A real-time magnetic resonance image sequence reconstruction method, system and medium

Abstract

The application discloses a real-time nuclear magnetic image sequence reconstruction method and system and a medium. The model training step comprises the following steps: 1) data preprocessing; 2) k-space data generation; and 3) data constraint through a loss function, and training and back propagation training of the model. The model inference step comprises the following steps: I. setting sampling parameters of the nuclear magnetic image sequence under full sampling conditions; II. real-time undersampling shooting, selecting center region data of a sampling region, and obtaining real-time undersampled k-space data; III. zero padding operation on the real-time undersampled k-space data; IV. generating a network framework, and generating 'pseudo' full-sampled k-space data according to the existing k-space center region data; and V. obtaining a final reconstructed image through IFFT transformation. For different types of target objects, the model is adjusted based on an individualized attention mechanism; and the conventional non-real-time single-frame multi-layer nuclear magnetic k-space data can also be reconstructed, and the reconstruction effect is more outstanding because the spatial relationship is considered.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a method, system, and medium for real-time nuclear magnetic resonance image sequence reconstruction. Background Technology

[0002] Magnetic Resonance Imaging (MRI) boasts high contrast, high resolution, and is radiation-free. Compared to other imaging methods, its acquisition speed is relatively slow, thus necessitating improvements in MRI imaging speed. However, while increasing speed, it's crucial to maintain good image quality. This corresponds to two key MRI metrics: temporal resolution and spatial resolution, which are also critical indicators of MRI instrument performance. In reality, due to the high cost and slow pace of instrument upgrades, improving these performance metrics from a hardware perspective is extremely difficult. Therefore, enhancing imaging performance from the perspective of sequence acquisition or algorithms, and integrating this improvement into the system, is currently the focus of research. It is generally believed that, given a fixed MRI sequence type and other sampling parameters, undersampling the phase direction of the original k-space data is the most effective and direct method to improve temporal resolution.

[0003] The existing MRI reconstruction protocols are as follows: 1) Real-time NMR reconstruction method based on keyhole acquisition: This method employs partial k-space filling, acquiring only one complete k-space magnetic resonance imaging (MRI) signal from scans at different times on the same plane. Subsequent scans acquire only data from the central portion of the k-space phase direction, while the boundary regions utilize the data from the first k-space boundary region. This approach improves scanning speed without sacrificing image contrast or spatial resolution and is commonly used in dynamic imaging. (Reference 1: *Medical Imaging Technology Terminology*, 1st Edition, 2020; Reference 2: https: / / mp.weixin.qq.com / s / Xff932sqJczjRRUQH_3R-g) 2) NMR reconstruction method based on non-Cartesian sampling acquired through sliding window update: In MRI acquisition, phase encoding in Cartesian sampling affects the sampling rate, while in non-Cartesian sampling, it affects the number of sampling lines. For example, in radial sampling, this translates to the number of strips, and in helical sampling, it's the number of cantilever lines. For non-Cartesian sampling, the angular difference between adjacent sampling lines is used as the granularity for sliding window updates. Taking radial sampling as an example, assuming a minimum of 15 sampling strips are required for real-time images, if 5 strips are sampled per frame (e.g., ... Figure 1 If T=2), then combining the sampled data of the k-space from the previous two frames (such as...) Figure 1The three frames (at times T=0 and T=1) are combined to form a data set of 15 sampling strips, which is used as the reconstruction result for the current frame. This reduces the acquisition time by 2 / 3. The data at time T=3 is then used to update the data from time T=0.

[0004] Note that in this example, the interval between the single-frame bars is 360 / 5°, and then the sampling for the next frame needs to be rotated by 360 / 15°. This is to ensure that the k-space data from the sliding window fusion is evenly distributed throughout the k-space. Then, at the current moment, only the k-space data at that moment is updated, and combined with the data from the previous two frames, it forms the 15 sampled bars for the current frame. (Reference: https: / / www.zhihu.com / question / 506617462) Both of the above schemes are k-space filling methods designed based on the correlation between information from previous and subsequent frames.

[0005] 3) NMR reconstruction method based on compressed sensing (CS) technology: Compressed sensing technology leverages the assumption that biological signals are sparse in some transform domain, such as the wavelet domain. This means that most of the information in the signal can be represented with fewer measurements, thus reducing sampling requirements. Compressed sensing-based MRI algorithms employ iterative algorithms, such as compressed sensing reconstruction (CS-MRI) or variational methods, to recover high-quality images from sparsely sampled data. These algorithms solve an optimization problem to find the sparse representation that best approximates the original image, which is then inversely transformed back to the image domain. Despite the lower sampling rate, compressed sensing-based MRI algorithms can often deliver image quality comparable to traditional high-density sampling, especially with appropriate algorithm and parameter settings. This makes it possible to maintain image quality while saving time.

[0006] 4) Deep learning-based nuclear magnetic resonance reconstruction methods: Deep learning-based MRI reconstruction algorithms typically employ deep learning architectures such as convolutional neural networks or variational autoencoders. These networks are capable of learning complex image features and structures, contributing to improved image quality. Deep learning models can learn structural information from large amounts of MRI image data, thus better reconstructing missing or undersampled data. This allows them to better handle noise and artifact problems.

[0007] In summary, current technical methods for reconstructing MRI image sequences have the following problems: i) As described in the first two schemes of the prior art, the reconstruction data of subsequent frames depends on the sequence content of the previous frames. If the speed is to be further improved, the reconstruction data needs to rely more on the information of the previous frames. If the shooting content is large and changes a lot, it is not possible to reconstruct subsequent frames well.

[0008] ii) The third approach, as described in the prior art, typically requires iterative optimization, leading to high computational complexity, especially for high-resolution and 3D image reconstruction. Algorithm performance is highly dependent on the chosen compressed sensing model, sparse transform domain, and parameter settings. Different choices can result in different reconstruction outcomes, requiring careful parameter tuning and selection, which increases algorithm complexity and poses a significant obstacle to achieving real-time NMR reconstruction tasks.

[0009] iii) As described in the fourth approach of the prior art, its associated deep learning algorithms require a large amount of labeled data to train the model, which may be limited in some cases. Furthermore, the quality and diversity of the labeled data have a significant impact on algorithm performance. Additionally, most algorithms simulate undersampling under different sampling masks, while the low-frequency information in the k-space (i.e., the phase direction center portion data described in the first approach of the prior art) is the key to determining image content. For non-periodic images, the related algorithms lack personalized settings for different types of scenes, resulting in poor generalization ability. Summary of the Invention

[0010] To address the aforementioned shortcomings, the technical solution of this invention aims to: reconstruct NMR sequences and generate near-realistic, high-quality NMR images while maintaining real-time performance. Furthermore, it employs a model adjustment based on a personalized attention mechanism for different target object categories.

[0011] The technical solution of this invention solves the problems of existing reconstruction technologies, such as the inability to perform personalized design and the convergence of reconstruction data sources. Considering the above factors, this invention provides a real-time MRI image sequence reconstruction method, system, and medium. Specifically, as follows: A real-time MRI image sequence reconstruction method, the model training process of which includes the following steps: Step 1: Data preprocessing, including: 1.1) Obtain the sampling parameters of the NMR image sequence under full sampling conditions, and ensure that the sampling parameters of the NMR image sequence are consistent; wherein, for multi-layer NMR image sequences acquired at the same time, the number of sampling matrices under full sampling is phase × frequency × number of layers; 1.2) Perform undersampling and select the central region data of the sampling area to obtain undersampling k-space data; 1.3) Perform zero-padding on the undersampled k-space data; Step 2: k-space data generation, including: By generating a network framework, “pseudo” fully sampled k-space data is generated based on existing k-space center region data; wherein, the k-space data is a complex number data type, and the k-space data is separated into real and imaginary channels for convolution operation; Step 3: Use the loss function to constrain the data, and complete the training of the model and backpropagation training.

[0012] Optionally, in step 1: the central region data of the sampling area includes 10%-20% of the data from the phase center of the sampling area.

[0013] Optionally, in step 2, the generative network framework includes, but is not limited to, the following generative models: generative adversarial network, Transformer, or diffusion model network.

[0014] Optionally, in step 3: The data constrained includes labeled data and unlabeled data; the labeled data and unlabeled data refer to the existence and non-existence of corresponding fully sampled k-space true reference data for each frame of undersampled k-space data, respectively. The unlabeled data includes at least one or more of the following groups: a) Zero-filled k-space frequency domain data and the corresponding spatial domain undersampled image; b) The generated “pseudo” fully sampled k-space frequency domain data and the corresponding spatial domain reconstructed image; c) Fully sampled k-space frequency domain data and corresponding spatial domain real images of the same type of NMR image sequence, excluding a) or b).

[0015] Optionally, for the unlabeled data: model training and backpropagation training are completed using a loss function, specifically including: In the frequency domain: The Gaussian Mixture Model (GMM) is used to iteratively solve the parameters of the frequency domain data to obtain the probability distribution of the frequency domain. The infoNCE loss function is used to constrain the three sets of data a), b), and c) so that the multivariate Gaussian distributions of the data in sets a) and b) are far apart, and the multivariate Gaussian distributions of the data in sets b) and c) are close. In the spatial domain: Inverse Fast Fourier Transform (IFFT) is performed on the frequency domain rate data to obtain spatial domain data, which is then passed through an encoder and a projector to obtain a one-dimensional representation. Then, a normalized temperature scale cross-entropy loss function is used to impose spatial domain-based constraints so that the one-dimensional representation distance of the data in groups a) and b) increases, and the one-dimensional representation distance of the data in groups b) and c) decreases. The normalized temperature-scale cross-entropy loss function includes, but is not limited to, NT-Xent Loss.

[0016] Optionally, for the labeled data, a multi-scale constraint learning strategy is introduced, specifically including: At the pixel level, three additional types of information are introduced: saliency map, edge information, and subject recognition, resulting in three types of images: saliency map, edge information based on the Sobel gradient operator, and target object subject. Attention-based pixel-level constraints are applied to the bright pixel regions in the above three types of images, and a loss function is constructed to give higher weight to the bright pixel regions. Among them, the loss function based on the target object is... L main_area The calculation formula is as follows: Formula (1) In the above formula (1), N The number of training batches; M main_area,i For the first i The mask of the target object in the image is set to 1 if the number of pixels exceeds a certain threshold, and 0 otherwise, in order to adjust the pixel-level weights of the loss function. λ main_area,i For the first i A regional mask of the target object in each image is used to adjust the region-level weights of the loss function; I rec,i For the first training batch i One reconstruction result data; I tru,i For the first training batch i A real image; The loss function based on the saliency map is calculated similarly using formula (1). L saliency and loss functions based on edge information L edge Among them, mask M saliency ,as well as M edge The results were obtained using logarithmic spectroscopy and the Sobel gradient operator, respectively. Thus, the final loss function is... L total The calculation formula is as follows: Formula (2) In the above formula (2): α, β, γ All of these are pre-set coefficients.

[0017] Optionally, for the labeled data: focus on the structural information of the inter-layer data to standardize and constrain the inter-layer information relationships, specifically including: During training, multi-layer data is used as a single input sample, and key location information points of the three-dimensional target area are obtained using point cloud methods. In conjunction with the target object in the three-dimensional target area, focus on the geometric positional relationship between the key positional information points within the three-dimensional target area; Minimize the geometric positional relationship to constrain geometric invariance, and update the network parameters in reverse, thereby completing the training of the model and backpropagation training.

[0018] Optionally, the model inference process of this method includes the following steps: Step 1: Set the sampling parameters for the MRI image sequence under full sampling conditions; Step II: Perform real-time undersampling imaging, select the data in the central region of the phase direction of the sampling area, and obtain real-time undersampling k-space data; Step III: Perform zero-padding on the real-time undersampled k-space data; Step IV: k-space data generation, including: generating "pseudo" fully sampled k-space data based on existing k-space central region data using a generative network framework; Step V: Obtain the final reconstructed image through Inverse Fast Fourier Transform (IFFT).

[0019] A system for real-time MRI image sequence reconstruction includes: Memory, used to store computer programs; and A processor is configured to execute the steps of the real-time nuclear magnetic resonance image sequence reconstruction method when executing the computer program.

[0020] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the real-time MRI image sequence reconstruction method.

[0021] Beneficial effects

[0022] The technical solution of the present invention has the following advantages: The technical solution of this invention aims to reconstruct NMR sequences and generate near-realistic, high-quality NMR images while meeting real-time performance requirements. For different target object categories, model adjustments based on a personalized attention mechanism are applied. This invention can also reconstruct conventional non-real-time single-frame multi-layer NMR k-space data; due to consideration of spatial relationships, the reconstruction results are even better. Attached Figure Description

[0023] Figure 1 This is an example diagram of a radial sampling method for real-time NMR reconstruction in the prior art; Figure 2 This is a flowchart of a real-time MRI image sequence reconstruction method according to an embodiment of the present invention. Detailed Implementation

[0024] The prior art and the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] Figure 1 This image shows an example of a radial sampling method for real-time NMR reconstruction in the prior art. This method is a non-Cartesian sampling NMR reconstruction method based on sliding window update acquisition. Figure 1 As shown, phase encoding in Cartesian sampling affects the sampling rate in MRI acquisition, while in non-Cartesian sampling it is the number of sampling lines. For example, in radial sampling, it is the number of strips, and in helical sampling, it is the number of cantilever arms. For non-Cartesian sampling, the angular difference between each adjacent sampling line is used as the granularity of the sliding window update. Taking radial sampling as an example, assuming that a real-time image requires at least 15 sampling strips, if 5 strips are sampled per frame (e.g.... Figure 1 If T=2), then combining the sampled data of the k-space from the previous two frames (such as...) Figure 1 The three frames (at times T=0 and T=1) are combined to form a data set of 15 sampling bars, which is used as the reconstruction result for the current frame. This reduces the acquisition time by 2 / 3. The data at time T=3 updates the data from time T=0. Note that in this example, the interval between the single-frame bars is 360 / 5°, and the next frame sampling requires a 360 / 15° rotation. This is to ensure that the k-space data from the sliding window fusion is evenly distributed throughout the k-space. Then, at the current time, only the k-space data at the current time is updated, and combined with the data from the previous two frames, it forms the 15 sampling bars for the current frame (Reference: https: / / www.zhihu.com / question / 506617462).

[0026] Figure 1 The method shown is a k-space filling method designed based on the correlation between the information of preceding and following frames. Its disadvantage is that the reconstruction data of subsequent frames depends on the sequence content of the previous frames. If the speed is to be further improved, the reconstruction data needs to rely more on the information of the previous frames. If the shooting content is large and changes a lot, it cannot reconstruct subsequent frames well.

[0027] Figure 2This is a flowchart of a real-time MRI image sequence reconstruction method according to an embodiment of the present invention.

[0028] The technical solution of this invention addresses the need for high spatial and temporal resolution in guided images. Based on the central k-space frequency domain sequence data acquired at each time step (frame) by a magnetic resonance imaging (MRI) instrument, it completes the global reconstruction of the k-space data. The solution employs a multi-domain contrastive learning and multi-scale constraint approach. First, it performs joint contrastive learning in the spatial and frequency domains to ensure the convergence of similar data distributions in both domains, taking into account the characteristics of MRI images. Second, it designs pixel-level constraints and spatial structure constraints based on a dynamic attention mechanism.

[0029] The key point of this invention is: 1. Pixel-level learning based on personalized attention mechanisms; 2. Geometric relationship learning based on personalized attention mechanisms; 3. Multi-domain contrastive learning joint constraints.

[0030] The technical solution of the present invention is as follows Figure 2 As shown: (I) Overview of the Plan: This invention attempts to address the following issues in model reasoning: First, the task needs to meet the major premise of real-time MRI reconstruction. When adopting this scheme, it is necessary to determine the sequence acquired by the MRI scanner. Under the condition of full sampling, it is based on the MRI sequence of about 200ms per frame. For example, taking the 2D MRI sequence, the number of sampling matrices for full sampling is 128×128 (phase×frequency) Echo Planar Imaging (EPI) sequence.

[0031] Secondly, undersampling is performed, with the sampling area selected from 10%-20% of the data at the phase center. At this time, the number of sampling matrices in the undersampling sequence is (12-24)×128×1, and the single-frame acquisition time is approximately 20-40ms.

[0032] Next, based on the real-time undersampled sequence obtained above, we pre-fill the undersampled k-space data using a zero-filling scheme. "Pseudo" fully sampled k-space data is obtained through a generative network framework; Finally, the reconstructed image is obtained through inverse fast fourier transform (IFFT).

[0033] Thus, the model reasoning process of the proposed scheme was completed, and a real-time NMR sequence reconstruction method was realized, achieving the reconstruction and presentation of high-quality NMR images under real-time conditions.

[0034] (II) Specific Model Training Details (1) Data preprocessing like Figure 2 The training process of the entire model is shown in detail. First, the data acquisition method is consistent with that of the inference part. However, since the data in the training process does not need to be real-time, while ensuring the consistency of the NMR sequence parameters, multiple layers (slices) of data are collected at the same time. For example, in a 2D NMR sequence, the number of sampling matrices under full sampling is 128×128×24 (phase×frequency×number of layers). 10%-20% of the data at the phase center is selected for the sampling region. At this time, the number of k-space sampling matrices of the sampling sequence is (12-24)×128×24. After zero-padding, the number of sampling matrices of the sampling sequence is padded to 128×128×24.

[0035] (2) k-space data generation By generating a generative network framework, pseudo-fully sampled k-space data is generated based on existing k-space central region data. This generative network framework can generally employ any generative model such as Generative Adversarial Networks (GANs), Transformers, or Diffusion Model Networks. Because k-space data is a complex number data type, this scheme separates the complex number into real and imaginary channels for convolution operations. That is, the final data becomes 128×128×24×2 (phase×frequency×number of layers×channels).

[0036] (3) Loss function constraints and backpropagation This part is the core of the overall solution. Regardless of the generation network used, constraints on the generated data are required.

[0037] This model considers two types of data: labeled data and unlabeled data. The label refers to the existence of a corresponding fully sampled k-space true reference NMR data for each frame of undersampled k-space NMR data.

[0038] (3.1) For unlabeled data (e.g.) Figure 2 As shown, in the backpropagation of training 1), this scheme adopts a multi-domain similarity contrastive learning strategy for the generated "pseudo" fully sampled k-space data. At this time, although there is no one-to-one corresponding real data reference, there are three types of data, namely: a) Zero-filled k-space frequency domain data and corresponding spatial domain undersampled NMR images; b) The generated “pseudo” full-sample k-space frequency domain data and the corresponding spatial domain reconstruction results; c) Fully sampled k-space frequency domain data and corresponding spatial domain real images of other similar NMR sequences.

[0039] For frequency domain data, we choose to use a Gaussian Mixture Model (GMM) to iteratively solve the parameters of the data to obtain the probability distribution in the frequency domain. We then use a loss function similar to infoNCE to constrain the three sets of data a, b, and c, so that the multivariate Gaussian distributions of a and b are far apart, while the distributions of b and c are close together.

[0040] In the image domain (spatial domain), the frequency domain data is directly transformed using IFFT to obtain image domain data. This data is then processed by an encoder and a projector to obtain a one-dimensional representation. A normalized temperature-scaled cross-entropy loss (NT-Xent Loss) is used to impose constraints based on the image domain, increasing the distance between the one-dimensional representations of a and b, and decreasing the distance between b and c. This loss function is used to complete the training and backpropagation of the unlabeled data model.

[0041] (3.2) Next is the treatment of labeled data (such as...) Figure 2 As shown in Figure 3, the backpropagation of training... The technical solution of this invention introduces a multi-scale constraint learning strategy based on training with unlabeled data. At the pixel level, it proposes to introduce three additional types of information: saliency map, edge information, and subject recognition. For the image, it obtains its saliency map, edge information based on the Sobel gradient operator, and the target object. For the three types of highlighted pixels, it proposes attention-based pixel-level constraints... L 1 The basic loss function, which assigns higher weights to highlighted areas, is the main loss function. L main_area The calculation formula is as follows: Formula (1) In the above formula (1): N is the number of training batches; M main_area,i For the first i A mask for identifying the subject in an image, such as for identifying MRI images, is set to 1 if the number of pixels exceeds a certain threshold, and 0 otherwise, in order to adjust the specific weights of the loss function at the pixel level. λ main_area,i For the firsti In each image, regional masks designed based on the characteristics of the target population, such as gender and age, and different brain structures, are used to adjust the weights of the loss function at the region level. For example, the gray matter structures of the caudate nucleus, putamen, and nucleus accumbens decrease with age, and the gray and white matter in the frontal and temporal lobes of men is thinner. If attention needs to be paid to these areas, then compared to younger people and women, the relevant regions need to be represented in a more detailed way. λ main_area,i The values ​​have been further improved. Different data sources necessitate adjustments to relevant parameters to meet the diverse design requirements of the model; I rec,i For the first training batch i One reconstruction result data; I tru,i For the first training batch i A real image.

[0042] Others, such as those based on saliency maps L saliency and edge information L edge The loss function is obtained similarly to that in equation (1). Its mask... M saliency and M edge The results were obtained using logarithmic spectroscopy and the Sobel gradient operator, respectively. Therefore, the final loss function... L total The calculation formula is as follows: Formula (2) In the above formula (2): α , β , γ All of these are pre-set coefficients.

[0043] Having completed the above design, for labeled data, further attention is paid to the structural information of the inter-layer data to better standardize and constrain the information relationships between layers. During training, multiple layers of data are treated as single input samples, and point cloud methods are used to obtain key location information points of the 3D target region, such as... Figure 2 As shown in the lower left region's 'Backpropagation during Training 2', and combined with the target object in the target region, the focus shifts to the geometric constraints between points within the target region. This minimizes the geometric positional relationships, thereby constraining geometric invariance, and updates the network parameters in reverse. This completes all training and backpropagation.

[0044] In the technical solution of this invention, the generating network part is not limited to the network mentioned above, that is, it only needs to be able to extract information between data. In addition, the feature extraction part of the deep network can perform feature extraction in the form of complex convolution, complex domain and filter product.

[0045] Furthermore, the design of personalized attention mechanisms is not limited to Sobel operators, logarithmic spectra, and point clouds; other methods can replace related components to achieve the relevant functions.

[0046] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for real-time reconstruction of nuclear magnetic resonance image sequences, characterized in that, The model training process of this method includes the following steps: Step 1: Data preprocessing, including: 1.1) Obtain the sampling parameters of the NMR image sequence under full sampling conditions, and ensure that the sampling parameters of the NMR image sequence are consistent; wherein, for multi-layer NMR image sequences acquired at the same time, the number of sampling matrices under full sampling is phase × frequency × number of layers; 1.2) Perform undersampling and select the central region data of the sampling area to obtain undersampling k-space data; 1.3) Perform zero-padding on the undersampled k-space data; Step 2: k-space data generation, including: By generating a network framework, "pseudo" fully sampled k-space data is generated based on existing k-space center region data; wherein, the k-space data is a complex number data type, and the k-space data is separated into real and imaginary channels for convolution operation; Step 3: Use the loss function to constrain the data, and complete the training of the model and backpropagation training; In step 3: The data constrained includes labeled data and unlabeled data; the labeled data and unlabeled data refer to the existence and non-existence of corresponding fully sampled k-space true reference data for each frame of undersampled k-space data, respectively. The unlabeled data includes at least one or more of the following groups: a) Zero-filled k-space frequency domain data and the corresponding spatial domain undersampled image; b) Generated "pseudo" fully sampled k-space frequency domain data and corresponding spatial domain reconstructed image; c) Fully sampled k-space frequency domain data and corresponding spatial domain real images of the same type of NMR image sequence, excluding a) or b); Specifically, for the unlabeled data: model training and backpropagation training are completed using a loss function, including: In the frequency domain: The Gaussian Mixture Model (GMM) is used to iteratively solve the parameters of the frequency domain data to obtain the probability distribution of the frequency domain. The infoNCE loss function is used to constrain the three sets of data a), b), and c) so that the multivariate Gaussian distributions of the data in sets a) and b) are far apart, and the multivariate Gaussian distributions of the data in sets b) and c) are close. In the spatial domain: Inverse Fast Fourier Transform (IFFT) is performed on the frequency domain rate data to obtain spatial domain data. Then, it is passed through the encoder and the projection head to obtain a one-dimensional representation. Then, the normalized temperature scale cross-entropy loss function is used to perform spatial domain-based constraints so that the one-dimensional representation distance of the data in groups a) and b) increases, and the one-dimensional representation distance of the data in groups b) and c) decreases. The normalized temperature-scale cross-entropy loss function includes, but is not limited to, NT-Xent Los.

2. The method according to claim 1, characterized in that, In step 1: the central region data of the sampling area includes 10%-20% of the data of the phase center of the sampling area.

3. The method according to claim 2, characterized in that, In step 2: the generative network framework includes, but is not limited to, the following generative models: generative adversarial network, Transformer, or diffusion model network.

4. The method according to claim 3, characterized in that, For the labeled data: a multi-scale constraint learning strategy is introduced, specifically including: At the pixel level, three additional types of information are introduced: saliency map, edge information, and subject recognition, resulting in three types of images: saliency map, edge information based on the Sobel gradient operator, and target object subject. Attention-based pixel-level constraints are applied to the bright pixel regions in the above three types of images, and a loss function is constructed to give higher weight to the bright pixel regions. Among them, the loss function L based on the target object is... main_area The calculation formula is as follows: Official (1); In formula (1) above, N is the number of samples in a training batch; M main_area,i λ is a mask for the main object in the i-th image. It is set to 1 if the number of pixels exceeds a certain threshold, and 0 otherwise, used to adjust the pixel-level weights of the loss function; main_area,i I is a regional mask for the main object in the i-th image, used to adjust the region-level weights of the loss function; rec,i For the i-th reconstruction result data of the training batch; I tru,i The i-th real image in the training batch; The loss function L based on the saliency map is calculated similarly using formula (1). saliency and the loss function L based on edge information edge Among them, mask M saliency and M edge The results were obtained using logarithmic spectroscopy and the Sobel gradient operator, respectively. Thus, the final loss function L total The calculation formula is as follows: Official (2); In the above formula (2): α, β, and γ are all pre-set coefficients.

5. The method according to claim 4, characterized in that, Regarding the labeled data: focus on the structural information of the data between layers to standardize and constrain the relationships between information between layers, specifically including: During training, multi-layer data is used as a single input sample, and key location information points of the three-dimensional target area are obtained using point cloud methods. In conjunction with the target object in the three-dimensional target area, focus on the geometric positional relationship between the key positional information points within the three-dimensional target area; Minimize the geometric positional relationship to constrain geometric invariance, and update the network parameters in reverse, thereby completing the training of the model and backpropagation training.

6. The method according to claim 5, characterized in that, The model inference process of this method includes the following steps: Step 1: Set the sampling parameters for the MRI image sequence under full sampling conditions; Step II: Perform real-time undersampling imaging, select the data in the central region of the phase direction of the sampling area, and obtain real-time undersampling k-space data; Step III: Perform zero-padding on the real-time undersampled k-space data; Step IV: k-space data generation, including: generating "pseudo" fully sampled k-space data based on existing k-space central region data using a generative network framework; Step V: Obtain the final reconstructed image through Inverse Fast Fourier Transform (IFFT).

7. A system for real-time nuclear magnetic resonance image sequence reconstruction, characterized in that, include: Memory, used to store computer programs; as well as A processor, configured to execute the steps of the real-time nuclear magnetic resonance image sequence reconstruction method as described in any one of claims 1-6 when executing the computer program.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the real-time nuclear magnetic resonance image sequence reconstruction method as described in any one of claims 1-6.

Images