mouse, a voice control input device, a touch-sensitive input device (eg, a touch screen), etc., for receiving user input of the system 600 .

[0034] It should be noted that the system 600 can operate as a standalone device or can be connected (e.g., networked or clustered) with other computing devices to perform the functions described herein. Figure 6 Only one example of each component is shown in the figure. Those skilled in the art will also understand that system 600 may include multiple examples of one or more components shown in the figure.

[0035] The motion artifact removal techniques described herein can be used in standalone systems (e.g., as part of an image post-processing process) to remove motion artifacts from MR images and improve image quality. The techniques can also be deployed as part of an imaging pipeline on a scanner. For example, the techniques can be used to reconstruct (e.g., correct) an MR image before the reconstructed image is presented to an end user, so that the end user has direct access to the motion artifact removed image (e.g., in real time). The techniques can also be deployed outside the imaging pipeline on a scanner and serve as a tool for the end user (e.g., the end user can choose to apply the techniques to examine images with and without motion removal). The techniques can also be deployed in an integrated pipeline, for example, as a regularization step in an iterative reconstruction process.

[0036] Although the present disclosure has been described in terms of certain embodiments and generally associated methods, variations and transformations of the embodiments and methods will be apparent to those skilled in the art. Therefore, the above description of exemplary embodiments does not limit the present disclosure. Other changes, substitutions, and variations are also possible without departing from the spirit and scope of the present disclosure. In addition, unless otherwise specifically stated, discussions utilizing terms such as "analyze," "determine," "enable," "identify," "modify," etc. refer to the actions and processes of a computer system or similar electronic computing device, which manipulate and transform data represented as physical (e.g., electronic) quantities within the registers and memories of a computer system into other data represented as physical quantities within the computer system memory or other such information storage, transmission, or display device.

[0037] It should be understood that the above description is intended to be illustrative, rather than restrictive. After reading and understanding the above description, many other embodiments will be apparent to those skilled in the art. Therefore, the scope of the present disclosure should be determined with reference to the full scope of equivalents to which the appended claims and such claims are given.

Claims

1. A computer-implemented method for removing motion artifacts from a magnetic resonance (MR) image, the method comprising: receiving a source MR image, wherein the source MR image is associated with an anatomical structure and includes one or more motion artifacts; and The source MR image is processed by an artificial neural network (ANN) to generate a target MR image without the one or more motion artifacts, wherein the ANN comprises parameters configured to remove the one or more motion artifacts from the source MR image, and wherein the parameters are learned by a training process comprising: obtaining a plurality of MR image pairs associated with the anatomical structure, wherein each MR image pair comprises a first MR image and a second MR image, the first MR image comprising a first motion artifact, and the second MR image comprising a second motion artifact; for each MR image pair, causing a training network to generate an output image based on the first MR image in the MR image pair to be similar to the second MR image in the MR image pair, and further causing the training network to adjust one or more parameters of the training network to minimize a difference between the output image and the second MR image in the MR image pair; wherein, a motion artifact removal network is used to decompose an input image into corresponding motion-free images and motion artifact images, a motion artifact synthesis network is used to synthesize the motion-free images and motion artifact images to obtain a motion-contaminated image, the motion artifact removal network is used to further decompose the synthesized image to obtain a decomposed motion-free image and a motion artifact image, the motion artifact synthesis network is used to generate an additional motion-contaminated image containing motion-free object information from the decomposed motion-free image and motion information from the decomposed motion artifact image, so as to calculate a processing result, and parameters of the motion artifact removal network are adjusted based on one or more differences between the processing results; And, storing the one or more parameters of the training network as the parameters of the ANN.

2. The computer-implemented method of claim 1 , wherein: The first motion artifact and the second motion artifact respectively included in the first MR image and the second MR image of each of the plurality of MR image pairs are randomly correlated.

3. The computer-implemented method of claim 2, wherein: At least one of the first motion artifact or the second motion artifact is caused by patient motion.

4. The computer-implemented method of claim 2, wherein: At least one of the first motion artifact or the second motion artifact is computer simulated.

5. The computer-implemented method of claim 4, wherein: The at least one of the first motion artifact or the second motion artifact is simulated based on k-space data.

6. The computer-implemented method of claim 5, wherein: The at least one of the first motion artifact or the second motion artifact is simulated by manipulating an order in which the k-space data is acquired.

7. The computer-implemented method of claim 1 , wherein: The first MR image and the second MR image of a first pair of the plurality of MR image pairs are associated with an anatomical structure of a first patient, and wherein the first MR image and the second MR image of a second pair of the plurality of MR image pairs are associated with an anatomical structure of a second patient.

8. The computer-implemented method of claim 1 , wherein: The parameters of the ANN are learned without reference to a motion-free MR image corresponding to the first MR image or the second MR image of each of the plurality of MR image pairs, wherein the ANN comprises a convolutional neural network.

9. An apparatus configured to remove motion artifacts from a magnetic resonance (MR) image, comprising: One or more processors configured to: receiving a source MR image, wherein the source MR image is associated with an anatomical structure and includes one or more motion artifacts; and The source MR image is processed by an artificial neural network (ANN) to generate a target MR image without the one or more motion artifacts, wherein the ANN comprises parameters configured to remove the one or more motion artifacts from the source MR image, and wherein the parameters are learned by a training process comprising: obtaining a plurality of MR image pairs associated with the anatomical structure, wherein each MR image pair comprises a first MR image and a second MR image, the first MR image comprising a first motion artifact, and the second MR image comprising a second motion artifact; for each MR image pair, causing a training network to generate an output image based on the first MR image in the MR image pair to be similar to the second MR image in the MR image pair, and further causing the training network to adjust one or more parameters of the training network to minimize a difference between the output image and the second MR image in the MR image pair; wherein, a motion artifact removal network is used to decompose an input image into corresponding motion-free images and motion artifact images, a motion artifact synthesis network is used to synthesize the motion-free images and motion artifact images to obtain a motion-contaminated image, the motion artifact removal network is used to further decompose the synthesized image to obtain a decomposed motion-free image and a motion artifact image, the motion artifact synthesis network is used to generate an additional motion-contaminated image containing motion-free object information from the decomposed motion-free image and motion information from the decomposed motion artifact image, so as to calculate a processing result, and parameters of the motion artifact removal network are adjusted based on one or more differences between the processing results; And, storing the one or more parameters of the training network as the parameters of the ANN.

10. A method for training an artificial neural network (ANN) to remove motion artifacts from magnetic resonance (MR) images, the method comprising: obtaining a plurality of MR image pairs associated with the anatomical structure, wherein each MR image pair comprises a first MR image and a second MR image, the first MR image comprising a first randomly generated motion artifact, and the second MR image comprising a second random motion artifact; and For each MR image pair in the plurality of MR image pairs: causing the ANN to generate an output image based on the first MR image in the MR image pair to be similar to the second MR image in the MR image pair; determining a difference between the output image and the second MR image; causing the ANN to adjust one or more parameters of the ANN to reduce the difference between the output image and the second MR image; wherein, a motion artifact removal network is used to decompose an input image into corresponding motion-free images and motion artifact images, a motion artifact synthesis network is used to synthesize the motion-free images and motion artifact images to obtain a motion-contaminated image, the motion artifact removal network is used to further decompose the synthesized image to obtain a decomposed motion-free image and a motion artifact image, the motion artifact synthesis network is used to generate an additional motion-contaminated image containing motion-free object information from the decomposed motion-free image and motion information from the decomposed motion artifact image, so as to calculate a processing result, and parameters of the motion artifact removal network are adjusted based on one or more differences between the processing results; And, in response to determining that one or more training criteria have been met, storing the one or more parameters of the ANN.

Images