Motion artifact simulator and corresponding method

Abstract

MR images for training artificial neural networks to remove motion artifacts from motion-contaminated magnetic resonance (MR) images are difficult to obtain. Systems, methods, and apparatus for injecting motion artifacts into clean MR images and using the artificially contaminated images for machine learning and neural network training are described herein. Motion-contaminated MR images can be created based on clean source MR images associated with multiple physiological cycles of a scanned object and by deriving MR data segments for multiple physiological cycles based on the source MR images. The MR data segments thus derived can be combined to obtain a simulated MR data set from which one or more target MR images can be generated to exhibit motion artifacts. Motion artifacts can be created by manipulating the source MR images and / or controlling the manner in which the MR data sets or target MR images are generated.

Description

Technical Field

[0001] The present application relates to magnetic resonance imaging (MRI) procedures, and in particular to procedures involving dynamic scanning of an object (eg, a heart). Background Art

[0002] Motion artifacts (such as those caused by the patient's respiratory motion) are often encountered during magnetic resonance imaging (MRI) procedures, especially those involving dynamic scanning of objects (such as the heart). 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. On the other hand, even though deep learning-based techniques have brought great progress to MRI image analysis and post-processing, the removal of motion artifacts remains a challenging task. The main obstacle is the lack of training data. After scanning, 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] Therefore, systems, methods and apparatus are desired to generate magnetic resonance (MR) images with simulated motion artifacts. These motion-contaminated MR images can then be used to solve a variety of problems, including, for example, training artificial neural networks to automatically remove motion artifacts from MR images. Summary of the invention

[0004] Described herein are systems, methods, and apparatus for injecting motion artifacts into magnetic resonance (MR) images to create artificially contaminated MR images that can be used for various purposes. A motion-contaminated MR image can be created based on a set of source MR images of a scanned object that are substantially free of motion artifacts. The source MR images can include multiple image subsets, and each subset can be associated with one of multiple physiological cycles of the scanned object (e.g., one of multiple cardiac cycles). From a subset of the source MR images associated with the physiological cycle or a modified version of a subset of the source MR images associated with the physiological cycle, (e.g., at least one) MR data segment can be derived for the physiological cycle. At this stage, motion artifacts can be introduced based on a motion pattern (e.g., an artificially created motion pattern) that represents motion artifacts. For example, artifacts can be introduced by deforming a subset of the source MR images based on a motion pattern, by reordering at least a portion of the source images, by interpolating a portion of the source images to derive one or more additional source MR images, and the like. MR data segments derived for multiple physiological cycles can be combined to obtain an MR data set similar to a data set acquired from an actual MR procedure (e.g., a simulated MR data set). Using the MR data set (e.g., a selected portion of the MR data set), one or more target MR images can be generated to include specific motion artifacts. For example, to simulate no-trigger artifacts, one or more target MR images can be generated so that they can be associated with two or more physiological cycles and can appear to have been captured with a missed trigger. The one or more processors are configured to generate the one or more target MR images further based on one or more coil sensitivity maps. The set of source MR images includes cardiac MR images, and the multiple physiological cycles include one or more cardiac intervals.

[0005] In an example, MR data segments associated with multiple physiological cycles can be combined (e.g., using a row-by-row Cartesian coordinate method) to obtain an MR data set (e.g., a digitized MR signal) from which a target image can be generated. In an example, combining the MR data segments can include interpolating the MR data segments to simulate top-down sampling of the MR signal. In an example, the MR data segments derived for multiple physiological cycles can include k-space data obtained by applying a fast Fourier transform (FFT) to a source MR image (or a modified version thereof). Such k-space data can represent, for example, spatial frequency and / or phase information of an object captured during an actual MRI procedure.

[0006] In an example, generating one or more target MR images based on a simulated MR dataset (e.g., k-space data) may include applying an inverse FFT (iFFT) to at least a portion of the MR dataset to derive the target MR image. In an example, generating one or more target MR images based on a simulated MR dataset may include interpolating a portion of the k-space data to cover a portion or all of the k-space. In an example, the one or more target MR images may be further generated based on one or more coil sensitivity maps.

[0007] MR images generated using the systems, methods, and devices described herein can be used for different purposes. For example, the MR images can include cardiac MR images and can be used to train a machine learning model to remove motion artifacts from cardiac MR images. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] A more detailed understanding of the examples disclosed herein may 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 framework for simulating motion artifacts in magnetic resonance (MR) images according to one or more embodiments described herein.

[0010] Figure 2 is a block diagram illustrating an example of simulating MR signals based on segmented MR data according to one or more embodiments described herein.

[0011] Figure 3 is a block diagram illustrating an example of incorporating motion dynamics (eg, motion artifacts) into MR images according to one or more embodiments described herein.

[0012] Figure 4 is a block diagram illustrating an example of simulating motion artifacts in MR images by controlling the manner in which the MR images are generated.

[0013] Figure 5A is an example MR image including no triggering artifacts simulated according to one or more embodiments described herein.

[0014] Figure 5B is an example MR image including respiratory artifacts simulated according to one or more embodiments described herein.

[0015] Figure 6 is a flow chart illustrating an example process for simulating motion artifacts in MR images according to one or more embodiments described herein.

[0016] Figure 7 is a block diagram illustrating example components of a motion simulator according to one or more embodiments described herein.

[0017] Figure 8 is a flow chart illustrating an example of training a neural network to perform motion artifact removal using simulated MR images. DETAILED DESCRIPTION

[0018] The present disclosure is illustrated by way of example and not limitation in the figures of the accompanying drawings.

[0019] Figure 1 The block diagram illustrates an example framework for simulating motion artifacts in magnetic resonance (MR) images. As shown, the simulation can be performed based on a set of source MR images 102 that are substantially free of motion artifacts, and can involve one or more of the following operations: deriving at least one MR data segment 104 for each of a plurality of physiological cycles (e.g., C1, C2, C3, etc.) of a scanned object based on the source MR images 102 (or a modified version of the source MR images 102) associated with the physiological cycle; combining the MR data segments 104 derived for the plurality of physiological cycles (e.g., C1, C2, C3, etc.) to obtain a simulated MR data set 106; and / or generating a target MR image 108 based on the simulated MR data set (e.g., using a selected portion of the MR data set) such that the MR target image 108 can exhibit one or more motion artifacts injected during the aforementioned operations.

[0020] The source MR images 102 may include a single image, a set of images, a movie video (e.g., including multiple MR images), etc. The images may be obtained from various sources, including, for example, from a magnetic resonance imaging (MRI) device, from a database storing patient imaging records, etc. Each source MR image 102 may be associated with a time point (e.g., a position along the time axis), and may represent the state of the scanned object at that time point (e.g., a motion state M). For example, the source images 102 may include N images captured during a physiological cycle of the scanned object (e.g., a complete cycle of cardiac contraction or an RR interval), from which it can be seen that the corresponding motion states (M0, M1, M2...MN) of the scanned object in the physiological cycle may be represented (e.g., the time interval between adjacent motion states may reflect the temporal resolution of the source images). As another example, the set of source MR images 102 may include multiple subsets of source MR images, wherein each subset of the source MR images may be associated with one of multiple physiological cycles (e.g., C1, C2, C3, etc.). These subsets of source MR images may be part of the original set of images acquired to facilitate motion simulation, or they may be created (e.g., copied) during the motion simulation process. For example, the original set of images acquired for motion simulation may include MR images associated only with physiological cycle C1, and as part of the motion simulation process, the original MR images may be further associated with physiological cycles C2, C3, etc. (e.g., copied for these physiological cycles).

[0021] exist Figure 1 During one or more of the stages shown, motion artifacts may be introduced into the target image 108. These motion artifacts may include, for example, overall patient motion, respiratory motion, irregular heartbeats, acquisition-related artifacts (e.g., hardware errors), etc. As an example, artifacts caused by respiratory motion may be introduced by deforming the source MR image 102 based on artificially created motion events (e.g., motion patterns) representing respiration before deriving the MR data segments 104 based on the deformed source MR images (e.g., modified versions of the source MR images). As another example, the duration and / or number of physiological cycles used in the motion simulation may be manipulated to simulate motion that may occur during an actual MRI procedure. Such motion may include, for example, ECG non-triggering, ECG false triggering, irregular heartbeats (e.g., atrial fibrillation and premature ventricular contractions), etc. The following will refer to, for example, Figure 3 and Figure 4 Provides more details on these artifact injection techniques.

[0022] The MR data segments 104 may include subsets of MR data segments, each subset being derived for a corresponding one of a plurality of physiological cycles (e.g., C1, C2, C3, etc.), and including k-space information representing spatial frequency information of a scanned object in two or three dimensions (e.g., k-space may be covered by phase and frequency encoding data). The relationship between the k-space data segments and the source image may be defined by a Fourier transform, such as a fast Fourier transform (FFT). It can be seen that a subset of MR data segments associated with a particular physiological cycle may be derived by applying an FFT to a subset of a source MR image associated with the particular physiological cycle or a modified version of a subset of a source MR image. The data segments thus obtained may be similar to data segments acquired during an actual MRI procedure. For example, using a 2-dimensional (2D) Fourier transform, a row in the derived MR data segment may correspond to a digitized MR signal at a particular phase encoding level, and each digitized data point of the MR signal may be represented by a complex number having a real part and an imaginary part, or may be defined as having an amplitude and a phase. The position of the digitized data point in the k-space may be directly related to the gradient field of the imaged object. Thus, as the gradient varies over time, the k-space data may be sampled in a trajectory through Fourier space.

[0023] Because only a limited amount of k-space data of the scanned object can be obtained from each subset of the source MR image (e.g., the number of k-space data per segment (NVS) may be limited), two or more of the MR data segments 104 derived for multiple physiological cycles can be combined into a simulated MR data set 106 to fill a portion (e.g., a desired portion) or all of the k-space. The k-space data described herein may correspond to MRI readout rows or rows in a 2D MR data space, which may be vertical, horizontal, or oblique. Figure 2 An example is illustrated in which MR data segments S1, S2, and S3 corresponding to physiological cycles C1, C2, and C3 are combined to obtain a set S of simulated MR data covering (e.g., substantially) the entire k-space. Various techniques can be used to perform the combination (e.g., to fill a desired portion or all of k-space), including, for example, techniques based on column-by-column or row-by-row Cartesian coordinate methods or techniques based on spiral and radial oriented trajectories. The timing of the MR data segments 104 can be simulated (e.g., when the MR data segments are acquired), and the MR data segments can be interpolated (e.g., based on the simulated timing of the MR data segments and / or the timing of the source images 102) to simulate actual k-space signal acquisition (e.g., the number of target images generated using the framework described herein can be different from the number of source images).

[0024] Back to Figure 1, a target MR image 108 may be generated by applying an inverse FFT to the simulated MR data 106 (e.g., to a selected portion of the MR data 106) and / or incorporating coil sensitivities (e.g., via one or more coil sensitivity maps) into the image generation process. For example, image generation (e.g., image reconstruction) may include obtaining the product of the complex conjugate of the coil sensitivity of each coil and the MR image obtained via the coil (e.g., via multiplication), and summing the products obtained for all coils, for example, as illustrated below:

[0025]

[0026] Wherein, m may represent an image of a coil combination, n may represent the number of coils, i may represent a coil index, csm_i may represent a coil sensitivity map of an i-th coil, and m_i may represent an MR image obtained via an i-th coil.

[0027] The coil sensitivities described herein may be similar to the sensitivities of a single coil (e.g., multiple coils combined to derive a coil combination image) or multiple coils (such as multiple phased array coils used in an actual MRI procedure). In examples (e.g., when the MR data includes complex-valued data elements), a coil sensitivity map may be determined based on the source image 102 and / or the simulated MR data 108 (e.g., k-space information) using various techniques. For example, a coil sensitivity map csm_i for the i-th coil may be determined using a sum of squares method based on the following equation: csm_i = m_i / m_sos, where m_i may represent an MR image captured using the i-th coil, and m_sos may be equal to and n may represent the number of coils being simulated. In an example (e.g., when the MR data includes only amplitude values), the coil sensitivity map may be simulated as a function of the distance from the coil center to the pixel location, for example. For example, if an image pixel is at location (x, y, z) and the coil center (e.g., in the same coordinate system) is at (a, b, c), then the coil sensitivity at (x, y, z) may be determined as Where r = sqrt((xa) 2 +(yb) 2 +(zc) 2 ). In an example, portions of the k-space data may be interpolated to derive additional target MR images, thereby increasing the temporal resolution of the target images (eg, continuous rather than discrete image frames may be generated).

[0028] As described herein, motion artifacts may be introduced by deforming the source MR image 102 based on artificially created motion events, by simulating the timing of physiological cycles, and by manipulating parameters of physiological cycles (e.g., duration) to simulate artifacts caused by ECG no triggering, ECG false triggering, and / or irregular heartbeats, among others. Figure 3 An example of incorporating motion dynamics (e.g., motion artifacts) into MR images is illustrated. A simulation may be performed based on a set of source MR images 302 that are substantially free of motion artifacts. These source MR images may be captured within a physiological cycle C1 of the scanned subject (e.g., a complete cycle of cardiac contraction), and each image may be associated with a corresponding time point within the physiological cycle. In order to inject motion artifacts into these source images, a motion pattern 304 may be created in the time direction of the source images 302. In an example, the motion pattern may include a breathing pattern (e.g., Figure 3 ), the breathing pattern represents the patient's breathing motion along the time axis. In an example, the motion pattern can represent irregular heartbeats (e.g., when the heart does not complete a contraction cycle), which can be simulated by reordering the source images in the time direction (e.g., the reordered images can correspond to motion states M0, M1, MN instead of M0, M1, M2...MN) or by artificially inserting additional motion states into the source image set (e.g., by interpolating two or more adjacent source images).

[0029] By deforming the source image according to the motion pattern, the motion dynamics or artifacts represented by the motion pattern 304 may be incorporated into the source image 302. For example, Figure 3 The respiratory pattern shown can be represented by a parameterized motion model, and the motion pattern 304 can include a set of motion curves, each of which gives the value of a parameter of the motion model as a function of time (e.g., matching a time point or time position of the source image 302). In an example, the motion model can be a translation displacement model that mimics respiratory motion. Such a model can utilize a sine function that includes a frequency component f and an amplitude component a, for example, in the form of displacement = a*sin(ft). In this equation, the frequency component f can represent the respiratory frequency, the amplitude component a can represent the degree of respiratory motion, t can represent a specific time point, and the displacement can represent the distance that the image (e.g., the entire image) is displaced in the head-to-foot direction (e.g., the direction can be the main motion direction of breathing during MR short-axis imaging) at time t (e.g., different values ​​can be set for a and f for different physiological settings or applications).

[0030] Based on the model and the motion curve, various image deformation and / or registration techniques may be employed to modify the source image 302 so that the motion dynamics associated with the motion pattern 304 may be reflected in the modified version of the source image 302. These image deformation and / or registration techniques may include, for example, affine transformations, rigid translations, non-rigid translations, or combinations thereof.

[0031] Once the source image 302 has been modified to incorporate the desired motion artifact, the modified version of the source image can be used to derive the corresponding MR data segments 306 for the relevant physiological cycle C1. As described herein, the derivation can be performed by applying a fast Fourier transform to the modified source image. In addition, the source image 302 can be combined with additional physiological cycles (e.g., Figure 3 C2 shown in the figure, and can be deformed based on the motion patterns 304 of these physiological cycles. Correspondingly, the corresponding MR data segments can be derived for the additional physiological cycles before merging (e.g., row by row or column by column) the data segments to obtain the simulated MR data set (e.g., digitized MR signals). The simulated MR data set (e.g., k-space data) can then be used to generate target MR images 308 that demonstrate the motion artifacts injected by the motion pattern 304. Since the MR data used to generate the target image can be assembled from k-space segments across multiple physiological cycles, and the k-space segments can be further interpolated to fill part or all of the k-space, the target MR image 308 can be generated at a desired time resolution (e.g., at any time resolution or any time interval) to demonstrate the artifacts indicated by the motion pattern 304.

[0032] Figure 4 Another example of motion artifacts in simulated MR images is illustrated. Figure 3In the example shown, motion simulation can be performed based on a set of source MR images 402 that are substantially free of motion artifacts. The source MR images can be grouped into a plurality of image subsets, each of which is associated with a physiological cycle. A physiological cycle (e.g., C1, C2, C3, etc.) can be created to simulate the timing of the R wave during the ECG process, whereby the physiological cycles C1, C2, C3, etc. can correspond to the RR interval in the ECG graph (e.g., the RR interval can represent the time elapsed between two consecutive R waves of the QRS signal on the electrocardiogram). Based on the image subsets associated with the plurality of physiological cycles, corresponding MR data segments 404 can be derived for the plurality of physiological cycles and combined to obtain a simulated MR data set (e.g., a digitized MR signal) that covers (e.g., substantially) a portion (e.g., representing an undersampling situation) or all (e.g., representing a full sampling situation) of the k-space. The simulated MR data set can then be used to generate a sequence of target MR images 406, and by controlling the manner in which the target MR images 406 are generated (e.g., controlling which portion of the MR data set is used to generate the target image), certain motion artifacts can be injected into the target image. For example, to simulate a no-trigger artifact such as an ECG no-trigger, the target images 406 can be distributed across (e.g., associated with) multiple physiological cycles (e.g., across RR intervals C1 and C2) rather than one such cycle, so that the sequence of target images appears to be taken with a missed trigger between C1 and C2.

[0033] Figure 5A and Figure 5B Example MR images with simulated untriggered artifacts and simulated respiratory artifacts are shown, respectively. It should be noted that although cardiac MRI (cMRI) is used herein to describe the proposed motion simulation technique, cMRI (e.g., retrospectively triggered cine cMRI) merely represents an exemplary field in which the proposed technique can be deployed. Other MRI procedures (e.g., brain, abdomen, etc.) may also benefit from the proposed technique. It should also be noted that what is specifically implemented in the examples provided herein is a framework for simulating motion artifacts in MR images. The types of artifacts that can be generated using the framework are not limited to the types described in the examples, and may include other types that are not specifically listed herein.

[0034] Figure 66 is a flow chart illustrating an example process 600 for simulating motion artifacts in MR images. The process may begin at 602, and at 604, a set of source MR images (e.g., a single MR image, multiple MR images, a movie video, etc.) that are substantially free of motion artifacts may be obtained. The source images may be obtained, for example, from an MRI process, a database storing patient imaging records, etc. Each source image may be associated with a time point or a time position, and the entire set of images may span one or more physiological cycles (e.g., RR intervals). For example, the set of source MR images may be initially associated with a first physiological cycle (e.g., captured within the first physiological cycle), and as part of the motion simulation process 600, may be associated with additional physiological cycles at 606. At 608, the source MR images may be manipulated to incorporate a motion pattern (e.g., to represent certain motion artifacts caused by respiratory motion, irregular heartbeat, etc.). In an example, motion artifacts may be injected into the source MR images by deforming the source images based on the motion pattern (e.g., via an affine, rigid, or non-rigid transformation). In an example, the artifact may be simulated by reordering at least a portion of the source images in time or by artificially inserting additional MR images into the set of source images (eg, based on interpolation of adjacent source images).

[0035] At 610, MR data segments can be derived for each physiological cycle based on a source MR image associated with the physiological cycle or a modified version of the source MR image (e.g., when motion artifacts are injected into the source image). For example, the MR data segments can be derived by applying a Fourier transform (e.g., FFT) to the source MR image to obtain k-space information, which k-space information represents spatial frequency and / or phase information of an object scanned in the source MR image. At 612, the MR data segments derived for the physiological cycle can be combined to obtain a simulated MR data set (e.g., digitized MR signals) that simulates a data set acquired from an actual MRI procedure. In an example, the combination of the MR data segments can include merging the MR data segments column by column or based on spiral and / or radial directional trajectories. In an example, the combination of the MR data segments can include interpolating the MR data segments to simulate top-down sampling of the MR signals.

[0036] Once obtained, at 614, the simulated MR data set can be used to generate one or more target MR images including specific motion artifacts. As described herein, the motion artifact can be incorporated into a modified version of the source MR image, from which one or more MR data segments can be derived. For example, motion artifacts can also be created by controlling the manner in which target MR images are generated. For example, each target MR image can be generated using a selected portion of the MR data set based on the desired motion artifact. As another example, in order to simulate ECG untriggered artifacts, target images can be generated (e.g., distributed) across multiple physiological cycles so that the images appear to be taken with a missed trigger between physiological cycles. The motion simulation process 600 can then end at 616.

[0037] For simplicity of illustration, the operations involved in the motion simulation process 600 are depicted and described herein in a particular order. However, it should be noted that these operations may occur in various orders, simultaneously, and / or with other operations not presented or described herein. In addition, it should be noted that the operations that may be included in the motion simulation process 600 may not be depicted and described in full herein, and not all illustrated operations need to be performed.

[0038] 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 7 is a block diagram illustrating an example motion simulator 700 that can be configured to perform one or more functions described herein. As shown, the motion simulator 700 may include a processor (e.g., one or more processors) 702, 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. The motion simulator 700 may also include a communication circuit 704, a memory 706, a mass storage device 708, an input device 710, and / or a communication link 712 (e.g., a communication bus), through which one or more components shown in the figure can exchange information.

[0039] The communication circuit 704 may 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 706 may include a storage medium (e.g., a non-transitory storage medium) configured to store machine-readable instructions, which, when executed, causes the processor 702 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 708 may include one or more disks, such as one or more built-in 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 702. The input device 710 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 motion simulator 700 .

[0040] It should be noted that the motion simulator 700 can operate as a stand-alone device or can be connected (e.g., networked or clustered) with other computing devices to perform the functions described herein. Figure 7 Only one example of each component is shown in the figure, and those skilled in the art will also understand that the motion simulator 700 may include multiple examples of one or more components shown in the figure.

[0041] Motion-contaminated MR images simulated using the techniques described herein can be used to improve the quality and efficiency of various clinical applications. For example, the simulated MR images can be used to train an artificial neural network (ANN) to learn a model (e.g., a machine learning (ML) model) to automatically remove motion artifacts from MR images acquired during an actual MRI procedure. Such pre-trained ANNs can be deployed (e.g., implemented) on one or more computing devices, which can be configured to receive MR images of scanned objects including motion artifacts, and process the MR images through the pre-trained ANNs to obtain output images that are substantially free of motion artifacts.

[0042] Figure 8A process 800 is illustrated that can be used to train a neural network (e.g., a convolutional neural network (CNN)) to learn a motion artifact removal model based on an MR image generated using the techniques described herein. The process 800 can begin at 802, and at 804, initial parameters of the neural network (e.g., weights associated with various filters or kernels of the neural network) can be initialized. The parameters can be initialized, for example, based on samples collected from one or more probability distributions or parameter values ​​of another neural network with a similar architecture. At 806, the neural network can receive a motion-contaminated MR image generated using the simulation techniques described herein, and process the image through the layers of the neural network to remove motion artifacts from the image (e.g., predict a corresponding clean MR image without motion artifacts). At 808, the results of the processing can be compared to a gold standard to determine adjustments that need to be made to the currently assigned neural network parameters. The gold standard can be a source MR image used to generate the motion-contaminated image (e.g., the source MR image 102, 302, or 402 described herein), and adjustments to the network parameters can be determined based on a loss function (e.g., based on a mean squared error (MSE)) and a gradient descent (e.g., stochastic gradient descent) associated with the loss function.

[0043] At 810, the neural network may apply adjustments to the currently assigned network parameters, such as via a back-propagation process. At 812, the neural network may determine whether one or more training termination criteria are met. For example, if the neural network has completed a predetermined number of training iterations, if the difference between the processed result and the gold standard value 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 are met. If it is determined at 812 that the training termination criteria are not met, the neural network may return to 806. If it is determined at 812 that the training termination criteria are met, the neural network may end the training process 800 at 814.

[0044] For simplicity of explanation, the training steps are depicted and described herein in a particular order. However, it should be understood that the training operations may occur in various orders, simultaneously, and / or with other operations not presented or described herein. In addition, 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.

[0045] Although the present disclosure has been described according to certain embodiments and generally associated methods, the changes 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 using 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 registers and memories of a computer system into other data represented as physical quantities within a computer system memory or other such information storage, transmission or display device.

[0046] 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 appended claims and the full scope of equivalents to which such claims are assigned.

Claims

1. A motion artifact simulator configured to inject motion artifacts associated with a medical procedure into a medical magnetic resonance (MR) image, comprising: One or more processors configured to: obtaining a set of source MR images of a scanned object, the set of source MR images being free of motion artifacts and comprising a plurality of subsets of the source MR images, each subset of the source MR images being associated with one of a plurality of physiological cycles of the scanned object; deriving at least one MR data segment for each physiological cycle of the plurality of physiological cycles based on a subset of the source MR images associated with the physiological cycle or a modified version of the subset of the source MR images associated with the physiological cycle; obtaining an MR data set based on corresponding MR data segments derived for the plurality of physiological cycles; and One or more target MR images are generated based on the MR data set such that the one or more target MR images include motion artifacts.

2. The motion artifact simulator according to claim 1, wherein: The at least one MR data segment for the physiological cycle is derived based on the modified version of the subset of the source MR images associated with the physiological cycle, and wherein the one or more processors are configured to obtain the modified version of the subset of the source MR images by modifying the subset of the source MR images to incorporate the motion pattern associated with the motion artifact.

3. The motion artifact simulator according to claim 2, wherein: The one or more processors being configured to modify a subset of the source MR images associated with each of the plurality of physiological cycles to incorporate the motion pattern includes the one or more processors being configured to deform the subset of the source MR images based on a model representing the motion pattern.

4. The motion artifact simulator according to claim 2, wherein: The one or more processors are configured to modify a subset of the source MR images associated with each of the multiple physiological cycles to incorporate the motion pattern, including: the one or more processors are configured to reorder at least a portion of the set of source MR images or interpolate a portion of the set of source MR images to derive one or more additional source MR images.

5. The motion artifact simulator according to claim 1, wherein: The one or more processors being configured to generate the one or more target MR images based on the MR dataset includes the one or more processors being configured to select a portion of the MR dataset and generate a target MR image based on the selected portion of the MR dataset.

6. The motion artifact simulator according to claim 1, wherein: The one or more processors are configured to obtain the MR data set based on the corresponding MR data segments derived for the multiple physiological cycles, including: the one or more processors are configured to combine the MR data segments derived for the multiple physiological cycles, wherein the one or more processors are configured to combine the MR data segments derived for the multiple physiological cycles, including: the one or more processors are configured to interpolate at least a portion of the MR data segments to obtain the MR data set.

7. The motion artifact simulator according to claim 1, wherein: The at least one MR data segment derived for each of the multiple physiological cycles includes k-space data, and wherein the one or more processors are configured to obtain the k-space data by applying a fast Fourier transform (FFT) to a subset of the source MR images associated with each of the multiple physiological cycles or a modified version of the subset of the source MR images associated with each of the multiple physiological cycles, wherein the one or more processors are configured to generate the one or more target MR images based on the MR data set, including: the one or more processors are configured to generate each of the one or more target MR images by applying an inverse fast Fourier transform (FFT) to at least a portion of the MR data set.

8. A method for injecting motion artifacts associated with a medical procedure into a medical magnetic resonance (MR) image, the method comprising: obtaining a set of source MR images of a scanned object, the set of source MR images being free of motion artifacts and comprising a plurality of subsets of the source MR images, each subset of the source MR images being associated with one of a plurality of physiological cycles of the scanned object; deriving at least one MR data segment for each physiological cycle of the plurality of physiological cycles based on a subset of the source MR images associated with the physiological cycle or a modified version of the subset of the source MR images associated with the physiological cycle; obtaining an MR data set based on corresponding MR data segments derived for the plurality of physiological cycles; as well as One or more target MR images are generated based on the MR data set such that the one or more target MR images include motion artifacts.

9. The method according to claim 8, further comprising: The one or more target MR images are used to train a machine learning model for removing motion artifacts from MR images.

10. A computer-implemented method for processing a medical magnetic resonance (MR) image, the method comprising: receiving an MR image of a scanned object, wherein the MR image includes motion artifacts; and The MR images are processed by an artificial neural network to obtain an output image without the motion artifact, wherein the artificial neural network is pre-trained using computer-generated MR training images to remove the motion artifact from the MR images, and the MR training images include motion artifacts created by at least: obtaining a set of source MR images of the scanned object, the set of source MR images being free of the motion artifact and comprising a plurality of subsets of source MR images, each subset of the source MR images being associated with one of a plurality of physiological cycles of the scanned object; deriving at least one MR data segment for each physiological cycle of the plurality of physiological cycles based on a subset of the source MR images associated with the physiological cycle or a modified version of the subset of the source MR images associated with the physiological cycle; obtaining an MR data set based on the corresponding MR data segments derived for the plurality of physiological cycles; and The MR training image is generated based on the MR dataset, wherein the generating comprises manipulating at least a portion of the MR dataset to inject the motion artifact into the MR training image.

Images