Magnetic resonance imaging general reconstruction method, system and terminal based on generative model

Abstract

This application provides a universal magnetic resonance imaging (MRI) reconstruction method, system, and terminal based on a generative model. The method includes: inputting fully sampled MRI k-space data into a generative large model to train a universal MRI reconstruction model; inputting clinically acquired undersampled MRI k-space data into the universal MRI reconstruction model as constraints to control the model for image reconstruction, obtaining an initial MRI reconstructed image, and performing post-processing operations to obtain the final MRI reconstructed image. This application solves the technical problems of high model development and usage costs, poor model versatility, and severe motion artifacts in existing technologies. It is applicable to different organs, different MRI contrasts, arbitrary acceleration ratios, and arbitrary sampling masks, and maintains accurate and clear reconstruction results in flexible and varied clinical scenarios and under high-acceleration imaging conditions, thus having practical application value in clinical diagnosis and treatment and industry technology development.

Description

Technical Field

[0001] This application relates to the field of magnetic resonance imaging, and in particular to a general reconstruction method, system and terminal for magnetic resonance imaging based on a generative model. Background Technology

[0002] In clinical magnetic resonance imaging (MRI), to improve imaging efficiency and shorten patient scan time, accelerated acquisition is typically achieved through multi-coil K-space undersampling. Specifically, K-space data is first undersampled and then the unsampled data is filled in using an MRI reconstruction algorithm, ultimately completing the reconstruction of the MRI image.

[0003] In the field of magnetic resonance imaging (MRI) reconstruction algorithms, compared to traditional MRI reconstruction algorithms such as GRAPPA, SENSE, and Compressed Sensing, deep learning-based MRI reconstruction algorithms typically offer better reconstruction results and faster reconstruction speeds at high acceleration magnifications, and have gradually become the preferred approach in clinical applications. However, existing deep learning-based MRI reconstruction algorithms require separate training of corresponding reconstruction models for different imaging magnifications, sampling masks, MRI contrast ratios, and different scanned organs. This situation directly leads to high development and usage costs for the models, and their lack of flexibility makes them unable to effectively cope with the diverse and changing imaging needs in clinical scenarios. In addition, motion artifacts are also a significant factor limiting MRI imaging performance. Although mainstream MRI manufacturers have equipped their advanced MRI scanners with AI-based reconstruction algorithms, these AI reconstruction algorithms still have significant shortcomings in practical applications: when faced with higher speedups, their reconstruction quality deteriorates significantly, and they are prone to aliasing artifacts, noise enhancement, and other adverse phenomena; when faced with sampling masks not seen during model training, the reconstruction quality drops drastically; and when the patient moves during the scan, the reconstructed image produces severe motion artifacts, making it difficult to meet the clarity requirements for clinical diagnosis. Summary of the Invention

[0004] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a general magnetic resonance imaging reconstruction method, system and terminal based on generative models, so as to solve the technical problems of high model development and use costs and poor model universality in the prior art.

[0005] To achieve the above and other related objectives, a first aspect of this application provides a general MRI reconstruction method based on a generative model. The method includes: inputting fully sampled MRI K-space data into a generative model to train a general MRI reconstruction model; inputting clinically acquired undersampled MRI K-space data into the general MRI reconstruction model as a constraint to control the general MRI reconstruction model to perform image reconstruction, thereby obtaining an initial MRI reconstruction image; and performing post-processing operations on the initial MRI reconstruction image to obtain a final MRI reconstruction image.

[0006] In one embodiment of the present invention, the step of inputting clinically acquired undersampled magnetic resonance k-space data into the general MRI reconstruction model as a constraint condition to control the general MRI reconstruction model to perform image reconstruction and obtain an initial MRI reconstructed image includes: performing probability fitting processing on the undersampled magnetic resonance k-space data using the maximum likelihood method to obtain a data fidelity constraint term; fusing the data fidelity constraint term with the prior knowledge of the distribution of fully sampled magnetic resonance k-space data learned by the general MRI reconstruction model to construct an objective function; and iteratively solving the objective function to obtain the initial MRI reconstructed image.

[0007] In one embodiment of the present invention, the generative model includes a diffusion model.

[0008] In one embodiment of the present invention, the step of fusing the data fidelity constraint term with the prior knowledge of the full-sample magnetic resonance K-space data distribution learned by the general MRI reconstruction model to construct the objective function includes: in each step of the inverse diffusion of the diffusion model, calculating the initial estimate of the target image to be optimized in the current step using the Tweedie formula; and fusing the data fidelity constraint term with the prior knowledge of the full-sample magnetic resonance K-space data distribution learned by the diffusion model to construct the objective function for optimizing the initial estimate of the target image.

[0009] In one embodiment of the present invention, the objective function includes a data fidelity constraint term and an image prior constraint term; wherein, the iterative solution of the objective function to obtain the initial MRI reconstructed image includes: in each step of the inverse diffusion of the diffusion model, the initial estimate of the target image is iteratively optimized and updated based on the objective function to dynamically and adaptively adjust the fusion method and ratio of the data fidelity constraint term and the image prior constraint term in the objective function, and the updated initial estimate of the target image is superimposed with the corresponding inverse diffusion step noise to match the forward diffusion process, so as to complete the calculation of the current inverse diffusion step; the above optimization update and superposition of the corresponding inverse diffusion step noise operation are repeated until all steps of inverse diffusion are completed to obtain the initial MRI reconstructed image.

[0010] In one embodiment of the present invention, the post-processing operation on the initial MRI reconstructed image to obtain the final MRI reconstructed image includes: performing edge smoothing processing and signal intensity normalization processing on the initial MRI reconstructed image to obtain the final MRI reconstructed image.

[0011] In one embodiment of the present invention, prompt words are selectively input into the generative model to guide the general MRI reconstruction model to learn the prior knowledge of the distribution of the full-sample magnetic resonance k-space data. When in use, prompt words are selectively input into the general MRI reconstruction model to guide the general MRI reconstruction model to output an initial MRI reconstruction image matching the imaging scene corresponding to the prompt words. The prompt words include one or a combination of organ type, imaging parameters, age, gender, initial diagnosis information, gene test indicators, and blood test indicators.

[0012] In one embodiment of the present invention, during the iterative solution of the objective function, an accelerated inference strategy, estimation of patient motion parameters, and / or optimization of coil sensitivity maps are selectively employed; wherein, the accelerated inference strategy includes one or a combination of inference sampling method optimization, model distillation, model pruning, and model quantization.

[0013] To achieve the above and other related objectives, a second aspect of this application provides a universal magnetic resonance imaging reconstruction system based on a generative model. The system includes: a model training module for inputting fully sampled magnetic resonance k-space data into a generative model to train a universal MRI reconstruction model; an image reconstruction module for inputting clinically acquired undersampled magnetic resonance k-space data into the universal MRI reconstruction model as constraints to control the universal MRI reconstruction model to perform image reconstruction, thereby obtaining an initial MRI reconstructed image; and a post-processing module for performing post-processing operations on the initial MRI reconstructed image to obtain a final MRI reconstructed image.

[0014] To achieve the above and other related objectives, a third aspect of this application provides an electronic terminal, including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the method.

[0015] As described above, this application discloses a general magnetic resonance imaging reconstruction method, system, and terminal based on a generative model. It inputs fully sampled MRI K-space data into a large generative model to train a general MRI reconstruction model. Clinically acquired undersampled MRI K-space data is then input into the general MRI reconstruction model as constraints to control image reconstruction, resulting in an initial MRI reconstructed image. Post-processing is then performed on the initial MRI reconstructed image to obtain the final MRI reconstructed image. This method offers the following advantages:

[0016] (1) A generative model learning framework is adopted, which avoids the limitations of the traditional discriminative paradigm by learning prior knowledge of the distribution of full-sample magnetic resonance K-space data. At the same time, prompt words can be selectively input into the generative model during the training phase, enabling the model to learn the distribution characteristics of full-sample data more comprehensively and accurately, improving the model's adaptability to different clinical scenarios, laying a solid foundation for subsequent general reconstruction, and effectively solving the problems of single learning dimension and insufficient adaptability of existing models.

[0017] (2) In the model training stage, the model training is driven only by the full-sampled magnetic resonance K-space data, so that the generative model can learn the true distribution law of the K-space data corresponding to the full-sampled MRI images and learn the general prior knowledge of the distribution of the full-sampled magnetic resonance K-space data. This makes the trained general MRI reconstruction model adaptable to a variety of MRI reconstruction tasks, improves the general adaptability of the model, eliminates the need to train the model separately for different situations, reduces the workload and cost of model development, simplifies the clinical use process, improves the flexibility of the model, and solves the technical problems of poor universality and high cost of existing models.

[0018] (3) By introducing fine-grained undersampled magnetic resonance K-space data as the fidelity condition of the reconstruction model, the obtained data fidelity constraint term is fused with the prior knowledge of the distribution of fully sampled magnetic resonance K-space data learned by the general MRI reconstruction model to construct an optimization objective function for image reconstruction, which guides the generative model to complete image reconstruction. The reconstruction result obtained in this way not only has a high consistency (fidelity) with the acquired data, but can also adapt to different speed-up factors and sampling masks, and has excellent flexibility and versatility.

[0019] (4) During the generative model usage phase, by estimating motion parameters and updating the coil sensitivity map, it is possible to correct motion artifacts of different degrees and types, effectively suppressing the impact of motion artifacts on the quality of the reconstructed image. Furthermore, based on the objective function, it strategically optimizes the reconstructed image, motion parameters, and coil sensitivity map simultaneously, for example, by estimating parameters from coarse-grained to fine-grained. Simultaneously, based on inference acceleration strategies such as sampling method optimization, model distillation, model pruning, and model quantization, the inference process is further accelerated, improving the model's inference speed while ensuring reconstruction fidelity and image quality.

[0020] In summary, this application solves the technical problems of high model development and usage costs, poor model versatility, and severe motion artifacts in the prior art; it is applicable to different organs, different magnetic resonance contrasts, arbitrary acceleration factors and arbitrary sampling masks, and can maintain accurate and clear reconstruction results in flexible and varied clinical scenarios and under high-acceleration imaging conditions, and has practical application value in clinical diagnosis and treatment and industry technology development. Attached Figure Description

[0021] Figure 1 The diagram shown is a flowchart of a general magnetic resonance imaging reconstruction method based on a generative model, according to one embodiment of this application.

[0022] Figure 2 The diagram shown is a flowchart illustrating the image reconstruction process in one embodiment of this application.

[0023] Figure 3 The diagram shown is a flowchart illustrating the construction of the objective function in one embodiment of this application.

[0024] Figure 4 The diagram shown is a schematic of an adaptive universal magnetic resonance reconstruction model based on a generative model in one embodiment of this application.

[0025] Figure 5 The diagram shown is a schematic block diagram of a generative model-based general reconstruction system for magnetic resonance imaging in one embodiment of this application.

[0026] Figure 6 The diagram shown is a structural schematic of an electronic terminal according to an embodiment of this application. Detailed Implementation

[0027] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0028] In the embodiments of this application, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" do not necessarily imply that they are different.

[0029] It should be noted that, in the embodiments of this application, the words "exemplary" or "for example" indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0030] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0031] Before providing a further detailed description of the present invention, the nouns and terms used in the embodiments of the present invention are explained, and the nouns and terms used in the embodiments of the present invention are subject to the following interpretations:

[0032] <1> GRAPPA: Generalized self-calibration part parallel acquisition principle, which utilizes the sensitivity difference of multi-channel phased array coils to perform undersampling in K space, while simultaneously acquiring a small amount of self-calibration signal; through linear fitting, interpolation is used to complete the unsampled K space data.

[0033] <2> SENSE: Sensitivity coding, which reconstructs images in the image domain. Based on the different sensitivity distributions of each coil, it performs dealiasing calculations on the aliased images caused by undersampling, and directly reconstructs a complete image without aliasing from the undersampled data.

[0034] <3> Compressed Sensing utilizes the sparsity of a signal in a transform domain (such as wavelet or Fourier) to perform undersampling at a rate much lower than the Nyquist sampling rate. Through sparse constraint optimization algorithms, the original signal is accurately reconstructed, achieving a significant acceleration in scanning speed.

[0035] <4> MRI: Magnetic Resonance Imaging, based on the magnetic resonance phenomenon generated by atomic nuclei (hydrogen protons) under the action of static magnetic fields, radio frequency pulses and gradient fields, acquires spatially encoded K-space signals, and reconstructs tomographic images of human tissues through Fourier transform, for use in non-invasive clinical examinations.

[0036] In some embodiments of this application, the technical solutions of this application can be applied to various magnetic resonance imaging-related scenarios and systems, such as: low field strength MRI systems, 1.5T MRI systems, 3.0T MRI systems, ultra-high field strength MRI systems, portable MRI systems, bedside MRI systems for intensive care units, intraoperative MRI systems, routine MRI diagnosis and treatment systems for all departments, precision tumor imaging and efficacy follow-up systems, MRI imaging systems for neurological and psychiatric diseases, MRI imaging systems for musculoskeletal and sports medicine, MRI imaging systems for pediatric non-sedation, rapid MRI imaging systems for acute and critical illnesses, original factory-matched algorithm upgrade systems for MRI equipment, intelligent transformation systems for old MRI equipment, and standardized processing systems for multi-center clinical research images. The general MRI diagnostic system can include cranial MRI imaging, spinal MRI imaging, abdominal MRI imaging, bone and joint MRI imaging, breast MRI imaging, pelvic MRI imaging, etc.; the MRI imaging process can include undersampling accelerated MRI imaging, free breathing motion-corrected MRI imaging, high-fidelity reconstruction of low signal-to-noise ratio images, and conditional MRI imaging guided by prompts, such as organ type, imaging parameters, age, gender, initial diagnosis information, lesion type, genes, and blood indicators; it can also be applied to pre-processing scenarios for AI-assisted diagnosis of medical images, radiomics and quantitative imaging research scenarios, rare disease and niche MRI sequence imaging research scenarios, MRI image reconstruction scenarios for telemedicine, and edge MRI image processing scenarios for machine communication and the Internet of Things.

[0037] To facilitate understanding of the embodiments of this application, firstly, in conjunction with Figure 1 Detailed explanation. Figure 1 This document illustrates a flowchart of a general magnetic resonance imaging reconstruction method based on a generative model, as described in an embodiment of the present invention. The method in this embodiment mainly includes the following steps:

[0038] Step S11: Input the fully sampled MRI K-space data into the generative model to train a general MRI reconstruction model.

[0039] In one embodiment, the fully sampled magnetic resonance K-space data includes:

[0040] (1) Real full sampling data, such as clinical full sampling data.

[0041] (2) High-quality K-space data after low-speed undersampling data is filled by reconstruction algorithms, such as high-quality K-space data filled by GRAPPA or Compressed Sensing under low-speed undersampling scanning.

[0042] In one embodiment, the general MRI reconstruction model is trained by inputting pre-collected full-sample MRI K-space data into a generative large model, and completing multiple rounds of iterative training through a preset self-supervised learning strategy, ultimately obtaining a general MRI reconstruction model with generalization and adaptation capabilities across all scenarios. In this embodiment, the full-sample MRI K-space data used for model training is clinically compliant full-sample MRI image data covering multiple organs, multiple contrasts, multiple field strength devices, multiple scanning parameters, and multiple population dimensions. The multiple organs include, but are not limited to, all parts involved in routine clinical MRI examinations, such as the brain, spine, abdomen, bones and joints, breast, and pelvis. The multiple contrasts include at least the image contrast features corresponding to routine clinical sequences such as T1-weighted imaging, T2-weighted imaging, liquid attenuation inversion recovery sequence, diffusion-weighted imaging, and dynamic contrast-enhanced imaging. Through a multi-dimensional and broadly covered training dataset, the model can fully learn the common features and distribution patterns of full-sample MRI data in different clinical scenarios, laying a complete data foundation for the model's general adaptability.

[0043] In one embodiment, this application employs a generative model learning paradigm, which is fundamentally different from the discriminative learning paradigm commonly used in the current field of AI-assisted MRI reconstruction. Current mainstream AI-based MRI reconstruction models mostly use undersampled-full sampled paired datasets for supervised training, forcing the model to learn a fixed mapping relationship from undersampled images to full sampled images. Under this paradigm, the training phase requires pre-specifying an undersampled mask and constructing paired undersampled MRI K-space data based on the mask. The resulting model can only adapt to the fixed undersampled patterns, imaging sites, and scanning parameters used during training, exhibiting technical drawbacks such as poor generalization, high cross-scenario adaptation costs, and limited clinical application.

[0044] The training scheme of this application does not require pre-defining any undersampling mask or constructing any undersampling MRI K-space data during the model training stage. It drives model training solely through full-sample MRI K-space data, enabling the generative model to autonomously learn the true distribution rules of K-space data corresponding to full-sample MRI images and acquire general prior knowledge of the distribution of full-sample MRI K-space data. This completely eliminates the strong dependence of existing supervised learning on paired training data, allowing the trained general MRI reconstruction model to be adapted to MRI reconstruction tasks with any undersampling mode, any imaging location, and any scanning parameters. It eliminates the need to train models separately for different situations, reducing the workload and cost of model development, simplifying the clinical use process, improving the model's flexibility, and enhancing its general adaptability, clinical application value, and anti-interference performance.

[0045] Step S12: Input the clinically acquired undersampled MRI K-space data into the general MRI reconstruction model as a constraint condition to control the general MRI reconstruction model to perform image reconstruction and obtain the initial MRI reconstruction image.

[0046] In one embodiment, based on the aforementioned trained general MRI reconstruction model, image reconstruction is performed on clinically acquired undersampled MRI K-space data to obtain an initial MRI reconstructed image. Instead of using the existing mapping mode of directly outputting reconstruction results from end-to-end input undersampled data, the clinically measured undersampled data is used as a constraint to control the reconstruction process. Through a dual-constraint optimization system of data fidelity and pre-trained full-sample distribution prior, universal, high-fidelity MRI undersampled reconstruction is achieved across all scenarios.

[0047] In one embodiment, such as Figure 2 The diagram illustrates the image reconstruction process in an embodiment of the present invention. Clinically acquired undersampled MRI K-space data is input into the general MRI reconstruction model as a constraint to control the model for image reconstruction, resulting in an initial MRI reconstructed image including:

[0048] Step S121: Perform probability fitting on the undersampled magnetic resonance K-space data using the maximum likelihood method to obtain data fidelity constraint terms.

[0049] In one embodiment, the data fidelity constraint is a hard constraint in the reconstruction process, used to quantify and limit the statistical consistency between the K-space data corresponding to the reconstructed image and the clinically measured undersampled magnetic resonance K-space data, ensuring that the reconstruction result always matches the real scan information collected clinically. At the same time, the data fidelity constraint can be directly adapted to any undersampled mask and any speedup ratio without the need for pre-adaptation adjustments for specific scenarios.

[0050] In one embodiment, the maximum likelihood method is a probabilistic statistical method used to fit the probability distribution of these undersampled data, quantifying the difference between the model-generated K-space data and the actual collected data.

[0051] Step S122: The data fidelity constraint term is fused with the prior knowledge of the full-sample magnetic resonance K-space data distribution learned by the general MRI reconstruction model to construct the objective function.

[0052] In one embodiment, the aforementioned data fidelity constraint is fused with prior knowledge of the distribution of full-sampled magnetic resonance k-space data covering multiple organs and contrast levels, learned during the training phase of a general MRI reconstruction model, to construct an optimization objective function for image reconstruction. The data fidelity constraint serves as the data consistency term in the optimization objective function, while the prior knowledge of the full-sampled magnetic resonance k-space data distribution learned by the general MRI reconstruction model serves as the image prior constraint. The organic fusion of these two constraints not only ensures the matching degree between the reconstruction results and clinically measured data but also provides high-fidelity completion of undersampled and missing k-space information that meets clinical diagnostic criteria, based on the real full-sampled data distribution patterns learned during pre-training.

[0053] Step S123: Iteratively solve the objective function to obtain the initial MRI reconstructed image.

[0054] In one embodiment, the constructed optimization objective function is then iteratively solved, i.e., maximum a posteriori estimation is performed. Once the objective function converges to a preset threshold, the initial MRI reconstructed image is obtained. This iterative solution process can seamlessly adapt to any type of generative large model architecture, such as diffusion models, generative adversarial networks, and variational autoencoders. It does not require modification of the model structure, retraining, or fine-tuning of the model for different imaging scenarios. High-fidelity reconstruction can be completed simply by iteratively optimizing the objective function.

[0055] It should be noted that, in the embodiments of this application, the generative model can be any type of artificial intelligence model with the ability to learn from the true distribution of data, including but not limited to diffusion models, generative adversarial networks (GANs), and variational autoencoders (VAEs). The inventive concept of this application does not depend on a specific type of generative model architecture. As long as the model can learn the distribution patterns of fully sampled MRI K-space data and output MRI reconstruction results that meet clinical diagnostic criteria, it can be adapted to the technical solution of this application. To clearly and completely illustrate the technical details of this application, this application uses a diffusion model (diffusion model) as a preferred embodiment to describe in detail the training and inference implementation methods of the generative model. This preferred embodiment is only used to explain this application and is not intended to limit the scope of protection of this application.

[0056] In one embodiment, the generative model is a diffusion model. The process of fusing data fidelity constraints with prior knowledge of the distribution of fully sampled magnetic resonance K-space data to construct the objective function is deeply coupled with the inverse diffusion iteration process of the diffusion model. Through the generation of initial estimation step by step and the fusion of dual constraints, the reconstruction process is refined and high-fidelity constrained, which is fully compatible with the inverse diffusion denoising logic of the diffusion model.

[0057] In one embodiment, such as Figure 3 The diagram illustrates the process of constructing the objective function in an embodiment of the present invention. The objective function is constructed by fusing the data fidelity constraint term with the prior knowledge of the full-sample magnetic resonance k-space data distribution learned by the general MRI reconstruction model, and includes:

[0058] Step S1221: In each step of the inverse diffusion of the diffusion model, the initial estimate of the target image to be optimized in the current step is calculated using the Tweedie formula.

[0059] Step S1222: The data fidelity constraint term is fused with the prior knowledge of the full-sample magnetic resonance K-space data distribution learned by the diffusion model to construct an objective function for optimizing the initial estimation of the target image.

[0060] In one embodiment, the Tweedie formula is a mathematical tool in the diffusion model that uses the noise predicted by the model, combined with the noise level of the current iteration step, to calculate an "initial estimate of the denoised image". This estimate is the starting point for subsequent optimization and is closer to the real image than random initialization.

[0061] In one embodiment, at each step of the inverse diffusion of the diffusion model, two actions are performed simultaneously: initial estimation generation and objective function construction. First, the initial estimate of the target image to be optimized in the current inverse diffusion step is calculated using the Tweedie formula. Second, the data fidelity constraint term constructed using the maximum likelihood method is fused with the prior knowledge of the full-sampled magnetic resonance K-space data distribution covering multiple organs and contrasts learned by the diffusion model during the pre-training phase to construct an objective function specifically for optimizing the current inverse diffusion step. The initial estimate of the current target image is iteratively optimized and updated using this objective function. The data fidelity constraint term serves as a hard constraint term for data consistency in the objective function, while the prior knowledge of the full-sampled magnetic resonance K-space data distribution serves as a prior image constraint term (soft constraint term) in the objective function. The weighted fusion of these two constraints provides a clear quantitative optimization criterion for the image update in the current inverse diffusion step.

[0062] In one embodiment, the objective function includes a data fidelity constraint term and an image prior constraint term; wherein, iteratively solving the objective function to obtain the initial MRI reconstructed image includes:

[0063] (1) In each step of the inverse diffusion of the diffusion model, the initial estimate of the target image is iteratively optimized and updated based on the objective function, so as to dynamically and adaptively adjust the fusion method and ratio of the data fidelity constraint term and the image prior constraint term in the objective function, and superimpose the updated initial estimate of the target image with the corresponding inverse diffusion step noise to match the forward diffusion process, so as to complete the calculation of the current inverse diffusion step.

[0064] (2) Repeat the above optimization update and superimpose corresponding inverse diffusion step noise operation until all steps of inverse diffusion are completed, and the initial MRI reconstruction image is obtained.

[0065] In one embodiment, at each step of the inverse diffusion, the initial estimate of the target image generated by the Tweedie formula is optimized and updated based on the dual-constraint objective function constructed in the current iteration step. During the update process, the model automatically identifies the characteristics of the currently acquired data, including the noise intensity and sampling mask pattern, to dynamically and adaptively adjust the fusion method and ratio of the data fidelity constraint term and the image prior constraint term in the objective function. For example, if you want to trust the acquired data more (lower noise and higher quality), you can adaptively increase the ratio of the data fidelity constraint term; if the quality of the acquired data is poor, you can selectively trust the acquired data and perform pixel-weighted data fidelity constraint terms based on the sampling mask pattern, noise intensity, etc. The optimal solution of the objective function is obtained by numerical optimization methods such as gradient descent. With data fidelity constraints as hard boundary guarantees, the updated image always fits the clinically measured undersampled K-space data. The prior knowledge of the distribution of fully sampled magnetic resonance K-space data is used as the optimization benchmark and as the image prior constraint to complete the missing K-space information of undersampled data. At the same time, the patient's motion parameters can be selectively estimated synchronously and the coil sensitivity map can be optimized. The data fidelity constraints and objective function are corrected in real time to eliminate image artifacts and biases caused by physiological motion and coil signal inhomogeneity.

[0066] In one embodiment, after completing the optimization update of the initial estimate of the target image at each step, the model automatically superimposes the corresponding backdiffusion step noise to match the forward diffusion process. The noise intensity is consistent with the noise intensity of the corresponding step in the forward diffusion process, ensuring the symmetry between the backdiffusion and forward diffusion processes, thereby completing the calculation of the current backdiffusion step and entering the next iteration. Based on the aforementioned dual-constraint objective function constructed for each step of backdiffusion, the above optimization update and superposition of corresponding backdiffusion step noise operations are repeatedly executed. The initial estimate of the target image is iteratively solved for all backdiffusion steps. After reaching the preset total number of iterations, the iteration stops, and the reconstructed initial MRI image is obtained.

[0067] It is worth noting that the initial estimate of the target image is iteratively optimized and updated based on the objective function to dynamically and adaptively adjust the fusion method and ratio of the data fidelity constraint term and the image prior constraint term in the objective function. The data fidelity constraint term ensures that the reconstruction result fits the clinically acquired undersampled magnetic resonance K-space data, avoiding deviation from the real lesion information. The image prior constraint term relies on the prior knowledge of the full sampling distribution to complete the missing information of the undersampled data and restore tissue details. This avoids both diagnostic bias caused by insufficient fidelity and detail distortion caused by excessive prior knowledge. The reconstruction result obtained in this way not only has a high consistency (fidelity) with the acquired data, but can also adapt to different acceleration factors and sampling masks, and has excellent flexibility and versatility.

[0068] Step S13: Perform post-processing operations on the initial MRI reconstructed image to obtain the final MRI reconstructed image. This includes:

[0069] Edge smoothing and signal intensity normalization are performed on the initial MRI reconstructed image to obtain the final MRI reconstructed image.

[0070] In one embodiment, standardized post-processing operations are performed on the initial MRI reconstructed image to further optimize image quality, eliminate non-diagnostic minor interferences, and standardize the clinical diagnostic criteria of the image, ultimately outputting a final MRI reconstructed image that conforms to the clinical interpretation standards of radiology. The post-processing operations specifically include edge smoothing and signal intensity normalization of the initial MRI reconstructed image. These two operations achieve refined optimization of image quality without altering or damaging diagnostic information.

[0071] Specifically, the edge smoothing process preferentially employs edge-preserving filtering algorithms, including but not limited to bilateral filtering, guided filtering, and anisotropic diffusion filtering. This processing addresses the minute Gaussian noise and edge ringing artifacts that may remain in the initial MRI reconstructed image from undersampled K-space data or the inverse diffusion iteration process. It smooths unstructured random noise while fully preserving core diagnostic information such as organ anatomical boundaries, lesion fine structures, and gray-white matter boundaries, avoiding the edge blurring and loss of diagnostic details that are common with traditional global smoothing algorithms. This significantly improves the visual readability and diagnostic friendliness of the image. The signal intensity normalization process is based on standardized grayscale intervals for clinical MRI diagnosis. Linear or nonlinear mapping of signal intensity is performed on the image after edge smoothing, uniformly calibrating the image's signal intensity to a preset standardized numerical range. This eliminates signal intensity deviations and grayscale distribution differences caused by different field strength devices, different scanning sequences, and different acquisition parameters, ensuring that the contrast and grayscale levels of the final output MRI reconstructed image perfectly match the clinical reading habits of radiologists. It also provides standardized image input for subsequent radiomics feature extraction and AI-assisted diagnostic analysis, further expanding the clinical application scenarios of this solution.

[0072] In one embodiment, to further enhance the scene adaptation flexibility and refined reconstruction capability of the general MRI reconstruction model, this solution introduces a selectable cue word conditional guidance mechanism. This mechanism can adapt to both model training and inference reconstruction stages. Specifically, cue words can be selectively input into the generative model to guide the general MRI reconstruction model in learning the prior knowledge of the full-sample MRI K-space data distribution. In use, cue words are selectively input into the general MRI reconstruction model to guide it in outputting an initial MRI reconstruction image matching the imaging scene corresponding to the cue word. The cue words include one or more of the following: organ type, imaging parameters, age, gender, initial diagnosis information, genetic testing indicators, and blood testing indicators.

[0073] Specifically, during the model training phase, prompts corresponding to the training samples can be selectively input into the generative model in pairs with multi-organ, multi-contrast full-sample MRI K-space data to guide the model in learning the distribution characteristics of full-sample MRI K-space data under different subdivided imaging scenarios. These prompts include, but are not limited to, organ type, imaging parameters, age, gender, initial diagnosis information, gene testing indicators, and blood test indicators. For example, organ type can correspond to routine clinical examination sites such as the brain, spine, abdomen, bones and joints, breast, and pelvis; imaging parameters can correspond to acquisition-related parameters such as MRI equipment field strength, scanning sequence, contrast, resolution, and organ positioning; age and gender can correspond to differences in anatomical structure and physiological characteristics among different populations; initial diagnosis information can correspond to personalized information such as examination purpose, suspected lesion type, and clinical diagnostic needs; gene testing indicators can correspond to specific gene parameters related to MRI examination; and blood test indicators can correspond to routine and specific clinical blood parameters.

[0074] In one embodiment, multi-dimensional prompts guide the reconstruction model to learn the data distribution of full-sampled magnetic resonance k-space data for different organs, different contrasts, and different acquisition sequences. This allows the model to learn the differentiated data distribution patterns under different sub-scenarios on the basis of learning the global general full-sampled data distribution. There is no need to develop independent reconstruction models for different imaging conditions, which improves the model's adaptability to different clinical scenarios and effectively solves the problems of single learning dimensions and insufficient adaptability of existing models.

[0075] In one embodiment, during the usage phase, fine-grained undersampled magnetic resonance k-space data is introduced as a fidelity condition for the reconstruction model, complementing the prior knowledge of the subdivided scenes learned by the prompt-guided model. Under the synergistic effect of the two, the generative model is guided to complete image reconstruction. The resulting reconstruction not only has a high degree of consistency (fidelity) with the acquired data, but also reduces the reconstruction illusion of the generative model by relying on the scene difference rules learned from the prompts, giving it excellent flexibility and versatility, and adapting to different acquisition scenarios in clinical practice.

[0076] It is worth emphasizing that the prompts in this training phase still follow the generative learning paradigm of this application. There is no need to pre-specify the undersampling mask or construct an undersampling-full sampling paired dataset. Conditional prior learning can be completed using only fully sampled magnetic resonance K-space data labeled with prompts, which simplifies the training process and reduces training costs.

[0077] In one embodiment, during the inference reconstruction phase, prompts corresponding to clinical examination needs can be selectively input into a general MRI reconstruction model that performs data fidelity constraints and full-sample distribution prior calculations. This guides the model to prioritize matching the full-sample data distribution characteristics of the imaging scene corresponding to the prompts during the iterative solution of the dual-constraint objective function, and outputs an initial MRI reconstruction image that highly matches the clinical examination needs. Through the learned global general priors, the prompts during the inference phase can more effectively guide the model to achieve zero-sample scenario adaptation.

[0078] In one embodiment, for example, when a clinical examination is conducted for the initial diagnosis of a liver lesion in an elderly male patient, the prompt words "abdomen, T2WI sequence, 3.0T, 68 years old, male, initial diagnosis of liver lesion" can be entered to guide the model to optimize the anatomical boundaries of the liver region and the fidelity of the lesion's fine structure during the reconstruction process, while always using the data fidelity constraints constructed from clinically measured undersampled magnetic resonance k-space data as a hard boundary.

[0079] In one embodiment, during the iterative solution of the objective function, accelerated inference strategies, estimation of patient motion parameters, and / or optimization of coil sensitivity maps are selectively employed; wherein, the accelerated inference strategies include one or a combination of inference sampling method optimization, model distillation, model pruning, and model quantization.

[0080] In one embodiment, to further improve the clinical applicability and scenario adaptability of the reconstruction scheme during the iterative solution of the dual-constraint optimization objective function, optimization methods adapted to the reconstruction logic of this application can be selectively introduced, including accelerated inference strategies, patient motion parameter estimation, and coil sensitivity map optimization. These three optimization methods can be implemented individually or in any combination. Specifically, during each iterative solution of the objective function, patient motion parameter estimation and coil sensitivity map optimization can be selectively performed simultaneously. Both optimizations are deeply coupled with the iterative process and update constraints in real time, enabling the correction of motion artifacts of different degrees and types, effectively suppressing the impact of motion artifacts on the quality of the reconstructed image. Furthermore, based on the objective function, the reconstructed image, motion parameters, and coil sensitivity map are strategically optimized simultaneously, for example, from coarse-grained to fine-grained parameter estimation. Simultaneously, accelerated inference strategies such as inference sampling method optimization, model distillation, model pruning, and model quantization further accelerate the inference process, improving the model's inference speed while ensuring reconstruction fidelity and image quality.

[0081] The patient motion parameter estimation is based on the target image estimation result of the current iteration step and clinically measured undersampled MRI K-space data. It estimates the spatial displacement parameters caused by the patient's breathing, heartbeat, and involuntary body movement using rigid or non-rigid registration algorithms. Based on these parameters, it corrects the motion offset of the K-space data in real time, synchronously updating the data fidelity constraints and optimization objective function. This eliminates the accumulation of motion artifacts at the root of the iteration, eliminating the need for the patient to hold their breath. It is suitable for physiologically sensitive areas such as the abdomen and heart, as well as for special populations such as children, agitated patients, and critically ill patients who cannot cooperate with long-term scans. The coil sensitivity map optimization is based on the target image estimation result of the current iteration step, synchronously updating the sensitivity distribution map of the multi-channel phased array coils. Based on the optimized coil sensitivity map, it corrects signal inhomogeneity and gain deviation in the multi-channel K-space data, synchronously optimizing the data fidelity constraints. This addresses the problems of low signal-to-noise ratio and uneven brightness distribution caused by coil sensitivity decay in low-field-strength equipment and older equipment, further improving the quality stability and diagnostic consistency of the reconstructed images.

[0082] To address the needs of real-time clinical imaging and edge device deployment, accelerated inference strategies can be selectively adopted during the iterative solution of the objective function. These accelerated inference strategies include, but are not limited to, at least one of inference sampling optimization, model distillation, model pruning, and model quantization, and can be flexibly selected and adapted according to the computing power conditions and timeliness requirements of the clinical scenario. Among these, the inference sampling method is optimized to be an acceleration strategy adapted to the inverse diffusion iteration process of the diffusion model. This strategy can significantly reduce the total number of inverse diffusion iterations. Combined with the accurate initial estimation of the target image generated by the Tweedie formula in each iteration of this scheme, the hundreds of inverse diffusion iterations required by the traditional diffusion model can be compressed to dozens of steps while maintaining reconstruction fidelity. This significantly reduces the amount of computation and improves the reconstruction speed. Model distillation, model pruning, and model quantization are general acceleration strategies adapted to all types of generative large models. Model distillation uses a pre-trained high-precision general MRI reconstruction model as the teacher model to distill a lightweight student model, which significantly reduces the number of model parameters while retaining the core reconstruction capabilities and generalization across all scenarios. Model pruning simplifies the model structure and reduces the computational cost of a single inference by removing redundant parameters, invalid convolutional kernels, and non-critical network layers from the backbone network. Model quantization reduces the numerical precision of model parameters, thereby reducing the consumption of computational resources and memory overhead and improving the inference speed of the model on low-power devices. By balancing the constraints of data fidelity with the prior constraints of the full sampled data distribution learned by the model, the inference speed of the model is improved. This allows the reconstruction speed of this solution to meet the clinical needs of intraoperative real-time image guidance, rapid imaging of critical care patients at the ICU bedside, and efficient processing of large batches of samples in outpatient settings. At the same time, it can be adapted to the deployment requirements of low-computing-power devices such as portable MRI and edge imaging workstations, realizing adaptive and universal magnetic resonance reconstruction. There is no need to develop different reconstruction models for different imaging conditions, which greatly expands the application boundaries of the solution.

[0083] In one embodiment, such as Figure 4 The diagram illustrates an adaptive universal magnetic resonance reconstruction model based on a generative model, as shown in this embodiment of the invention.

[0084] (1) A structured text prompt is constructed based on the clinical scanning scenario. The content includes: scanning organ (Organ: HEAD), scanning view (View:TRA), imaging contrast (Contrast: T1), voxel spacing (Spacing:(0.8, 0.8, 6.0)), field strength (Field strength: 1.5), equipment manufacturer (Vendor: UIH), patient gender (Gender: M), patient age (Age: 29), and scanning sequence (Sequence: SE). The text prompt is input into the text encoder, and the text information is converted into text embeddings through a pre-trained encoding network as scene guidance conditions.

[0085] (2) Generate a random noise tensor x that follows a standard normal distribution. T -N(0,I) is used as the initial input for the inverse process of the diffusion model. Simultaneously, undersampled magnetic resonance K-space data is used as the data fidelity constraint (inference input).

[0086] (3) The diffusion model adopts a multi-step denoising method to gradually generate high-quality magnetic resonance images from random noise. Step T (initial denoising step): The random noise tensor xT, text embedding and data fidelity constraints are input into the diffusion model. The image detail features are preserved through skip connection, and the first denoising is completed under text guidance and data constraints.

[0087] (4) Step T-1 to Step 1 (Iterative Denoising): The output of the previous step is used as the input of the current step, and the denoising process is repeated. In each iteration, text embeddings and data constraints are continuously input, and the optimized inference strategy, the data fidelity constraints, and the model's learned full-sample data distribution prior are used. As the iteration number progresses from Step T-1 to Step 1, the noise in the image is continuously reduced, and the anatomical structure and detailed texture gradually become clear.

[0088] (5) Output reconstruction results After all iterations are completed, the model outputs high-quality magnetic resonance reconstruction images adapted to the current clinical scenario, which can support adaptive reconstruction under multiple organs, multiple contrasts, arbitrary acceleration factor and sampling mask.

[0089] Figure 5 This is a schematic block diagram of a general magnetic resonance imaging reconstruction system based on a generative model, provided in an embodiment of this application. Figure 5 As shown, the generative model-based universal magnetic resonance imaging reconstruction system 500 includes:

[0090] The model training module 501 is used to input fully sampled magnetic resonance K-space data into the generative model and train it to obtain a general MRI reconstruction model.

[0091] Image reconstruction module 502 is used to input clinically acquired undersampled magnetic resonance k-space data into the general MRI reconstruction model as a constraint condition to control the general MRI reconstruction model to perform image reconstruction and obtain an initial MRI reconstruction image;

[0092] The post-processing module 503 is used to perform post-processing operations on the initial MRI reconstructed image to obtain the final MRI reconstructed image.

[0093] It should be understood that the specific process of each module performing the above-mentioned steps has been described in detail in the above method embodiments, and will not be repeated here for the sake of brevity.

[0094] It should also be understood that the module division in the embodiments of this application is illustrative and only represents a logical functional division; in actual implementation, there may be other division methods. Furthermore, the functional modules in the various embodiments of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0095] Figure 6 This is a schematic block diagram of the electronic terminal provided in an embodiment of this application. Figure 6 As shown, the electronic terminal 600 includes at least one processor 601, a memory 602, at least one network interface 603, and a user interface 605. The various components in the electronic terminal 600 are coupled together via a bus system 604. It is understood that the bus system 604 is used to implement communication between these components. In addition to a data bus, the bus system 604 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 6 The general will label all buses as bus systems.

[0096] The user interface 605 may include a monitor, keyboard, mouse, trackball, clicker, button, touchpad, or touch screen.

[0097] It is understood that memory 602 can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM) or programmable read-only memory (PROM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM) and synchronous static random access memory (SSRAM). The memories described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable categories of memory.

[0098] In this embodiment of the invention, the memory 602 is used to store various types of data to support the operation of the electronic terminal 600. Examples of this data include: any executable program for operation on the electronic terminal 600, such as the operating system 6021 and application programs 6022; the operating system 6021 contains various system programs, such as the framework layer, core library layer, driver layer, etc., for implementing various basic services and handling hardware-based tasks. The application program 6022 may contain various applications, such as a media player, browser, etc., for implementing various application services. The implementation of the generative model-based magnetic resonance imaging general reconstruction method provided in this embodiment of the invention can be included in the application program 6022.

[0099] The methods disclosed in the above embodiments of the present invention can be applied to processor 601, or implemented by processor 601. Processor 601 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in processor 601 or by instructions in the form of software. The processor 601 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 601 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. General-purpose processor 601 may be a microprocessor or any conventional processor, etc. The steps of the accessory optimization method provided in the embodiments of the present invention can be directly reflected as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software module may be located in a storage medium, which is located in memory. The processor reads the information in the memory and combines it with its hardware to complete the steps of the aforementioned method.

[0100] In an exemplary embodiment, the electronic terminal 600 may be used by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), or complex programmable logic devices (CPLDs) to perform the aforementioned method.

[0101] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes: computer program code, which, when run on a computer, causes the computer to perform the method of any of the embodiments described above.

[0102] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium storing program code, which, when run on a computer, causes the computer to perform the method of any of the embodiments described above.

[0103] As used in this specification, the terms "component," "module," "system," etc., are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process and / or an execution thread, and components may be located on a single computer and / or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0104] Those skilled in the art will recognize that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0105] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0106] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0107] 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 units can be selected to achieve the purpose of this embodiment according to actual needs.

[0108] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0109] In the above embodiments, the functions of each functional unit can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0110] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, 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 of the various embodiments of this application. 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.

[0111] In summary, this application provides a universal MRI reconstruction method, system, and terminal based on a generative model. This application inputs fully sampled MRI K-space data into a generative model to train a universal MRI reconstruction model. Clinically acquired undersampled MRI K-space data is then input into the universal MRI reconstruction model as constraints to control image reconstruction, resulting in an initial MRI reconstructed image. Post-processing is then performed to obtain the final MRI reconstructed image. This application solves the technical problems of high model development and usage costs, poor model versatility, and severe motion artifacts in existing technologies. It is applicable to different organs, different MRI contrasts, arbitrary acceleration ratios, and arbitrary sampling masks, and maintains accurate and clear reconstruction results in flexible and varied clinical scenarios and under high-acceleration imaging conditions, thus having practical application value in clinical diagnosis and treatment and industry technology development.

[0112] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A general reconstruction method for magnetic resonance imaging based on a generative model, characterized in that, include: Fully sampled magnetic resonance k-space data is input into a generative model to train a general MRI reconstruction model. The clinically acquired undersampled magnetic resonance k-space data is input into the general MRI reconstruction model as a constraint condition to control the general MRI reconstruction model to perform image reconstruction, thereby obtaining an initial MRI reconstruction image. Post-processing operations are performed on the initial MRI reconstructed image to obtain the final MRI reconstructed image.

2. The general reconstruction method for magnetic resonance imaging based on a generative model according to claim 1, characterized in that, The step of inputting clinically acquired undersampled MRI K-space data into the general MRI reconstruction model as a constraint to control the general MRI reconstruction model to perform image reconstruction, and obtaining the initial MRI reconstructed image includes: The undersampled magnetic resonance K-space data is subjected to probability fitting using the maximum likelihood method to obtain data fidelity constraint terms. The objective function is constructed by fusing the data fidelity constraint term with the prior knowledge of the full-sample magnetic resonance k-space data distribution learned by the general MRI reconstruction model. The objective function is solved iteratively to obtain the initial MRI reconstructed image.

3. The general reconstruction method for magnetic resonance imaging based on a generative model according to claim 2, characterized in that, The generative model includes a diffusion model.

4. The general reconstruction method for magnetic resonance imaging based on a generative model according to claim 3, characterized in that, The process of fusing the data fidelity constraint term with the prior knowledge of the full-sample magnetic resonance k-space data distribution learned by the general MRI reconstruction model to construct the objective function includes: In each step of the inverse diffusion of the diffusion model, the initial estimate of the target image to be optimized in the current step is calculated using the Tweedie formula; The data fidelity constraint term is fused with the prior knowledge of the full-sample magnetic resonance K-space data distribution learned by the diffusion model to construct an objective function for optimizing the initial estimation of the target image.

5. The general magnetic resonance imaging reconstruction method based on a generative model according to claim 4, characterized in that, The objective function includes data fidelity constraints and image prior constraints; wherein, the iterative solution of the objective function to obtain the initial MRI reconstructed image includes: In each step of the inverse diffusion of the diffusion model, the initial estimate of the target image is iteratively optimized and updated based on the objective function to dynamically and adaptively adjust the fusion method and ratio of the data fidelity constraint term and the image prior constraint term in the objective function. The updated initial estimate of the target image is then superimposed with the corresponding inverse diffusion step noise to match the forward diffusion process, thereby completing the calculation of the current inverse diffusion step. Repeat the above optimization and update process, and superimpose the corresponding inverse diffusion step noise operation until all steps of inverse diffusion are completed, and obtain the initial MRI reconstructed image.

6. The general reconstruction method for magnetic resonance imaging based on a generative model according to claim 1, characterized in that, The post-processing operation performed on the initial MRI reconstructed image to obtain the final MRI reconstructed image includes: Edge smoothing and signal intensity normalization are performed on the initial MRI reconstructed image to obtain the final MRI reconstructed image.

7. The general reconstruction method for magnetic resonance imaging based on a generative model according to claim 1, characterized in that, Selectively inputting prompts into the generative model to guide the general MRI reconstruction model to learn the prior knowledge of the distribution of the full-sample magnetic resonance k-space data; selectively inputting prompts into the general MRI reconstruction model during use to guide the general MRI reconstruction model to output an initial MRI reconstruction image that matches the imaging scene corresponding to the prompt; The prompts include one or a combination of organ type, imaging parameters, age, gender, initial diagnosis information, genetic test indicators, and blood test indicators.

8. The general reconstruction method for magnetic resonance imaging based on a generative model according to claim 2, characterized in that, During the iterative solution of the objective function, strategies for accelerating inference, estimating patient motion parameters, and / or optimizing coil sensitivity maps are selectively employed. The accelerated inference strategy includes one or a combination of inference sampling optimization, model distillation, model pruning, and model quantization.

9. A universal magnetic resonance imaging reconstruction system based on a generative model, characterized in that, The system includes: The model training module is used to input fully sampled magnetic resonance K-space data into the generative model and train it to obtain a general MRI reconstruction model. The image reconstruction module is used to input clinically acquired undersampled magnetic resonance k-space data into the general MRI reconstruction model as a constraint condition to control the general MRI reconstruction model to perform image reconstruction and obtain an initial MRI reconstruction image. The post-processing module is used to perform post-processing operations on the initial MRI reconstructed image to obtain the final MRI reconstructed image.

10. An electronic terminal, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method as described in any one of claims 1 to 8.

Images