A zero-shot dynamic MRI reconstruction method based on global and local diffusion models

Abstract

The present invention provides a zero-sample dynamic MRI reconstruction method based on global and local diffusion models, comprising the following steps: generating an image data set based on original two-dimensional dynamic full-sampling cardiac magnetic resonance multi-coil data; designing global and local time-interleaved sampling strategies to construct two types of training data sets for generating network models using a global and local cascade learning network with a two-stage learning mechanism; using an iterative solution algorithm that quickly extracts prior information of the network model to alternately predict, correct, constrain low rank, and ensure data consistency; applying the iterative solution algorithm to the trained network model and testing it, outputting a dynamic magnetic resonance reconstruction image, and evaluating the image quality. The present invention constructs fully encoded data through different time-interleaved sampling strategies, combines the image prior information of the network model, and realizes fast and accurate reconstruction of images under undersampling conditions, significantly improving the quality and efficiency of the reconstructed images.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a zero-sample dynamic MRI reconstruction method based on global and local diffusion models. Background Art

[0002] Dynamic MRI is an advanced medical imaging technology that uses the principle of MRI (Magnetic Resonance Imaging) to obtain a series of magnetic resonance images reflecting physiological processes in a short period of time. Unlike static MRI, dynamic MRI captures dynamic processes such as heartbeat, breathing and blood flow in real time through multiple rapid imaging. It is widely used in medical diagnosis of the heart, brain, tumors and other parts. Its advantage is that it can provide temporal and spatial information of patients' physiological activities, providing richer and more detailed image data for clinical diagnosis.

[0003] However, the inherent physical characteristics and physiological limitations of the dynamic MRI acquisition process restrict the acquisition speed, and the extended scanning time makes the imaging of dynamic structures more complicated. In order to solve this problem, dynamic MRI reconstruction technology came into being. Dynamic MRI reconstruction technology aims to restore high-quality images from undersampled data. Although early traditional image reconstruction algorithms, such as parallel imaging and compressed sensing, have achieved certain results, they can reduce imaging time and increase image acquisition speed while maintaining image quality. However, traditional image reconstruction algorithms usually assume that the signal distribution in the image has sparse or periodic characteristics, while dynamic MRI reconstruction requires effective processing of temporal correlation. In this case, the limitations of traditional image reconstruction algorithms gradually become apparent, and it is difficult to fully capture the complex dynamic changes in physiological processes, especially when the imaging time is shortened and the amount of data is reduced, the image quality is easily affected. Summary of the invention

[0004] The purpose of the present invention is to propose a practical, high-performance, and environmentally adaptable zero-sample dynamic MRI reconstruction method based on global and local diffusion models. The method does not require a large number of fully sampled dynamic magnetic resonance image samples. For the reconstruction of missing parts of under-sampled images, it can not only effectively fill and repair, but also retain the global features and local details of the image, obtain clear image edges and textures, and improve the image quality.

[0005] To achieve the above object, the technical solution provided by the present invention is as follows: A zero-sample dynamic MRI reconstruction method based on global and local diffusion models comprises the following steps:

[0006] Step A: generating an image data set based on the original two-dimensional dynamic full-sampling cardiac magnetic resonance multi-coil data, wherein the image data set is divided into an image training data set and an image test data set;

[0007] Step B: Based on the image training dataset, a global and local time-interleaved sampling strategy is designed to construct two fully encoded training datasets containing global information and local information respectively, which are used to generate a network model using a global and local cascade learning network with a two-stage learning mechanism;

[0008] Step C: Use an iterative solution algorithm that quickly extracts prior information from the network model to alternately predict and correct the undersampled data, and use a multi-level iterative optimization strategy to improve image quality. In each iteration, low-rank constraints and data consistency optimization strategies are used to correct outliers introduced by noise or prediction errors.

[0009] Step D: Apply the iterative solution algorithm to the trained network model and test the image test data set, output the final dynamic magnetic resonance reconstruction image, and test the peak signal-to-noise ratio value, structural similarity and mean square error value of the dynamic magnetic resonance reconstruction image to evaluate the image quality.

[0010] Furthermore, the step A specifically includes:

[0011] Using an adaptive coil combination method to generate two-dimensional dynamic full-sampling cardiac magnetic resonance single-coil data from original two-dimensional dynamic full-sampling cardiac magnetic resonance multi-coil data;

[0012] Then, different time-interleaved sampling strategies were implemented on the two-dimensional dynamic full-sampling cardiac magnetic resonance single-coil data used for network model training to obtain an image training data set;

[0013] The two-dimensional dynamic fully sampled cardiac magnetic resonance single coil data used for testing were undersampled using radial and Cartesian masks respectively to obtain the image testing dataset.

[0014] Furthermore, the step B specifically includes:

[0015] (1) Design a global and local time-staggered sampling strategy

[0016] The time interleaved sampling strategy is set to global and local. According to the different acceleration factors R set by the global and local time interleaved sampling strategies, R adjacent undersampled frames in the time dimension are merged into one frame, and then the central phase encoding of the merged frames is used for fidelity, forming two fully encoded training data sets containing global information and local information respectively, namely the global time interleaved sampling data set and the local time interleaved sampling data set;

[0017] (2) Global and local cascade learning networks

[0018] ① Two-stage learning mechanism:

[0019] Two fully encoded training data sets containing global information and local information respectively processed by global and local time interleaved sampling strategies are input into a two-stage training network; in the first stage, the global diffusion model is trained using the global time interleaved sampling data set; in the second stage, the local diffusion model is trained using the local time interleaved sampling data set;

[0020] ② The learning process of prior information of global and local diffusion models:

[0021] According to the different characteristics of the data sets obtained using the global and local time-interleaved sampling strategies, the diffusion model is used in the cascade network to learn the prior information. The global diffusion model and the local diffusion model use a formalized diffusion process to gradually transform the sample data in the K space into a standard prior distribution. The diffusion process is represented by the following stochastic differential equation:

[0022] ,

[0023] in, is a standard Wiener process. is the drift coefficient function, is the diffusion coefficient function;

[0024] The diffusion process is further optimized in the form of VE-SDE. The forward process of VE-SDE is used to transform the undersampled data into a prior distribution that is easy to handle. Its form is:

[0025] ,

[0026] in, is a Gaussian noise function in continuous time, and the continuous time variable ;

[0027] Through this process, the global diffusion model and the local diffusion model can gradually introduce noise in the K-space domain and map the original data distribution to a standard Gaussian distribution;

[0028] Once the forward diffusion is completed, the global diffusion model and the local diffusion model begin the reverse operation, transforming the prior distribution back to the original data distribution in K space through the reverse VE-SDE, and its mathematical form is:

[0029] ,

[0030] in, It's time from Backflow to The standard Wiener process when is an infinitesimal negative time step, is t The distribution of time, For each time The edge distribution gradient of

[0031] The global diffusion model and the local diffusion model predict the noise distribution of the undersampled data by learning the estimated score function of the noise, and then restore it to a state close to the real fully sampled data;

[0032] During the optimization process, the global diffusion model and the local diffusion model will dynamically adjust the weights of the loss function according to the time step to ensure that global and local information can be fully learned at different stages.

[0033] Further, the step C comprises:

[0034] Alternately perform prediction, correction, low-rank constraint, and data consistency steps;

[0035] In each iteration, the current undersampled data is updated using a predictor based on an inverse stochastic differential equation solver that estimates the current state by discretizing time steps. , the predictor will update the next state according to the following equation :

[0036] ,

[0037] in, is a drift term used to estimate the deterministic change in the data, is the change in noise variance, represents the injected noise, is the diffusion coefficient function, is standard Gaussian noise;

[0038] The corrector corrects the prediction results. The correction step uses improved annealing Langevin dynamics and introduces noise to simulate the randomness of the diffusion process. The correction formula is:

[0039] ,

[0040] in, is the estimated noise scoring function, is the correction step size;

[0041] After each iteration, a low-rank constraint is introduced for optimization. The low-rank constraint extracts the low-rank characteristics in the data by constructing the Hankel matrix of the K space and performing singular value decomposition on it. The low-rank constraint optimization problem is expressed as:

[0042] ,

[0043] in, is a linear operator, is the reconstruction matrix, is the sampled data, is the Hankel pseudo-inverse operator, is a low-rank matrix after hard threshold singular value decomposition, is the matrix Rank operation, is a matrix rank, are constraints that need to be satisfied;

[0044] The sample value is restored using data consistency, and data consistency is expressed as:

[0045] ,

[0046] in, represents the index set of acquired K-space samples, Represents the frequency domain position and time point The reconstruction data, is a tuning parameter used to control the raw data and update value The weight ratio of .

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The present invention proposes a novel global and local time-interleaved sampling strategy, and designs a dynamic magnetic resonance image reconstruction network architecture based on a diffusion model. In particular, the present invention does not need to use fully sampled data for training. The present invention uses a global and local time-interleaved sampling strategy to accelerate the acquisition of training data, thereby making full use of the time redundancy inherent in the data set. At the same time, a two-stage learning mechanism is adopted to ensure that the diffusion model can fully learn global and local information. In the reconstruction process, through the alternating execution of prediction and correction, coupled with the joint action of low-rank constraints and data consistency optimization strategies, the global diffusion model and the local diffusion model can gradually extract and optimize prior information in each iteration. Zero-shot learning not only reduces the dependence on large-scale data sets, but also avoids overfitting caused by insufficient samples, and has a significant improvement in generalization ability. Finally, it is achieved in data-scarce scenarios, under undersampling conditions, fast and accurate reconstruction of images, and significantly improves the quality and efficiency of reconstructed images. A large number of experimental results show that the present invention can produce the best or equivalent effects in both qualitative and quantitative aspects. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1It is a flow chart of a zero-sample dynamic MRI reconstruction method based on global and local diffusion models of the present invention;

[0050] Figure 2 It is a network model framework diagram of zero-sample dynamic MRI reconstruction based on global and local diffusion models of the present invention;

[0051] Figure 3 It is the reconstruction image of the present invention under the condition of acceleration factor R=8 and radial undersampling, wherein (a) is the original image, (b)(c)(d)(e)(f)(g) are the reconstruction result images of BCS, 3D-CSC, Tgvnn, DLTG, DD-UGM and the present invention respectively;

[0052] Figure 4 It is the reconstructed image of the present invention under the condition of acceleration factor R=10 and Radial undersampling, wherein (a) is the original image, (b)(c)(d)(e)(f)(g) are the reconstruction result images of BCS, 3D-CSC, Tgvnn, DLTG, DD-UGM and the present invention respectively. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The specific embodiments described here are only used to explain the technical solution of the present invention and are not limited to the present invention.

[0054] At present, existing methods use more excellent preprocessing methods to extract prior information of images under the framework of network models to reconstruct better results. Inspired by these good performances, the purpose of the present invention is to provide a zero-sample dynamic MRI reconstruction method based on global and local diffusion models: two types of training data sets containing global information and local information are constructed for network training through different time-staggered sampling strategies, and the image prior information in the diffusion process is combined. The global diffusion model is used to capture macroscopic structural information, and the local diffusion model is used to process complex details. Multiple network structures are optimized under different noise levels to form a global and local joint diffusion mechanism that adapts to different imaging conditions. The optimization strategy of adaptive diffusion step size and the refined iterative generation network are adopted to achieve accurate and efficient reconstruction of dynamic magnetic resonance images under zero-sample conditions.

[0055] Specifically, the present invention uses a diffusion model based on deep learning for medical image reconstruction. The basic idea of ​​the diffusion model is to gradually eliminate the uncertainty caused by noise and undersampling by simulating the diffusion process of data, thereby restoring high-quality images. The advantage of the diffusion model is that it can effectively restore the details and overall contours of the image by relying on the inherent geometric structure and random process when the data is incomplete. Especially in dynamic magnetic resonance imaging, the diffusion model can better balance the requirements of space and time, while maintaining spatial resolution, enhancing temporal resolution, and providing more accurate dynamic information.

[0056] In addition, due to the scarcity of dynamic magnetic resonance samples, in the field of medical imaging, it is often very difficult and costly to obtain a large amount of high-quality annotated data. Especially for some rare pathological types or special patient groups, the lack of data further increases the difficulty of image reconstruction. In order to solve the problem of data scarcity, the present invention applies zero-sample learning technology to image reconstruction to demonstrate its unique advantages. Zero-sample learning can perform high-quality image reconstruction through transfer learning or pre-trained models in the absence of a large amount of annotated data. This method not only reduces the dependence on large-scale data sets, but also avoids the overfitting phenomenon caused by insufficient samples, and has a significant improvement in generalization ability. For dynamic magnetic resonance imaging, zero-sample learning technology can provide reliable reconstruction results in data-scarce scenarios.

[0057] In summary, the diffusion model can reduce image reconstruction time while maintaining good image quality. Based on this, the development of more advanced zero-shot learning strategies can better cope with the common data scarcity problem in medical images. The technical combination of the diffusion model and the zero-shot learning strategy provides strong technical support for the future development of dynamic magnetic resonance imaging, and is expected to play a greater role in actual clinical applications. In the future, with the continuous optimization of algorithms and the improvement of hardware computing power, this technical combination will play a more important role in the field of medical imaging.

[0058] (one) Figure 1 It is a flow chart of the present invention, now combined with the attached Figure 1 The specific steps of the method of the present invention are described as follows:

[0059] Step A: Generate an image dataset based on the original two-dimensional dynamic full-sampling cardiac magnetic resonance multi-coil data. The image dataset is divided into an image training dataset and an image test dataset.

[0060] Step A specifically includes:

[0061] Using an adaptive coil combination method to generate two-dimensional dynamic full-sampling cardiac magnetic resonance single-coil data from original two-dimensional dynamic full-sampling cardiac magnetic resonance multi-coil data;

[0062] Then, different time-interleaved sampling strategies were implemented on the two-dimensional dynamic full-sampling cardiac magnetic resonance single-coil data used for network model training to obtain an image training data set;

[0063] The two-dimensional dynamic fully sampled cardiac magnetic resonance single coil data used for testing were undersampled using radial and Cartesian masks respectively to obtain the image testing dataset.

[0064] In the specific implementation of step A, for example, the preparation of A set of original two-dimensional dynamic fully sampled cardiac magnetic resonance multi-coil data.

[0065] First of all, The original two-dimensional dynamic full-sampling cardiac magnetic resonance multi-coil data are converted into the format of The original two-dimensional dynamic full-sampling cardiac magnetic resonance multi-coil data is combined in the format of Two-dimensional dynamic fully sampled cardiac magnetic resonance single coil data.

[0066] Then, the time-interleaved sampling strategy is used to process A set of two-dimensional dynamic full-sampling cardiac magnetic resonance single coil data was obtained. The format is The under-sampled data forms an image training dataset, which is used to train the network model or determine the network model parameters.

[0067] Finally, radial and Cartesian masks are used to mask the remaining The two-dimensional dynamic full-sampling cardiac magnetic resonance single-coil data is under-sampled to form an image test dataset, which is used to test the generalization ability of the trained network model.

[0068] Step B: Based on the image training dataset, a global and local time-interleaved sampling strategy is designed to construct two fully encoded training datasets containing global information and local information respectively, which are used to generate a network model using a global and local cascade learning network with a two-stage learning mechanism.

[0069] Step B specifically includes:

[0070] (1) Design a global and local time-staggered sampling strategy

[0071] The time interleaved sampling strategy is set to global and local. According to the different acceleration factors R set by the global and local time interleaved sampling strategies, R adjacent undersampled frames in the time dimension are merged into one frame, and then the central phase encoding of the merged frames is used for fidelity, forming two fully encoded training data sets containing global information and local information respectively, namely the global time interleaved sampling data set and the local time interleaved sampling data set;

[0072] (2) Global and local cascade learning networks

[0073] ① Two-stage learning mechanism:

[0074] Two fully encoded training data sets containing global information and local information respectively processed by global and local time interleaved sampling strategies are input into a two-stage training network; in the first stage, the global diffusion model is trained using the global time interleaved sampling data set; in the second stage, the local diffusion model is trained using the local time interleaved sampling data set;

[0075] ② The learning process of prior information of global and local diffusion models:

[0076] According to the different characteristics of the data sets obtained using the global and local time-interleaved sampling strategies, the diffusion model is used in the cascade network to learn the prior information. The global diffusion model and the local diffusion model use a formalized diffusion process to gradually transform the sample data in the K space into a standard prior distribution. The diffusion process is represented by the following stochastic differential equation:

[0077] ,

[0078] in, is a standard Wiener process. is the drift coefficient function, is the diffusion coefficient function;

[0079] The diffusion process is further optimized in the form of VE-SDE. The forward process of VE-SDE is used to transform the undersampled data into a prior distribution that is easy to handle. Its form is:

[0080] ,

[0081] in, is a Gaussian noise function in continuous time, and the continuous time variable ;

[0082] Through this process, the global diffusion model and the local diffusion model can gradually introduce noise in the K-space domain and map the original data distribution to a standard Gaussian distribution;

[0083] Once the forward diffusion is completed, the global diffusion model and the local diffusion model begin the reverse operation, transforming the prior distribution back to the original data distribution in K space through the reverse VE-SDE, and its mathematical form is:

[0084] ,

[0085] in, It's time from Backflow to The standard Wiener process when is an infinitesimal negative time step, is the distribution at time t, For each time The edge distribution gradient of

[0086] The global diffusion model and the local diffusion model predict the noise distribution of the undersampled data by learning the estimated score function of the noise, and then restore it to a state close to the real fully sampled data;

[0087] During the optimization process, the global diffusion model and the local diffusion model will dynamically adjust the weights of the loss function according to the time step to ensure that global and local information can be fully learned at different stages.

[0088] In the specific implementation process of step B, for example:

[0089] (1) Two fully encoded training datasets containing global information and local information respectively, namely the global time-interleaved sampling dataset and the local time-interleaved sampling dataset, and their construction methods are as follows:

[0090] Based on the image training dataset, set the acceleration factor of the global time-interleaved sampling strategy for First, the adjacent time dimension The under-sampled frames are merged into one frame, where the cardiac phase encoding contains the information of all under-sampled frames, and the remaining phase encodings contain the information of different under-sampled frames. Then, in order to make full use of the information contained in the under-sampled frames, these phase encodings are used on the merged frame. The central phase encoding of the undersampled frames is used to maintain data fidelity to ensure that its contrast information is similar to that of the full sampling. After batch processing, a fully encoded K-space training data set containing global information is finally obtained. , that is, a global time-interleaved sampling dataset.

[0091] Similarly, based on the image training dataset, set the acceleration factor of the local time staggered sampling strategy for , obtain a fully encoded K-space training data set containing local information , that is, a local time-staggered sampling data set.

[0092] (2) According to the different characteristics of the data sets obtained using the global and local time staggered sampling strategies, a two-stage iterative refinement diffusion network is used to learn the prior information. The prior learning process of the score-based diffusion model is to use SDE to transform the complex original data distribution into a known prior distribution. The goal is to construct a continuous time variable Index diffusion process ,in Represents a fully encoded K-space training dataset containing global information obtained by the global time-interleaved sampling strategy The original data distribution of Represents an easy-to-handle prior distribution for efficient sample generation. The mathematical expression of the global diffusion model learning process is as follows:

[0093] ,

[0094] in, is a standard Wiener process. is the drift coefficient function, which is used to describe the deterministic changes of the system. is the diffusion coefficient function, which is used to describe the random changes of the system. The VE-SDE form is used in the global diffusion model learning process, which usually leads to higher sample quality. and select , the forward VE-SDE is expressed as:

[0095] ,

[0096] in, is a Gaussian noise function in continuous time, and the continuous time variable .

[0097] By introducing noise in the K-space domain, VE-SDE gradually transforms the original data distribution into a known prior distribution. Inverse VE-SDE restores the prior distribution to the original data distribution by denoising. Inverse VE-SDE is expressed as:

[0098] ,

[0099] in, It's time from Backflow to The standard Wiener process when is an infinitesimal negative time step, is t The distribution of time. For each time The marginal distribution gradient , simulating it into the original data distribution The distribution sampling of K-space data is carried out by training the scoring network based on VE-SDE. The fully encoded K-space training dataset containing global information is obtained by using VE-SDE in a global time-interleaved sampling strategy. The global diffusion model is trained on . The optimization objectives of the global diffusion model are as follows:

[0100] ,

[0101] in, It's time The positive weight coefficient is used to adjust the loss value at different time steps. It is used to predict the time and status Estimated score function for Gaussian noise under . is the gradient of the exact log-probability function. The optimization goal is By minimizing the loss function To optimize the performance of the global diffusion model. Score Network After training, it can accurately approximate the gradient of the data distribution, which enables the global diffusion model to extract the gradient from the prior distribution during reconstruction. Denoising and generating new data sample distribution while effectively capturing and utilizing global prior information.

[0102] Learning local prior information is similar to learning global prior information, but the focus is shifted to capturing local details in the data. The goal of this stage is to improve the quality and details of the generated samples in the local region by refining the local diffusion model. The diffusion process is also modeled in the K-space domain. The following is a representation of the VE-SDE of the local diffusion model learning process:

[0103] ,

[0104] Similarly, the fully encoded K-space training dataset containing local information is obtained using the local time staggered sampling strategy of VE-SDE The local diffusion model is trained on . The optimization objectives of the local diffusion model are as follows:

[0105] ,

[0106] in, It's time The positive weight coefficient of It is used to predict the time and status The estimated score function under Gaussian noise, is the gradient of the exact logarithmic probability function, optimizing the objective By minimizing the loss function To optimize the performance of the local diffusion model and ensure that the scoring network can accurately estimate the local data distribution.

[0107] Step C: Use an iterative solution algorithm that quickly extracts prior information of the network model to alternately predict and correct the undersampled data, and adopt a multi-level iterative optimization strategy to improve image quality. In each iteration, low-rank constraints and data consistency optimization strategies are used to correct outliers introduced by noise or prediction errors.

[0108] Step C includes:

[0109] Alternately perform prediction, correction, low-rank constraint, and data consistency steps;

[0110] In each iteration, the current undersampled data is updated using a predictor based on an inverse stochastic differential equation solver that estimates the current state by discretizing time steps. , the predictor will update the next state according to the following equation :

[0111] ,

[0112] in, is a drift term used to estimate the deterministic change in the data, is the change in noise variance, represents the injected noise, is the diffusion coefficient function, is standard Gaussian noise;

[0113] The corrector corrects the prediction results. The correction step uses improved annealing Langevin dynamics and introduces noise to simulate the randomness of the diffusion process. The correction formula is:

[0114] ,

[0115] in, is the estimated noise scoring function, is the correction step size;

[0116] After each iteration, a low-rank constraint is introduced for optimization. The low-rank constraint extracts the low-rank characteristics in the data by constructing the Hankel matrix of the K space and performing singular value decomposition on it. The low-rank constraint optimization problem is expressed as:

[0117] ,

[0118] in, is a linear operator, is the reconstruction matrix, is the sampled data, is the Hankel pseudo-inverse operator, is a low-rank matrix after hard threshold singular value decomposition, is the matrix Rank operation, matrix rank, are constraints that need to be satisfied;

[0119] The sample value is restored using data consistency, and data consistency is expressed as:

[0120] ,

[0121] in, represents the index set of acquired K-space samples, Represents the frequency domain position and time point The reconstruction data, is a tuning parameter used to control the raw data and update value The weight ratio of .

[0122] In the specific implementation process of step C, for example:

[0123] In dynamic MRI reconstruction, models trained on specific datasets are of unique value, and cascaded models targeting different data features are significantly better than using a single model. During the reconstruction process, predictors and correctors are used alternately to extract prior information about the network model. The predictor is any inverse stochastic differential equation solver with a fixed discretization strategy. The corrector uses annealed Langevin dynamics, but its framework is slightly modified for better interpretability and performance. The method is based on the inverse process of the diffusion model. Given the initial data , global diffusion model denoising and is generated by the following update formula :

[0124] ,

[0125] Among them, the first behavior predictor calculation formula, the second behavior corrector calculation formula. Represents the global diffusion model function, which is used to predict or estimate the gradient in the diffusion process. is the change in noise variance, used to adjust the amplitude of the global diffusion model output, and Represents injected noise, which is used to simulate the randomness in the diffusion process. Represents the prediction revision of the global diffusion model at the current time step. Represents the noise added during the correction process to ensure the randomness and diversity of the generation process. After generating the data, in order to further optimize the sample, the local diffusion model is used to make the generated data more accurately conform to the target distribution. The specific update formula is:

[0126] ,

[0127] Among them, the first behavior predictor calculation formula, the second behavior corrector calculation formula. Represents the local diffusion model function. Through multiple iterations, the noise is gradually reduced, and finally a result close to the original data is obtained.

[0128] After applying the prediction and correction steps of the two-stage model, a series of data processing operations are also introduced to ensure the consistency and rationality of the reconstructed data. A multi-level iterative optimization strategy is adopted, and the optimization of each level is based on the results of the previous level, gradually improving the quality of the image until convergence. In each iteration, low-rank constraints and data consistency optimization strategies are used to correct outliers introduced by noise or prediction errors. The low-rank constraint constructs the Hankel matrix based on the prior information generated in the K-space domain, and the hard threshold singular value analysis of the Hankel matrix is ​​used to recover the target contrast data from the undersampled data. The low-rank constraint optimization process is described as follows:

[0129] ,

[0130] in, is a linear operator, is the reconstruction matrix, is the sampled data, is the Hankel pseudo-inverse operator, is a low-rank matrix after hard threshold singular value decomposition, is the matrix Rank operation, is a matrix rank, It is the abbreviation of "subject to", which means the constraints that need to be satisfied. Then, the sampled values ​​are restored using data consistency. Data consistency is expressed as:

[0131] ,

[0132] in, Represents the index set of acquired K-space samples. Represents the frequency domain position and time point The reconstruction data, is a tuning parameter used to control the raw data and update value The weight ratio of .

[0133] Step D: Apply the iterative solution algorithm to the trained network model and test the image test data set, output the final dynamic magnetic resonance reconstruction image, and test the peak signal-to-noise ratio value, structural similarity and mean square error value of the dynamic magnetic resonance reconstruction image to evaluate the image quality.

[0134] (two) Figure 2 This is the overall framework diagram of the time-staggered multi-diffusion model of the present invention. Figure 2 The specific steps of the method of the present invention are described as follows:

[0135] Two fully encoded training datasets containing global information and local information respectively are input into the two-stage training network. In the first stage, the global diffusion model is trained using the fully encoded training dataset containing global information (i.e., the global time-interleaved sampling dataset), which helps the network model learn a wider range of information in the initial stage, especially the broad features related to the overall distribution of the data. In the second stage, the local diffusion model is trained using the fully encoded training dataset containing local information (i.e., the local time-interleaved sampling dataset), which further improves the accuracy of the network model and its ability to handle complex details by processing more refined data. Figure 2 The stages 1 and 2 of represent the training stages of the fully encoded data constructed by the time-interleaved sampling strategy. Stage 1 merges global time frames to train the global diffusion model (GM), and stage 2 merges local time frames to train the local diffusion model (LM). In both stages of training, the U-Net framework is used to learn the prior distribution of the data. Before reconstruction, the trained global diffusion model and local diffusion model are loaded. These models can generate high-quality images during the reconstruction process by learning the prior distribution of the data, and the reconstructed image at any time can be viewed through the inverse Fourier transform (IFFT) during the reconstruction process. The adaptive diffusion step optimization strategy is adopted in the reconstruction process to approximate the real image through refinement iteration. Specifically, the inverse diffusion predictor and the Langevin dynamics corrector are used alternately to reduce noise and generate high-quality image data. At the same time, a multi-level iterative optimization unit (OU) is used, and the optimization of each level is based on the results of the previous level, gradually improving the quality of the image until convergence. In each iteration, the data is subject to a low-rank prior (LR) constraint. The low-rank prior is to convert the currently generated data matrix Decomposed into the product of three matrices: ,in Contains data matrix The left singular vectors of is the data matrix The singular value matrix of Contains data matrix The low-rank prior can generate the data matrix It is approximated by fewer singular values ​​and corresponding singular vectors, which can reduce the complexity of the matrix while retaining the most important feature information. It is then combined with a data consistency (DC) check to ensure that the generated image is consistent with the input K-space data. This helps reduce errors in the reconstruction process and improves the reliability of the final image.

[0136] (III) This paper proposes a zero-sample dynamic MRI reconstruction method based on global and local diffusion models. The implementation configuration requirements of this experiment are as follows:

[0137] The training and testing phases of this experiment were implemented on the Ubuntu 14.04 LTS (64-bit) operating system. The system is equipped with an Intel Core i7-4790K CPU and an NVIDIA GTX1080 GPU with 8GB of graphics card memory. The experiment was conducted using the open source framework PyTorch that supports CUDA and CUDNN. The data used in the experiment are 210 two-dimensional dynamic full-sampling cardiac magnetic resonance multi-coil data provided by the Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences. All data were acquired on a 3.0T scanner equipped with 20 coils, using a balanced steady-state free precession (BSSFP) sequence. The BSSFP scan uses the following parameters: field of view 330 330 square millimeters, acquisition matrix 256 256, slice thickness 6 mm, time resolution 40.0 ms. Each volunteer was required to hold his breath for 15-20 seconds on each slice. Each scan covered the entire cardiac dynamic process with one slice, for a total of 25 frames. The scanning parameters of all patients remained unchanged.

[0138] Figure 3 Neutralization Figure 4 The following are the reconstruction images of different methods under the acceleration factor R=8 and R=10, where (a) is the original image, (b) and (d) are the reconstruction images of the low-order constraint methods BCS and Tgvnn, (c) is the reconstruction image of the sparse coding method 3D-CSC, (e) and (f) are the reconstruction images of the deep learning-based methods DLTG and DD-UGM, and (g) is the reconstruction image of the method proposed in this invention. Figure 3 Neutralization Figure 4It can be seen that the BCS method relies on low-rank constraints and has a weak ability to suppress noise and artifacts in the spatial domain. Therefore, there is obvious blurring in the reconstructed image, especially in the details. Although Tgvnn improves the edge fidelity of the image to a certain extent through total variation regularization, it still cannot fully capture the dynamic changes in the image, resulting in the loss of texture details in some areas. Although the 3D-CSC method can capture local details, the artifacts are more obvious, especially in the case of high-multiple accelerated sampling. The deep learning methods DLTG and DD-UGM perform better in the reconstruction task. DLTG and DD-UGM effectively model the high-dimensional features of the image through multi-layer neural networks. However, these methods may still have local blurring when processing complex details, resulting in certain limitations in the fine recovery of dynamic information. Compared with other methods, the residual of the method proposed in the present invention is significantly smaller, which can not only effectively restore the global structure of the image at a low sampling rate, but also maintain the integrity of local details. This shows that it has higher accuracy in restoring high-frequency details and complex dynamic changes, and can make full use of the potential information in the sample, especially in the scene where the sample is scarce, showing a strong generalization ability.

[0139] The above only expresses the preferred embodiments of the present invention, and the description is relatively detailed and specific, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications, improvements and substitutions can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A zero-shot dynamic MRI reconstruction method based on global and local diffusion models, characterized in that: The following steps are involved: Step A: generating an image data set based on the original two-dimensional dynamic full-sampling cardiac magnetic resonance multi-coil data, wherein the image data set is divided into an image training data set and an image test data set; Step B: Based on the image training dataset, a global and local time-interleaved sampling strategy is designed to construct two fully encoded training datasets containing global information and local information respectively, which are used to generate a network model using a global and local cascade learning network with a two-stage learning mechanism; Step C: Use an iterative solution algorithm that quickly extracts prior information from the network model to alternately predict and correct the undersampled data, and use a multi-level iterative optimization strategy to improve image quality. In each iteration, low-rank constraints and data consistency optimization strategies are used to correct outliers introduced by noise or prediction errors. Step D: Apply the iterative solution algorithm to the trained network model and test the image test data set, output the final dynamic magnetic resonance reconstruction image, and test the peak signal-to-noise ratio value, structural similarity and mean square error value of the dynamic magnetic resonance reconstruction image to evaluate the image quality.

2. The method according to claim 1, characterized in that The step A specifically comprises: Using an adaptive coil combination method to generate two-dimensional dynamic full-sampling cardiac magnetic resonance single-coil data from original two-dimensional dynamic full-sampling cardiac magnetic resonance multi-coil data; Then, different time-interleaved sampling strategies were implemented on the two-dimensional dynamic full-sampling cardiac magnetic resonance single-coil data used for network model training to obtain an image training data set; The two-dimensional dynamic fully sampled cardiac magnetic resonance single coil data used for testing were undersampled using radial and Cartesian masks respectively to obtain the image testing dataset.

3. The method according to claim 1, characterized in that The step B specifically comprises: (1) Design a global and local time-staggered sampling strategy The time interleaved sampling strategy is set to global and local. According to the different acceleration factors R set by the global and local time interleaved sampling strategies, R adjacent undersampled frames in the time dimension are merged into one frame, and then the central phase encoding of the merged frames is used for fidelity, forming two fully encoded training data sets containing global information and local information respectively, namely the global time interleaved sampling data set and the local time interleaved sampling data set; (2) Global and local cascade learning networks ① Two-stage learning mechanism: Two fully encoded training data sets containing global information and local information respectively processed by global and local time interleaved sampling strategies are input into a two-stage training network; in the first stage, the global diffusion model is trained using the global time interleaved sampling data set; in the second stage, the local diffusion model is trained using the local time interleaved sampling data set; ② The learning process of prior information of global and local diffusion models: According to the different characteristics of the data sets obtained using the global and local time-interleaved sampling strategies, the diffusion model is used in the cascade network to learn the prior information. The global diffusion model and the local diffusion model use a formalized diffusion process to gradually transform the sample data in the K space into a standard prior distribution. The diffusion process is represented by the following stochastic differential equation: ， in, is a standard Wiener process. is the drift coefficient function, is the diffusion coefficient function; The diffusion process is further optimized in the form of VE-SDE. The forward process of VE-SDE is used to transform the undersampled data into a prior distribution that is easy to handle. Its form is: ， in, is a Gaussian noise function in continuous time, and the continuous time variable ; Through this process, the global diffusion model and the local diffusion model can gradually introduce noise in the K-space domain and map the original data distribution to a standard Gaussian distribution; Once the forward diffusion is completed, the global diffusion model and the local diffusion model begin the reverse operation, transforming the prior distribution back to the original data distribution in K space through the reverse VE-SDE, and its mathematical form is: ， in, It's time from Backflow to The standard Wiener process when is an infinitesimal negative time step, is t The distribution of time, For each time The edge distribution gradient of The global diffusion model and the local diffusion model predict the noise distribution of undersampled data by learning the estimated score function of the noise, and then restore it to a state close to the real fully sampled data; During the optimization process, the global diffusion model and the local diffusion model will dynamically adjust the weights of the loss function according to the time step to ensure that global and local information can be fully learned at different stages.

4. The method according to claim 1, characterized in that The step C comprises: Alternately perform prediction, correction, low-rank constraint, and data consistency steps; In each iteration, the current undersampled data is updated using a predictor based on an inverse stochastic differential equation solver that estimates the current state by discretizing time steps. , the predictor will update the next state according to the following equation : ， in, is a drift term used to estimate the deterministic change in the data, is the change in noise variance, represents the injected noise, is the diffusion coefficient function, is standard Gaussian noise; The corrector corrects the prediction results. The correction step uses improved annealing Langevin dynamics and introduces noise to simulate the randomness of the diffusion process. The correction formula is: ， in, is the estimated noise scoring function, is the correction step size; After each iteration, a low-rank constraint is introduced for optimization. The low-rank constraint extracts the low-rank characteristics in the data by constructing the Hankel matrix of the K space and performing singular value decomposition on it. The low-rank constraint optimization problem is expressed as: ， in, is a linear operator, is the reconstruction matrix, is the sampled data, is the Hankel pseudo-inverse operator, is a low-rank matrix after hard threshold singular value decomposition, is the matrix Rank operation, is a matrix rank, are constraints that need to be satisfied; The sample value is restored using data consistency, and data consistency is expressed as: ， in, represents the index set of acquired K-space samples, Represents the frequency domain position and time point The reconstruction data, is a tuning parameter used to control the raw data and update value The weight ratio of .

Images