Magnetic resonance image reconstruction method, apparatus, device, storage medium, and program product

Abstract

The application relates to the technical field of image reconstruction, and provides a magnetic resonance image reconstruction method, device, equipment, storage medium and program product, wherein the method comprises the following steps: acquiring undersampled non-Cartesian K-space data of magnetic resonance imaging; inputting the undersampled non-Cartesian K-space data into a magnetic resonance image reconstruction model to obtain a reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model; the magnetic resonance image reconstruction model is obtained by training sample undersampled non-Cartesian K-space data; the magnetic resonance image reconstruction model is used for mapping the undersampled non-Cartesian K-space data into undersampled Cartesian K-space data and reconstructing the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data. The data mapping in the magnetic resonance image reconstruction model replaces time-consuming physical mathematical operators, the calculation amount is reduced, and the reconstruction efficiency of the magnetic resonance image is improved, so that the real-time requirement of the clinic is met.

Description

Technical Field

[0001] This invention relates to the field of image reconstruction technology, and in particular to a magnetic resonance image reconstruction method, apparatus, device, storage medium, and program product. Background Technology

[0002] Magnetic resonance imaging (MRI), as a non-invasive medical imaging technique, plays an irreplaceable role in clinical diagnosis. Dynamic MRI, such as Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) and cardiac cine imaging, captures the dynamic physiological processes of target organs by adding a temporal dimension encoding on top of the spatial dimension encoding. However, the inherent physical limitations of MRI result in a relatively slow data acquisition speed, which severely restricts the temporal resolution of dynamic imaging, easily leading to the loss of information at key time points or the generation of motion artifacts.

[0003] To accelerate data acquisition, undersampled non-Cartesian sampling trajectories (such as golden angle radial sampling and spiral sampling) are often used. Non-Cartesian sampling has strong resistance to motion artifacts and supports flexible retrospective time window reconstruction. However, since the acquired K-space data is non-uniformly distributed, its image reconstruction process faces significant computational challenges. Therefore, how to perform magnetic resonance image reconstruction on undersampled non-Cartesian K-space data is a pressing need.

[0004] Currently, reconstruction of undersampled non-Cartesian K-space data typically relies on iterative model-based reconstruction methods (such as combining parallel imaging and compressed sensing). However, existing techniques require repeated execution of the Non-Uniform Fast Fourier Transform (NUFFT) and its conjugate transpose operations across the entire time series during reconstruction. The computational load of these operators, based on fixed mathematical and physical formulas, is extremely large, resulting in very lengthy image reconstruction times that cannot meet the clinical demand for rapid, real-time image output.

[0005] In summary, how to achieve faster magnetic resonance image reconstruction is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes a magnetic resonance image reconstruction method that replaces time-consuming physical and mathematical operators with data mapping within the magnetic resonance image reconstruction model, completely eliminating the most time-consuming computational bottleneck in the reconstruction process, thus reducing the computational load and improving the reconstruction efficiency of magnetic resonance images to meet the real-time requirements of clinical practice.

[0007] The present invention also provides a magnetic resonance image reconstruction apparatus, an electronic device, a non-transitory computer-readable storage medium, and a computer program product.

[0008] A magnetic resonance image reconstruction method according to a first aspect of the present invention includes: Acquire undersampled non-Cartesian K-space data for magnetic resonance imaging; The undersampled non-Cartesian K-space data is input into the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model; The magnetic resonance image reconstruction model is trained based on undersampled non-Cartesian K-space data. The magnetic resonance image reconstruction model is used to map the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, and to reconstruct the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data.

[0009] According to the magnetic resonance image reconstruction method of this invention, undersampled non-Cartesian K-space data of magnetic resonance imaging is acquired, and the undersampled non-Cartesian K-space data is input into a magnetic resonance image reconstruction model to obtain a reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model. The magnetic resonance image reconstruction model is used to map the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data. This data mapping within the magnetic resonance image reconstruction model replaces time-consuming physical and mathematical operators, completely eliminating the most time-consuming computational bottleneck in the reconstruction process, thus reducing the computational load and improving the reconstruction efficiency of magnetic resonance images to meet clinical real-time requirements. Simultaneously, the magnetic resonance image reconstruction model is trained based on sample undersampled non-Cartesian K-space data, i.e., joint optimization of the magnetic resonance image reconstruction model achieves an end-to-end deep learning model. This avoids the impact of errors in the preceding modules on the processing accuracy of subsequent modules in traditional modular processing, thereby improving the reconstruction quality of magnetic resonance images to meet clinical image quality requirements.

[0010] According to an embodiment of the present invention, the step of inputting the undersampled non-Cartesian K-space data into a magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model includes: The undersampled non-Cartesian K-space data is input into the non-Cartesian interpolation weight estimation layer in the magnetic resonance image reconstruction model to obtain the interpolation weights output by the non-Cartesian interpolation weight estimation layer; the non-Cartesian interpolation weight estimation layer is used to generate interpolation weights for non-Cartesian coordinate to Cartesian coordinate mapping based on the undersampled non-Cartesian K-space data. The interpolation weights and the undersampled non-Cartesian K-space data are input into the meshing layer of the magnetic resonance image reconstruction model to obtain the undersampled Cartesian K-space data output by the meshing layer; the meshing layer is used to mesh the undersampled non-Cartesian K-space data into undersampled Cartesian K-space data. The undersampled Cartesian K-space data is input into the image reconstruction layer of the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the image reconstruction layer.

[0011] According to one embodiment of the present invention, the step of inputting the undersampled Cartesian K-space data into the image reconstruction layer of the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the image reconstruction layer includes: The undersampled non-Cartesian K-space data is input into the coil sensitivity estimation layer in the magnetic resonance image reconstruction model to obtain the coil sensitivity map output by the coil sensitivity estimation layer. The undersampled Cartesian K-space data and the coil sensitivity map are input into the image reconstruction layer to obtain the reconstructed magnetic resonance image output by the image reconstruction layer.

[0012] According to one embodiment of the present invention, the step of inputting the undersampled non-Cartesian K-space data into a non-Cartesian interpolation weight estimation layer in a magnetic resonance image reconstruction model to obtain the interpolation weights output by the non-Cartesian interpolation weight estimation layer includes: Extract the data from the central region of the K-space of the undersampled non-Cartesian K-space data to obtain self-calibrated signal data; The self-calibrated signal data is input into the non-Cartesian interpolation weight estimation layer to obtain the interpolation weights output by the non-Cartesian interpolation weight estimation layer.

[0013] According to one embodiment of the present invention, the undersampled non-Cartesian K-space data includes non-Cartesian K-space data of a dynamic magnetic resonance imaging sequence, the dynamic magnetic resonance imaging sequence includes multiple consecutive frames of magnetic resonance imaging, and the undersampled non-Cartesian K-space data is acquired by a golden angle non-Cartesian acquisition method; The step of extracting data from the central region of the K-space of the undersampled non-Cartesian K-space data to obtain self-calibration signal data includes: The undersampled non-Cartesian K-space data is rearranged in the time dimension to obtain rearranged K-space data. Data from the central region of the rearranged K-space data is extracted to obtain self-calibration signal data.

[0014] According to one embodiment of the present invention, the image reconstruction layer includes a plurality of cascaded image sub-reconstruction layers, each of which integrates a denoising unit and a data fidelity unit; The noise reduction unit is used to remove noise in magnetic resonance imaging; The data fidelity unit is used to ensure the consistency between the reconstructed magnetic resonance image and the undersampled Cartesian K-space data.

[0015] According to one embodiment of the present invention, the magnetic resonance image reconstruction model is trained in the following manner: The undersampled non-Cartesian K-space data of the sample is input into the model to be trained to obtain the reconstructed magnetic resonance image of the sample output by the model to be trained; The loss function value is determined based on the magnetic resonance image labels corresponding to the sample magnetic resonance images and the sample undersampled non-Cartesian K-space data; the magnetic resonance image labels are reconstructed based on fully sampled K-space data. Based on the loss function value, the model parameters of the model to be trained are updated.

[0016] According to one embodiment of the present invention, the undersampled non-Cartesian K-space data includes non-Cartesian K-space data of a dynamic magnetic resonance imaging sequence; the dynamic magnetic resonance imaging sequence includes multiple consecutive frames of magnetic resonance imaging.

[0017] A magnetic resonance image reconstruction apparatus according to a second aspect of the present invention includes: The data acquisition module is used to acquire undersampled non-Cartesian K-space data from magnetic resonance imaging. The image reconstruction module is used to input the undersampled non-Cartesian K-space data into the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model. The magnetic resonance image reconstruction model is trained based on undersampled non-Cartesian K-space data. The magnetic resonance image reconstruction model is used to map the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, and to reconstruct the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data.

[0018] An electronic device according to a third aspect of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the magnetic resonance image reconstruction method as described above.

[0019] According to a fourth aspect of the present invention, a non-transitory computer-readable storage medium is provided thereon storing a computer program that, when executed by a processor, implements the magnetic resonance image reconstruction method as described above.

[0020] A computer program product according to a fifth aspect of the present invention includes a computer program that, when executed by a processor, implements the magnetic resonance image reconstruction method as described above.

[0021] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: Undersampled non-Cartesian K-space data from magnetic resonance imaging (MRI) is acquired and input into an MRI image reconstruction model. The resulting reconstructed MRI image is generated from this model. The MRI image reconstruction model maps the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data. This data mapping within the MRI image reconstruction model replaces time-consuming physical and mathematical operators, completely eliminating the most computationally intensive bottleneck in the reconstruction process. This reduces computational load and improves the reconstruction efficiency of MRI images to meet real-time clinical requirements. Furthermore, the MRI image reconstruction model is trained based on sample undersampled non-Cartesian K-space data, meaning it undergoes joint optimization to achieve an end-to-end deep learning model. This avoids the impact of errors in earlier modules on the processing accuracy of subsequent modules in traditional modular processing, thereby improving the reconstruction quality of MRI images to meet clinical image quality requirements.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is one of the flowcharts of the magnetic resonance image reconstruction method provided in the embodiments of the present invention.

[0025] Figure 2 This is the second schematic flowchart of the magnetic resonance image reconstruction method provided in the embodiments of the present invention.

[0026] Figure 3 This is the third schematic flowchart of the magnetic resonance image reconstruction method provided in the embodiments of the present invention.

[0027] Figure 4 This is a schematic diagram of the magnetic resonance image reconstruction device provided in an embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] Currently, for magnetic resonance image reconstruction using undersampled non-Cartesian K-space data, traditional iterative algorithms (such as compressed sensing and parallel imaging) heavily rely on the Non-Uniform Fourier Transform (NUFFT) operator for repeated forward and backward operations, resulting in enormous computational costs and extremely slow reconstruction speeds. For example, combining parallel imaging with coil sensitivity information and compressed sensing utilizing temporal sparsity for image reconstruction, the optimized model is shown below: ; In the formula, This represents the dynamic magnetic resonance imaging sequence to be reconstructed. This indicates undersampled non-Cartesian K-space data. Indicates an undersampling mask. Indicates the relationship with dynamic magnetic resonance imaging sequences The corresponding non-uniform Fourier transform operator, Represents the coil sensitivity operator. The first-order difference operator represents the time dimension. This represents the regularization coefficient.

[0031] To address the aforementioned speed bottleneck, the inventors of this invention began exploring the use of deep learning technology. However, their research revealed that deep learning-based magnetic resonance image reconstruction schemes, when processing non-Cartesian data, typically assume that the transformation from non-Cartesian coordinates to Cartesian coordinates (i.e., the meshing process) must rely on explicit operators with rigorous mathematical and physical definitions (such as standard NUFFT or pre-computed density compensation functions). Therefore, this deep learning scheme often treats the neural network merely as a local module within the traditional reconstruction process (e.g., only for image domain denoising or as a regularization term). Before the data is input into the neural network, extensive preprocessing computations using traditional fixed operators are still required. This non-end-to-end architecture, separating "fixed preprocessing mapping" from "post-processing network reconstruction," introduces significant drawbacks: firstly, the failure to eliminate fixed operator computations does not fundamentally solve the problem of lengthy reconstruction times; secondly, the fixed preprocessing mapping process cannot perceive the needs of the subsequent reconstruction network, and the meshing errors inevitably introduced in the preprocessing stage are directly transmitted and amplified, limiting the potential for improving the final image quality.

[0032] Based on a profound understanding of the aforementioned technical shortcomings, this invention proposes a novel inventive concept: completely breaking the technical bias that "grid mapping must rely on explicit mathematical-physical operators," and constructing a fully data-driven, end-to-end magnetic resonance image reconstruction model. Specifically, the "mapping process from non-Cartesian data to Cartesian data" and the "image reconstruction process" are integrated into a single deep learning framework for joint optimization. This invention utilizes a pre-trained magnetic resonance image reconstruction model to directly receive raw, undersampled non-Cartesian K-space data; this magnetic resonance image reconstruction model implicitly learns and performs non-Cartesian to Cartesian coordinate mapping internally, and then directly outputs a high-quality reconstructed magnetic resonance image based on the mapped undersampled Cartesian K-space data.

[0033] Based on the above-described inventive concept, the present invention proposes the following embodiments. The magnetic resonance image reconstruction method provided by the embodiments of the present invention is described below with reference to the accompanying drawings. The executing entity of this magnetic resonance image reconstruction method can be a magnetic resonance image reconstruction system, a server, a desktop computer, a laptop computer, or a user's terminal, including but not limited to mobile phones, tablet computers, medical devices, vehicle terminals, and smart home appliances.

[0034] Figure 1 This is one of the flowcharts illustrating the magnetic resonance image reconstruction method provided in this embodiment of the invention, such as... Figure 1 As shown, the magnetic resonance image reconstruction method includes steps 110 and 120.

[0035] Step 110: Obtain undersampled non-Cartesian K-space data from magnetic resonance imaging.

[0036] Here, magnetic resonance imaging (MRI) is a non-invasive medical imaging technique that utilizes the principles of nuclear magnetic resonance to acquire radio frequency signals emitted by internal tissues of the human body and reconstruct images by applying gradient magnetic fields and radio frequency pulses. Furthermore, multi-dynamic magnetic resonance imaging (MMRI) adds a time dimension encoding to this process, enabling the capture of dynamic processes in target organs (such as the heartbeat).

[0037] Here, undersampled non-Cartesian K-space data refers to K-space data obtained through non-Cartesian sampling. K-space data is the raw spatial frequency domain data directly acquired by magnetic resonance imaging (MRI) equipment. MRI images are actually the spatial domain representation of K-space data after Fourier transform or similar mathematical transformation. Non-Cartesian sampling refers to the sampling trajectory not following a regular, mutually perpendicular grid (Cartesian grid) when acquiring K-space data. Instead, it uses irregular trajectories such as golden angle radial sampling or spiral sampling. This sampling method has strong resistance to motion artifacts and supports flexible retrospective time window reconstruction. Undersampling refers to acquiring less data than the required fully sampled data in order to accelerate the data acquisition process.

[0038] It should be understood that by acquiring undersampled non-Cartesian K-space data, the data acquisition speed of magnetic resonance imaging can be significantly improved, reducing patient scanning time and discomfort. At the same time, data acquired using non-Cartesian trajectories can provide better resistance to motion artifacts while accelerating acquisition, providing basic data support for subsequent high-quality image reconstruction.

[0039] Step 120: Input the undersampled non-Cartesian K-space data into the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model.

[0040] The magnetic resonance image reconstruction model is trained based on undersampled non-Cartesian K-space data. The magnetic resonance image reconstruction model is used to map the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, and to reconstruct the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data.

[0041] Here, the magnetic resonance image reconstruction model is an end-to-end deep learning model. End-to-end means that the magnetic resonance image reconstruction model directly uses the original undersampled non-Cartesian K-space data as input, without any manual intervention or traditional independent explicit mathematical calculation modules (such as the non-uniform Fourier transform operator or offline coil sensitivity calculation step that must be set up separately in the traditional process). The output of the magnetic resonance image reconstruction model is the final reconstructed magnetic resonance image.

[0042] In one specific embodiment, by pre-constructing a training set containing a large number of undersampled non-Cartesian K-space data and their corresponding high-quality fully sampled magnetic resonance image labels, the network parameters are updated using backpropagation with a loss function, enabling the magnetic resonance image reconstruction model to learn the optimal mapping rule from undersampled non-Cartesian K-space data to high-quality magnetic resonance images.

[0043] Here, the undersampled Cartesian K-space data refers to undersampled Cartesian-sampled K-space data. It should be noted that in traditional model-based iterative reconstruction, non-Cartesian data cannot be directly converted using the Fast Fourier Transform (FFT). It requires computationally intensive non-uniform Fourier Transform operators and density compensation functions (DCF) for repeated iterative transformations. Therefore, in this embodiment of the invention, the magnetic resonance image reconstruction model implicitly learns the data mapping from non-Cartesian coordinates to Cartesian regular grid coordinates within its internal network structure.

[0044] Furthermore, the magnetic resonance image reconstruction model is also used to: estimate coil sensitivity based on undersampled non-Cartesian K-space data to obtain a coil sensitivity map, and then reconstruct the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data and the coil sensitivity map. In a specific embodiment, the undersampled non-Cartesian K-space data is input into the coil sensitivity estimation layer in the magnetic resonance image reconstruction model to obtain the coil sensitivity map output by the coil sensitivity estimation layer; the undersampled Cartesian K-space data and the coil sensitivity map are input into the image reconstruction layer in the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the image reconstruction layer.

[0045] It should be understood that by directly mapping non-Cartesian to Cartesian data using a trained magnetic resonance imaging (MRI) image reconstruction model, the traditional iterative reconstruction method completely replaces and avoids the repeated execution of the NUFFT forward and inverse transform operators involving the entire time series, significantly reducing computational load and thus substantially shortening the MRI image reconstruction time, thereby improving MRI image reconstruction efficiency. Simultaneously, the end-to-end direct output of reconstructed images eliminates cumbersome intermediate preprocessing steps, greatly enhancing the efficiency of clinical image processing.

[0046] It should be understood that the magnetic resonance image reconstruction model is trained based on undersampled non-Cartesian K-space data. During the training process, the magnetic resonance image reconstruction model automatically learns the optimal mapping method and artifact removal regularization strategy. This avoids the defects of tedious manual parameter tuning for different speed-up factors and different sparsity characteristics in traditional iterative methods. As a result, the embodiments of the present invention have natural compatibility and strong robustness to different speed-up factors and sampling densities, that is, improve the robustness of magnetic resonance image reconstruction.

[0047] The magnetic resonance image reconstruction method provided in this invention acquires undersampled non-Cartesian K-space data from magnetic resonance imaging, inputs this undersampled non-Cartesian K-space data into a magnetic resonance image reconstruction model, and obtains the reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model. The magnetic resonance image reconstruction model maps the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, thereby replacing time-consuming physical and mathematical operators with data mapping within the magnetic resonance image reconstruction model. This completely eliminates the most time-consuming computational bottleneck in the reconstruction process, reducing computational load and improving the reconstruction efficiency of magnetic resonance images to meet real-time clinical needs. Furthermore, the magnetic resonance image reconstruction model is trained based on sample undersampled non-Cartesian K-space data, meaning it undergoes joint optimization to achieve an end-to-end deep learning model. This avoids the error of the preceding module affecting the processing accuracy of subsequent modules in traditional modular processing, thereby improving the reconstruction quality of magnetic resonance images to meet clinical image quality requirements.

[0048] Based on any of the above embodiments Figure 2 This is a second schematic flowchart of the magnetic resonance image reconstruction method provided in this embodiment of the invention, as shown below. Figure 2 As shown, step 120 above includes steps 121, 122 and 123.

[0049] Step 121: Input the undersampled non-Cartesian K-space data into the non-Cartesian interpolation weight estimation layer in the magnetic resonance image reconstruction model to obtain the interpolation weights output by the non-Cartesian interpolation weight estimation layer.

[0050] The non-Cartesian interpolation weight estimation layer is used to generate interpolation weights for mapping non-Cartesian coordinates to Cartesian coordinates based on the undersampled non-Cartesian K-space data.

[0051] Here, the non-Cartesian interpolation weight estimation layer is used to address the coordinate inhomogeneity problem caused by non-Cartesian sampling. Specifically, the non-Cartesian interpolation weight estimation layer is configured to implicitly learn mapping weights containing sampling trajectory information and density compensation information from the input undersampled non-Cartesian K-space data.

[0052] It should be noted that existing iterative reconstruction methods require explicit design and computation of a density compensation function (DCF) to correct the energy inhomogeneity caused by non-Cartesian sampling (such as radial sampling with dense centers and sparse edges). However, in this embodiment of the invention, a non-Cartesian interpolation weight estimation layer directly extracts this physical information from the intrinsic structure of the data and integrates it into the interpolation weights.

[0053] It should be understood that by implicitly learning mapping weights containing sampling trajectory information and density compensation information through non-Cartesian interpolation weight estimation layers, the traditional complex explicit non-uniform Fourier transform (NUFFT) calculation and explicit density compensation function (DCF) design are completely replaced. This data-driven interpolation weight generation method not only avoids the errors introduced by manually designing DCF, but also obtains customized interpolation weights that are optimal and most suitable for subsequent reconstruction tasks, thereby improving the reconstruction quality of magnetic resonance images.

[0054] Step 122: Input the interpolation weights and the undersampled non-Cartesian K-space data into the meshing layer of the magnetic resonance image reconstruction model to obtain the undersampled Cartesian K-space data output by the meshing layer.

[0055] The gridding layer is used to grid the undersampled non-Cartesian K-space data into undersampled Cartesian K-space data.

[0056] Here, the gridding layer acts as a bridge connecting the non-Cartesian and Cartesian domains. Gridding refers to the process of redistributing and interpolating K-space data points, originally scattered on irregular trajectories (such as golden-angle radial spokes), onto regular, orthogonal Cartesian grid points using interpolation weights. Since the interpolation weights have already been optimized by a deep learning network, this gridding process is an efficient one-way mapping process.

[0057] It should be understood that combining pre-calculated interpolation weights with undersampled non-Cartesian K-space data for meshing achieves a one-time, efficient conversion from non-Cartesian to Cartesian data format. This implicit meshing module replaces the traditional explicit NUFFT / DCF operation, completely eliminating the most time-consuming process of repeatedly performing NUFFT forward and inverse transform calculations in the traditional iterative reconstruction process, greatly improving the computational efficiency of front-end data processing, i.e., improving the reconstruction efficiency of magnetic resonance images.

[0058] Step 123: Input the undersampled Cartesian K-space data into the image reconstruction layer of the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the image reconstruction layer.

[0059] Here, the image reconstruction layer is the core structure responsible for image reconstruction in the magnetic resonance image reconstruction model. For example, the image reconstruction layer is used to remove undersampling artifacts and restore image details. It receives the Cartesian domain undersampling data after it has been processed by the gridding layer (at this point, the data is already on a regular grid, and standard deep learning operations such as convolution can be directly applied), and maps it directly to the final magnetic resonance image through the network's forward propagation.

[0060] In one embodiment, the image reconstruction layer may employ an expanded network architecture comprising multiple cascaded reconstruction modules, expanding the traditional physical iterative optimization process into a fixed-depth network forward inference. Specifically, the image reconstruction layer includes multiple cascaded image sub-reconstruction layers.

[0061] It should be understood that the image reconstruction layer directly maps the gridded, undersampled Cartesian K-space data to the final reconstructed image, avoiding the repeated calculation of conjugate gradients and other optimization solutions in traditional iterative reconstruction, thereby reducing computational load and improving the reconstruction efficiency of magnetic resonance images. Furthermore, the data only needs to undergo one forward pass in the image reconstruction layer to complete artifact removal and detail restoration, significantly shortening the overall image reconstruction time.

[0062] The magnetic resonance image reconstruction method provided in this invention comprises a non-Cartesian interpolation weight estimation layer, a meshing layer, and an image reconstruction layer connected in sequence. The method first outputs interpolation weights, then performs meshing, and finally reconstructs the image, thus constructing an end-to-end deep learning model. This avoids the impact of errors in the preceding modules on the processing accuracy of subsequent modules in traditional modular processing, thereby improving the reconstruction quality of magnetic resonance images to meet clinical image quality requirements. Simultaneously, the sequentially connected interpolation weight estimation and meshing layers directly and instantly solve the coordinate mapping problem within the model, transforming time-consuming iterations into efficient single-pass forward mapping, reducing computational load, and thus improving the reconstruction efficiency of magnetic resonance images to meet real-time clinical needs. Furthermore, unifying these three layers within an end-to-end model, the generation of interpolation weights is guided by the backpropagation of the final image reconstruction error, ensuring that the meshing process serves to obtain the optimal reconstructed image. This effectively reduces traditional meshing errors and prevents irreversible transmission of interpolation errors from the front-end processing to subsequent reconstruction modules, thereby improving the reconstruction quality of magnetic resonance images to meet clinical image quality requirements.

[0063] Based on any of the above embodiments, in this method, step 123 includes: The undersampled non-Cartesian K-space data is input into the coil sensitivity estimation layer in the magnetic resonance image reconstruction model to obtain the coil sensitivity map output by the coil sensitivity estimation layer. The undersampled Cartesian K-space data and the coil sensitivity map are input into the image reconstruction layer to obtain the reconstructed magnetic resonance image output by the image reconstruction layer.

[0064] Here, the coil sensitivity estimation layer is a network layer in the magnetic resonance image reconstruction model specifically designed to extract prior information about the parallel imaging space. This coil sensitivity estimation layer is configured to infer and output a coil sensitivity map directly from the input undersampled non-Cartesian K-space data.

[0065] Here, the Coil Sensitivity Map (CSM) is used in multi-channel parallel magnetic resonance imaging. Because multiple receiving coils are used, each coil has varying sensitivities to signal reception at different spatial locations within the scanning field of view. This distribution of spatial sensitivity is what we call the Coil Sensitivity Map. A high-quality Coil Sensitivity Map is crucial prior information for eliminating spatial aliasing artifacts in parallel imaging.

[0066] It should be understood that traditional iterative reconstruction frameworks typically require pre-calculation of the coil sensitivity map, which often relies on additional calibration scans or extremely complex offline calculation algorithms (such as the ESPIRiT algorithm), which is not only time-consuming but also computationally complex. Therefore, this invention introduces a coil sensitivity estimation layer to directly infer the coil sensitivity map from the input undersampled non-Cartesian K-space data, completely avoiding the reliance on additional calibration scans or complex offline algorithms found in traditional methods. This achieves fast and robust estimation of the coil sensitivity map, thereby improving the reconstruction efficiency and quality of magnetic resonance images.

[0067] It should be understood that by synchronously inputting the prior information in the spatial domain (coil sensitivity map) and the frequency domain data (undersampled Cartesian K-space data) after coordinate unification into the image reconstruction layer, the image reconstruction layer can make full use of the parallel imaging acceleration capability of the multi-channel, and then directly map the fused data to the final high-quality reconstructed image, that is, effectively perform multi-channel parallel reconstruction, and ultimately improve the reconstruction efficiency and reconstruction quality of magnetic resonance images.

[0068] The magnetic resonance image reconstruction method provided in this invention integrates the completely independent coil sensitivity calculation in the traditional upstream process into a single deep learning model. This achieves a fully networked operation from undersampled non-Cartesian K-space data to CSM estimation, meshing, and image reconstruction, significantly simplifying the system complexity and processing flow of magnetic resonance image reconstruction. Specifically, it constructs an end-to-end deep learning model, avoiding the impact of errors in earlier modules on the processing accuracy of subsequent modules in traditional modular processing, thereby improving the reconstruction quality of magnetic resonance images to meet clinical image quality requirements. Simultaneously, in the end-to-end model, the coil sensitivity estimation layer... The coil sensitivity map is jointly optimized with the subsequent image reconstruction layers under the same loss function constraint. Therefore, the coil sensitivity map no longer merely pursues absolute physical accuracy, but rather the sensitivity map that is most conducive to the current image reconstruction layer in eliminating artifacts and improving image quality. This significantly improves the model's generalization ability and robustness at different speedup factors, thereby improving the reconstruction quality of magnetic resonance images to meet clinical image quality requirements. At the same time, since the inference of coil sensitivity becomes a forward propagation process of the network, the computation time of traditional complex algorithms is eliminated, thereby reducing the computational load and improving the reconstruction efficiency of magnetic resonance images to meet the real-time requirements of clinical practice.

[0069] Based on any of the above embodiments, in this method, step 121 includes: Extract the data from the central region of the K-space of the undersampled non-Cartesian K-space data to obtain self-calibrated signal data; The self-calibrated signal data is input into the non-Cartesian interpolation weight estimation layer to obtain the interpolation weights output by the non-Cartesian interpolation weight estimation layer.

[0070] Here, the data in the central region of K-space refers to the low-frequency signals of the image in the K-space (frequency domain) of magnetic resonance imaging. The low-frequency signals contain most of the image's energy and determine the overall outline, contrast, and main basic structure of the image; while the outer region of K-space represents high-frequency signals, which mainly determine the image's edges and details.

[0071] Here, the Auto-Calibration Signal (ACS) refers to a small subset of high-density sampled data extracted specifically from the central region of the K-space obtained from the scan. It serves as a reliable reference source for network learning and estimation of key reconstruction parameters (such as interpolation weights and coil sensitivity) without the need for a separate calibration scan.

[0072] It should be understood that by extracting only the data from the central region of K-space as self-calibration signal data, it is possible to greatly filter out the sparse high-frequency data in the periphery while preserving the low-frequency, high-density information that determines the core structure of the image. This significantly reduces the data dimension and volume that subsequent network layers need to process, thereby reducing the computational load and improving the reconstruction efficiency of magnetic resonance images.

[0073] It's important to note that the non-Cartesian interpolation weight estimation layer no longer receives complete, large-scale undersampled non-Cartesian K-space data. Instead, it only receives small-volume, high-information-density self-calibration signal data (ACS data). The network utilizes this data, which contains densely sampled trajectory information and core contrast information, to implicitly learn and output interpolation weights for the non-Cartesian-to-Cartesian mapping. Based on this, replacing the input data from complete undersampled non-Cartesian K-space data with lightweight self-calibration signal data significantly reduces the number of computational parameters and the computational load of the non-Cartesian interpolation weight estimation layer, significantly accelerating the generation of interpolation weights while ensuring the accuracy of weight estimation, thereby improving the reconstruction efficiency and quality of magnetic resonance images.

[0074] The magnetic resonance image reconstruction method provided in this invention, by extracting the core data with the highest information density (ACS data) to drive the non-Cartesian interpolation weight estimation layer, greatly reduces the computational load and memory usage of the network front-end, making the entire end-to-end deep learning model more lightweight. This further meets the clinical needs of rapid, real-time magnetic resonance imaging, i.e., reducing computational load and thus improving the reconstruction efficiency of magnetic resonance images to meet the real-time requirements of clinical practice. At the same time, it directly extracts data from the central region of the K-space from the acquired undersampled non-Cartesian K-space data as self-calibration signal data, eliminating the need to insert any additional, time-consuming calibration pre-scanning sequences in the clinical scanning process, thereby improving the utilization efficiency of the scanning equipment and the patient's examination experience.

[0075] Based on any of the above embodiments, in this method, the step of inputting the undersampled non-Cartesian K-space data into the coil sensitivity estimation layer of the magnetic resonance image reconstruction model to obtain the coil sensitivity map output by the coil sensitivity estimation layer includes: Extract the data from the central region of the K-space of the undersampled non-Cartesian K-space data to obtain self-calibrated signal data; The self-calibration signal data is input to the coil sensitivity estimation layer to obtain the coil sensitivity map output by the coil sensitivity estimation layer.

[0076] Here, the data in the central region of K-space refers to the low-frequency signals of the image in the K-space (frequency domain) of magnetic resonance imaging. The low-frequency signals contain most of the image's energy and determine the overall outline, contrast, and main basic structure of the image; while the outer region of K-space represents high-frequency signals, which mainly determine the image's edges and details.

[0077] Here, the Auto-Calibration Signal (ACS) refers to a small subset of high-density sampled data extracted specifically from the central region of the K-space obtained from the scan. It serves as a reliable reference source for network learning and estimation of key reconstruction parameters (such as interpolation weights and coil sensitivity) without the need for a separate calibration scan.

[0078] It should be noted that since the spatial sensitivity distribution of the coil (i.e., the coil sensitivity map) is essentially a very flat spatial low-frequency information, the self-calibration signal data can also be input to the coil sensitivity estimation layer to obtain the coil sensitivity map output by the coil sensitivity estimation layer.

[0079] Furthermore, by having the non-Cartesian interpolation weight estimation layer and the coil sensitivity estimation layer share the same set of self-calibration signal data (ACS data) as input, the overall computational architecture of the end-to-end model can be further simplified.

[0080] It should be understood that by extracting only the data from the central region of K-space as self-calibration signal data, it is possible to greatly filter out the sparse high-frequency data in the periphery while preserving the low-frequency, high-density information that determines the core structure of the image. This significantly reduces the data dimension and volume that subsequent network layers need to process, thereby reducing the computational load and improving the reconstruction efficiency of magnetic resonance images.

[0081] It's important to note that the coil sensitivity estimation layer no longer receives complete, large-scale undersampled non-Cartesian K-space data, but instead only receives small-volume, high-information-density self-calibration signal data (ACS data). The network uses this data, which contains densely sampled trajectory information and core contrast information, to implicitly learn and output a sensitivity map. Based on this, replacing the input data from complete undersampled non-Cartesian K-space data with lightweight self-calibration signal data significantly reduces the number of computational parameters and the computational load of the coil sensitivity estimation layer, significantly accelerating the generation speed of the coil sensitivity map while ensuring its accuracy, thereby improving the reconstruction efficiency and quality of magnetic resonance images.

[0082] The magnetic resonance image reconstruction method provided in this invention, by extracting the core data with the highest information density (ACS data) to drive the coil sensitivity estimation layer, greatly reduces the computational load and memory usage of the network front-end, making the entire end-to-end deep learning model more lightweight. This further meets the clinical needs of rapid, real-time magnetic resonance imaging, i.e., reducing computational load and thus improving the reconstruction efficiency of magnetic resonance images to meet the real-time requirements of clinical practice. At the same time, it directly extracts data from the central region of the K-space from the acquired undersampled non-Cartesian K-space data as self-calibration signal data, eliminating the need to insert any additional, time-consuming calibration pre-scanning sequences in the clinical scanning process, thereby improving the utilization efficiency of the scanning equipment and the patient's examination experience.

[0083] Based on any of the above embodiments, in this method, the undersampled non-Cartesian K-space data includes non-Cartesian K-space data of a dynamic magnetic resonance imaging sequence, the dynamic magnetic resonance imaging sequence includes multiple consecutive frames of magnetic resonance imaging, and the undersampled non-Cartesian K-space data is acquired by a golden angle non-Cartesian acquisition method.

[0084] Here, dynamic magnetic resonance imaging sequence refers to a series of images acquired sequentially over time at the same anatomical location. It includes multiple consecutive frames of magnetic resonance imaging and is used to reflect the physiological or pathological processes of tissues and organs over time.

[0085] Here, the golden angle non-Cartesian acquisition method is a specific irregular sampling trajectory generation strategy. During the acquisition process, the angle between two consecutive acquisition trajectories (such as spokes in radial acquisition) is set as the golden angle (approximately 137.5 degrees). This acquisition method does not rely on a fixed time window or period.

[0086] It should be understood that by limiting the acquisition method to the golden angle and not using Cartesian methods, the magnetic resonance imaging process becomes insensitive to physiological movements (such as breathing and heartbeat), thus possessing extremely strong resistance to motion artifacts. At the same time, due to the irrational nature of the golden angle, the distribution of data acquired over any time span in the K-space is ensured to be nearly uniform, providing a theoretical and physical basis for subsequent flexible retrospective data reconstruction.

[0087] Accordingly, the step of extracting data from the central region of the K-space of the undersampled non-Cartesian K-space data to obtain self-calibration signal data includes: The undersampled non-Cartesian K-space data is rearranged in the time dimension to obtain rearranged K-space data. Data from the central region of the rearranged K-space data is extracted to obtain self-calibration signal data.

[0088] In one specific embodiment, rearranging the undersampled non-Cartesian K-space data in the time dimension means temporarily ignoring the temporal order of multiple frames of data during acquisition, and aggregating and merging the undersampled non-Cartesian K-space data belonging to different time frames (e.g., radial spoke data acquired from all frames) along the time dimension to form an aggregated data set without distinguishing the frame order.

[0089] For example, the readout size corresponding to full-scale radial acquisition is M, and the number of spokes is M. pi / 2; Set the downsampling ratio to n, and the number of spokes sampled in each frame to M. pi / 2 / n; Using the forward operator of NUFFT, T frames of spoke data X{M, M can be generated according to the acquisition order of the golden angle. pi / 2 / n, T, C}, i.e., the undersampled non-Cartesian K-space data is the spoke data X; then, the radially sampled data X at the golden angle is rearranged into {M, M}. pi / 2 / n T, C} and select the central portion as ACS data. {ACSread, ACSspoke, C}, that is, rearranged K-space data into {M, M} pi / 2 / n T, C}, self-calibration signal data is ACS data The data acquisition matrix is ​​M. M, with T frames and C channels. Correspondingly, rearrange the K-space data {M, M... pi / 2 / n When T, C} and interpolation weights G are input together into the gridding layer, the downsampled multi-frame data {M, M, T, C} after gridding can be obtained.

[0090] It should be understood that the fact that undersampled non-Cartesian K-space data includes multiple consecutive frames of magnetic resonance imaging provides the data foundation for cross-frame rearrangement in the time dimension. More importantly, because undersampled non-Cartesian K-space data is acquired using a golden-angle non-Cartesian acquisition method, even when sampling trajectories (spokes) acquired at different time points are mixed together, as long as they are continuously acquired data segments, the reassembled K-space data maintains a high degree of uniformity in the K-space frequency domain. This not only effectively fills the data gaps caused by single-frame undersampling but also avoids excessively uneven frequency domain sampling density or aliasing interference caused by time-dimensional merging.

[0091] In one specific embodiment, after obtaining rearranged K-space data with significantly increased sampling density after time-dimensional rearrangement, low-frequency sampling data located in and around the geometric center of K-space is extracted. This part of the data constitutes self-calibrated signal data (ACS data). It should be understood that extracting data from the central region of the K-space from the rearranged high-density aggregated data can extract a static reference dataset with extremely high signal-to-noise ratio and extremely rich structural information from the dynamically acquired data itself without increasing the additional scanning time.

[0092] The magnetic resonance image reconstruction method provided in this invention fully utilizes the unique uniform distribution properties and temporal redundancy of the golden angle acquisition method. By cleverly rearranging the time dimension, it transforms multiple frames of sparse undersampled data into a single frame of dense static calibration data. This method completely avoids the additional low-resolution full sampling scans required in traditional methods to obtain calibration signals, maximizing the saving of clinical scanning time for magnetic resonance imaging equipment, improving the utilization efficiency of scanning equipment and the patient's examination experience. At the same time, the obtained self-calibration signal data not only has a small data volume and low computational burden, but also possesses extremely high signal-to-noise ratio and accurate basic contrast due to the fusion of center frequency information from multiple time frames. Inputting it into the non-Cartesian interpolation weight estimation layer (and coil sensitivity estimation layer) can significantly improve the accuracy and robustness of the network in learning interpolation weights and coil sensitivity maps, thereby ensuring the ultra-high quality and high fidelity of the final end-to-end image reconstruction from the source, that is, improving the reconstruction quality and efficiency of magnetic resonance images.

[0093] Based on any of the above embodiments, in this method, the image reconstruction layer includes multiple cascaded image sub-reconstruction layers, and each of the image sub-reconstruction layers integrates a denoising unit and a data fidelity unit.

[0094] It's important to note that the traditional model-based iterative optimization process (such as proximal gradient descent in compressed sensing) is expanded with a fixed number of iterations, forming multiple cascaded deep learning network sub-layers. Data undergoes unidirectional forward propagation within these cascaded sub-layers. Based on this, by constructing cascaded image sub-reconstruction layers to form an expanded network, the traditional iterative optimization process, which requires setting stopping conditions and involves repeated iterations, is transformed into a forward mapping process with a fixed-depth neural network. This not only avoids the time-consuming process of repeatedly calculating conjugate gradients in traditional iterative reconstruction but also allows the network to learn the optimal parameters for each iteration end-to-end through backpropagation, significantly improving reconstruction efficiency.

[0095] The denoising unit is used to remove noise in magnetic resonance imaging. In one specific embodiment, the denoising unit includes a low-rank processing unit and a sparse processing unit.

[0096] The low-rank processing unit is used to suppress noise by utilizing the temporal correlation of multi-frame magnetic resonance imaging. In multi-dynamic magnetic resonance imaging, since the human anatomical background remains largely static and organ movements (such as heartbeats) have a certain periodicity and continuity, when the dynamic image sequence is arranged into a spatiotemporal matrix, this matrix often exhibits low-rank characteristics. The low-rank processing unit is a network module designed using this physical prior characteristic. Furthermore, this low-rank processing unit can be embedded in the network through a singular value thresholding (SVT) operation, or approximated and replaced by a deep neural network module with similar functionality.

[0097] It should be understood that by explicitly integrating low-rank processing units, it is possible to effectively capture and utilize the high correlation and redundancy of multiple frames in the temporal dimension, thereby strongly suppressing random noise and motion artifacts in the temporal dimension without blurring dynamic details.

[0098] The sparse processing unit utilizes the sparsity of magnetic resonance imaging in the transform domain to suppress spatial artifacts. While medical images are not sparse in the pixel domain, most coefficients in specific mathematical transform domains (such as wavelet or gradient domains) are close to zero, exhibiting sparsity. The sparse processing unit enhances this sparsity to eliminate spatial aliasing artifacts caused by undersampling. Furthermore, this sparse processing unit can be implemented by applying a soft-thresholding shrinkage operation in a specific transform domain (such as the wavelet domain), or by learning this sparse prior feature through a sub-network composed of convolutional layers, nonlinear activation functions, etc.

[0099] It should be understood that through sparse processing units, the network can finely process and eliminate spatial domain aliasing artifacts caused by non-Cartesian undersampling, while preserving and sharpening the spatial structural features and edge details of the image to the greatest extent.

[0100] The data fidelity unit ensures the consistency between the reconstructed magnetic resonance image and the undersampled Cartesian K-space data. After low-rank and sparse processing, the image data may become overly smoothed or deviate from the actual acquired signal. The data fidelity unit converts the intermediate reconstruction result back to K-space and compares and forces correction with the actually acquired measurement data (i.e., the undersampled Cartesian K-space data output by the meshing layer).

[0101] It should be understood that by introducing a data fidelity unit, it is possible to ensure that the image features output by each sub-reconstruction layer are always constrained by real physical observation data, preventing the deep learning model from generating false image structures, thereby ensuring the authenticity, basic accuracy and reliability of the final reconstructed image and clinical diagnosis.

[0102] The magnetic resonance image reconstruction method provided in this invention explicitly decouples and embeds the rigorously mathematically proven physical priors (low rank, sparsity, and data consistency) from traditional compressed sensing into the internal modules of a deep learning network. This design not only endows the black-box neural network with clear physical interpretability but also obtains the optimal combination of each prior module through data-driven training. Meanwhile, traditional regularization iterative methods often require manual adjustment of regularization parameters, which can easily lead to excessive smoothing of dynamic images in the time dimension, resulting in the loss of key time point information such as the instant of heartbeat or the peak of rapid contrast agent perfusion. The image reconstruction layer of this invention adaptively learns the optimal low-rank and sparse prior strengths and their combinations based on a large amount of training data through a deep learning model. While effectively removing artifacts, it perfectly preserves high-frequency dynamic details, achieving high-quality reconstruction with excellent spatiotemporal resolution, thus improving the reconstruction quality and efficiency of magnetic resonance images.

[0103] Based on any of the above embodiments, in this method, the magnetic resonance image reconstruction model is trained in the following manner: The undersampled non-Cartesian K-space data of the sample is input into the model to be trained to obtain the reconstructed magnetic resonance image of the sample output by the model to be trained; Based on the magnetic resonance image labels corresponding to the sample magnetic resonance images and the sample undersampled non-Cartesian K-space data, the loss function value is determined. Based on the loss function value, the model parameters of the model to be trained are updated.

[0104] Here, undersampled non-Cartesian K-space data refers to input data used for model training that has known corresponding true labels. In one specific embodiment, this undersampled non-Cartesian K-space data can be generated by artificially downsampling existing high-quality fully sampled data, for example, retaining only a portion (e.g., 1 / n) of the spoke data according to the golden angle radial trajectory.

[0105] Here, the model to be trained refers to the initial magnetic resonance image reconstruction model for which the optimal weights and bias parameters have not yet been determined. Furthermore, it internally includes the aforementioned non-Cartesian interpolation weight estimation layer, meshing layer, coil sensitivity estimation layer, and image reconstruction layer.

[0106] The magnetic resonance image labels are obtained by reconstruction based on fully sampled K-space data.

[0107] It should be noted that full sampling refers to a state where the amount of data in the K-space satisfies the Nyquist theorem (for example, in radial sampling, the number of spokes reaches M). The image sequence reconstructed using the fully sampled data (typically M×M×T×C) exhibits extremely high signal-to-noise ratio, accurate contrast, and clear anatomical structures, representing the ideal imaging effect under this scanning sequence. The data acquisition matrix is ​​M. M, with T frames and C channels.

[0108] Here, the loss function value is used to measure the degree of difference between the sample magnetic resonance image output by the model and the magnetic resonance image label. In a specific embodiment, the loss function can be the mean squared error (MSE) to ensure pixel-level accuracy, or it can simultaneously select the structural similarity index (SSIM) as a loss function term to ensure a high degree of consistency in the perceptual structure of the images.

[0109] It should be understood that by using image labels generated based on full-sample data to calculate the loss function value, the model can be provided with an accurate and objective direction for optimization, ensuring that the model continuously approaches the real, artifact-free physical target during training, thereby guaranteeing the high fidelity and diagnostic value of the final output image of the model.

[0110] In one specific embodiment, using optimization algorithms such as gradient descent, the gradient information calculated based on the loss function value is used to adjust the weights and biases of each neuron in the magnetic resonance image reconstruction model layer by layer through the backpropagation mechanism.

[0111] It should be understood that, since the model is designed end-to-end, this training process achieves joint optimization of the parameters of the non-Cartesian interpolation weight estimation layer, the coil sensitivity estimation layer, and the image reconstruction layer, enabling the layers to work together to minimize the reconstruction error, thereby improving the overall reconstruction accuracy and generalization ability of the model.

[0112] The magnetic resonance image reconstruction method provided in this invention, through the aforementioned method, allows the gradient information of the loss function, calculated based on the final output image, to be backpropagated to every level of the model's front end. This enables steps such as interpolation weight estimation and coil sensitivity estimation to adaptively learn with the goal of achieving optimal final reconstruction quality. This overcomes the shortcomings of traditional piecewise methods where inconsistent optimization objectives at each step lead to overall performance degradation, thereby improving the reconstruction quality of magnetic resonance images. Simultaneously, by using the gold standard generated from full sampling as a label, it ensures that the magnetic resonance image reconstruction model learns objective laws of magnetic resonance physics evolution, rather than simple image smoothing or spurious enhancement, thus further improving the reconstruction quality of magnetic resonance images.

[0113] Based on any of the above embodiments, in this method, the undersampled non-Cartesian K-space data includes non-Cartesian K-space data of a dynamic magnetic resonance imaging sequence; the dynamic magnetic resonance imaging sequence includes multiple consecutive frames of magnetic resonance imaging.

[0114] Here, dynamic magnetic resonance imaging (MRI) sequence refers to an imaging method that adds a temporal dimension encoding on top of traditional two-dimensional or three-dimensional spatial dimension encoding. It aims to capture and record the dynamic physiological evolution of target organs or tissues in the human body over a period of time.

[0115] It should be noted that by limiting the processed data to dynamic magnetic resonance imaging sequence data, the magnetic resonance image reconstruction method of the present invention can provide functional information beyond a single static anatomical structure for clinical diagnosis, significantly improving the clinical diagnostic value of medical images.

[0116] Accordingly, the magnetic resonance image reconstruction model outputs the reconstructed dynamic magnetic resonance imaging sequence, that is, the reconstructed multi-frame continuous magnetic resonance imaging.

[0117] The magnetic resonance image reconstruction method provided in this invention addresses the inherent physical limitations of MRI that result in slow imaging speeds, severely restricting the temporal resolution of dynamic imaging. By processing undersampled non-Cartesian K-space data and combining it with an end-to-end magnetic resonance image reconstruction model, this invention significantly shortens data acquisition time while ensuring image quality at high temporal resolution. It effectively avoids the loss of key dynamic time point information and motion artifacts caused by patient discomfort, thereby improving the reconstruction quality and efficiency of multi-dynamic magnetic resonance images.

[0118] To facilitate understanding of the above embodiments, a specific embodiment will be described here. For example... Figure 3 As shown, taking dynamic magnetic resonance imaging acquired using a golden angle radial trajectory as an example, the overall process of a fully end-to-end magnetic resonance image reconstruction model provided in this embodiment of the invention is explained. First, the magnetic resonance scanning device uses multi-channel receiving coils to acquire continuous K-space raw data according to a golden angle non-Cartesian radial trajectory, such as... Figure 3 As shown in the upper left corner, this data is represented as a three-dimensional data block (containing readout dimension, spoke dimension, and channel dimension); subsequently, as... Figure 3 As indicated by the blue arrow at the top center, "Rearranged into Multiple Frames," the continuously acquired scattered radial spokes are rearranged in chronological order and divided into T consecutive time frames. This results in a sequence of undersampled, non-Cartesian multi-frame data (frame 1, frame 2, ..., frame T); and as shown... Figure 3 The blue box "Center ACS" illustrates the downward extraction process. After all radially acquired data from the golden angle are aggregated and rearranged in the time dimension, the central portion of the data is directly extracted from the low-frequency core region of the K-space to obtain self-calibration signal data (ACS data). Subsequently, the extracted ACS data is synchronously fed into two parallel deep learning front-end modules: an implicit interpolation weight estimation module (i.e., a non-Cartesian interpolation weight estimation layer, which receives ACS data, implicitly learns from the data's intrinsic structure, and outputs interpolation weights for non-Cartesian to Cartesian mapping) and a coil sensitivity estimation module (i.e., a coil sensitivity estimation layer). This coil sensitivity estimation module also receives ACS data and directly infers and outputs a multi-channel spatial coil sensitivity map (e.g., ...). Figure 3 (The color spatial distribution map shown in the lower middle); then, as... Figure 3 As shown in the central position, the interpolation module (i.e., the meshing layer) receives the aforementioned rearranged "undersampled non-Cartesian multi-frame data (non-Cartesian data from frames 1 to T)" and, in conjunction with the "interpolation weights" output by the implicit interpolation weight estimation module, performs meshing processing on the non-Cartesian data. Through this step, the scattered radial data is mapped onto a regular Cartesian grid in one go, resulting in undersampled multi-frame data in the Cartesian domain. Figure 3 (A square image sequence with radial artifacts in the middle section); then, after obtaining undersampled multi-frame data in the Cartesian domain and the coil sensitivity map, both are fed as joint input to... Figure 3 The image reconstruction module (i.e., the image reconstruction layer) on the right side of the image, such as... Figure 3 As shown, the image reconstruction module adopts an expanded network architecture, containing K cascaded reconstruction modules (the first reconstruction module, the second reconstruction module, ..., the Kth reconstruction module). According to... Figure 3The detailed instructions in the enlarged box at the bottom center show that each sub-reconstruction module explicitly integrates three physical prior units: a low-rank module, a sparse module, and a data consistency module (i.e., a data fidelity unit). After single forward propagation processing through these K cascaded modules, the model ultimately outputs a clear, artifact-free reconstructed multi-frame dynamic magnetic resonance image sequence (e.g., ...). Figure 3 (As shown on the far right).

[0119] The magnetic resonance image reconstruction apparatus provided by the present invention is described below. The magnetic resonance image reconstruction apparatus described below can be referred to in correspondence with the magnetic resonance image reconstruction method described above.

[0120] Figure 4 This is a schematic diagram of the magnetic resonance image reconstruction device provided in an embodiment of the present invention, as shown below. Figure 4 As shown, the magnetic resonance image reconstruction apparatus includes: Data acquisition module 410 is used to acquire undersampled non-Cartesian K-space data of magnetic resonance imaging; Image reconstruction module 420 is used to input the undersampled non-Cartesian K-space data into the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model; The magnetic resonance image reconstruction model is trained based on undersampled non-Cartesian K-space data. The magnetic resonance image reconstruction model is used to map the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, and to reconstruct the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data.

[0121] The magnetic resonance image reconstruction apparatus provided in this invention acquires undersampled non-Cartesian K-space data from magnetic resonance imaging, inputs the undersampled non-Cartesian K-space data into a magnetic resonance image reconstruction model, and obtains the reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model. The magnetic resonance image reconstruction model maps the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, thereby replacing time-consuming physical and mathematical operators through data mapping within the magnetic resonance image reconstruction model. This completely eliminates the most time-consuming computational bottleneck in the reconstruction process, reducing the computational load and improving the reconstruction efficiency of magnetic resonance images to meet clinical real-time requirements. Simultaneously, the magnetic resonance image reconstruction model is trained based on sample undersampled non-Cartesian K-space data, meaning the magnetic resonance image reconstruction model is jointly optimized to achieve an end-to-end deep learning model. This avoids the error of the preceding module affecting the processing accuracy of subsequent modules in traditional modular processing, thereby improving the reconstruction quality of magnetic resonance images to meet clinical image quality requirements.

[0122] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a magnetic resonance image reconstruction method. This method includes: acquiring undersampled non-Cartesian K-space data from magnetic resonance imaging; inputting the undersampled non-Cartesian K-space data into a magnetic resonance image reconstruction model to obtain a reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model; wherein the magnetic resonance image reconstruction model is trained based on sample undersampled non-Cartesian K-space data; the magnetic resonance image reconstruction model is used to map the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, and to reconstruct the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data.

[0123] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0124] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the magnetic resonance image reconstruction method provided by the above methods. The method includes: acquiring undersampled non-Cartesian K-space data of magnetic resonance imaging; inputting the undersampled non-Cartesian K-space data into a magnetic resonance image reconstruction model to obtain a reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model; wherein the magnetic resonance image reconstruction model is trained based on sample undersampled non-Cartesian K-space data; the magnetic resonance image reconstruction model is used to map the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, and to reconstruct the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data.

[0125] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the magnetic resonance image reconstruction method provided by the above methods. The method includes: acquiring undersampled non-Cartesian K-space data of magnetic resonance imaging; inputting the undersampled non-Cartesian K-space data into a magnetic resonance image reconstruction model to obtain a reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model; wherein the magnetic resonance image reconstruction model is trained based on sample undersampled non-Cartesian K-space data; the magnetic resonance image reconstruction model is used to map the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, and to reconstruct the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data.

[0126] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0127] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0129] The above embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Although the invention has been described in detail with reference to the embodiments, those skilled in the art should understand that various combinations, modifications, or equivalent substitutions of the technical solutions of the invention do not depart from the spirit and scope of the invention and should be covered within the scope of the claims of the invention.

Claims

1. A magnetic resonance image reconstruction method, characterized in that, include: Acquire undersampled non-Cartesian K-space data for magnetic resonance imaging; The undersampled non-Cartesian K-space data is input into the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model; The magnetic resonance image reconstruction model is trained based on undersampled non-Cartesian K-space data. The magnetic resonance image reconstruction model is used to map the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, and to reconstruct the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data.

2. The magnetic resonance image reconstruction method according to claim 1, characterized in that, The step of inputting the undersampled non-Cartesian K-space data into the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model includes: The undersampled non-Cartesian K-space data is input into the non-Cartesian interpolation weight estimation layer in the magnetic resonance image reconstruction model to obtain the interpolation weights output by the non-Cartesian interpolation weight estimation layer; the non-Cartesian interpolation weight estimation layer is used to generate interpolation weights for non-Cartesian coordinate to Cartesian coordinate mapping based on the undersampled non-Cartesian K-space data. The interpolation weights and the undersampled non-Cartesian K-space data are input into the meshing layer of the magnetic resonance image reconstruction model to obtain the undersampled Cartesian K-space data output by the meshing layer; the meshing layer is used to mesh the undersampled non-Cartesian K-space data into undersampled Cartesian K-space data. The undersampled Cartesian K-space data is input into the image reconstruction layer of the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the image reconstruction layer.

3. The magnetic resonance image reconstruction method according to claim 2, characterized in that, The step of inputting the undersampled Cartesian K-space data into the image reconstruction layer of the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the image reconstruction layer includes: The undersampled non-Cartesian K-space data is input into the coil sensitivity estimation layer in the magnetic resonance image reconstruction model to obtain the coil sensitivity map output by the coil sensitivity estimation layer. The undersampled Cartesian K-space data and the coil sensitivity map are input into the image reconstruction layer to obtain the reconstructed magnetic resonance image output by the image reconstruction layer.

4. The magnetic resonance image reconstruction method according to claim 2, characterized in that, The step of inputting the undersampled non-Cartesian K-space data into the non-Cartesian interpolation weight estimation layer of the magnetic resonance image reconstruction model to obtain the interpolation weights output by the non-Cartesian interpolation weight estimation layer includes: Extract the data from the central region of the K-space of the undersampled non-Cartesian K-space data to obtain self-calibrated signal data; The self-calibrated signal data is input into the non-Cartesian interpolation weight estimation layer to obtain the interpolation weights output by the non-Cartesian interpolation weight estimation layer.

5. The magnetic resonance image reconstruction method according to claim 4, characterized in that, The undersampled non-Cartesian K-space data includes non-Cartesian K-space data from dynamic magnetic resonance imaging sequences, which include multiple consecutive frames of magnetic resonance imaging. The undersampled non-Cartesian K-space data is acquired using a golden angle non-Cartesian acquisition method. The step of extracting data from the central region of the K-space of the undersampled non-Cartesian K-space data to obtain self-calibration signal data includes: The undersampled non-Cartesian K-space data is rearranged in the time dimension to obtain rearranged K-space data. Data from the central region of the rearranged K-space data is extracted to obtain self-calibration signal data.

6. The magnetic resonance image reconstruction method according to claim 2, characterized in that, The image reconstruction layer includes multiple cascaded image sub-reconstruction layers, and each of the image sub-reconstruction layers integrates a denoising unit and a data fidelity unit. The noise reduction unit is used to remove noise in magnetic resonance imaging; The data fidelity unit is used to ensure the consistency between the reconstructed magnetic resonance image and the undersampled Cartesian K-space data.

7. The magnetic resonance image reconstruction method according to any one of claims 1 to 6, characterized in that, The magnetic resonance image reconstruction model is trained based on the following method: The undersampled non-Cartesian K-space data of the sample is input into the model to be trained to obtain the reconstructed magnetic resonance image of the sample output by the model to be trained; The loss function value is determined based on the magnetic resonance image labels corresponding to the sample magnetic resonance images and the sample undersampled non-Cartesian K-space data; the magnetic resonance image labels are reconstructed based on fully sampled K-space data. Based on the loss function value, the model parameters of the model to be trained are updated.

8. The magnetic resonance image reconstruction method according to any one of claims 1 to 6, characterized in that, The undersampled non-Cartesian K-space data includes non-Cartesian K-space data from dynamic magnetic resonance imaging sequences; the dynamic magnetic resonance imaging sequences include multiple consecutive frames of magnetic resonance imaging.

9. A magnetic resonance image reconstruction device, characterized in that, include: The data acquisition module is used to acquire undersampled non-Cartesian K-space data from magnetic resonance imaging. The image reconstruction module is used to input the undersampled non-Cartesian K-space data into the magnetic resonance image reconstruction model to obtain the reconstructed magnetic resonance image output by the magnetic resonance image reconstruction model. The magnetic resonance image reconstruction model is trained based on undersampled non-Cartesian K-space data. The magnetic resonance image reconstruction model is used to map the undersampled non-Cartesian K-space data to undersampled Cartesian K-space data, and to reconstruct the reconstructed magnetic resonance image based on the undersampled Cartesian K-space data.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the magnetic resonance image reconstruction method as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the magnetic resonance image reconstruction method as described in any one of claims 1 to 8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the magnetic resonance image reconstruction method as described in any one of claims 1 to 8.

Images