Method for mri motion artifact correction based on physically guided k-vit with implicit decoupling

Abstract

The present application relates to the field of medical image processing, and provides a method for correcting MRI motion artifacts based on physical guided k-ViT and implicit decoupling. The method first decomposes the phase disturbance component caused by motion from the original signal containing motion artifacts based on the k-space physical model; then, the k-ViT network is used to suppress the phase disturbance through its phase-aware attention calculation, and the preliminary corrected k-space signal is obtained; then, the preliminary corrected signal is decoupled into static tissue structure and dynamic motion field through the implicit neural representation network, and the artifact-free k-space signal is reconstructed only according to the static structure; finally, the reconstructed signal is inverse Fourier transformed to obtain the corrected MRI image. Through the physical guided decomposition at the data source, the global suppression in the frequency domain and the decoupling reconstruction of the signal components, the present application can accurately and completely remove the motion artifacts and effectively improve the image quality, providing a more reliable image basis for clinical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and more specifically, to a method for MRI motion artifact correction based on physically guided k-ViT and implicit decoupling. Background Technology

[0002] Magnetic resonance imaging (MRI) is an indispensable non-invasive imaging tool in clinical diagnosis and scientific research. However, its data acquisition process is time-consuming. The unavoidable physiological movements (such as heartbeat, breathing, and intestinal peristalsis) and involuntary movements (such as limb swaying) of patients during the scanning process can introduce phase errors and signal inconsistencies in the k-space (frequency space) data. Ultimately, this produces severe motion artifacts in the reconstructed images, which manifest as ghosting, blurring, and stripes. This significantly reduces image quality and affects the diagnostic confidence of radiologists and the accuracy of subsequent quantitative analysis.

[0003] To address this challenge, existing technical solutions can be broadly categorized into three types, but all have inherent limitations. The first type is image post-processing-based methods, which directly suppress artifacts in the image domain through filtering or algorithms. These methods often operate as a "black box," lacking consideration for the physical imaging mechanisms of MRI. While removing artifacts, they are prone to losing image details or introducing unrealistic smoothing effects, and their effectiveness heavily relies on empirical parameter adjustments, resulting in weak generalization ability. The second type is k-space data compensation-based methods, which attempt to directly detect and correct anomalous data points in k-space. However, the core challenge of this method lies in accurately identifying motion-contaminated data. Currently, it largely relies on manually or semi-automatically setting thresholds, which are highly sensitive and prone to misjudgment in complex motion scenarios, leading to under- or over-correction. The third type is the recently emerging supervised deep learning-based methods, which rely on a large number of "artifact-containing-clean" image pairs to train the model. Although some progress has been made, its application is limited by the high cost of acquiring paired data, and the model's interpretability is poor, posing a risk of unclear physical meaning in the output results, making it difficult to meet the stringent reliability requirements of clinical applications.

[0004] In summary, existing technologies generally face three major bottlenecks: insufficient physical interpretability, reliance on empirical thresholds or difficult-to-obtain paired supervision signals, and poor adaptability to complex motion patterns. Therefore, there is an urgent need in this field for a novel correction scheme that can deeply integrate MRI physical mechanisms into model design, achieve efficient training without paired labels, and accurately decouple the root causes of motion artifacts, in order to overcome the bottlenecks of existing technologies and improve the robustness, reliability, and universality of the method. Summary of the Invention

[0005] This invention addresses the technical problems existing in the prior art by providing an MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling. This method directly and accurately separates and eliminates phase perturbations caused by motion at the k-space data source, thereby recovering high-quality MRI images without motion artifacts.

[0006] According to a first aspect of the present invention, a method for MRI motion artifact correction based on physically guided k-ViT and implicit decoupling is provided, comprising:

[0007] S1, based on the k-space physical model, decomposes the phase perturbation component caused by motion from the k-space signal containing motion artifacts;

[0008] S2, the phase perturbation component caused by motion is suppressed by the trained k-ViT network to obtain the preliminary corrected k-space signal;

[0009] S3, using a trained implicit neural representation network, the preliminary corrected k-space signal is decoupled into a static structure and a dynamic motion field, and an artifact-free k-space signal is reconstructed based on the static structure;

[0010] S4. Perform an inverse Fourier transform on the artifact-free k-space signal to obtain the corrected MRI image.

[0011] Based on the above technical solution, the present invention can also be improved as follows.

[0012] Optionally, step S1 includes:

[0013] S101, Obtain a k-space physical model containing motion artifacts. The k-space physical model containing motion artifacts expresses the k-space signal containing motion artifacts as a quantitative mapping relationship between a pure structure signal, a motion displacement field, and a phase perturbation factor.

[0014] S102, based on the k-space physical model containing motion artifacts, the k-space signal containing motion artifacts is decomposed into a tissue structure component after displacement distortion and a phase perturbation component caused by motion.

[0015] Optionally, step S2 includes:

[0016] S201, normalize the k-space signal containing motion artifacts and perform adaptive block processing based on the k-space radius, extract amplitude and phase features from each data block obtained by block division, and concatenate them into a complex feature vector;

[0017] S202, the complex feature vector is input into the phase-aware multi-head self-attention module of the k-ViT network, and global features are obtained by jointly modeling the global dependency relationship between amplitude and phase;

[0018] S203, the global features are processed by the encoder of the k-ViT network to suppress motion artifact-related features and enhance tissue structure features, and the preliminary corrected k-space signal is reconstructed through inverse block operation.

[0019] Optionally, in step S201, the normalization and adaptive block segmentation processing of the k-space signal containing motion artifacts includes:

[0020] The k-space signal containing motion artifacts is normalized so that its amplitude range is mapped to the [0,1] interval;

[0021] Calculate the radius value of each data point in the k-space, and divide the k-space into a central region and an edge region based on a preset radius threshold;

[0022] The data in the central region is divided into blocks using a first-size block strategy, while the data in the edge region is divided into blocks using a second-size block strategy that is larger than the first size.

[0023] Optionally, step S202 includes:

[0024] S2021, the complex feature vector is input into the phase-aware multi-head self-attention module of the k-ViT network, and linear transformations are performed on the amplitude and phase features in the complex feature vector to generate a query, key, and value matrix of amplitude features. And phase feature queries, key matrix ;

[0025] S2022, calculate the phase-aware attention score, which is composed of the dot product of the amplitude query and the key matrix, plus the dot product of the phase query and the key matrix after being weighted by a weighting factor;

[0026] S2023, apply the phase-aware attention score to the amplitude value matrix, and output the global feature through multi-head stitching and linear projection.

[0027] Optionally, step S203 includes:

[0028] S2031, The global feature input is nonlinearly transformed by the encoder composed of a multi-layer feedforward network, wherein the feedforward network uses the GELU activation function;

[0029] S2032, in the encoder, through multi-layer stacked nonlinear transformation, the feature response related to the phase perturbation factor is attenuated, and the feature response related to the static tissue structure is enhanced;

[0030] S2033, perform an inverse block operation on the feature vector output by the encoder to reconstruct it into a preliminary corrected k-space signal with the same size as the k-space signal containing motion artifacts.

[0031] Optionally, step S3 includes:

[0032] S301, the spatial coordinates of the preliminarily corrected k-space signal are input into the static structure branch of the implicit neural representation network, and the intrinsic magnetization intensity distribution characterizing the static tissue structure is obtained by mapping through a multilayer perceptron;

[0033] S302, the spatial coordinates and sampling time of the preliminary corrected k-space signal are input into the dynamic motion field branch of the implicit neural representation network, and the dynamic motion field representing the motion displacement is obtained by mapping through another multilayer perceptron.

[0034] S303, based on the inherent magnetization distribution of the static structural branch output, an artifact-free k-space signal is reconstructed by integral calculation using the k-space physical model.

[0035] Optionally, prior to step S1, the method further includes:

[0036] We constructed a k-space physical model with motion artifacts, a k-ViT network, and a hidden neural representation network;

[0037] The k-space physical model containing motion artifacts is constructed by introducing a motion displacement field and a phase perturbation factor into the pure k-space signal model. The pure k-space signal model is obtained by Fourier transform based on the magnetization distribution of spin protons. The phase perturbation factor is generated by the coupling effect of the gradient field and the motion displacement field.

[0038] The k-ViT network includes an input preprocessing layer, a phase-aware multi-head self-attention module, and a multi-layer encoder. The input preprocessing layer is configured to perform adaptive block division and complex feature extraction of k-space signals, and the phase-aware multi-head self-attention module is configured to jointly model the global dependency between amplitude and phase.

[0039] The implicit neural representation network includes parallel static and dynamic branches, and the independence of the static and dynamic branches is constrained by a decoupling loss function. The static branch includes a multilayer perceptron, which takes spatial coordinates as input and static organizational structure as output. The dynamic branch includes another multilayer perceptron, which takes spatial coordinates and sampling time as input and dynamic motion field as output.

[0040] A self-supervised loss function is constructed based on k-space physical constraints and a decoupling loss function, and the k-ViT network and the implicit neural representation network are jointly trained based on the self-supervised loss function.

[0041] Optionally, the step of constructing a self-supervised loss function based on k-space physical constraints and a decoupling loss function, and jointly training the k-ViT network and the implicit neural representation network based on the self-supervised loss function, includes:

[0042] Construct a self-supervised loss function, expressed as:

[0043]

[0044] in, For self-supervised loss function, This is a symmetric loss based on the conjugate symmetry of k-space. The sampling consistency loss is based on the consistency of adjacent sampling points. Let be the decoupling loss function. , , These are the weighting coefficients corresponding to each loss term;

[0045] With minimizing the self-supervised loss function as the optimization objective, the k-ViT network and the implicit neural representation network are jointly trained end-to-end.

[0046] The network parameters are updated through an iterative optimization process until the self-supervised loss function converges.

[0047] According to a second aspect of the present invention, an MRI motion artifact correction system based on physically guided k-ViT and implicit decoupling is provided, comprising:

[0048] The signal input module is configured to decompose the phase perturbation component caused by motion from the k-space signal containing motion artifacts based on the k-space physical model.

[0049] The phase perturbation suppression module includes a trained k-ViT network, which is configured to suppress the motion-induced phase perturbation components decomposed from the k-space signal and output a pre-corrected k-space signal.

[0050] The signal decoupling and reconstruction module includes a trained implicit neural representation network, which is configured to decouple the pre-corrected k-space signal into a static structure and a dynamic motion field, and reconstruct an artifact-free k-space signal based on the static structure.

[0051] The image generation module is configured to perform an inverse Fourier transform on the artifact-free k-space signal to generate and output a corrected MRI image.

[0052] According to a third aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the processor is configured to implement the steps of the above-described MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling when executing a computer management program stored in the memory.

[0053] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer management program stored thereon, which, when executed by a processor, implements the steps of the above-described MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling.

[0054] This invention provides a method, system, electronic device, and storage medium for MRI motion artifact correction based on physically guided k-ViT and implicit decoupling. It eliminates motion artifacts by modeling in k-space (the original data domain). First, based on a physical model, the phase perturbation component caused by motion is separated from the k-space signal containing motion artifacts. Next, using a k-ViT network specifically designed for frequency domain data, this phase perturbation component is globally suppressed through its phase-aware attention mechanism, resulting in a pre-corrected k-space signal. Then, an implicit neural representation network is used to decouple the pre-corrected k-space signal, decomposing it into a static tissue structure that does not change over time and a dynamic motion field describing the motion. A clean k-space signal is reconstructed based solely on the static structure. Finally, the final corrected MRI image is obtained through inverse Fourier transform. This invention's decomposition and correction, starting from a physical model, are more targeted and improve accuracy; processing at the k-space source avoids information loss during image domain post-processing; through the two-step synergy of "suppressing phase perturbation" and "decoupling static structure," motion artifacts can be separated and removed more thoroughly; the constructed k-ViT network model and implicit neural representation network support self-supervised training based on k-space physical properties (such as symmetry), reducing the dependence on hard-to-obtain "clean-artifact" image pairs. Attached Figure Description

[0055] Figure 1 A flowchart of an MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling provided by the present invention;

[0056] Figure 2 A block diagram of an MRI motion artifact correction system based on physically guided k-ViT and implicit decoupling provided by the present invention;

[0057] Figure 3 A schematic diagram of a possible hardware structure of an electronic device provided by the present invention;

[0058] Figure 4 This is a schematic diagram of the hardware structure of a possible computer-readable storage medium provided by the present invention. Detailed Implementation

[0059] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0060] Figure 1 A flowchart of an MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling provided by this invention is shown below. Figure 1 As shown, the method includes steps S1 to S4:

[0061] S1, based on the k-space physical model, decomposes the phase perturbation component caused by motion from the k-space signal containing motion artifacts.

[0062] This step, based on the physical principles of magnetic resonance imaging (MRI), mathematically establishes a definitive expression for the signal containing motion artifacts. This allows for the separation of the phase error component directly caused by body movement from the acquired raw k-space signal. This step provides a clear mathematical object and physical target for the entire correction process, enabling subsequent processing to precisely target the root cause of motion artifacts, rather than performing blurry estimations and patching on the image.

[0063] S2, the phase perturbation component caused by motion is suppressed by the trained k-ViT network to obtain the preliminary corrected k-space signal.

[0064] This step employs a Transformer network (i.e., a k-ViT network) specifically designed for processing complex signals in k-space to perform global signal analysis. For example, it extracts the amplitude and phase features of the signal and utilizes an improved attention computation method to model and suppress phase perturbation correlations caused by motion. This step performs global correction directly at the level of the raw data from which artifacts are generated, effectively reducing coherent motion errors throughout the entire k-space while preserving the true tissue structure signal, thus preparing for subsequent fine decoupling.

[0065] S3, using a trained implicit neural representation network, the preliminary corrected k-space signal is decoupled into a static structure and a dynamic motion field, and an artifact-free k-space signal is reconstructed based on the static structure.

[0066] This step uses a network capable of learning continuous representations. The signal processed in the previous step is input into its two parallel pathways: one pathway learns the static anatomical structure that does not change over time, and the other pathway learns the dynamic displacement field caused by motion, thereby achieving complete separation of signal components. Finally, only the learned static structural components are used to reconstruct a clean k-space signal. This step completely decouples real tissue information from motion interference at the signal representation level, ensuring that the reconstructed signal physically corresponds to a clean, motion-free scan result, fundamentally avoiding artifact residue.

[0067] S4. Perform an inverse Fourier transform on the artifact-free k-space signal to obtain the corrected MRI image.

[0068] This step performs the inverse Fourier transform, a deterministic mathematical operation, which transforms the k-space data, which has been successfully corrected and has had motion errors eliminated in the previous steps, back into anatomical images that can finally be used for clinical diagnosis, thus completing the entire generation process from contaminated raw data to high-quality result images.

[0069] Understandably, given the shortcomings in the background technology, this invention proposes an MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling. This method, based on the physical laws of MRI imaging, first locates and separates the phase perturbation term directly generated by motion in the original k-space data; then, using a k-ViT network specifically designed for frequency domain data, it globally suppresses this phase perturbation term within the k-space; next, through an implicit neural representation network, it decouples the pre-processed signal into a static tissue structure and a dynamic motion field, and reconstructs a perturbation-free k-space signal based solely on the static structure; finally, it obtains the final MRI image through inverse Fourier transform. This invention addresses the source of artifact generation (k-space) and, through the synergistic processing of physical model decomposition, global frequency domain suppression, and signal component decoupling, can accurately and thoroughly eliminate motion artifacts while preserving the true tissue structure information to the maximum extent, thus obtaining a high-quality corrected image. The method of this invention has advantages such as strong physical interpretability, high detail fidelity, no need for ground truth values ​​for model training, and excellent generalization ability, making it suitable for various clinical MRI scenarios.

[0070] Based on the above technical solution, this embodiment can be further improved as follows.

[0071] In one possible embodiment, step S1 includes sub-steps S101~S102:

[0072] S101, Obtain a k-space physical model containing motion artifacts. The k-space physical model containing motion artifacts expresses the k-space signal containing motion artifacts as a quantitative mapping relationship between a pure structure signal, a motion displacement field, and a phase perturbation factor.

[0073] S102, based on the k-space physical model containing motion artifacts, the k-space signal containing motion artifacts is decomposed into a tissue structure component after displacement distortion and a phase perturbation component caused by motion.

[0074] It is understandable that motion artifacts are essentially caused by the abnormal accumulation of spin proton phase due to the movement of human tissue during imaging, which in turn introduces a perturbation signal into k-space. In this embodiment, the k-space physical model containing motion artifacts, also known as the motion artifact phase perturbation model, is the physical basis and mathematical core of the entire correction method of this invention. Essentially, it establishes a quantitative mapping relationship of "artifact-free signal → motion interference → artifact-containing signal" by reconstructing the physical process of MRI imaging. Based on this principle, the modeling approach of the k-space physical model containing motion artifacts in this embodiment is as follows: first, define the "pure signal without motion," then introduce the "position + phase changes caused by motion," and finally derive the "signal model containing artifacts." The entire process follows the physical imaging principles of MRI, ensuring that each step has physical meaning.

[0075] The construction process of a k-space physical model containing motion artifacts is illustrated with an example.

[0076] (1) Original MRI Space signal model

[0077] In MRI imaging, The space signal is generated by the Fourier transform of the magnetization of spin protons. In an ideal, motionless scenario, Spatial signals can be represented as:

[0078] (1)

[0079] in, for Spatial wave vector, Sampling time, For the imaging area, For position The intrinsic magnetization of a spin proton (reflecting information about tissue structure). The imaginary unit, It is the dot product of the wave vector and the spatial position.

[0080] (2) Phase perturbation model introduced by motion

[0081] When there is organized motion, the actual position of the spin proton changes with time, denoted as the displacement field. (i.e., time) Location The displacement of the proton at that point), then the instantaneous position of the proton after the motion is Simultaneously, motion causes abnormal phase accumulation, introducing additional phase perturbation terms. Ultimately, including motion artifacts The spatial signal model is:

[0082] (2)

[0083] Among them, the phase perturbation term Generated by the coupling effect of gradient field and motion, and combined with MRI gradient field sequences (slice selection, phase encoding, frequency encoding gradients), its quantitative expression is:

[0084] (3)

[0085] In equation (3), The proton gyromagnetic ratio (a constant, approximately) ), For a moment The synthetic gradient field intensity. Substituting equation (3) into equation (2) allows for further decomposition of the artifact-containing area. The spatial signal is in the form of the product of the structure term and the disturbance term:

[0086] (4)

[0087] in, This is the global phase perturbation factor. This represents the structural signal after displacement and distortion.

[0088] Equation (4) expresses the k-space signal containing motion artifacts as a quantitative mapping relationship between the pure structural signal, the motion displacement field and the phase perturbation factor. Through Equation (4), the k-space signal containing motion artifacts can be decomposed into the tissue structure component after displacement distortion and the phase perturbation component caused by motion.

[0089] This model clarifies the physical origin of motion artifacts, provides interpretable mathematical constraints for subsequent correction steps, and overcomes the shortcomings of the "black box" correction in traditional methods.

[0090] Traditional Transformers primarily model in the image domain (spatial domain), neglecting the frequency domain characteristics of MRI signals. This embodiment designs a frequency-domain dedicated Transformer (k-ViT) that directly models global artifact structures in k-space, while introducing a phase-aware attention mechanism to accurately capture the coupling relationship between phase perturbations and structural signals. The characteristics of the k-ViT network include: frequency domain adaptation, dual-feature modeling, and global capture. Its workflow consists of three steps: k-space input preprocessing and block division, phase-aware multi-head self-attention mechanism, and encoder stacking and preliminary correction. Each step closely adapts to the frequency domain physical characteristics of MRI.

[0091] In one possible embodiment, the k-ViT network includes an input preprocessing layer, a phase-aware multi-head self-attention module, and a multi-layer encoder. The input preprocessing layer is configured to perform adaptive block division and complex feature extraction of the k-space signal, and the phase-aware multi-head self-attention module is configured to jointly model the global dependencies between amplitude and phase.

[0092] In this embodiment, based on the above-described k-ViT network architecture, step S2 includes sub-steps S201 to S203.

[0093] S201, executed in the input preprocessing layer Spatial input preprocessing and block partitioning strategies, data preparation adapted to k-space characteristics

[0094] Considering Due to the conjugate symmetry and non-uniform sampling characteristics of spatial signals (high signal strength in the central region and weak signal strength in the edge region), an adaptive non-uniform block partitioning strategy is adopted:

[0095] (1) For those containing motion artifacts Space signals Normalization is performed to map the amplitude range to the [0,1] interval, thus eliminating the influence of amplitude differences.

[0096]

[0097] (2) Based on spatial radius Blocks, central area for The space is half the height and width (1 / 2). Small blocks, edge areas use Large blocks, balancing details and overall information;

[0098] (3) MRI signals are complex numbers; their amplitude reflects signal intensity (corresponding to image brightness), and their phase reflects the phase shift of the signal (corresponding to the root cause of artifacts). Both are indispensable. Therefore, dual features are extracted from each block and stitched together:

[0099] Extract amplitude features for each block Phase characteristics Concatenate them into a complex eigenvector , The block index is used as the input to the k-ViT network.

[0100] S202, Phase-Aware Multi-Head Self-Attention Mechanism

[0101] Traditional self-attention mechanisms focus only on amplitude features, neglecting crucial information about phase perturbations. This embodiment designs a phase-aware attention mechanism that jointly models amplitude and phase features. The attention score is calculated using the following formula:

[0102] (5)

[0103] Attention score consists of two parts: amplitude attention. and phase attention ( By using weighted summation, dual-feature joint modeling is achieved.

[0104] In equation (5), , , The query, key, and value matrices for amplitude features are arranged sequentially, and the matrix is ​​composed of complex feature vectors. The amplitude component is generated through a linear transformation and used to capture the intensity correlation between different blocks (such as the intensity coupling between the central structure and edge artifacts):

[0105] ;

[0106] In equation (5), , The sequence is: phase feature query, key matrix, and so on. The phase component is generated through a linear transformation and is used to capture the phase correlation between different blocks (such as the propagation law of phase perturbations caused by motion in the global context).

[0107] ;

[0108] In equation (5), The dimension of the key matrix is... These are the phase weighting coefficients. Through adaptive adjustment, the value range is: This is used to balance the contributions of amplitude and phase;

[0109] Multi-head attention through parallel computation Each single-head attention node is concatenated to output global features:

[0110]

[0111] in, This is used to output the projection matrix. Multiple single-head attention can simultaneously capture various information such as local phase perturbations, global artifact distribution, and amplitude-phase coupling, avoiding the one-sidedness of a single attention and enabling the model to more comprehensively model complex artifacts (such as mixed artifacts of rigid body motion and non-rigid body motion).

[0112] S203, the encoder and preliminary correction of the k-ViT network, realizes the process from feature modeling to signal reconstruction.

[0113] The encoder function of the k-ViT network is to process the global features obtained from S202 through a multi-layer stack of "attention + feedforward networks". Deep processing is performed to suppress artifact-related features, enhance tissue structure features, and finally reconstruct a preliminarily corrected k-space signal. For example, in this embodiment, the encoder of the k-ViT network consists of four layers of "phase-aware attention + feedforward network" stacked together. This is because the complexity of k-space artifacts is moderate (mainly phase perturbations caused by motion, rather than random noise), and the four-layer stack can fully model the artifacts while avoiding overfitting.

[0114] The feedforward network employs a dual-channel structure to process amplitude and phase characteristics separately:

[0115] (6)

[0116] in, For activation function, This is the weight matrix. For bias. The encoder output is reconstructed through inverse block operation into a pre-corrected k-space signal with the same size as the k-space signal containing motion artifacts. This achieves initial suppression of global phase perturbation.

[0117] In this step, multi-layered stacked nonlinear transformations are used to attenuate the characteristic responses related to the phase perturbation factor and enhance the characteristic responses related to the static tissue structure, thereby suppressing the global phase perturbation factor. The impact, among which, Corresponding to the phase perturbation term in the physical model of step S1, the structural signal after displacement and distortion is retained. The core information.

[0118] Experimental verification showed that the preliminarily corrected k-space signal Compared to the original k-space signal containing motion artifacts The artifact intensity can be significantly reduced, but slight displacement distortion still exists, requiring further decoupling in subsequent steps.

[0119] In step S2, the output of the k-ViT network is a high-quality intermediate result, not the final correction result. Its value is to reduce the pressure on the subsequent "structure-motor decoupling" of the implicit neural representation network and avoid decoupling failure caused by the implicit neural representation network directly processing strong artifact signals.

[0120] To address the shortcomings of traditional methods that rely on dictionaries and thresholds, step S3 constructs a two-branch implicit neural representation network (INR) to decouple tissue structure and motion perturbations into two continuous implicit functions, achieving accurate separation without dictionaries or thresholds. Step S3, based on the initial correction by the k-ViT network, completely separates the "structural signals with slight displacement and distortion" from the "residual motion perturbations" (corresponding to the physical model in...). and Ultimately, it outputs a clean MRI signal without artifacts.

[0121] In one possible embodiment, step S3 includes sub-steps S301 to S303.

[0122] S301, the spatial coordinates of the preliminarily corrected k-space signal are input into the static structure branch of the implicit neural representation network, and the intrinsic magnetization distribution characterizing the static tissue structure is obtained through a multilayer perceptron mapping.

[0123] Understandably, the Implicit Neural Representation Network (INR) comprises two independent, learnable continuous neural network (MLP) branches: one responsible for learning the "static organizational structure" and the other responsible for learning the "dynamic motion field." By constraining the independence of the two through a decoupling loss function, precise separation is ultimately achieved.

[0124] The input to the implicit neural representation network INR is the pre-corrected k-space signal output by the k-ViT network. It has suppressed 80% of artifacts, preserved the core structure and slight displacements, and simultaneously inputs k-space position vectors. and sampling time This provides position and time information for the branches of the continuous neural network MLP to adapt to the dynamic characteristics of motion, and the outputs are the intrinsic magnetization. With displacement field .in:

[0125] (1) Static branching (static structure modeling): A 6-layer MLP is used, with k spatial position vectors as input. Output intrinsic magnetization Add L2 regularization constraints to ensure structural smoothness: ;

[0126] (2) Dynamic branching (dynamic motion field modeling): 8-layer MLP is used, with input k-space position vector and sampling time. Output displacement field Add displacement continuity constraints: This ensures that the sports field conforms to the laws of human tissue movement.

[0127] S302, the spatial coordinates and sampling time of the preliminary corrected k-space signal are input into the dynamic motion field branch of the implicit neural representation network, and the dynamic motion field representing the motion displacement is obtained through mapping by another multilayer perceptron.

[0128] S303, based on the inherent magnetization distribution of the static structural branch output, an artifact-free k-space signal is reconstructed by integral calculation using the k-space physical model.

[0129] It is understandable that steps S302 and S303 are based on the k-space physical model of step S1 to realize signal reconstruction, comparison, and optimization, that is, to perform decoupling constraints and fine correction.

[0130] Specifically, based on the k-space physical model in step S1, the artifact-free k-space signal is reconstructed using the output of the implicit neural representation network INR. And the initial corrected k-space signal output in step S2 Construct a decoupling loss function, and iteratively optimize the INR parameters:

[0131] (7)

[0132] The decoupling loss function is:

[0133] (8)

[0134] in, It is the Frobenius norm. Regularization weights (take respectively) Displacement field output by dynamic branch It is obtained by calculation using formula (3).

[0135] To minimize this decoupling loss, the Adam optimizer can be used. Iteratively update the MLP parameters of the two branches:

[0136] During the optimization process, the static structure branch will gradually focus on "pure structures that are independent of motion," because Dynamic branches are modeled separately, while structural branches do not need to learn motion information;

[0137] The dynamic motion field branch will gradually focus on "motion perturbations that are independent of structure," because Static branches are modeled separately, while motion branches do not need to learn structural information.

[0138] After final convergence, the static branch outputs an artifact-free k-space. That is, a completely artifact-free pure k-space signal, the displacement field output by the dynamic branch. This is the root cause of motion artifacts, thus achieving precise decoupling between the two.

[0139] In one possible embodiment, before performing steps S1-S4, the present invention further includes a model construction and training process, mainly including:

[0140] We constructed a k-space physical model with motion artifacts, a k-ViT network, and a hidden neural representation network;

[0141] The k-space physical model containing motion artifacts is constructed by introducing a motion displacement field and a phase perturbation factor into the pure k-space signal model. The pure k-space signal model is obtained by Fourier transform based on the magnetization distribution of spin protons. The phase perturbation factor is generated by the coupling effect of the gradient field and the motion displacement field.

[0142] The k-ViT network includes an input preprocessing layer, a phase-aware multi-head self-attention module, and a multi-layer encoder. The input preprocessing layer is configured to perform adaptive block division and complex feature extraction of k-space signals, and the phase-aware multi-head self-attention module is configured to jointly model the global dependency between amplitude and phase.

[0143] The implicit neural representation network includes parallel static and dynamic branches, and the independence of the static and dynamic branches is constrained by a decoupling loss function. The static branch includes a multilayer perceptron, which takes spatial coordinates as input and static organizational structure as output. The dynamic branch includes another multilayer perceptron, which takes spatial coordinates and sampling time as input and dynamic motion field as output.

[0144] A self-supervised loss function is constructed based on k-space physical constraints and a decoupling loss function, and the k-ViT network and the implicit neural representation network are jointly trained based on the self-supervised loss function.

[0145] Understandably, this step is the core of the optimization loop in the entire MRI artifact correction framework. Its goal is to achieve accurate end-to-end model training by designing a loss function based on k-space physical properties and decoupling constraints, without clean MRI images (ground truth labels), thus completely solving the shortcomings of traditional methods that rely on paired clean labels and have poor generalization. Specifically, a total loss function is constructed based on three major constraints: k-space physical properties (conjugate symmetry), inherent signal regularity (sampling consistency), and decoupling accuracy (structure-motion separation). Through end-to-end optimization, the model automatically learns the mapping "from k-space signals containing artifacts to clean signals".

[0146] The process includes constructing a self-supervised loss function based on k-space physical constraints and a decoupling loss function, and jointly training the k-ViT network and the implicit neural representation network based on this self-supervised loss function. This mainly includes:

[0147] (1) Construct a multi-constraint self-supervised loss function

[0148] The total loss function consists of three terms, which respectively constrain signal consistency, physical rationality, and decoupling accuracy:

[0149] (9)

[0150] in, For self-supervised loss function, This is a symmetric loss based on the conjugate symmetry of k-space. The sampling consistency loss is based on the consistency of adjacent sampling points. Let be the decoupling loss function. , , These are the weighting coefficients corresponding to each loss term.

[0151] 1) Conjugate symmetric loss

[0152] Utilizing the k-space conjugate symmetry. One of the core physical properties of MRI imaging is the conjugate symmetry of k-space signals: for any k-space wave vector k, its corresponding signal... The signal corresponding to the wave vector −k Satisfying the "complex conjugate relation", that is:

[0153] , here To represent complex conjugation, the reconstructed signal is constrained to satisfy this physical property:

[0154] (10)

[0155] Conjugate symmetric loss The smaller the value, the more the reconstructed signal conforms to the conjugate symmetry characteristic.

[0156] 2) Sampling consistency loss

[0157] For MRI undersampling scenarios, consistency constraints between adjacent sampling points are introduced to reduce noise and artifact interference.

[0158] Let the number of sampling points be , and If the adjacent sampled wave vectors are defined as follows:

[0159] (11)

[0160] Sampling consistency loss The smaller the value, the more continuous the adjacent signals are, and the less noise and artifact interference there is.

[0161] 3) Decoupling loss

[0162] Decoupling loss Consistent with the decoupling loss function shown in equation (8) in the aforementioned embodiments, it is used to ensure the decoupling accuracy between the static structure and the dynamic motion field.

[0163] Loss weights in equation (9) Adaptable to Through the Adam optimizer (learning rate can be set to...) The attenuation coefficient can be set to Minimize the total loss to achieve end-to-end self-supervised training of the model.

[0164] (2) Model training process

[0165] The k-space physical model from step S1, the k-ViT network from step S2, the implicit neural representation network INR from step S3, and the constructed self-supervised loss function are integrated into a complete network. The input is a k-space signal containing artifacts, and the output is a corrected image. The entire process relies solely on the total loss function. Backpropagation updates all module parameters, eliminating the need for phased training.

[0166] 1) Preparation before training

[0167] Data input: The training set is "k-space signals with motion artifacts" (no need to pair clean signals), covering different MRI scenes (brain, chest, abdomen) and different motion patterns (breathing, heartbeat, limb swaying);

[0168] Parameter initialization: Weight matrices of the k-ViT network (4-layer encoder, phase-aware attention parameters) and the implicit neural representation network INR (6-layer MLP with static branch, 8-layer MLP with dynamic branch) (e.g.) , , ) and bias (e.g. , All are initialized using a random normal distribution;

[0169] Optimizer settings: Adam optimizer used, learning rate... (Balancing convergence speed and stability), decay coefficient 0.99 (to avoid excessive learning rate in the later stage leading to parameter oscillation).

[0170] 2) Forward propagation: from the artifact-containing signal to the corrected image

[0171] Input layer: k-space signal containing artifacts Input k-ViT network;

[0172] k-ViT preliminary correction: After preprocessing (normalization, non-uniform block partitioning), phase-aware attention, and a 4-layer encoder, the output is a preliminarily corrected k-space signal. ;

[0173] INR demodulation and reconstruction: converting the initially corrected k-space signal Spatial coordinates Sampling time The input is INR with two branches, and the output is a k-space signal without artifacts. With sports field ;

[0174] Image output: for artifact-free k-space signals Perform inverse Fourier transform To obtain the corrected image , is represented as:

[0175] (12)

[0176] 3) Backpropagation: Loss Calculation and Parameter Update

[0177] Loss calculation: based on the artifact-free k-space signal Preliminary correction of k-space signal Displacement field Calculate the loss item separately , , Then, sum them according to the weights to obtain the self-supervised loss function. The value;

[0178] Gradient backpropagation: Calculating the self-supervised loss function using the chain rule. For k-ViT (such as attention weights) ), INR (such as MLP weights) The gradient of all parameters;

[0179] Parameter Update: The Adam optimizer adjusts all parameters based on the gradient to minimize the self-supervised loss function. ;

[0180] Iterative convergence: Repeat the "forward propagation-backward propagation-parameter update" process (usually training for 100-200 epochs) until the self-supervised loss function is reached. Convergence (loss decrease over 10 consecutive epochs < Training ended.

[0181] The k-space signal containing motion artifacts to be corrected is input into the trained model, and steps S1 to S4 are executed sequentially to obtain the final corrected image. .

[0182] The trained model of this invention has the advantages of strong physical interpretability, high detail fidelity, no need for ground truth values ​​for training, and excellent generalization ability, and is suitable for various clinical MRI scenarios.

[0183] Figure 2 A structural diagram of an MRI motion artifact correction system based on physically guided k-ViT and implicit decoupling, provided for an embodiment of the present invention, is shown below. Figure 2 As shown, an MRI motion artifact correction system based on physically guided k-ViT and implicit decoupling includes a signal input module, a phase perturbation suppression module, a signal decoupling and reconstruction module, and an image generation module, wherein:

[0184] The signal input module is configured to decompose the phase perturbation component caused by motion from the k-space signal containing motion artifacts based on the k-space physical model.

[0185] The phase perturbation suppression module includes a trained k-ViT network, which is configured to suppress the motion-induced phase perturbation components decomposed from the k-space signal and output a pre-corrected k-space signal.

[0186] The signal decoupling and reconstruction module includes a trained implicit neural representation network, which is configured to decouple the pre-corrected k-space signal into a static structure and a dynamic motion field, and reconstruct an artifact-free k-space signal based on the static structure.

[0187] The image generation module is configured to perform an inverse Fourier transform on the artifact-free k-space signal to generate and output a corrected MRI image.

[0188] It is understood that the MRI motion artifact correction system based on physically guided k-ViT and implicit decoupling provided by this invention corresponds to the MRI motion artifact correction methods based on physically guided k-ViT and implicit decoupling provided in the foregoing embodiments. The relevant technical features of the MRI motion artifact correction system based on physically guided k-ViT and implicit decoupling can be referred to the relevant technical features of the MRI motion artifact correction methods based on physically guided k-ViT and implicit decoupling, and will not be repeated here.

[0189] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating an embodiment of the electronic device provided in this invention. For example... Figure 3As shown, this embodiment of the invention provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, it performs the following steps:

[0190] S1, based on the k-space physical model, decomposes the phase perturbation component caused by motion from the k-space signal containing motion artifacts;

[0191] S2, the phase perturbation component caused by motion is suppressed by the trained k-ViT network to obtain the preliminary corrected k-space signal;

[0192] S3, using a trained implicit neural representation network, the preliminary corrected k-space signal is decoupled into a static structure and a dynamic motion field, and an artifact-free k-space signal is reconstructed based on the static structure;

[0193] S4. Perform an inverse Fourier transform on the artifact-free k-space signal to obtain the corrected MRI image.

[0194] Please see Figure 4 , Figure 4 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided by the present invention. (See diagram below.) Figure 4 As shown, this embodiment provides a computer-readable storage medium 400 on which a computer program 311 is stored. When the computer program 311 is executed by a processor, it performs the following steps:

[0195] S1, based on the k-space physical model, decomposes the phase perturbation component caused by motion from the k-space signal containing motion artifacts;

[0196] S2, the phase perturbation component caused by motion is suppressed by the trained k-ViT network to obtain the preliminary corrected k-space signal;

[0197] S3, using a trained implicit neural representation network, the preliminary corrected k-space signal is decoupled into a static structure and a dynamic motion field, and an artifact-free k-space signal is reconstructed based on the static structure;

[0198] S4. Perform an inverse Fourier transform on the artifact-free k-space signal to obtain the corrected MRI image.

[0199] This invention provides a method, system, and storage medium for MRI motion artifact correction based on physically guided k-ViT and implicit decoupling.

[0200] First, based on the physical imaging principles of MRI, the phase perturbation component directly generated by motion is established and separated in the k-space signal. Then, using a k-ViT network specifically designed for processing complex signals in the frequency domain, the perturbation component is globally suppressed through its phase-aware attention calculation, resulting in a pre-corrected k-space signal. Next, an implicit neural representation network with parallel branches is used to decouple the pre-corrected k-space signal into a static tissue structure function and a dynamic motion field function, and a pure k-space signal is reconstructed based solely on the static structure. Finally, the final image is obtained through inverse Fourier transform. This invention, by performing physically guided decomposition, global frequency domain suppression, and signal component decoupling at the data source (k-space), can more accurately and thoroughly eliminate motion artifacts while preserving the true anatomical structure information to the greatest extent, thereby generating high-quality corrected MRI images.

[0201] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0202] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0203] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0204] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0205] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0206] In this embodiment of the invention, when collecting, processing, and storing user personal information (such as images, behavioral characteristics, etc.), the implementation of the technical solution strictly adheres to the principles of legality, legitimacy, and necessity, as well as the core rule of "notification-consent." Specifically, before information collection, the system clearly informs the user of the purpose, method, scope, and usage rules of information collection through an interactive interface, and requires the user's active authorization and consent. The entire information processing process employs data encryption, access control, and other technical measures to ensure information security, and establishes mechanisms to facilitate users' exercise of their rights (such as querying, correcting, withdrawing consent, and deleting information). For exceptions stipulated by law (such as those necessary for fulfilling statutory duties or responding to public health emergencies), their application is strictly limited to the scope and limits authorized by law, ensuring that the technical solution does not contain any content that violates the law, social morality, or harms the public interest.

[0207] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0208] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for correcting motion artifacts in MRI based on physically guided k-ViT and implicit decoupling, characterized in that, include: S1, based on the k-space physical model, decomposes the phase perturbation component caused by motion from the k-space signal containing motion artifacts; S2, the phase perturbation component caused by motion is suppressed by the trained k-ViT network to obtain the preliminary corrected k-space signal; S3, using a trained implicit neural representation network, the preliminary corrected k-space signal is decoupled into a static structure and a dynamic motion field, and an artifact-free k-space signal is reconstructed based on the static structure; S4. Perform an inverse Fourier transform on the artifact-free k-space signal to obtain the corrected MRI image.

2. The MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling according to claim 1, characterized in that, Step S1 includes: S101, Obtain a k-space physical model containing motion artifacts. The k-space physical model containing motion artifacts expresses the k-space signal containing motion artifacts as a quantitative mapping relationship between a pure structure signal, a motion displacement field, and a phase perturbation factor. S102, based on the k-space physical model containing motion artifacts, the k-space signal containing motion artifacts is decomposed into a tissue structure component after displacement distortion and a phase perturbation component caused by motion.

3. The MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling according to claim 1, characterized in that, Step S2 includes: S201, normalize the k-space signal containing motion artifacts and perform adaptive block processing based on the k-space radius, extract amplitude and phase features from each data block obtained by block division, and concatenate them into a complex feature vector; S202, the complex feature vector is input into the phase-aware multi-head self-attention module of the k-ViT network, and global features are obtained by jointly modeling the global dependency relationship between amplitude and phase; S203, the global features are processed by the encoder of the k-ViT network to suppress motion artifact-related features and enhance tissue structure features, and the preliminary corrected k-space signal is reconstructed through inverse block operation.

4. The MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling according to claim 3, characterized in that, In step S201, the normalization and adaptive block segmentation based on the k-space radius of the k-space signal containing motion artifacts includes: The k-space signal containing motion artifacts is normalized so that its amplitude range is mapped to the [0,1] interval; Calculate the radius value of each data point in the k-space, and divide the k-space into a central region and an edge region based on a preset radius threshold; The data in the central region is divided into blocks using a first-size block strategy, while the data in the edge region is divided into blocks using a second-size block strategy that is larger than the first size.

5. The MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling according to claim 3, characterized in that, Step S202 includes: S2021, the complex feature vector is input into the phase-aware multi-head self-attention module of the k-ViT network, and linear transformations are performed on the amplitude and phase features in the complex feature vector to generate a query, key, and value matrix of amplitude features. And phase feature queries, key matrix ; S2022, calculate the phase-aware attention score, which is composed of the dot product of the amplitude query and the key matrix, plus the dot product of the phase query and the key matrix after being weighted by a weighting factor; S2023, apply the phase-aware attention score to the amplitude value matrix, and output the global feature through multi-head stitching and linear projection.

6. The MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling according to claim 3, characterized in that, Step S203 includes: S2031, The global feature input is nonlinearly transformed by the encoder composed of a multi-layer feedforward network, wherein the feedforward network uses the GELU activation function; S2032, in the encoder, through multi-layer stacked nonlinear transformation, the feature response related to the phase perturbation factor is attenuated, and the feature response related to the static tissue structure is enhanced; S2033, perform an inverse block operation on the feature vector output by the encoder to reconstruct it into a preliminary corrected k-space signal with the same size as the k-space signal containing motion artifacts.

7. The MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling according to claim 1, characterized in that, Step S3 includes: S301, the spatial coordinates of the preliminarily corrected k-space signal are input into the static structure branch of the implicit neural representation network, and the intrinsic magnetization intensity distribution characterizing the static tissue structure is obtained by mapping through a multilayer perceptron; S302, the spatial coordinates and sampling time of the preliminary corrected k-space signal are input into the dynamic motion field branch of the implicit neural representation network, and the dynamic motion field representing the motion displacement is obtained by mapping through another multilayer perceptron. S303, based on the inherent magnetization distribution of the static structural branch output, an artifact-free k-space signal is reconstructed by integral calculation using the k-space physical model.

8. A method for MRI motion artifact correction based on physically guided k-ViT and implicit decoupling according to any one of claims 1 to 7, characterized in that, Before step S1, the following are also included: We constructed a k-space physical model with motion artifacts, a k-ViT network, and a hidden neural representation network; The k-space physical model containing motion artifacts is constructed by introducing a motion displacement field and a phase perturbation factor into the pure k-space signal model. The pure k-space signal model is obtained by Fourier transform based on the magnetization distribution of spin protons. The phase perturbation factor is generated by the coupling effect of the gradient field and the motion displacement field. The k-ViT network includes an input preprocessing layer, a phase-aware multi-head self-attention module, and a multi-layer encoder. The input preprocessing layer is configured to perform adaptive block division and complex feature extraction of k-space signals, and the phase-aware multi-head self-attention module is configured to jointly model the global dependency between amplitude and phase. The implicit neural representation network includes parallel static and dynamic branches, and the independence of the static and dynamic branches is constrained by a decoupling loss function. The static branch includes a multilayer perceptron, which takes spatial coordinates as input and static organizational structure as output. The dynamic branch includes another multilayer perceptron, which takes spatial coordinates and sampling time as input and dynamic motion field as output. A self-supervised loss function is constructed based on k-space physical constraints and a decoupling loss function, and the k-ViT network and the implicit neural representation network are jointly trained based on the self-supervised loss function.

9. The MRI motion artifact correction method based on physically guided k-ViT and implicit decoupling according to claim 8, characterized in that, The process involves constructing a self-supervised loss function based on k-space physical constraints and a decoupling loss function, and then jointly training the k-ViT network and the implicit neural representation network based on this self-supervised loss function, including: Construct a self-supervised loss function, expressed as: in, For self-supervised loss function, This is a symmetric loss based on the conjugate symmetry of k-space. The sampling consistency loss is based on the consistency of adjacent sampling points. Let be the decoupling loss function. , , These are the weighting coefficients corresponding to each loss term; With minimizing the self-supervised loss function as the optimization objective, the k-ViT network and the implicit neural representation network are jointly trained end-to-end. The network parameters are updated through an iterative optimization process until the self-supervised loss function converges.

10. An MRI motion artifact correction system based on physically guided k-ViT and implicit decoupling, characterized in that, include: The signal input module is configured to decompose the phase perturbation component caused by motion from the k-space signal containing motion artifacts based on the k-space physical model. The phase perturbation suppression module includes a trained k-ViT network, which is configured to suppress the motion-induced phase perturbation components decomposed from the k-space signal and output a pre-corrected k-space signal. The signal decoupling and reconstruction module includes a trained implicit neural representation network, which is configured to decouple the pre-corrected k-space signal into a static structure and a dynamic motion field, and reconstruct an artifact-free k-space signal based on the static structure. The image generation module is configured to perform an inverse Fourier transform on the artifact-free k-space signal to generate and output a corrected MRI image.

Images