Magnetic resonance interventional image reconstruction method and equipment

Abstract

The invention discloses a magnetic resonance intervention image reconstruction method. The method comprises the following steps: S1, collecting non-uniform frequency domain data in real time by using a magnetic resonance scanner; s2, performing data consistency updating on the non-uniform frequency domain data; and S3, inputting the updated data into the neural network to obtain a reconstruction result. According to the invention, a multi-scale online re-parameterized convolutional recurrent neural network MSOR-CRNN is provided, and two core innovations are introduced on the basis of a traditional CRNN architecture: an online re-parameterized module OR; and a multi-scale feature propagation mechanism MSFP is adopted, so that the whole system optimizes the reconstruction effect and realizes performance improvement on the premise of not increasing the calculation burden.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging, and more particularly to a method and apparatus for reconstructing interventional magnetic resonance images. Background Technology

[0002] Magnetic Resonance Imaging (MRI) is a non-invasive medical imaging technique with excellent soft tissue contrast, multi-parameter imaging capabilities, and no ionizing radiation. It is widely used for the diagnosis and treatment guidance of diseases in multiple organs, including the brain, heart, and abdomen. In recent years, with advancements in MRI hardware and imaging technology, MRI has been increasingly introduced into interventional radiology for real-time guidance of minimally invasive procedures such as intravascular catheterization, including targeted drug delivery, embolization, angioplasty, and ablation. Compared to traditional X-ray fluoroscopy, MRI provides high tissue contrast while avoiding radiation exposure and contrast agent toxicity, thus being considered one of the important imaging directions for future interventional procedures.

[0003] However, the real-time imaging capability of MRI is limited by its inherent sampling mechanism. Compared with computed tomography (CT) or ultrasound imaging, MRI has a longer data acquisition and reconstruction time, resulting in lower imaging efficiency and insufficient temporal resolution, which is not conducive to capturing dynamic processes and accurately positioning surgical instruments. To achieve rapid imaging, researchers have proposed various acceleration techniques, including parallel imaging and compressed sensing. The former expands the sampling bandwidth by utilizing the sensitivity differences of multiple coil receiving channels, thereby reducing the number of sampling points; the latter utilizes the sparsity characteristics of images to achieve reconstruction at sampling rates far below the Nyquist rate. The introduction of non-uniform sampling has also further shortened the single-frame scan time. However, these highly accelerated sampling methods often lead to severe aliasing artifacts and noise amplification, while making the reconstruction algorithm highly complex and computationally intensive, making it difficult to meet the low latency requirements of real-time intervention.

[0004] Traditional MRI image reconstruction methods are typically based on iterative optimization processes, requiring repeated solutions to regularization terms and parameter tuning, such as low-rank models, sparse constraints, and dictionary learning. While these methods can achieve high image quality, they incur significant computational costs and reconstruction delays, limiting their real-time applications. In dynamic MRI, low-rank and sparse decompositions (such as the L+S model) can further utilize temporal redundancy to improve image consistency, but still require numerous iterative steps, making it difficult to balance speed and accuracy.

[0005] With the development of deep learning, neural network-based reconstruction methods have become an important direction for accelerating MRI. These methods learn the mapping relationship from undersampled data to high-quality images in a data-driven manner, significantly reducing the number of iterations and improving reconstruction efficiency. Some studies employ end-to-end reconstruction networks to directly generate images from k-space data; others introduce learnable prior models while maintaining data consistency by unfolding traditional iterative algorithms. Representative methods include MoDL, ISTANet, and VANet, which have demonstrated excellent performance in both static and dynamic MRI reconstruction. However, most of these methods are geared towards offline or latency-tolerant scenarios, and still have limitations for interventional imaging requiring low latency and real-time frame-by-frame updates.

[0006] In dynamic and interventional MRI scenarios, convolutional recurrent neural networks (CRNNs) are considered a suitable architecture for real-time reconstruction due to their ability to capture short-term dependencies in time series and balance latency and quality. Previous studies have applied basic CRNN models to some interventional tasks, but their reconstruction performance remains limited, especially under high acceleration conditions where spatiotemporal information loss and detail blurring are common.

[0007] To enhance the representational power and reconstruction quality of networks, various convolutional branching structures and multi-scale feature fusion strategies have emerged in recent years. Online re-parameterization (OR) introduces multi-branch convolutional structures during training to enhance model expressiveness and then folds them into a single convolutional layer during inference, thereby improving performance without increasing inference computation. This idea has been validated in models such as Inception, ACNet, DBB, and RepVGG. On the other hand, multi-scale feature propagation (MSFP) expands the effective receptive field by transferring information across different spatial scales, improving the model's ability to comprehensively express local details and global structure. This strategy is widely used in network architectures such as U-Net, HRNet, and DeepLab.

[0008] While existing studies have validated the effectiveness of these methods in single-frame or dynamic MRI reconstruction, challenges remain in meeting the low latency, high frame rate, and high fidelity requirements of interventional scenarios. Current network architectures struggle to simultaneously ensure real-time performance while also managing the modeling of complex spatiotemporal dependencies, and performance improvements are limited by scarce computational resources. Therefore, improving reconstruction quality while maintaining low latency remains a critical issue that urgently needs to be addressed in real-time interventional MRI imaging technology.

[0009] In summary, existing real-time MRI imaging technologies have the following drawbacks: 1. Existing deep learning-based MRI reconstruction methods often suffer from high computational latency during the inference phase, making it difficult to meet the requirements of real-time imaging.

[0010] 2. In interventional MRI scenarios, the sampling mode is complex and the amount of data is limited. Traditional depth models are prone to overfitting and the reconstruction performance is unstable under different acceleration rates and trajectory conditions.

[0011] 3. In interventional MRI settings, while 3D imaging provides complete spatial information, its acquisition and reconstruction delays are too high, making it difficult to meet real-time guidance requirements. 2D imaging, although offering higher temporal resolution, is limited by slice thickness and slice selection: thick slices can obscure small catheters, while thin slices can easily cause instruments to deviate from the imaging plane, increasing positioning uncertainty. Existing single-plane imaging cannot simultaneously achieve both timeliness and spatial coverage.

[0012] Therefore, those skilled in the art are dedicated to developing a magnetic resonance imaging reconstruction method. Summary of the Invention

[0013] To achieve the above objectives, the present invention provides a method for reconstructing interventional magnetic resonance images, comprising: S1: Real-time acquisition of non-uniform frequency domain data using a magnetic resonance scanner; S2: Perform data consistency updates on non-uniform frequency domain data; S3: Input the updated data into the neural network to obtain the reconstruction results.

[0014] Furthermore, step S2, which updates the data consistency of non-uniform frequency domain data, includes: performing a gradient descent correction on the reconstruction result of the current frame based on the principle of near-end gradient descent to ensure that it is consistent with the original sampled data.

[0015] Furthermore, the neural network in step S3 adopts a convolutional recurrent neural network structure, which includes an iterative recurrent branch and a temporal recurrent branch. The iterative recurrent branch is used to pass feature states across iterations within the same frame, and the temporal recurrent branch is used to propagate temporal context information between adjacent frames.

[0016] Furthermore, the convolutional recurrent neural network adopts a unidirectional recursive design and is divided into five layers, namely tiCRNN, iCRNN, iCRNN, iCRNN, and CNN. Each recursive layer receives the features from the previous layer and updates them by combining the hidden states of the current frame, the previous frame, and the previous iteration. Finally, it outputs the reconstruction result through residual connections. The first layer includes a recurrent structure with time and iteration dimensions, any one of the second to fourth layers includes a recurrent structure with iteration dimensions, and the fifth layer is implemented with simple convolution.

[0017] Furthermore, each convolutional operation unit in the convolutional recurrent neural network is an online reparameterization module (OR). Each convolutional operator in the OR is expanded into multiple parallel branches. Each parallel branch consists of multiple sequentially connected small convolutional kernels, which are weighted and fused using learnable scaling factors. During the training phase, each parallel branch independently participates in feature extraction and parameter update. During the inference phase, algebraic operations are used to fuse multi-branch convolutional kernels within the same scale into an equivalent single convolutional kernel, achieving an output that is completely equivalent to the training structure. The batch normalization layer of the OR is an affine scaling layer, and the bias term is stored independently.

[0018] Furthermore, the third layer of the convolutional recurrent neural network includes a first-order multi-scale expansion mechanism (MS) module. The MS module adds parallel paths of different resolutions on the basis of the main convolutional branch to extract multi-scale spatial features.

[0019] Furthermore, there are three parallel paths. The first parallel path performs a standard 3×3 convolution at the original resolution. The second parallel path performs a 2x downsampling through average pooling. After performing convolution in the low-resolution space, bilinear interpolation is used to restore the original resolution. The third parallel path uses a 4x downsampling and the same "convolution-interpolation" process to extract structural features at a larger scale.

[0020] Furthermore, the third layer of the convolutional recurrent neural network has 64 channels for any convolution and a kernel size of 3×3.

[0021] This invention also provides a method for cross-view imaging magnetic resonance interventional image reconstruction: Define a first tangent and a second tangent, with the first tangent and the second tangent being orthogonal to each other; Image reconstruction was performed in the first section using the aforementioned magnetic resonance imaging reconstruction method. Image reconstruction was performed in the second section using the aforementioned magnetic resonance imaging reconstruction method. The spatial position of the instrument is estimated using the image reconstruction results from the first and second sections and aligned with the pre-acquired three-dimensional reference model of the instrument.

[0022] The present invention also provides a computer device, comprising: Memory; One or more processors, coupled to memory; One or more applications, wherein one or more applications are stored in memory and configured to be executed by one or more processors, and one or more applications are configured to perform the reconstruction method described above.

[0023] The present invention also provides a computer-readable storage medium storing program code, which can be called by a processor to execute the above-described reconstruction method.

[0024] Technical effect

[0025] 1. This invention proposes a Multi-Scale Online Reparameterized Convolutional Recurrent Neural Network (MSOR-CRNN), introducing two core innovations on the traditional CRNN architecture: an online reparameterization module (OR) and a multi-scale feature propagation mechanism (MSFP). The OR module enhances feature representation by introducing multi-branch convolutions during training and fuses them into a single equivalent convolutional kernel during inference, achieving a structural equivalence transformation of "training a large model and inferring a small model," thus improving training stability and inference efficiency. The MSFP module introduces a multi-scale path under an iterative-temporal dual recursive structure, expanding the effective receptive field (ERF) through downsampling, upsampling, and feature aggregation, enhancing the spatial hierarchy and temporal consistency of the model. MSOR-CRNN significantly improves reconstruction quality while maintaining low-latency inference, achieving higher PSNR and SSIM metrics than the baseline CRNN, and maintaining stable temporal consistency across consecutive frames. The fused OR structure does not increase inference complexity, and the MSFP module only introduces lightweight interpolation and pooling operations, achieving overall system performance improvement without increasing computational burden.

[0026] 2. MSOR-CRNN achieves adaptive updates in both temporal and iterative dimensions through a recursive fusion structure. Combined with multi-scale and online reparameterization mechanisms, it enhances cross-task generalization capabilities. Iterative recursion enables adaptive refinement of intra-frame features, while temporal recursion handles inter-frame information propagation. These two mechanisms form a closed-loop iterative optimization mechanism within a unified network, ensuring reconstruction stability across time steps. By sharing fusion weights under different sampling rates and angles, the model automatically adapts to the signal characteristics of different sampling modes. This mechanism demonstrates consistent performance improvement on both CMR and IMR datasets, proving its robustness to changes in acceleration rate and sampling mode. Experiments show that it maintains stable reconstruction results even under high acceleration conditions, providing a generalizable online reconstruction framework for practical intervention scenarios.

[0027] 3. This invention is also applicable to cross-view imaging scenarios using orthogonal cross-view technology. This technique utilizes two orthogonal planes to lock onto the target region, alternately exciting and acquiring signals within the imaging area to obtain complementary viewpoint information, thereby enhancing spatial positioning and obtaining three-dimensional position estimation within a two-dimensional acquisition framework. Through the MSOR-CRNN reconstruction module with strong spatiotemporal consistency, it can stably restore high-quality images of the signal intersection region, reducing blurring caused by motion artifacts and signal overlap, thus providing low-latency, high-precision spatial awareness capabilities for MRI navigation in real-world interventional scenarios.

[0028] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of a magnetic resonance imaging reconstruction method according to a preferred embodiment of the present invention; Figure 2 This is a schematic diagram of a CRNN network architecture according to a preferred embodiment of the present invention; Figure 3 This is a comparison diagram of the reconstruction result and the zero-fill reconstruction result of a preferred embodiment of the present invention; Figure 4 This is a cross-view imaging effect diagram of a preferred embodiment of the present invention. Detailed Implementation

[0030] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0031] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0032] like Figure 1 , 2 As shown, this embodiment provides a method and apparatus for real-time reconstruction of magnetic resonance images based on a multi-scale online reparameterized convolutional recurrent neural network, used to achieve high-quality, low-latency dynamic image reconstruction during magnetic resonance acquisition. This method is based on the Neural Proximal Gradient Descent (NPGD) algorithm, and by introducing a convolutional recurrent structure, a multi-scale expansion mechanism, and an online reparameterization module into the traditional reconstruction framework, it achieves the goal of significantly improving reconstruction quality while ensuring real-time performance.

[0033] During system operation, the non-uniform frequency domain data (golden angle radial trajectory) acquired by the magnetic resonance scanner is first input to the data consistency update module. This update module performs gradient descent correction on the reconstruction result of the current frame based on the principle of proximal gradient descent to ensure consistency with the original sampled data. This step uses a non-uniform fast Fourier transform algorithm to calculate the forward and backward projections. The backward projection term can be calculated once per frame and reused in multiple iterations, thereby reducing computational overhead and meeting the latency requirements of real-time reconstruction. Only the iteration step size is a trainable parameter.

[0034] After data consistency correction, the system enters the proximal operator network stage. This stage employs a convolutional recurrent neural network (CRNN) structure to model complex prior constraints and continuously model the dynamic process temporally. The network structure includes two branches: iterative recursion and temporal recursion. Iterative recursion is used to propagate feature states across iterations within the same frame, while temporal recursion is used to propagate temporal context information between adjacent frames, thus achieving both spatial consistency and temporal smoothness. The network as a whole adopts a unidirectional recursive design to adapt to the causal requirements of real-time imaging scenarios. Its backbone consists of multiple recursive layers stacked together, forming a hierarchical structure of "tiCRNN–iCRNN–iCRNN–iCRNN–CNN". Each recursive layer receives features from the previous layer and updates them by combining the hidden states of the current frame, the previous frame, and the previous iteration, thereby dynamically fusing information across the temporal domain and across iterations, ultimately outputting the reconstruction result through residual connections. The alternating iterative processes of the data consistency module and the proximal operator network optimization module constitute the overall reconstruction framework, specifically consisting of... Figure 1 As shown. The optimized design and specific structure of the network will be described later.

[0035] In this invention, each convolutional operation unit is replaced by an Online Re-parameterization (OR) module instead of a traditional convolutional structure. This module introduces a multi-branch parallel convolutional form during the training phase to enhance the diversity of feature representation and the network's nonlinear modeling capabilities. Each convolutional operator is expanded into several parallel branches, each branch consisting of multiple sequentially connected small convolutional kernels, which are weighted and fused using learnable scaling factors. During training, each branch independently participates in feature extraction and parameter updates; while during inference, the system uses algebraic operations to fuse multi-branch convolutional kernels within the same scale into an equivalent single convolutional kernel, achieving an output completely equivalent to the training structure. This fusion process does not rely on statistical normalization parameters; therefore, batch normalization layers are replaced by affine scaling layers in the design, ensuring the analytical feasibility of the fusion. Bias terms are stored independently to prevent fusion errors. This scheme significantly improves the network's expressive power during the training phase while maintaining inference complexity consistent with the standard single convolutional structure, providing higher accuracy and robustness for online reconstruction.

[0036] To further expand the network's effective receptive field and improve texture capture capabilities, this invention introduces a first-order multi-scale extension (MS) mechanism in the intermediate convolutional module. This module adds parallel paths at different resolutions to the main convolutional branch to extract multi-scale spatial features. Specifically, one path performs a standard 3×3 convolution at the original resolution to capture detailed information; the second path performs a 2x downsampling using average pooling, performs convolution in the low-resolution space, and then uses bilinear interpolation to restore the original resolution; the third path uses a 4x downsampling and the same "convolution-interpolation" process to extract structural features at a larger scale. The outputs of the three paths are fused within their respective scales to avoid the high-dimensional computational burden of cross-scale fusion. This design effectively expands the network's receptive field through lightweight linear downsampling and upsampling operations without using large-size or dilated convolutions, thereby enhancing the model's structural expressiveness while maintaining computational efficiency.

[0037] The overall network reconstruction framework is designed as follows: (1) Sequence: Radial trajectory sampling mode is used, and the contrast uses FLASH and bSSFP sequence structure to improve acquisition efficiency. (2) Reconstruction process: Data is continuously acquired. In the initial stage (1s), the data is used to calculate correction parameters, sensitivity, etc. After that, every 14 sampling lines are used to reconstruct one frame of data. First, a low-quality result is obtained by fast non-uniform Fourier transform. Then, the data is alternately optimized by gradient descent based on data consistency and image optimization based on the network. The number of iterations is changed according to the needs of this iterative process. (3) Image optimization network: The specific structure can be seen in the image optimization network. Figure 2 The network's backbone is a 5-layer convolutional recurrent neural network. The first layer introduces recurrent structures for both time and iteration dimensions, while layers two through four introduce recurrent structures for iteration dimensions. The final layer is a simple convolution. ReLU activation is used between convolutions, and all convolutions are replaced with the aforementioned reparameterized staged structure. The intermediate layers use the aforementioned lightweight multi-scale approach to enhance expressive power. During deployment, all intermediate convolutional layers have a fixed 64 channels, and the kernel size uses a 3×3 small kernel structure. The network input includes a 2-channel (real and imaginary) image from the forward convolution, 64-channel temporal hidden layer data, and four 64-channel iterative hidden layer data. All designs aim for the highest possible computational efficiency for reconstruction. Test results show that 8-coil data and an image scale of 208... Under 208 conditions, the reconstruction time is approximately 33ms (30fps).

[0038] Training data utilized multiple publicly available MRI datasets, with some non-radial sampling data simulated through manual sampling. Simulated instrument intervention processes were used to construct the dataset, along with datasets for upper-level interventional navigation within vascular models. To conserve GPU memory, continuous sequences were manually segmented into approximately 20-frame segments for training. The basic MAE loss function was used; however, to accommodate this segmented data format, an exponentially decaying weight penalty was applied to the first 10 frames of reconstruction to prevent optimization direction errors. The Adamw optimizer was used, optimizing for 20,000 epochs at a learning rate of 0.001, followed by 5,000 epochs at a learning rate of 0.0001 to obtain the trained network parameters. Further optimizations, such as pruning, can be performed during deployment to further improve computational efficiency.

[0039] In the actual reconstruction process, the system executes the above update process frame by frame. After each acquisition, the new frequency domain data first enters the data consistency module for correction, and then is input into the near-end network module to complete the nonlinear reconstruction. The network depends only on the feature states of the previous frame in time, without using future information, thus ensuring the causality and real-time response capability of the algorithm. The reconstruction result of each frame is combined with the previous result through residual connections to form a smooth and stable dynamic reconstruction sequence. The system can run continuously during magnetic resonance scanning, realizing an online imaging mode of simultaneous acquisition and reconstruction.

[0040] During the inference phase, after multi-branch fusion and scale simplification, the model contains only a few standard convolution operators and simple linear downsampling and interpolation operations. Its parameter count and computational complexity are basically equivalent to those of traditional single-convolutional CRNN models. Experiments show that the system can run stably under real-time magnetic resonance imaging (MRI) conditions, with a reconstruction latency of less than 100 milliseconds per frame, significantly improving image quality while maintaining the real-time frame rate. By introducing multi-scale paths, the system gains a larger effective receptive field and stronger detail recovery capabilities; the online reparameterization mechanism significantly improves the network's nonlinear expressive power and generalization ability without increasing the inference burden. The overall solution has a clear structure, strong deployability, and controllable computational load, and can be seamlessly embedded into existing MRI acquisition systems to achieve high-precision, low-latency dynamic reconstruction, providing an efficient and feasible technical approach for interventional MRI.

[0041] This technical solution, based on a multi-scale online reparameterized convolutional recurrent neural network (MSOR-CRNN) and its cross-view imaging strategy, achieves real-time, high-quality image reconstruction and equipment spatial localization in interventional magnetic resonance imaging (IMR). This method can be directly embedded into existing real-time imaging and reconstruction chains without altering the hardware structure of the MRI system, requiring only the addition of a software-based reconstruction module and image synchronization processing logic, thus exhibiting good compatibility and scalability.

[0042] Unlike existing imaging algorithms based on fixed structures or offline reconstruction, this technical solution introduces online re-parameterization (OR) and multi-scale feature propagation (MSFP) mechanisms. This allows for high-capacity representation during the training phase and the fusion and compression of computational graphs during deployment, thus balancing imaging accuracy and inference efficiency. This design significantly reduces MRI reconstruction latency, enabling biplane reconstruction to be completed in less than 60 milliseconds per frame, meeting the stringent real-time requirements of interventional procedures.

[0043] In terms of hardware system compatibility, this solution can be directly integrated into existing GPU or FPGA acceleration platforms without additional peripherals or data acquisition modifications, and can work collaboratively with mainstream non-Cartesian sampling (such as radial and helical trajectory) acquisition methods. Compared with traditional retrospective data processing or high-latency deep network-based solutions, this technology significantly reduces computational resource consumption, can run stably on edge computing nodes, and provides a feasible technical path for the deployment of real-time interventional imaging systems.

[0044] like Figure 3 As shown, compared to traditional zero-fill direct reconstruction methods, the method of this invention, by introducing an improved reconstruction algorithm, can significantly improve image resolution during image reconstruction, optimize the detail representation in the reconstructed image, and make the imaging results clearer and more accurate. The method of this invention also shows a significant advantage in reducing artifacts. Traditional zero-fill methods often produce artifacts due to insufficient image sampling, affecting image quality. However, this invention, through its improved reconstruction strategy, effectively reduces the generation of these artifacts, thereby further improving the reliability and accuracy of the image. While ensuring accelerated imaging, it significantly optimizes the visual effect of the image, providing higher-quality imaging data for clinical applications.

[0045] In addition, such as Figure 4 As shown, by combining the cross-view imaging mechanism of this scheme, signals can be acquired alternately between two sets of orthogonal planes, enabling real-time positioning of interventional instruments such as catheters and guidewires in three-dimensional space. Through the highly spatiotemporally consistent MSOR-CRNN reconstruction module, high-quality images of signal intersection areas can be stably restored, reducing blurring caused by motion artifacts and signal overlap. Experimental results show that, in conventional multi-planar short TR sequence acquisition, for a target reconstruction size of 192×192, the system can continuously output dual-view real-time images with a reconstruction delay of 54.4 milliseconds, achieving continuous monitoring of catheter path and spatial trajectory estimation, providing real-time navigation support for clinical interventional procedures.

[0046] In summary, this technical solution combines high efficiency, real-time performance, and system compatibility, and can significantly improve the practical performance of magnetic resonance interventional imaging without the need for additional hardware modifications.

[0047] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for reconstructing interventional magnetic resonance images, characterized in that, include, S1: Real-time acquisition of non-uniform frequency domain data using a magnetic resonance scanner; S2: Perform data consistency updates on non-uniform frequency domain data; S3: Input the updated data into the neural network to obtain the reconstruction results.

2. The magnetic resonance imaging reconstruction method as described in claim 1, characterized in that, The method for updating the data consistency of non-uniform frequency domain data in step S2 includes: performing a gradient descent correction on the reconstruction result of the current frame based on the principle of near-end gradient descent to ensure that it is consistent with the original sampled data.

3. The magnetic resonance imaging reconstruction method as described in claim 1, characterized in that, The neural network adopts a convolutional recurrent neural network structure, which includes an iterative recurrent branch and a temporal recurrent branch. The iterative recurrent branch is used to pass feature states across iterations within the same frame, and the temporal recurrent branch is used to propagate temporal context information between adjacent frames.

4. The magnetic resonance imaging reconstruction method as described in claim 3, characterized in that, The convolutional recurrent neural network adopts a unidirectional recursive design and is divided into five layers, namely tiCRNN, iCRNN, iCRNN, iCRNN, and CNN. Each recursive layer receives the features from the previous layer and updates them by combining the hidden states of the current frame, the previous frame, and the previous iteration. Finally, it outputs the reconstruction result through residual connections. The first layer includes a recurrent structure with time and iteration dimensions, any one of the second to fourth layers includes a recurrent structure with iteration dimensions, and the fifth layer is implemented with simple convolution.

5. The magnetic resonance imaging reconstruction method as described in claim 4, characterized in that, Each convolutional operation unit in the described convolutional recurrent neural network is an online reparameterization module (OR). Each convolutional operator in the OR is expanded into multiple parallel branches. Each parallel branch consists of multiple sequentially connected small convolutional kernels, which are weighted and fused using learnable scaling factors. During the training phase, each parallel branch independently participates in feature extraction and parameter update. During the inference phase, algebraic operations are used to fuse multi-branch convolutional kernels within the same scale into an equivalent single convolutional kernel, achieving an output completely equivalent to the training structure. The batch normalization layer of the OR is an affine scaling layer, and the bias term is stored independently.

6. The magnetic resonance imaging reconstruction method as described in claim 5, characterized in that, The third layer of the convolutional recurrent neural network includes a first-order multi-scale expansion mechanism (MS) module. The MS module adds parallel paths of different resolutions on the basis of the main convolutional branch to extract multi-scale spatial features.

7. The magnetic resonance imaging reconstruction method as described in claim 6, characterized in that, The parallel paths include three paths. The first parallel path performs a standard 3×3 convolution at the original resolution. The second parallel path performs a 2x downsampling through average pooling. After performing convolution in the low-resolution space, bilinear interpolation is used to restore the original resolution. The third parallel path uses a 4x downsampling and the same "convolution-interpolation" process to extract structural features at a larger scale.

8. The magnetic resonance imaging reconstruction method as described in claim 7, characterized in that, The convolutional recurrent neural network has a structure where any convolution in the third layer has 64 channels and a kernel size of 3×3.

9. A computer device, characterized in that, include: Memory; One or more processors are coupled to the memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by one or more processors, the one or more applications being configured to perform the reconstruction method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the reconstruction method as described in any one of claims 1 to 8.

Images