Method for motion artifact removal reconstruction applicable to real scenes, system, and terminal

Abstract

The present invention discloses a method for motion artifact removal reconstruction applicable to real scenes, a system, and a terminal. The method comprises: acquiring raw k-space data, and inputting the raw k-space data into a score matching diffusion model for sampling with a preset number of steps, so as to obtain an incompletely reconstructed image containing motion artifacts; the score matching diffusion model continuously adding noise of different scales to the incompletely reconstructed image and performing sampling generation, after each generation of an image from noise, performing motion field estimation-based one-shot motion correction with the generated image as a first frame of the motion correction, and when the motion correction is completed, adding noise to the corrected image and performing a next sampling; and after all steps of reverse diffusion sampling are completed, performing motion correction on a sampled image generated in a last step, and outputting a high-quality fully reconstructed image without artifacts. The present invention realizes accelerated reconstruction and removal of motion artifacts.

Description

A method, system, and terminal for motion artifact reconstruction in real-world scenarios. Technical Field

[0001] This invention relates to the field of magnetic resonance imaging technology, and in particular to a method, system, terminal, and computer-readable storage medium for motion artifact reconstruction in real-world scenarios. Background Technology

[0002] During magnetic resonance imaging (MRI), the subject needs to be scanned multiple times to acquire data. To obtain accurate images, the subject is required to remain still for extended periods during the scan. In practical applications, this is often impractical. Even minor movements of the patient within the scanning chamber, as well as physiological activities such as breathing and heartbeat, can cause motion artifacts in the final image, reducing image quality and affecting the doctor's diagnostic results.

[0003] To further improve clinical imaging quality, improvements are needed in two aspects: one is reconstructing MRI images from undersampled data to accelerate the scanning acquisition process; the other is eliminating motion artifacts in the reconstructed images. Current motion artifact removal reconstruction methods mainly fall into two categories: the first is based on traditional optimization algorithms, but these require manual adjustment of hyperparameters, easily leading to suboptimal solutions and high computational complexity; the other utilizes deep learning methods, but most are based on supervised learning. Due to the difficulty in obtaining paired motion-free and damaged data, these methods typically use simulated motion artifact images to train the network. Therefore, it is difficult to apply them to real motion-damaged data, and most motion artifact removal methods cannot complete reconstruction tasks at high speedup levels.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for motion artifact reconstruction applicable to real-world scenarios. This aims to solve the problems that existing motion artifact reconstruction methods are difficult to apply to real motion-damaged data and cannot complete reconstruction tasks at high speeds.

[0006] To achieve the above objectives, the present invention provides a motion artifact removal reconstruction method suitable for real-world scenes, the method comprising the following steps:

[0007] Obtain the original k-space data, input the original k-space data into the fractional matching diffusion model for sampling at a preset number of steps, and obtain an incompletely reconstructed image containing motion artifacts;

[0008] The fractional matching diffusion model continuously adds noise of different scales to the incompletely reconstructed image and samples it to generate an image. After each image is generated from the noise, the generated image is used as the first frame for motion correction and one-shot motion correction based on motion field estimation is performed. After the motion correction is completed, noise is added to the corrected image and the next sampling is performed.

[0009] After performing all steps of inverse diffusion sampling, motion correction is applied to the sampled image generated in the last step to output a high-quality, fully reconstructed image without artifacts.

[0010] Furthermore, to achieve the above objectives, the present invention also provides a motion artifact removal reconstruction system suitable for real-world scenes, wherein the motion artifact removal reconstruction system suitable for real-world scenes includes:

[0011] The data sampling module is used to acquire the original k-space data, input the original k-space data into the fractional matching diffusion model for sampling at a preset number of steps, and obtain an incompletely reconstructed image containing motion artifacts.

[0012] The image correction module is used to continuously add noise of different scales to the incompletely reconstructed image by the fractional matching diffusion model, and to sample and generate an image each time. After generating an image from the noise, the generated image is used as the first frame for motion correction and one-shot motion correction based on motion field estimation is performed. After the motion correction is completed, noise is added to the corrected image and the next sampling is performed.

[0013] The image reconstruction module is used to perform motion correction on the sampled image generated in the last step after all steps of inverse diffusion sampling have been completed, and output a high-quality fully reconstructed image without artifacts.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a motion artifact removal reconstruction program for real-world scenes stored in the memory and executable on the processor, wherein when the motion artifact removal reconstruction program for real-world scenes is executed by the processor, it implements the steps of the motion artifact removal reconstruction method for real-world scenes as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a motion artifact removal reconstruction program suitable for real-world scenes, and when the motion artifact removal reconstruction program suitable for real-world scenes is executed by a processor, it implements the steps of the motion artifact removal reconstruction method suitable for real-world scenes as described above.

[0016] In this invention, raw k-space data is acquired and input into a fractional-matching diffusion model for a preset number of sampling steps, resulting in an incompletely reconstructed image containing motion artifacts. The fractional-matching diffusion model continuously adds noise of different scales to the incompletely reconstructed image, generating a new image each time. After each image is generated from the noise, one-shot motion correction based on motion field estimation is performed, using the generated image as the first frame for motion correction. Once motion correction is complete, noise is added to the corrected image, and the next sampling is performed. After completing all steps of inverse diffusion sampling, motion correction is performed on the sampled image generated in the last step, outputting a high-quality, fully reconstructed image without artifacts. This invention combines a fractional-matching diffusion model with an unsupervised motion correction model based on motion field estimation, coupling the unsupervised motion correction model into the sampling process of the diffusion model. The sampled image serves as a prior constraint for the unsupervised motion correction model, and sampling continues from the corrected image. The two are each other's priors, ultimately enabling the generation of high-quality magnetic resonance images without motion artifacts at a high speedup, achieving accelerated reconstruction and removal of motion artifacts. Attached Figure Description

[0017] Figure 1 is a flowchart of a preferred embodiment of the motion artifact reconstruction method of the present invention applicable to real-world scenarios;

[0018] Figure 2 is a schematic diagram of the forward perturbation process and the sampling process of the reverse coupling motion correction model in a preferred embodiment of the motion artifact reconstruction method applicable to real scenes of the present invention.

[0019] Figure 3 is a structural diagram of a preferred embodiment of the motion artifact removal and reconstruction system of the present invention applicable to real-world scenarios;

[0020] Figure 4 is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] During the MRI acquisition process, acquiring a single 2D image requires multiple scans of the patient. During these multiple scans, the patient's micro-movements and physiological activities can interfere with the data acquisition process, ultimately manifesting as motion artifacts in the reconstructed images, which can hinder the doctor's diagnosis.

[0023] To remove motion artifacts, a physical model is used to describe the impact of patient motion on the acquired data. By incorporating motion operators into the forward model of MRI, a generalized matrix equation is established to describe the process of generating an interference-affected image from an ideal image, improving model consistency and removing motion artifacts from the reconstructed image. However, optimizing a large number of motion parameters places a significant computational burden on this method. A single 3–5 s reconnaissance scan and optimized sequence sequencing avoid time-consuming joint or alternating optimization. The reconnaissance scan provides sufficient motion artifact-free k-space data (k-space data has important applications in MRI; k-space is the storage space for MRI signal data, and each MRI image has its corresponding k-space data; k-space is an abstract three-dimensional or two-dimensional space, and imaging data is arranged in specific k-space positions according to different spatial frequencies and finally transformed into an image), thus determining accurate motion parameters for each scan of high-resolution images. However, the efficiency of this method only holds true when correcting rigid body motion. Furthermore, this type of method based on traditional optimization cannot achieve motion artifact-free reconstruction at high acceleration magnification.

[0024] Some deep learning methods can directly map images affected by motion artifacts to clean images using end-to-end training strategies without considering specific motion parameters. However, the performance of these methods is highly dependent on the data pairs used for training. They are usually generated by simulating motion to create paired data. However, the distribution of motion in real scenes is not the same as that of simulated motion. Therefore, these methods are difficult to achieve good results on real data.

[0025] Recently, diffusion models based on score matching have achieved some success in image artifact removal. As a generative model, diffusion models are not limited by specific degradation models, making the solution of inverse problems more flexible. Many studies are exploring the combination of diffusion models and artifact removal methods, such as inserting the estimation of rigid body motion parameters into the sampling process, and alternately optimizing the sampling process and motion parameters to achieve reconstruction without motion artifacts.

[0026] Existing motion artifact removal reconstruction methods mainly fall into two categories: those based on traditional optimization algorithms and those based on deep learning. Traditional optimization methods require optimizing a large number of motion parameters, resulting in a huge computational burden. Therefore, they can only remove artifacts caused by rigid body motion, and traditional methods cannot achieve motion artifact removal reconstruction at high speeds. Supervised deep learning methods have high requirements for training data, making them impractical in real-world applications. Artifact removal methods combined with diffusion models are also only effective for rigid motion, limiting their application in real-world clinical scenarios. Therefore, this invention aims to establish a one-shot (single fast spin echo) dynamic reconstruction diffusion model coupled with motion field estimation. By combining unsupervised motion field estimation methods with diffusion models, this invention achieves a high-speed motion artifact removal reconstruction method applicable to both rigid and non-rigid motion.

[0027] The preferred embodiment of the motion artifact removal reconstruction method for real-world scenes according to the present invention, as shown in Figures 1 and 2, includes the following steps:

[0028] Step S10: Obtain the original k-space data, input the original k-space data into the fractional matching diffusion model for sampling at a preset number of steps, and obtain an incompletely reconstructed image containing motion artifacts;

[0029] Step S20: The fractional matching diffusion model continuously adds noise of different scales to the incompletely reconstructed image and performs sampling to generate an image. After each generation of an image from the noise, the generated image is used as the first frame for motion correction and one-shot motion correction based on motion field estimation is performed. After the motion correction is completed, noise is added to the corrected image and the next sampling is performed.

[0030] Step S30: After completing all steps of inverse diffusion sampling, perform motion correction on the sampled image generated in the last step to output a high-quality, fully reconstructed image without artifacts.

[0031] Specifically, this invention aims to establish a one-shot dynamic reconstruction diffusion model coupled with motion field estimation. It combines unsupervised motion field estimation methods with a diffusion model to achieve a significantly faster motion artifact reconstruction method applicable to both rigid and non-rigid motion. This method mainly consists of two parts: the first part involves perturbing a clean magnetic resonance image using a Gaussian perturbation kernel (referred to as the forward process), transforming the difficult-to-learn data distribution into a noise distribution under conditional probability. Then, the gradient of the logarithm of the data distribution is learned through score matching, and the network is used to represent the conditional distribution information of the data. The second part is the sampling generation process (referred to as the backward process). In the forward pass, the conditional distribution information learned by the network is used as a priori to conditionally guide the Langevin dynamics Markov Monte Carlo sampling process to generate the reconstructed image. To remove motion artifacts, a one-shot motion correction model based on motion field estimation is coupled in this process. After sampling the image at time t to generate the image at time t-1, the motion field is generated using a neural network to perform motion correction on the sampled image, changing the conditional distribution of the sampled image to move it towards the distribution of the image without motion artifacts. Then, the Monte Carlo sampling process continues to generate the image of the previous time step. Sampling and motion correction are performed alternately, and finally a clean image without motion artifacts is generated.

[0032] The fractional matching diffusion model of the present invention mainly consists of two parts: fractional matching and denoising probability generation.

[0033] Define a neural network s with parameter θ. θ Neural networks θ It is the gradient of the logarithm of the probability density function p(x), that is, an approximation of the fractional function. in, The gradient of Langevin dynamics for sampled data x is represented by:

[0034] The above process is also known as the reverse process, where x i+1 x represents the sampled data generated for the next iteration step. i It is the sampled data of the current iteration step, g i The step size represents the standard Wiener process during iterative sampling. Represents a standard Gaussian distribution. It means that for x i The gradient of p i (x i ) represents the current iteration step x i The probability density function.

[0035] Solve directly in formula (1) It is difficult, therefore a neural network needs to be trained to approximate it. Training a neural network s θ To approximate the fractional function, the process of denoising and matching the fractional data involves perturbing the data using Gaussian perturbation kernels with varying variances, and then using a neural network. θ To approximate the logarithmic gradient of the perturbated data, the approximation form is:

[0036] Where, x t |x0 represents the perturbation data after adding the perturbation kernel, t represents the diffusion time step, x t Let s represent a random variable at time t. θ (x t Let ,t) represent the neural network used to approximate the fractional function at time step t, where θ represents the parameters of the neural network. This means that, given x0, for the variable x t The expected value of the condition is obtained. This represents the expected value obtained with respect to the variable x0. This represents the expected value obtained with respect to variable t. It means that for x t The gradient of p(x) t |x0) represents x t The conditional probability density distribution at x0, ||·|| l2 This represents the l2 norm, which is the square root of the square.

[0037] If we consider a series of noise scales σ min =σ1<σ2<…<σ L =σ max Noise is subject to Gaussian noise, where, This indicates that the mean is 0 and the standard deviation is σ. t Gaussian distribution, σ t σ represents the standard deviation of Gaussian noise. t The larger the value, the more severe the noise. During the diffusion process, the data needs to be perturbed with noise. The noise perturbation at time step t follows a certain pattern. σ min This indicates a pre-defined lower bound for the noise scale, which is relatively small, σ max This represents a pre-defined upper bound on the scale, which is relatively large, σ. L The standard deviation of the noise in the final step is equal to σ. max .

[0038] Treating the disturbance process as a continuous diffusion process, after a certain time T, the data... It is approximately standard Gaussian noise, where p(x) L|x0) represents x L The conditional probability density distribution under x0, where x0 represents the original data, x L This represents the spread data at step L, a process also known as the forward spread process, where s is learned. θ (x t ,t) is the conditional probability density function p(x t The approximation of the gradient of the logarithm of |x0) contains information about the conditional distribution.

[0039] Untrained Neural Networks (UNNs) are a method for image processing tasks that utilizes an untrained neural network structure as a prior. This reveals that the network structure itself is a prior, enabling the network to optimize based on only a single image without extensive training on large datasets, generating high-quality images from random variables. Furthermore, UNNs in the image domain, with their decoder-based structures, allow the network to extract information from low-dimensional latent variables, reconstructing high-dimensional images / information. This invention uses this method to generate corrected images and motion fields from random low-dimensional vectors.

[0040] This invention first utilizes a deep network Φ1 with a decoder structure to generate a baseline image from a random vector z0, which is also the first frame image x1 corresponding to the acquisition process. This setup is reasonable considering that subjects are more likely to remain still in the first few seconds after entering the scanning chamber. Then, deep networks Φ2 to Φ... n Generate sports field arrive Where n represents the total number of frames acquired, the motion field is used to deform the first frame image x1 to generate the remaining frame images x2 to x3. n This process can be represented as: x1=Φ1(z0); (3)

[0041] Where ° indicates the use of a sports field Register x1, and denote x as the distance from x1 to x2. n For ease of representation, the process of formulas (3) to (5) is expressed by the following formula: x=Φ(z0); (6)

[0042] Combining the coil sensitivity with Fourier transform of the generated image data x, multi-channel k-space data is obtained. The sampled data b from each frame is used for backfilling to obtain the k-space data y:

[0043] in, Let represent the multi-channel k-space data corresponding to the generated image data x, b represent the original measurement data, M represent the sampling mask of each frame, and y represent the self-consistency loss and the loss of x. i , The regularization constraints together constitute the loss function:

[0044] Where I represents the Id function, F and F -1 Let represent the Fourier and inverse Fourier transforms respectively, C represent the coil sensitivity mapping, and Conj(C) represent the conjugate of C. TV λ represents the TV norm, and λ1 and λ2 represent the penalty term coefficients, which are pre-defined constants; backpropagation is performed based on the loss function to optimize the weights of each depth network.

[0045] The motion correction process is shown in the table below:

[0046] Where Φ = bp(LOSS) represents updating {Φ through backpropagation based on LOSS. i The network weights are defined as follows. The final output x1 is the first frame image in the last updated array x.

[0047] In practice, assuming there is already From the posterior p(x) t |x0) is solved by performing the Langevin dynamics Markov Monte Carlo sampling method, which recovers the fully sampled image:

[0048] Where ∈ represents standard Gaussian noise, whose dimension is consistent with the dimension of the input sample, as shown in formula (9). Since the prior knowledge is usually unknown, this is why the network s is used through score matching, as mentioned earlier. θ (x t The distribution of data is approximated by a diffusion model (x,t), which transforms the data distribution into network parameters, thus indirectly obtaining the conditional distribution of the data. However, the sampled data used to achieve data consistency during the sampling process is subject to motion interference. If only the diffusion model is used, the final obtained image will be motion-disturbed.

[0049] Therefore, this invention measures x during the sampling process. t Motion correction is performed so that the sampled image at time t changes from x t |x′0 transforms into x t |x0, where x0 is a clean image without motion interference.

[0050] Specifically, after a preset number of steps in the diffusion sampling process, the following three processes are repeated: x′ t-1 =MotionRemove(x t-1 (11)

[0051] Among them, MotionRemove(x t-1 ) represents x t-1 For the prior motion correction process, x t-1 Indicates from The image generated by mid-sampling, x′ t-1 Indicates x t-1 The motion-corrected image, Represents noisy image data, δ t This represents the noise variance at a pre-defined time t.

[0052] In order to use the results of diffusion sampling as a priori for motion correction, the loss function expressed in Equation (8) is changed to:

[0053] Wherein, λ3 represents the penalty term coefficient, which is a pre-set constant.

[0054] By iteratively executing formulas (10) and (12), a high-quality, fully reconstructed image without artifacts can be obtained. (i.e., high-quality reconstructed magnetic resonance images without motion interference).

[0055] As shown in Figure 2, It is the standard Wiener process. Let x represent a standard Gaussian distribution. T It is the noisy image at time T at the end of the forward process. It is a noisy image that has undergone motion correction during the reverse diffusion process.

[0056] The innovative aspects of this invention are as follows:

[0057] (1) This invention combines motion field estimation with diffusion model for the first time, and realizes motion artifact reconstruction under high acceleration factor.

[0058] (2) The method proposed in this invention has greater flexibility and generalization. Unlike traditional deep learning methods, which require retraining when changing datasets and sampling templates, the proposed method does not require retraining. It only needs to perform acquisition coupled with motion field estimation.

[0059] (3) The method proposed in this invention is applicable to any form of motion artifact, including rigid body motion and non-rigid body motion.

[0060] This invention provides a high-performance motion artifact reconstruction method. Compared with traditional optimization methods, it is more efficient in estimating motion parameters and can achieve reconstruction at high acceleration. Compared with some previous depth methods, the method proposed in this invention is not limited to correcting rigid body motion, but also has good results for non-rigid body motion.

[0061] Furthermore, as shown in Figure 3, based on the above-described motion artifact removal reconstruction method suitable for real-world scenes, the present invention also provides a motion artifact removal reconstruction system suitable for real-world scenes, wherein the motion artifact removal reconstruction system suitable for real-world scenes includes:

[0062] Data sampling module 51 is used to acquire raw k-space data, input the raw k-space data into the fractional matching diffusion model for sampling at a preset number of steps, and obtain an incompletely reconstructed image containing motion artifacts;

[0063] The image correction module 52 is used to continuously add noise of different scales to the incompletely reconstructed image by the fractional matching diffusion model, and to sample and generate an image each time. After generating an image from the noise, the generated image is used as the first frame for motion correction and one-shot motion correction based on motion field estimation is performed. After the motion correction is completed, noise is added to the corrected image and the next sampling is performed.

[0064] The image reconstruction module 53 is used to perform motion correction on the sampled image generated in the last step after all steps of inverse diffusion sampling have been completed, and output a high-quality fully reconstructed image without artifacts.

[0065] Furthermore, as shown in Figure 4, based on the above-described motion artifact reconstruction method and system applicable to real-world scenarios, the present invention also provides a terminal, which includes a processor 10, a memory 20, and a display 30. Figure 4 only shows some components of the terminal; however, it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0066] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a motion artifact reconstruction program 40 suitable for real-world scenes, which can be executed by the processor 10 to implement the motion artifact reconstruction method suitable for real-world scenes in this application.

[0067] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the motion artifact reconstruction method applicable to real-world scenarios.

[0068] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The terminal's processor 10, memory 20, and display 30 communicate with each other via a system bus.

[0069] In one embodiment, when the processor 10 executes the motion artifact reconstruction program 40 for a real scene stored in the memory 20, it implements the steps of the motion artifact reconstruction method for a real scene as described above.

[0070] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a motion artifact removal reconstruction program suitable for real-world scenes, and the motion artifact removal reconstruction program suitable for real-world scenes, when executed by a processor, implements the steps of the motion artifact removal reconstruction method suitable for real-world scenes as described above.

[0071] In summary, this invention provides a method, system, terminal, and computer-readable storage medium for motion artifact removal reconstruction suitable for real-world scenarios. The method includes: acquiring raw k-space data; inputting the raw k-space data into a fractional matching diffusion model for sampling at a preset number of steps to obtain an incompletely reconstructed image containing motion artifacts; the fractional matching diffusion model continuously adds noise of different scales to the incompletely reconstructed image and performs sampling to generate an image; after each image is generated from the noise, one-shot motion correction based on motion field estimation is performed using the generated image as the first frame for motion correction; after motion correction is completed, noise is added to the corrected image, and the next sampling is performed; after performing all steps of inverse diffusion sampling, motion correction is performed on the sampled image generated in the last step to output a high-quality, fully reconstructed image without artifacts. This invention combines a fraction-matching-based diffusion model with an unsupervised motion correction model based on motion field estimation. The unsupervised motion correction model is coupled into the sampling process of the diffusion model. The sampled image is used as a prior constraint for the unsupervised motion correction model. Then, sampling continues from the corrected image. The two are each other's priors. Finally, high-quality magnetic resonance images with motion artifacts removed can be generated at a high speedup, achieving accelerated reconstruction and removal of motion artifacts.

[0072] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0073] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0074] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for motion artifact reconstruction suitable for real-world scenes, characterized in that, The motion artifact reconstruction method applicable to real-world scenarios includes: Obtain the original k-space data, input the original k-space data into the fractional matching diffusion model for sampling at a preset number of steps, and obtain an incompletely reconstructed image containing motion artifacts; The fractional matching diffusion model continuously adds noise of different scales to the incompletely reconstructed image and samples it to generate an image. After each image is generated from the noise, the generated image is used as the first frame for motion correction and one-shot motion correction based on motion field estimation is performed. After the motion correction is completed, noise is added to the corrected image and the next sampling is performed. After performing all steps of inverse diffusion sampling, motion correction is applied to the sampled image generated in the last step to output a high-quality, fully reconstructed image without artifacts.

2. The motion artifact removal and reconstruction method for real-world scenes according to claim 1, characterized in that, The score matching diffusion model is used for score matching and denoising probability generation.

3. The motion artifact removal and reconstruction method for real-world scenes according to claim 2, characterized in that, The process of acquiring the original k-space data and inputting it into a fractional matching diffusion model for sampling at a preset number of steps to obtain an incompletely reconstructed image containing motion artifacts specifically includes: Define a neural network s with parameter θ. θ Neural networks θ It is the gradient with respect to the logarithm of the probability density function p(x), and the fractional function is... in, The gradient of the sampled data x, expressed by Langevin dynamics Markov Monte Carlo sampling, is: Where, x i+1 x represents the sampled data generated for the next iteration step. i It is the sampled data of the current iteration step, g i The step size represents the standard Wiener process during iterative sampling. Represents a standard Gaussian distribution. It means that for x i The gradient of p i (x i ) represents the current iteration step x i The probability density function; Training a neural network s θ To approximate the fractional function, the process of denoising and matching the fractions involves perturbing the data using Gaussian perturbation kernels with varying variances, and then using a neural network. θ To approximate the logarithmic gradient of the perturbated data, the approximation form is: Where, x t |x0 represents the perturbation data after adding the perturbation kernel, t represents the diffusion time step, x t Let s represent a random variable at time t. θ (x t Let ,t) represent the neural network used to approximate the fractional function at time step t, where θ represents the parameters of the neural network. This means that, given x0, for the variable x t The expected value of the condition is obtained. This represents the expected value obtained with respect to the variable x0. This represents the expected value obtained with respect to variable t. It means that for x t The gradient of p(x) t |x0) represents x t The conditional probability density distribution at x0, ||·|| l2 This represents the l2 norm.

4. The motion artifact removal and reconstruction method for real-world scenes according to claim 3, characterized in that, The process of acquiring the original k-space data, inputting the original k-space data into a fractional matching diffusion model for sampling at a preset number of steps to obtain an incompletely reconstructed image containing motion artifacts, and then further includes: If we consider a series of noise scales σ min =σ1<σ2<…<σ L =σ max Noise is subject to Gaussian noise, where, This indicates that the mean is 0 and the standard deviation is σ. t Gaussian distribution, σ t The standard deviation of Gaussian noise is represented by the noise perturbation data during the diffusion process. The noise perturbation at time step t follows a certain pattern. σ min σ represents the pre-defined lower bound of the noise scale. max This represents a pre-defined upper bound on the scale, σ. L The standard deviation of the noise in the final step is equal to σ. max ; Treating the disturbance process as a continuous diffusion process, after a certain time T, the data... It is approximately standard Gaussian noise, where p(x) L |x0) represents x L The conditional probability density distribution under x0, where x0 represents the original data, x L This represents the diffusion data at step L, and the s learned during the forward diffusion process. θ (x t ,t) is the conditional probability density function p(x t An approximation of the gradient of the logarithm of |x0).

5. The motion artifact removal and reconstruction method for real-world scenes according to claim 4, characterized in that, The fractional matching diffusion model continuously adds noise of different scales to the incompletely reconstructed image, samples and generates an image. After each image is generated from the noise, one-shot motion correction based on motion field estimation is performed using the generated image as the first frame for motion correction. After motion correction is completed, noise is added to the corrected image, and the next sampling is performed. Specifically, this includes: A reference image is generated from a random vector z0 using a deep network Φ1 with a decoder structure, representing the first frame image x1 during the acquisition process. Then, a deep network Φ2 with a decoder structure is used to generate a reference image from the random vector z0. n Generate sports field arrive Where n represents the total number of frames acquired, the motion field is used to deform the first frame image x1 to generate the remaining frame images x2 to x3. n The process is as follows: x1=Φ1(z0);(3) in, Indicates use of sports field Register x1, and denote x as the distance from x1 to x2. n The union: x = Φ(z0); (6) Combining the coil sensitivity with Fourier transform of the generated image data x, multi-channel k-space data is obtained. The sampled data b from each frame is used for backfilling to obtain the k-space data y: in, Let represent the multi-channel k-space data corresponding to the generated image data x, b represent the original measurement data, M represent the sampling mask of each frame, and y represent the self-consistency loss and the loss of x. i , The regularization constraints together constitute the loss function: Where I represents the Id function, F and F -1 Let represent the Fourier and inverse Fourier transforms respectively, C represent the coil sensitivity mapping, and Conj(C) represent the conjugate of C. TV Let λt denote the TV norm, and λ1 and λ2 denote the coefficients of the penalty term; Backpropagation is performed based on the loss function to optimize the weights of each depth network.

6. The motion artifact removal and reconstruction method for real-world scenes according to claim 5, characterized in that, After performing all steps of inverse diffusion sampling, motion correction is applied to the sampled image generated in the last step to output a high-quality, fully reconstructed image without artifacts. Specifically, this includes: If already From the posterior p(x) t |x0) is solved by performing the Langevin dynamics Markov Monte Carlo sampling method to recover the fully sampled image: Where ε represents standard Gaussian noise, and the network s is used through a fractional matching method. θ (x t To approximate the data distribution, we transform the data distribution into network parameters, thereby indirectly obtaining the conditional distribution of the data. During the sampling process, x t Motion correction is performed so that the sampled image at time t changes from x t |x′0 transforms into x t |x0, x0 is a clean image without motion interference; After a preset number of steps in the diffusion sampling process, the following three processes are repeated: x′ t-1 =MotionRemove(x t-1 );(11) Among them, MotionRemove(x t-1 ) represents x t-1 For the prior motion correction process, x t-1 Indicates from The image generated by mid-sampling, x′ t-1 Indicates x t-1 Motion-corrected image, Represents noisy image data, δ t This represents the noise variance at a pre-defined time t; Using the results of diffusion sampling as a priori for motion correction, the loss function expressed in equation (8) is changed to: Where λ3 represents the coefficient of the penalty term; By iteratively executing formulas (10) and (12), a high-quality, fully reconstructed image without artifacts is obtained.

7. The method for motion artifact removal and reconstruction suitable for real-world scenes according to claim 1, characterized in that, The motion artifact reconstruction method applicable to real-world scenarios includes a forward process and a backward process; the forward process is used to perturb a clean magnetic resonance image using a Gaussian perturbation kernel; the backward process is used for sampling and generation.

8. A motion artifact removal reconstruction system suitable for real-world scenes, characterized in that, The motion artifact reconstruction system suitable for real-world scenarios includes: The data sampling module is used to acquire the original k-space data, input the original k-space data into the fractional matching diffusion model for sampling at a preset number of steps, and obtain an incompletely reconstructed image containing motion artifacts. The image correction module is used to continuously add noise of different scales to the incompletely reconstructed image by the fractional matching diffusion model, and to sample and generate an image each time. After generating an image from the noise, the generated image is used as the first frame for motion correction and one-shot motion correction based on motion field estimation is performed. After the motion correction is completed, noise is added to the corrected image and the next sampling is performed. The image reconstruction module is used to perform motion correction on the sampled image generated in the last step after all steps of inverse diffusion sampling have been completed, and output a high-quality fully reconstructed image without artifacts.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a motion artifact removal reconstruction program for real-world scenes stored in the memory and executable on the processor. When the motion artifact removal reconstruction program for real-world scenes is executed by the processor, it implements the steps of the motion artifact removal reconstruction method for real-world scenes as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a motion artifact removal reconstruction program suitable for real-world scenes, which, when executed by a processor, implements the steps of the motion artifact removal reconstruction method suitable for real-world scenes as described in any one of claims 1-7.

Images