Four-dimensional flow field magnetic resonance data synthesis method and system based on CT images

Abstract

The application discloses a four-dimensional flow field magnetic resonance data synthesis method and system based on CT images, and is applied to the technical field of medical images, and the method comprises the steps of: acquiring a CTA image; performing image segmentation and three-dimensional reconstruction processing on the CTA image to obtain a target three-dimensional model of a thoracic aorta system; performing fluid dynamics simulation processing based on the target three-dimensional model to obtain blood flow field simulation parameters of the thoracic aorta system; and performing four-dimensional flow field magnetic resonance data synthesis processing based on the blood flow field simulation parameters to obtain magnetic resonance synthesis data corresponding to the CTA image. The application can generate 4D Flow MRI data from a CT image, and effectively improves the generation accuracy of the 4D Flow MRI data.

Description

Technical Field

[0001] This application relates to the field of medical imaging technology, and in particular to a method and system for synthesizing four-dimensional flow field magnetic resonance data based on CT images. Background Technology

[0002] 4D Flow MRI (four-dimensional flow field magnetic resonance imaging) achieves non-invasive quantitative blood flow analysis through three-dimensional spatial encoding and dynamic capture in the temporal dimension, combined with phase contrast technology. This technique provides high-resolution visualization data of blood flow streamlines, eddy current distribution, and energy loss without the need for ionizing radiation or contrast agents. In related technologies, the 4DflowNet model is used to generate 4D Flow MRI data. However, the training of the 4DflowNet model is highly dependent on existing 4D Flow MRI data, which is difficult to obtain. This leads to insufficient training of the 4DflowNet model, resulting in low accuracy in the generation of 4D Flow MRI data. Summary of the Invention

[0003] This application provides a method and system for synthesizing four-dimensional flow field magnetic resonance data based on CT images. It can generate 4D flow MRI data from CT images, effectively improving the generation accuracy of 4D flow MRI data.

[0004] On the one hand, embodiments of this application provide a method for synthesizing four-dimensional flow field magnetic resonance data based on CT images, including the following steps:

[0005] Acquire CTA images;

[0006] The CTA images are segmented and reconstructed in three dimensions to obtain a target three-dimensional model of the thoracic aortic system.

[0007] Based on the target three-dimensional model, fluid dynamics simulation processing is performed to obtain the blood flow field simulation parameters of the thoracic aortic system;

[0008] Based on the blood flow field simulation parameters, four-dimensional flow field magnetic resonance data synthesis processing is performed to obtain magnetic resonance synthesized data corresponding to the CTA image.

[0009] On the other hand, embodiments of this application provide a four-dimensional flow field magnetic resonance data synthesis system based on CT images, including:

[0010] The acquisition module is used to acquire CTA images;

[0011] The first processing module is used to perform image segmentation and three-dimensional reconstruction processing on the CTA image to obtain a target three-dimensional model of the thoracic aortic system.

[0012] The second processing module is used to perform fluid dynamics simulation processing based on the target three-dimensional model to obtain the blood flow field simulation parameters of the thoracic aortic system.

[0013] The third processing module is used to perform four-dimensional flow field magnetic resonance data synthesis processing based on the blood flow field simulation parameters to obtain magnetic resonance synthesis data corresponding to the CTA image.

[0014] According to an embodiment of this application, a method and system for synthesizing four-dimensional flow field magnetic resonance data based on CT images is provided. First, CTA images are acquired; then, image segmentation and three-dimensional reconstruction processing are performed on the CTA images to obtain a target three-dimensional model of the thoracic aortic system; subsequently, fluid dynamics simulation processing is performed based on the target three-dimensional model to obtain blood flow field simulation parameters of the thoracic aortic system; finally, four-dimensional flow field magnetic resonance data synthesis processing is performed based on the blood flow field simulation parameters to obtain magnetic resonance synthesized data corresponding to the CTA images. The technical solution of this application embodiment can generate 4D Flow MRI data from CT images without relying on existing 4D Flow MRI data, thereby effectively improving the generation accuracy of 4D Flow MRI data.

[0015] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a method for synthesizing four-dimensional flow field magnetic resonance data based on CT images provided in this application;

[0017] Figure 2A This is an example diagram of the threshold segmentation method provided in this application;

[0018] Figure 2B This is an example diagram of the region growing method provided in this application;

[0019] Figure 3 This is a schematic diagram illustrating the principle of model optimization provided in this application;

[0020] Figure 4 This is a flowchart of the CFD simulation provided in this application;

[0021] Figure 5 This is an example diagram of the velocity inlet and pressure outlet provided in this application;

[0022] Figure 6 This is a structural diagram of the Windkessel model provided in this application;

[0023] Figure 7 This is an example diagram of a CFD simulation result provided in this application;

[0024] Figure 8 This is another example of the CFD simulation results provided in this application;

[0025] Figure 9 This is the effect verification diagram provided in this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0028] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0030] Cardiovascular diseases, characterized by high morbidity, high disability rate, and high mortality rate, have become a major global public health challenge. In the diagnosis and treatment of these diseases, hemodynamic parameters are a core assessment indicator, but their accurate acquisition still faces technical bottlenecks: Transthoracic echocardiography (TTE), while capable of real-time measurement of blood flow velocity using Doppler technology, carries a risk of systematic overestimation in pressure gradient calculation; Magnetic resonance imaging (MRI), while enabling four-dimensional flow imaging (4D flow) phase-encoded full-cycle blood flow tracking, still relies on complex phase contrast sequences for hemodynamic assessment, resulting in limitations such as time-consuming examinations, high costs, and high requirements for patient cooperation; CT angiography (CTA), although capable of clearly displaying the three-dimensional structure of blood vessels and the anatomical details of surrounding tissues, cannot perform hemodynamic analysis.

[0031] It is worth noting that currently no single imaging method can simultaneously meet the clinical needs of anatomical visualization and dynamic blood flow analysis. This makes multimodal image fusion and computational fluid dynamics (CFD)-assisted modeling a key direction for overcoming technical barriers: by integrating the anatomical accuracy of CTA, the dynamic blood flow data of 4D FlowMRI, and the hemodynamic simulation of CFD, a comprehensive assessment system for vascular morphology and function can be constructed.

[0032] As a revolutionary tool for hemodynamic assessment, 4D Flow MRI achieves non-invasive quantitative blood flow analysis through three-dimensional spatial encoding and dynamic capture in the temporal dimension, combined with phase contrast technology. This technology provides high-resolution visualization data on blood flow streamlines, eddy current distribution, and energy loss without the need for ionizing radiation or contrast agents. It is widely used in the diagnosis and preoperative planning of lesions such as valvular heart disease, congenital heart disease, aortic aneurysm, and cerebrovascular abnormalities, and is particularly adept at detecting complex hemodynamic abnormalities. Despite challenges such as long scan times, limited spatiotemporal resolution, and complex data processing, the clinical value of 4D Flow MRI continues to improve in precision medicine and personalized treatment, thanks to advancements in AI-accelerated reconstruction and multimodal image fusion.

[0033] The introduction of deep learning technology has brought about a breakthrough in the clinical application of 4D Flow MRI. Algorithms based on architectures such as convolutional neural networks and Transformers can significantly shorten image reconstruction time and improve the spatiotemporal resolution of images. For example, high-fidelity dynamic images of blood flow can be rapidly generated from sparsely sampled k-space data, while simultaneously achieving automated segmentation and quantification analysis of blood flow parameters. In addition, super-resolution models can reconstruct sub-millimeter-level high-resolution blood flow velocity fields by learning the spatiotemporal characteristics of low-resolution 4D Flow MRI data, accurately restoring key details such as thin vessel wall boundaries and microscale eddy features. At the same time, by combining the physical prior knowledge of hemodynamic equations, the spatial continuity and temporal dynamic consistency of the velocity vector field are ensured.

[0034] However, the training of such models is highly dependent on a large amount of patient data. The raw data of 4D Flow MRI contains detailed anatomical information of the patient's heart and hemodynamic sensitive information. Under privacy constraints, publicly shared standardized datasets are extremely limited. This leads to limited generalization ability of the model and the risk of data bias. In particular, the training of super-resolution models requires pairs of low- and high-resolution blood flow data. However, the cost of obtaining high-resolution real data in actual clinical practice is high. Super-resolution models based on spatiotemporal attention mechanisms often face the risk of overfitting due to data scarcity.

[0035] To address this issue, several methods exist for obtaining high-quality simulated 4D flow MRI data:

[0036] (1) 4DflowNet: This study generates synthetic 4D flow MRI data by combining CFD simulation with MRI physical models. First, CFD simulations were performed using three 3D aortic models, including one model generated from 4D flow MRI data of healthy volunteers through threshold segmentation and Paraview surface extraction, and two models of constricted aorta cases from the MICCAI-STACOM CFD Challenge. Boundary conditions such as inlet velocity waveform, outlet constant pressure, and branch pressure waveform were set, and the Navier-Stokes equations were solved to obtain a high-resolution (HR) blood flow velocity field. Subsequently, the CFD velocity components were mapped to MRI phase images, and corresponding amplitude images were generated. The HR data was converted to the frequency domain (k-space) using 3D Fourier transform, high-frequency components were truncated to achieve a 2x downsampling, and Gaussian noise was added to the k-space to simulate the noise distribution of real MRI, ultimately generating low-resolution (LR) noisy data.

[0037] However, the 4DflowNet model has the following drawbacks:

[0038] Firstly, the simplification assumptions of the CFD model: the construction of the CFD model in the 4DflowNet model often relies on multiple simplification assumptions, such as setting its outlet to constant static pressure ( This setting can affect the accuracy of downstream flow, causing the flow pattern in CFD simulations to deviate from the actual physiological state, and ultimately affecting the model's prediction accuracy for complex blood flow.

[0039] Secondly, the source of the 3D model is flawed: the 3D model of the aorta of healthy volunteers used to generate synthetic data in the 4DflowNet model is derived from the segmentation and reconstruction of 4D Flow MRI data. This results in the model's data acquisition being limited by the scarcity of 4D Flow MRI data. Furthermore, the low resolution of the original 4D Flow MRI will lead to a rough surface of the segmented 3D model, which in turn causes errors in the CFD simulation.

[0040] Third, insufficient validation data: The training and validation process of the 4DflowNet model relies heavily on existing 4D FlowMRI data, which results in insufficient training of the model. Moreover, this leads to the model over-adapting to idealized CFD simulation features and failing to fully capture the complex noise patterns, motion artifacts, and pathology-related blood flow heterogeneity in real 4D FlowMRI data.

[0041] (2) SynRealSeg-4DFlow: This study generates diverse three-dimensional models of the aorta using a statistical shape model (SSM), combines CFD to simulate hemodynamics, and generates phase-contrast magnetic resonance angiography (PCMRA) images based on time-averaged velocity fields. The low signal-to-noise ratio and partial volume effect of the simulated real 4D flow MRI are adjusted by spatial downsampling, adding Gaussian noise and VENC value, while randomizing the Magnitude signal intensity to enhance data diversity. Finally, by blurring the edges with Gaussian filtering, synthetic data with highly realistic anatomical morphology, blood flow patterns and imaging parameters are generated. The 3D U-Net model is trained with real data, which improves the DICE score of aortic segmentation from 0.65 to 0.83, effectively solving the problems of scarcity of real data and insufficient generalization.

[0042] However, SynRealSeg-4Dflow has the following drawbacks:

[0043] First, there is a lack of pathological models: the synthetic data generated in the SynRealSeg-4Dflow model is based only on healthy aortic models and does not include any disease models. This leads to segmentation failures or decreased accuracy in real clinical cases when the segmentation model trained on these data is used.

[0044] Secondly, it cannot perform individualized modeling: Although the SSM used in the SynRealSeg-4Dflow model can generate diverse aortic morphologies through principal component analysis, its essence is parametric modeling based on population statistical characteristics. This leads to significant deficiencies in individualized simulation of synthetic blood vessel models. Specifically, SSM constructs statistical distributions based on the morphological features of a limited sample. The generated new models can only cover common variations in the population, but cannot accurately simulate the unique anatomical details of real patients, thus failing to achieve modeling and simulation for a specific patient.

[0045] It is evident that although there are studies that use CFD simulation to obtain synthetic 4D Flow MRI data for training and validation of neural models, existing technologies suffer from problems such as simple background modeling, scarce data, lack of pathological models, and inability to personalize modeling, which leads to low accuracy in the generation of 4D Flow MRI data.

[0046] Therefore, this application provides a method and system for synthesizing four-dimensional flow field magnetic resonance data based on CT images. High-fidelity blood flow simulation is achieved through multimodal data fusion and physical constraint enhancement, used to generate 4D FlowMRI simulation data from CT images. Unlike traditional simplified models, this application employs dynamic Windkessel boundary conditions to simulate peripheral vascular impedance effects and rapidly generates individualized case models (such as aortic aneurysms and coronary artery stenosis) from CT images. By introducing device-specific noise models and physiological motion artifact simulation, the synthesized 4D Flow MRI data output by Windkessel boundary conditions significantly outperforms existing methods in terms of geometric fidelity and hemodynamic consistency. This provides a training data foundation for deep learning models that combines physiological realism with pathological diversity, promoting the clinical translation of precision medicine.

[0047] First, the implementation steps of a four-dimensional flow field magnetic resonance data synthesis method based on CT images provided in this application will be described in detail below with reference to the accompanying drawings.

[0048] It should be noted that in all specific embodiments of this disclosure, when processing based on data such as CTA images is required, the permission or consent of the subject is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this disclosure require the acquisition of data such as CTA images, separate permission or consent from the subject is obtained through pop-ups or redirection to a confirmation page. Only after obtaining the subject's separate permission or consent is the necessary data, such as CTA images, required for the proper functioning of the embodiments of this disclosure be acquired.

[0049] This application provides a method for synthesizing four-dimensional flow field magnetic resonance data based on CT images. This method can be applied to terminals, servers, or software running on either terminal or server. Terminals can be tablets, laptops, desktop computers, etc., but are not limited to these. Servers can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Furthermore, a server can be a node server in a blockchain network, but is not limited to these. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0050] Reference Figure 1 The above-mentioned four-dimensional flow field magnetic resonance data synthesis method may include the following steps S101-S104:

[0051] S101, acquire CTA image.

[0052] It should be noted that CTA images are image data pre-stored in a database, rather than image data obtained from the patient in real time.

[0053] S102, perform image segmentation and 3D reconstruction processing on CTA images to obtain the target 3D model of the thoracic aortic system.

[0054] In this step, firstly, image segmentation is performed on the CTA image to extract the relevant regions of the thoracic aortic system; then, three-dimensional reconstruction is performed based on the segmentation results to reconstruct a three-dimensional model of the thoracic aorta, thereby obtaining the target three-dimensional model of the thoracic aortic system.

[0055] S103, based on the target three-dimensional model, performs fluid dynamics simulation processing to obtain the blood flow field simulation parameters of the thoracic aortic system.

[0056] It should be noted that the blood flow field simulation parameters are the same as the hemodynamic parameters, which may include, but are not limited to, physical quantities in the blood flow field such as velocity, pressure, temperature, density, and viscosity.

[0057] In this step, the target 3D model is used as the baseline for CFD simulation. Simply put, the target 3D model is meshed, dividing the continuous blood flow field into multiple small discrete units. Dynamic boundary conditions for the CFD simulation are defined based on physiological parameters and the Windkessel principle of three elements. Hemodynamic parameters are simulated using the Navier-Stokes (NS) equations, and time-resolved blood flow field calculations are performed in conjunction with the cardiac cycle, thus obtaining the simulation parameters of the blood flow field in the thoracic aortic system.

[0058] S104. Based on blood flow field simulation parameters, four-dimensional flow field magnetic resonance data synthesis processing is performed to obtain magnetic resonance synthesis data corresponding to CTA images.

[0059] In this step, the CFD blood flow field simulation data based on time averaging and spatial downsampling is used to generate noisy 4D Flow MRI simulation data through phase encoding model, golden angle radial undersampling and multi-coil k-space synthesis. This data is the magnetic resonance synthesis data corresponding to the CTA images and can be used for downstream tasks such as model training.

[0060] In some implementations, refer to Figure 1In step S102 above, image segmentation and three-dimensional reconstruction processing of the CTA image are performed to obtain a target three-dimensional model of the thoracic aortic system, which may include:

[0061] Image segmentation processing is performed on CTA images to obtain the target tissue region associated with the thoracic aorta; three-dimensional reconstruction processing is performed based on the target tissue region to obtain an initial three-dimensional model of the thoracic aortic system; the initial three-dimensional model is optimized to obtain the target three-dimensional model.

[0062] In this embodiment, CT medical image data cannot be directly used for hemodynamic analysis due to the lack of fluid domain spatial topology information. It requires three-dimensional geometric reconstruction before it can be combined with CFD simulation technology for numerical simulation analysis. Therefore, the target tissue region associated with the thoracic aorta is segmented from the CTA images. After accurate segmentation of the region of interest of the thoracic aorta, a coarse three-dimensional model of the target region, i.e., the initial three-dimensional model of the thoracic aortic system, needs to be generated based on this using three-dimensional reconstruction technology to facilitate CFD simulation processing. However, due to mesh quality defects and other issues in the coarse three-dimensional model, it needs to be optimized to eliminate mesh quality defects and meet the finite element calculation requirements of CFD simulation processing. Optimization yields the target three-dimensional model of the thoracic aortic system.

[0063] Optionally, the image segmentation method can be set according to the actual situation, and this implementation method does not limit it. For example, deep learning models such as U-Net can be used to implement image segmentation, but it is not limited to this.

[0064] Optionally, the method of 3D reconstruction can be set according to the actual situation, and this embodiment does not limit it. For example, 3D reconstruction technology such as diffusion model can be used to achieve 3D reconstruction, but it is not limited to this. For another example, the MIMICS software interface is displayed on the terminal. The MIMICS software interface includes 3D reconstruction function controls. The MIMICS software interface imports and displays CTA images in DICOM format. In response to the trigger command of the 3D reconstruction function controls, the 3D reconstruction function window is displayed. The 3D reconstruction function window includes an algorithm selection area, a smoothing factor setting area, an iteration number setting area, and a first determination control. In response to the trigger command of the algorithm selection area, the 3D reconstruction algorithm is set to a continuous optimization algorithm. In response to the trigger command of the smoothing factor setting area, the smoothing factor is set. In response to the trigger command of the iteration number setting area, the iteration number is set. In response to the trigger command of the first determination control, the 3D reconstruction operation is performed based on the continuous optimization algorithm, the smoothing factor, and the iteration number to obtain an initial 3D model. Here, the 3D reconstruction function of the MIMICS software is called. By selecting "continuous optimization" as the benchmark algorithm, a smoothing factor is set and iterated multiple times to process the segmentation mask (i.e. the target tissue region) to obtain the initial 3D model. The smoothing factor and the number of iterations can be flexibly set according to the actual situation. For example, the smoothing factor can be 0.4 and the number of iterations can be 6, but it is not limited to these.

[0065] In some implementations, refer to Figure 1 The above-mentioned image segmentation processing of CTA images to obtain the target tissue region associated with the thoracic aorta may include:

[0066] The CTA images were segmented using a thresholding method to obtain the initial tissue region associated with the thoracic aorta; the initial tissue region was then refined using a region growing method to obtain the target tissue region.

[0067] In this embodiment, to improve the accuracy and efficiency of image segmentation, a threshold segmentation method is used to achieve image segmentation, such as... Figure 2A As shown, it classifies the pixels of CTA images into different anatomical structural regions by setting a grayscale threshold. When the grayscale gradient difference between the target structural region (i.e., the thoracic aorta region) and the background tissue is significant, tissue segmentation can be effectively achieved, thereby obtaining the initial tissue region associated with the thoracic aorta.

[0068] It should be understood that threshold segmentation is an existing technology. Simply put, threshold segmentation, as a fundamental theoretical method in the field of image segmentation, has the technical advantages of high computational efficiency and strong real-time performance. Its core relies on gray-level histogram analysis. When the target structural region (i.e., the thoracic aorta region) and the background tissue show a significant gray-level difference, the histogram exhibits a bimodal distribution characteristic. That is, the target structural region (i.e., the thoracic aorta region) and the background tissue each correspond to two probability density peaks. The probability density valley between the two peaks usually corresponds to the optimal segmentation threshold, at which point the best segmentation effect can be achieved.

[0069] Optionally, the segmentation threshold can be set according to the actual situation, and this embodiment does not limit it. For example, the segmentation threshold is set to a range of 320HU-1000HU, but it is not limited to this.

[0070] After tissue segmentation, although the thoracic aorta region in CTA images shows a significant difference in grayscale contrast compared to other structures such as bones, lungs, and liver, initial separation can be achieved using thresholding. However, due to the grayscale overlap between the thoracic aorta and cardiac anatomy, the thoracic aorta region cannot be directly segmented. In other words, the initial tissue region associated with the thoracic aorta obtained through thresholding contains some error. Therefore, a region growing method is used to remove the extra tissue region, thereby obtaining the target tissue region associated with the thoracic aorta, such as... Figure 2B As shown in the figure. This can effectively improve the segmentation accuracy and efficiency of the thoracic aorta region.

[0071] Furthermore, region growing is also an existing technology. Simply put, before using region growing, the operator needs to use a lasso tool through the front-end interface to isolate the aorta and adjacent blood vessels, thus forming an independent, closed anatomical region. After this, the region growing method is executed: any coordinate point within the lumen of the thoracic aorta is determined as the initial seed point, and the region growing method is performed based on a preset grayscale similarity threshold. When the growth boundary encounters a non-connected anatomical structure region, a growth termination mechanism is automatically triggered through three-dimensional topological connectivity detection, ultimately achieving fully automated extraction of the region of interest in the aorta.

[0072] In some implementations, refer to Figure 3 The above optimization process of the initial 3D model to obtain the target 3D model may include:

[0073] The initial 3D model is preprocessed to obtain a preprocessed initial 3D model; the preprocessed initial 3D model is decomposed to obtain multiple surface patches; multiple grid nodes are constructed based on the multiple surface patches; and the target 3D model is generated based on the multiple grid nodes.

[0074] In this embodiment, in order to eliminate the mesh quality defects in the initial 3D model and to ensure that the 3D model of the thoracic aortic system meets the finite element calculation requirements of CFD simulation, surface fitting needs to be achieved through multi-scale geometric optimization operations:

[0075] First, a curvature continuity preservation algorithm is used to perform global mesh smoothing on the initial 3D model to eliminate surface noise and non-physiological topological distortions. For regions with abrupt changes in curvature, an adaptive Gaussian filtering algorithm is used to perform gradient-constrained smoothing on the initial 3D model to balance geometric accuracy and smoothness. For complex anatomical junctions such as vascular bifurcation, a boundary constraint cutting algorithm is used to reconstruct the physiological flow field boundary of the initial 3D model. Specifically, the normal of the cutting plane is determined based on the tangential vector of the vessel centerline at the bifurcation port in the initial 3D model, thereby generating a reference plane perpendicular to the blood flow direction to define the outlet and inlet boundaries of the initial 3D model. Through the above preprocessing operations, noise and unreasonable distortions in the initial 3D model can be effectively eliminated, and the surface smoothness of the initial 3D model can be improved, thus effectively improving the modeling quality of the 3D model of the thoracic aortic system.

[0076] Then, based on the differential geometric topological features of the initial three-dimensional model (i.e., the umbilical point trajectory formed by the extreme points of Gaussian curvature and the feature network formed by the interweaving of the zero-value lines of average curvature in the initial three-dimensional model), the preprocessed initial three-dimensional model is decomposed into multiple quadrilateral parameterized surface patches, which together constitute the complete vascular surface.

[0077] Then, in the parameter space of the parametric surface patch ( In, along , Orthogonal parameter lines are generated at equal intervals. These equidistant coordinate lines interweave on the 3D surface to form a bottom-level grid framework. The bottom-level grid framework includes multiple grid nodes, each storing parameter coordinates, 3D spatial coordinates, and principal curvature tensors, collectively describing the anatomical morphology and geometric features of the thoracic aortic vessel surface. By iteratively adjusting the grid node coordinates, the surface morphology optimization and curvature distribution balance of the 3D model of the thoracic aortic system can be achieved.

[0078] Finally, the topologically continuous high-quality parametric grid is transformed into a finite element mesh dominated by structured quadrilaterals, thereby achieving surface fitting of the three-dimensional model of the thoracic aortic system. These finite element meshes dominated by structured quadrilaterals will together constitute the target three-dimensional model of the thoracic aortic system.

[0079] The above steps can effectively improve the modeling quality of the 3D model of the thoracic aortic system. While preserving the complete vascular anatomical features, it ensures that the 3D model of the thoracic aortic system, such as element orthogonality and aspect ratio, meets the requirements of subsequent CFD simulation for mesh convergence and computational stability, thereby ensuring the accuracy of subsequent CFD simulation.

[0080] In some implementations, refer to Figure 4 In step S103 above, fluid dynamics simulation processing is performed based on the target three-dimensional model to obtain the blood flow field simulation parameters of the thoracic aortic system, which may include:

[0081] The hemodynamic control equations and boundary conditions of the thoracic aortic system are obtained; the target three-dimensional model is meshed to obtain a finite element model; the finite element model includes multiple discrete elements, each of which is constrained by boundary conditions; the finite element model is solved based on the hemodynamic control equations to obtain the blood flow field simulation parameters.

[0082] In this embodiment, the flow field model used in CFD simulation is a target three-dimensional model of the thoracic aortic system, and the flow field refers to the blood flow field of the thoracic aorta. First, the hemodynamic control equations and boundary conditions of the thoracic aortic system are set, such as the inlet velocity waveform and outlet pressure, to correctly solve the algebraic equations. Simultaneously, the target three-dimensional model is meshed to divide the continuous vascular flow field model into multiple small discrete elements (volume meshes). These small discrete elements collectively constitute the finite element model, and they are constrained by the aforementioned boundary conditions.

[0083] It is understandable that meshing is an existing technology. Its essence is to transform a continuous flow field model into a finite element calculation model through spatial discretization, that is, to transform the continuous control equations into a set of algebraic equations for discrete nodes (elements). This process is based on the discretization principle of differential equations: when the spacing between mesh elements meets the convergence condition, the original fluid control equations can be transformed into a set of difference equations based on Taylor expansion. By solving the algebraic equations corresponding to each discrete element, the global flow field parameter distribution can be reconstructed, that is, the final output hemodynamic parameter field, including velocity field, pressure field, etc.

[0084] It is important to note that although the governing equations are unified, the coefficient matrices of the discrete equations differ at different locations: the internal discrete units follow the standard discretization scheme, while the boundary discrete units require the introduction of boundary condition correction terms. This can be understood as the equations for each discrete unit being "custom-made".

[0085] In the meshing process, firstly, a surface mesh is generated on the geometric surface. The boundary layer morphology of the flow field model is discretized using triangular or quadrilateral elements, accurately capturing the wall flow characteristics of the flow field model and providing a basis for assigning boundary conditions. Based on this, a volume mesh is constructed extending into the fluid domain. The three-dimensional space is topologically partitioned using tetrahedral, hexahedral, or prism-like layered elements, ensuring that the discretization solution conditions of the conservation law equations (i.e., governing equations) are satisfied within each volume mesh. This layered discretization strategy not only guarantees the analytical accuracy of key surface features such as wall curvature and flow separation in the fluid model, but also balances the computational efficiency and numerical stability of the complex geometric domain in the fluid model through structured / unstructured hybrid mesh technology. Ultimately, this lays the discretization foundation for the full-dimensional solution of physical quantities such as flow pressure, velocity, and turbulent kinetic energy.

[0086] By making this discretization assumption, the flow conditions within each small cell are relatively uniform, allowing the numerical solution to approximate the solution for the entire region on these cells. After meshing, the flow field model transforms from a continuous geometric region into an approximate field composed of a large number of discrete cells. This approximation is used iteratively to continuously approximate the true solution, with residuals set during the iteration process to measure the imbalance of the equations. Iteration stops when the residuals reach a set threshold, thus obtaining an accurate solution. Combining the solutions from these discrete cells approximates the solution for the entire flow field. The threshold and maximum iteration step can be flexibly set according to actual conditions; for example, the threshold could be set to 0.001 and the maximum iteration step to 100, but this is not a limitation.

[0087] Understandably, the iterative process of CFD simulation is a current technology. Its core lies in solving the hemodynamic control equations numerically, and the adaptability of the equation set depends on the reasonable definition of the boundary conditions. The residuals involved play a similar role to the loss function in guiding parameter optimization during neural network training. In the iterative solution process of CFD simulation, the residuals are used to quantify the imbalance of the hemodynamic control equations (such as mass conservation and momentum conservation). When the residuals converge to a set threshold, it is determined that the numerical solution has sufficiently approximated the real physical field. At this point, the iteration is terminated and the converged solution is output.

[0088] It is worth noting that these discrete elements serve as discretized representations of the flow field model's boundaries, acting as direct objects of wall boundary conditions (such as wall shear stress calculations), thus forming the closed boundary of the flow field model's solution domain. The solution of a discrete element literally refers to the solution of a system of partial differential equations, i.e., the physical quantity of each element in the entire flow field model. However, in numerical calculations, this implementation obtains approximate values ​​of the physical quantities on each discrete element. Therefore, strictly speaking, this implementation obtains an approximate solution of the entire flow field model on these discrete elements. The distribution of physical quantities in the entire flow field model can be approximately reconstructed using the values ​​of these discrete elements. Specifically, the solution of each discrete element is a vector set containing multiple physical quantities. .in, This refers to the vein value of the node (unit: Pa). It refers to a three-dimensional velocity vector (unit: m / s). It refers to the shear stress tensor (unit: ).

[0089] In some implementations, obtaining the hemodynamic control equations and boundary conditions of the thoracic aortic system may include:

[0090] Obtain the mass conservation equation, momentum conservation equation, energy conservation equation, state equation, and heat conduction equation of the target 3D model; simplify the mass conservation equation and momentum conservation equation to obtain simplified mass conservation equation and simplified momentum conservation equation; determine the energy conservation equation, state equation, heat conduction equation, simplified mass conservation equation, and simplified momentum conservation equation as the hemodynamic control equations.

[0091] In this embodiment, obtaining the hemodynamic parameters of the thoracic aortic system through CFD simulation essentially involves solving the fundamental equations of hemodynamics in the thoracic aortic system. Blood in the thoracic aortic system, as a fluid, follows the fundamental physical laws of fluid dynamics, namely the equations of mass conservation, momentum conservation, energy conservation, state equations, and heat conduction equations. The solutions to these equations are the hemodynamic parameters, including physical quantities such as velocity, pressure, temperature, density, and viscosity.

[0092] The mass equation satisfies the following formula (1):

[0093] (1);

[0094] In equation (1), This indicates the fluid density of the thoracic aortic system; This represents the fluid velocity vector of the thoracic aortic system; Indicates time.

[0095] The momentum equation satisfies the following formula (2):

[0096] (2);

[0097] In equation (2), This indicates the shear stress in the thoracic aortic system; It represents velocity and acceleration.

[0098] The energy equation satisfies the following formula (3):

[0099] (3);

[0100] In equation (3), This represents the fluid internal energy of the thoracic aortic system; This indicates the fluid pressure in the thoracic aortic system; Indicates the fluid temperature of the thoracic aortic system; Indicates thermal conductivity; This represents the viscous dissipation function.

[0101] The state equations satisfy the following formula (4):

[0102] (4).

[0103] The heat conduction (internal energy) equation satisfies the following formula (5):

[0104] (5).

[0105] However, the blood in the thoracic aortic system not only possesses the characteristics of a general fluid but also exhibits its own unique properties. This embodiment assumes that the blood flow in the thoracic aortic system is an incompressible viscous Newtonian fluid. This assumption has two implications: incompressibility implies a constant blood density, while the viscous Newtonian fluid assumption implies that the blood viscosity is constant and that shear stress is linearly related to the velocity gradient. Optionally, this embodiment considers the blood flow in the thoracic aortic system as laminar flow, for example, with a density of 10⁶ K⁻¹. The viscosity is 0.0035. However, it is not limited to this. Under this assumption, the mass conservation equation and the momentum conservation equation can be simplified to the Navier-Stokes equation, resulting in the simplified mass conservation equation shown in formula (6) and the simplified momentum conservation equation shown in formula (7):

[0106] (6);

[0107] (7);

[0108] In equation (7), The formula represents the dynamic viscosity of blood in the thoracic aortic system. The left side of the formula represents the inertial force, i.e., the acceleration term, of the fluid per unit volume in the thoracic aortic system, while the right side of the formula represents the external forces acting on a unit volume in the thoracic aortic system, such as pressure gradient, gravity, viscous force, and other physical quantities.

[0109] Finally, the energy conservation equation, state equation, heat conduction equation, simplified mass conservation equation, and simplified momentum conservation equation were determined as the hemodynamic governing equations. Constructing the hemodynamic governing equations based on the fundamental physical laws of fluids ensures that they closely match the flow field conditions of the thoracic aortic system, thereby improving their accuracy and contributing to the accuracy of CFD simulations. Furthermore, simplifying some of the governing equations makes them more concise, thus improving the efficiency of CFD simulations.

[0110] In some implementations, the finite element model may include a velocity inlet; the boundary conditions may include velocity inlet boundary conditions, which are used to limit the velocity inlet; obtaining the hemodynamic control equations and boundary conditions of the thoracic aortic system may include:

[0111] The velocity time-varying functions of the systolic cycle, isovolumetric relaxation cycle, and diastolic cycle in the thoracic aortic system are determined as the velocity inlet boundary conditions; among them, the velocity time-varying function of the systolic cycle adopts a positive exponential-sine composite function, the velocity time-varying function of the isovolumetric relaxation cycle adopts a negative decay function, and the value of the velocity time-varying function of the diastolic cycle is always zero.

[0112] In this implementation, boundary conditions serve as additional constraints in the partial differential equation boundary value problem. By defining the evolution of physical quantities (such as velocity and pressure distribution) at the boundary of the computational domain, they limit the degrees of freedom of the equation solution, thereby ensuring that the numerical iteration converges to a unique solution in physical reality. Accordingly, this implementation adopts a hybrid boundary condition strategy: a velocity boundary condition is set at the flow inlet of the finite element model, forcing the flow field model to satisfy the velocity distribution function given by experiment or theory; a pressure boundary condition is applied in the outlet region of the finite element model, and the energy balance relationship of the open boundary in the flow field model is constructed by fixing the static pressure reference value. This physical-numerical dual constraint mechanism not only strictly follows the basic principles of mass and momentum conservation in the flow field model, but also effectively suppresses numerical pseudo-diffusion and backflow instability in CFD simulation through the dynamic coupling of the inlet velocity gradient and outlet pressure relaxation in the flow field model, ultimately achieving a closed solution of the global flow field parameters, thereby effectively improving the accuracy of CFD simulation.

[0113] Specifically, such as Figure 5As shown, the finite element model of the thoracic aortic system can include one velocity inlet and four pressure outlets. In this embodiment, the boundary conditions of the velocity inlet in the CFD simulation are modeled with physiological drive to limit the velocity inlet: based on the characteristics of the cardiac cycle, the time-varying function of velocity is divided into three phases: systole (0~0.24s), isovolumetric relaxation (0.24~0.274s), and diastole (0.274~0.8s).

[0114] During the contraction period, a positive exponential-sine composite function is used, as shown in the following formula (8):

[0115] (8);

[0116] Equation (8) simulates the pulsatile blood flow driven by left ventricular contraction. The exponential term reflects the blood viscosity dissipation effect, while the sine term characterizes the pressure wave propagation characteristics.

[0117] The isovolumetric relaxation period switches to a negative decay function, as shown in the following formula (9):

[0118] (9);

[0119] Equation (9) describes the minute backflow phenomenon that may occur at the moment the thoracic aortic valve closes by using the negative velocity term.

[0120] Since no new blood enters the thoracic aorta during diastole, the value of the diastolic velocity time-varying function is always zero, corresponding to the physiological state of no new blood injection during the cardiac filling phase.

[0121] In some implementations, the finite element model may include four pressure outlets; the boundary conditions may include pressure outlet boundary conditions used to limit the pressure outlets; obtaining the hemodynamic control equations and boundary conditions of the thoracic aortic system may include:

[0122] Based on the attribute data of the thoracic aortic system, a first resistance, a second resistance, and a capacitance are set for each aortic branch in the thoracic aortic system. Based on the first resistance, second resistance, and capacitance of each aortic branch, each aortic branch is equivalent to a corresponding hemodynamic circuit model. The hemodynamic circuit model includes a first resistance, a second resistance, and a capacitance. One end of the first resistance is connected to one end of the capacitance and one end of the second resistance, respectively. The other end of the capacitance and the other end of the second resistance are both grounded. The capacitance and the second resistance are connected in parallel. According to the hemodynamic circuit model of each aortic branch, the mapping relationship between the coupled blood pressure and volumetric flow rate of each aortic branch is obtained as the pressure outlet boundary condition.

[0123] In this implementation, boundary conditions serve as additional constraints in the partial differential equation boundary value problem. By defining the evolution of physical quantities (such as velocity and pressure distribution) at the boundary of the computational domain, the degrees of freedom of the equation solution are limited, thereby ensuring that the numerical iteration converges to a unique solution in physical reality. Accordingly, this implementation adopts a hybrid boundary condition strategy: a velocity boundary condition is set at the flow inlet of the finite element model to force the flow field model to satisfy the velocity distribution function given by experiment or theory; a pressure boundary condition is applied in the outlet region of the finite element model, and the energy balance relationship of the open boundary in the flow field model is constructed by fixing the static pressure reference value. This physical-numerical dual constraint mechanism not only strictly follows the basic principles of mass and momentum conservation in the flow field model, but also effectively suppresses numerical pseudo-diffusion and backflow instability in CFD simulation through the dynamic coupling of the inlet velocity gradient and outlet pressure relaxation in the flow field model, ultimately achieving a closed solution of the global flow field parameters, thereby effectively improving the accuracy of CFD simulation.

[0124] Specifically, such as Figure 5 As shown, the finite element model of the thoracic aortic system can include one velocity inlet and four pressure outlets, with each of the four pressure outlets corresponding to one of the four aortic branches. For each pressure outlet (aortic branch), this embodiment uses a three-element Windkessel model to construct a hemodynamic equivalent circuit for pressure reconstruction. It should be understood that the Windkessel model is used to describe the pressure and flow relationship in the heart and aortic system. Its core idea is to treat the thoracic aortic system as an elastic container, simulating pressure and flow changes in the blood flow of the thoracic aortic system by mimicking components such as resistors and capacitors in a circuit. Accordingly, for each aortic branch, the following operations are performed:

[0125] Step 1: To achieve personalized blood flow simulation, a first resistor needs to be set. Second resistor and capacitor The attribute data of the thoracic aortic system may include the inlet volumetric flow rate, mean blood pressure, and total volumetric flow rate of the thoracic aortic system, as well as the volumetric flow rate and mean blood pressure of each aortic branch. First, the ratio of the inlet volumetric flow rate to the mean blood pressure is determined as the total electrical capacity of the thoracic aortic system, as shown in the following formula (10):

[0126] (10);

[0127] In equation (10), Indicates total electrical capacity; Indicates the inlet volumetric flow rate; This indicates average blood pressure.

[0128] Then, the ratio of the current aortic branch's volumetric flow rate to the total volumetric flow rate is determined as the current aortic branch's volumetric flow rate distribution ratio, and the product of the current aortic branch's volumetric flow rate distribution ratio and the total capacitance is calculated as the current aortic branch's capacitance, as shown in the following formula (11):

[0129] , (11);

[0130] In equation (11), This indicates the current volumetric flow rate distribution ratio of the aortic branches; This indicates the current volumetric flow rate of the aortic branch; This indicates the total volumetric flow rate.

[0131] Accordingly, the ratio of the mean blood pressure of the current aortic branch to the volumetric flow rate of the current aortic branch is determined as the total resistance of the current aortic branch, as shown in the following formula (12):

[0132] (12);

[0133] In equation (12), This indicates the total resistance of the current aortic branch; This indicates the average blood pressure of the current aortic branches.

[0134] Based on this, the first resistor is set according to the ratio of the first resistor to the total resistance, as shown in the following formula (13):

[0135] (13);

[0136] In equation (13), This indicates the ratio of the first resistor to the total resistance. It can be set according to the actual situation. For example, the value range of the ratio of the first resistor to the total resistance is 4% to 6%, but it is not limited to this.

[0137] The second resistance is the difference between the total resistance and the first resistance, as shown in formula (14) below:

[0138] (14).

[0139] Step 2: Based on the first resistor Second resistor and capacitor The current aortic branch is equivalent to the corresponding hemodynamic circuit model, such as... Figure 6 As shown, the first resistor One end is connected to the capacitor One end and the second resistor One end is connected to the capacitor. The other end and the second resistor The other end is grounded, the second resistor and capacitor in parallel.

[0140] This model simulates the impedance characteristics of the vascular system of the current aortic branch using a resistor-capacitor coupling network: the first resistor... The characteristic impedance characterizing proximal blood vessels reflects energy loss during pressure wave propagation; the second resistance Characterizing peripheral vascular resistance, it describes the dissipative properties of blood in the microcirculation of the thoracic aortic system; capacitance Characterizing vascular compliance, the transient blood retention effect caused by arterial wall elastic deformation was quantified. Based on this, in Figure 6 middle, Representing blood pressure, it indicates the pressure exerted by blood in blood vessels, similar to the voltage value in an electrical circuit, which is the potential difference that drives current through a conductor; The flow rate represents the volume of blood passing through the blood vessels per unit time, which is equivalent to the current value in a circuit, i.e., the amount of charge passing through a conductor per unit time. Through this analogy, the Windkessel model can be regarded as an equivalent model of a circuit, and therefore the balance equation can be established using Kirchhoff's voltage and current laws, as shown in the following formula (15):

[0141] , , , (15);

[0142] In equation (15), because the second resistor and capacitor They are connected in parallel, so their pressure is the same. . refers to capacitor Traffic flow at the location It refers to the second resistor The pressure can be obtained from the flow rates at the two locations. The expression for pressure. equal to the first resistance Pressure at the location plus capacitance Pressure at the point, first resistance The pressure at the point is Flowing into this point It is equivalent to flowing out of this point and The sum of these flows follows the flow balance equation.

[0143] Based on this, the mapping relationship between the coupled blood pressure and volumetric flow rate of the current aortic branch can be obtained, which can be used as the pressure outlet boundary condition for the current aortic branch (pressure outlet), as shown in the following formula (16):

[0144] (16);

[0145] Equation (16) dynamically couples blood pressure ( ) and volumetric flow rate ( The time-varying gradient of the pressure-flow phase relationship was used to achieve a closed-loop solution in the time domain.

[0146] Optionally, the condition shown in the above formula (16) is in derivative form. In order to improve the accuracy of the pressure outlet boundary condition, it is converted into differential form, that is, discretized, as shown in the following formula (17):

[0147] (17);

[0148] By transforming the above formula (17), we can obtain the pressure outlet boundary conditions shown in the following formula (18), which are the final conditions used for CFD simulation:

[0149] (18);

[0150] In equations (17)-(18), the concept of a time step is introduced, and the above equation (16) is transformed into a time step. Discretize into pressure at the current time step Pressure compared to the previous time step The difference will Discretized into the flow rate at the current time step Traffic compared to the previous time step The difference will Discretize into a very small time step Thus, the final pressure export boundary condition is obtained.

[0151] Through the above steps, independent circuit modeling and corresponding pressure outlet boundary conditions can be set for each aortic branch (pressure outlet). It should be understood that, for different aortic branches, only the parameter values ​​mentioned above need to be adjusted to set the corresponding pressure boundary conditions; the circuit modeling operation does not need to be repeated. This allows for the setting of boundary conditions for different aortic branches.

[0152] In some embodiments, the blood flow field simulation parameters may include velocity field parameters; in step S104, performing four-dimensional flow field magnetic resonance data synthesis processing based on the blood flow field simulation parameters to obtain magnetic resonance synthesized data corresponding to the CTA image may include:

[0153] Phase encoding is performed based on velocity field parameters to obtain a composite image of the phase amplitude of the thoracic aortic system. The composite image of the phase amplitude is then converted to the spatial domain and downsampled to obtain a downsampled composite image of the phase amplitude. Noise is injected into the downsampled composite image of the phase amplitude and then restored to obtain the magnetic resonance synthesis data.

[0154] In this embodiment, a high-precision CFD blood flow velocity field (i.e., velocity field parameters) is converted into equivalent 4D Flow MRI image data using a multiphysics coupling method. The core process includes three operations: phase encoding, frequency domain downsampling, and noise injection.

[0155] First, phase encoding is performed. The velocity field parameters are then mapped to [the appropriate parameters] based on the velocity encoding parameter (VENC) threshold. Phase intervals are used to obtain a phase map generated by mapping the velocity field parameters via VENC. The VENC value exceeds the maximum blood flow velocity in the thoracic aortic system, and this value can be pre-determined. Simultaneously, a non-zero amplitude image of the fluid region is constructed. Specifically, the amplitude (blood flow signal intensity) of the CTA image is multiplied element-wise by a preset fluid domain binary mask. The blood region of the fluid domain binary mask is set to 1, and the tissue region is set to 0. This operation generates spatially uniform non-zero amplitudes within the fluid domain, while forcibly setting them to zero in the tissue region, thus achieving a physical distinction between flowing signals and static tissue.

[0156] Then, frequency domain downsampling is performed. After completing phase mapping and amplitude image construction, the phase map and the non-zero amplitude image of the fluid region are superimposed to generate a phase-amplitude composite image. Subsequently, the phase-amplitude composite image is transformed to the k-space domain by Fast Fourier Transform (FFT), and downsampling is performed by truncating high-frequency components to simulate the low-pass filtering effect of MRI scans, thus obtaining the downsampled phase-amplitude composite image.

[0157] Finally, noise injection is performed. Zero-mean Gaussian noise is superimposed on the downsampled phase-amplitude composite image to reconstruct the signal-to-noise ratio characteristics of the real device, thus obtaining the noise-injected phase-amplitude composite image. Following this, an inverse Fourier transform is performed on the noise-injected phase-amplitude composite image to reconstruct a low-resolution phase / amplitude image in the spatial domain, obtaining 4D Flow MRI image data. This completes the cross-scale simulation from high-resolution CFD data to clinical-grade 4D Flow MRI images.

[0158] It is worth noting that the blood flow field simulation parameters can be obtained through the aforementioned CFD simulation processing, which include physical quantities such as velocity, pressure, temperature, density, and viscosity in the blood flow field of the thoracic aortic system. In this embodiment, only the velocity field parameters (i.e., the velocity vectors of blood flow in three directions) are selected to participate in the synthesis processing of 4D Flow MRI image data. The other parameters, although present, do not actually participate in data synthesis, but can be saved for use in downstream tasks such as subsequent neural network learning. For example, the wall shear stress can be saved as a label for learning, thereby recovering the wall shear stress from the velocity field and other information.

[0159] To facilitate understanding of the above-described method for synthesizing four-dimensional flow field magnetic resonance data based on CT images, an example of its practical application is provided below. (Refer to...) Figure 1 Multiple CTA images were retrieved from a database, including both healthy and diseased cases. For each CTA image, the following operations were performed:

[0160] S01, CTA Image Aortic Segmentation: The CTA image is segmented using a threshold segmentation method to obtain the initial tissue region associated with the thoracic aorta. The initial tissue region is then refined using a region growing method to obtain the target tissue region of the thoracic aortic system.

[0161] S02, 3D reconstruction and geometry optimization:

[0162] 3D Reconstruction: The MIMICS software interface is displayed on the terminal. The MIMICS software interface includes 3D reconstruction function controls. The MIMICS software interface imports and displays CTA images in DICOM format. In response to the trigger command of the 3D reconstruction function controls, the 3D reconstruction function window is displayed. The 3D reconstruction function window includes an algorithm selection area, a smoothing factor setting area, an iteration number setting area, and a first determination control. In response to the trigger command of the algorithm selection area, the 3D reconstruction algorithm is set to a continuous optimization algorithm. In response to the trigger command of the smoothing factor setting area, the smoothing factor is set to 0.4. In response to the trigger command of the iteration number setting area, the iteration number is set to 6. In response to the trigger command of the first determination control, the 3D reconstruction operation is performed based on the continuous optimization algorithm, smoothing factor, and iteration number to obtain the initial 3D model of the thoracic aortic system.

[0163] Geometric Optimization: The Geomagic software interface is displayed in the terminal. The Geomagic interface includes smoothing controls, filtering controls, constraint cutting controls, surface patching controls, gridding controls, and mesh processing controls. First, in response to the trigger command of the smoothing control, a curvature continuity preservation algorithm is used to perform global mesh smoothing on the initial 3D model. In response to the trigger command of the filtering control, an adaptive Gaussian filtering algorithm is used to perform gradient constraint smoothing on the initial 3D model for regions with abrupt curvature changes. In response to the trigger command of the constraint cutting control, for complex anatomical junctions such as blood vessel bifurcation, a boundary constraint cutting algorithm is used to reconstruct the physiological flow field boundary of the initial 3D model, thus obtaining the preprocessed initial 3D model. Then, in response to the trigger command of the surface patching control, based on the differential geometric topological features of the initial 3D model, the preprocessed initial 3D model is decomposed into multiple quadrilateral parametric surface patches. Afterwards, in response to the trigger command of the gridding control, in the parameter space of the parametric surface patches (… In, along , Orthogonal parametric lines are generated at equal intervals. These equidistant coordinate lines interweave on the 3D surface to form a bottom-level grid framework. The bottom-level grid framework includes multiple grid nodes, each storing parametric coordinates, 3D spatial coordinates, and principal curvature tensors. Finally, in response to trigger commands to the mesh processing controls, the topologically continuous, high-quality parametric grid is transformed into a finite element mesh dominated by structured quadrilaterals. These structured quadrilateral-dominated finite element meshes will collectively constitute the target 3D model of the thoracic aortic system.

[0164] S03, CFD simulation, the flow field model is the target 3D model:

[0165] Mesh generation: The target 3D model is meshed to divide the continuous blood vessel flow field model into multiple small discrete units (volume meshes), which together constitute the finite element model.

[0166] Simulation parameter settings: Set the hemodynamic control equations for CFD simulation as shown in formulas (3)-(7) above. At the same time, define the dynamic boundary conditions for CFD simulation based on the three-dimensional parameters and the three-element Windkessel principle, which include velocity inlet boundary conditions and pressure outlet boundary conditions.

[0167] Among them, the velocity inlet boundary condition is used to limit the velocity inlet of the finite element model, including the velocity time-varying function during the contraction period, the isochoric relaxation period and the relaxation period. The velocity time-varying function during the contraction period and the isochoric relaxation period is shown in the above formulas (8)-(9), and the velocity time-varying function during the relaxation period is always zero.

[0168] Pressure outlet boundary conditions are used to limit the pressure outlets of each aortic branch in the finite element model. A corresponding Windkessel model is constructed for each aortic branch, such as... Figure 6 As shown, the pressure outlet boundary conditions of each aortic branch are set accordingly, as shown in formula (16) or (18) above. Among them, the resistance and capacitance in formula (16) or (18) follow the formula (11)-(14) above.

[0169] Solution calculation: After meshing, the flow field model transforms from a continuous geometric region into an approximate field composed of a large number of discrete elements. This approximation is achieved through iterative methods, continuously approximating the true solution while setting residuals during the iteration process. Iteration stops when the residuals reach a set threshold, thus obtaining an accurate solution. Combining the solutions from these discrete elements approximates the solution for the entire flow field, i.e., the blood flow field simulation parameters. The threshold is set to 0.001, and the maximum number of iterations is 100.

[0170] Interactive display technology can be applied here, that is, displaying a CFD simulation interface on the terminal. The CFD simulation interface includes meshing controls, equation setting controls, condition setting controls, and solver controls. In response to the trigger command of the meshing controls, the meshing operation is performed; in response to the trigger command of the equation setting controls, the hemodynamic control equations are obtained; in response to the trigger command of the condition setting controls, the velocity inlet boundary conditions and pressure outlet boundary conditions are obtained; and in response to the trigger command of the solver controls, the solver calculation operation is performed.

[0171] Reference Figure 7 , Figure 7 The pressure distribution and dynamic changes in blood flow velocity during the systolic phase (inlet pressure ≈ 120 mmHg) and diastolic phase (inlet pressure ≈ 80 mmHg) were clearly presented: (1) Pressure distribution: During the systolic phase, the high pressure zone (red contour line) was transmitted from the inlet to the distal end, and the pressure gradient conformed to the physiological attenuation characteristics (120 mmHg → downstream attenuation). During the diastolic phase, the overall pressure dropped to 80 mmHg but still maintained a reasonable gradient (blue contour line), which accurately reproduced the propagation law of aortic pressure pulse wave and verified the ability of dynamic Windkessel boundary conditions to simulate physiological pressure waveforms; (2) Velocity streamline: During the systolic phase, the high-speed blood flow (red streamline) was concentrated in the center of the lumen, forming a parabolic velocity profile. The peak velocity reached the physiological range. During the diastolic phase, the flow velocity decreased significantly and micro vortices appeared near the wall. The flow field morphology conformed to the laminar-turbulent transition characteristics. The inlet pressure values ​​(systolic blood pressure 120 mmHg / diastolic blood pressure 80 mmHg) are in high agreement with the clinical data in the literature, and the spatiotemporal evolution of pressure / velocity is both physiologically reasonable and visually interpretable.

[0172] Reference Figure 8 , Figure 8This presentation showcases flow velocity visualizations obtained from CFD simulations using two pathological models. The model on the left depicts a patient with Marfan syndrome, where significant low-velocity eddies are observed in the aneurysmal dilatation region at the aortic root, creating localized blood flow stagnation. The model on the right shows a coactation of the aorta, where the streamlines accelerate and contract at the stenosis, indicating increased flow velocity due to the stenosis. Simultaneously, turbulent vortices arise downstream of the stenosis due to flow field separation. The simulation results validate the synergistic effectiveness of individualized boundary conditions and pathological geometric reconstruction, and the generated flow field parameters are clinically interpretable.

[0173] S04, 4D Flow MRI Synthesis: First, the velocity field parameters in the blood flow field simulation parameters are mapped to the VENC threshold. Phase maps generated from velocity field parameters via VENC mapping are obtained within the phase interval. Simultaneously, a non-zero amplitude image of the fluid region is constructed. Then, the phase map and the non-zero amplitude image of the fluid region are superimposed to generate a phase-amplitude composite image. This composite image is then transformed to the k-space domain using a Fast Fourier Transform (FFT), and downsampled by truncating high-frequency components, resulting in a downsampled phase-amplitude composite image. Finally, zero-mean Gaussian noise is superimposed on the downsampled phase-amplitude composite image to obtain a denoised phase-amplitude composite image. An inverse Fourier transform is then performed on the denoised phase-amplitude composite image to obtain the 4D Flow MRI image data.

[0174] By traversing the above CTA images, the corresponding 4D Flow MRI image data can be obtained. The 4D Flow MRI image data corresponding to these CTA images can be used to construct a 4D Flow MRI synthetic dataset for downstream tasks such as model training.

[0175] The effectiveness of the above-mentioned method for synthesizing four-dimensional flow field magnetic resonance data based on CT images will be verified below.

[0176] The 4D Flow MRI synthetic dataset obtained in this application was used to train the FlowVN (Deep Variational Network for Rapid 4D Flow MRI Reconstruction), a deep learning model specifically designed to accelerate 4D Flow MRI reconstruction. This model significantly improves computational efficiency while maintaining reconstruction quality by unfolding the traditional iterative optimization process into a 10-layer neural network structure. Each layer integrates a data consistency term (ensuring the reconstruction result matches the acquired k-space data) and a learnable 3D convolutional regularization term (capturing spatiotemporal correlations). Given the strong demand for large-scale training data for the FlowVN model as a deep learning architecture, this research fully leverages the unique advantages of the 4D Flow MRI synthetic data provided in this application.

[0177] To systematically evaluate the value of the 4D Flow MRI synthetic data presented in this application, a control experiment was designed. First, Golden angular radial undersampling (acceleration factor set to 20) was used to simulate clinical scanning conditions, and multi-channel coil sensitivity was fused to generate input data, which consisted of the original undersampled k-space complex matrix and its corresponding binary undersampled mask. The model output (label) was set to the reconstructed 4D hemodynamic complex image, which possesses full sampling resolution in both spatial and temporal dimensions and can be directly used for blood flow velocity field calculation and clinical parameter quantification. Subsequently, the following three models were trained:

[0178] (1) Group A was trained using only 2 real volunteer data;

[0179] (2) Group B combines the same two real data sets with two synthetic 4D Flow MRI data sets from this application for joint training;

[0180] (3) Group C is trained by combining these two real data with four synthetic 4D Flow MRI data from this application.

[0181] By comparing key indicators such as amplitude image error, relative velocity-flow field error, and velocity-flow field angle error of the three models in the hemodynamic reconstruction task, the quantification of the performance improvement effect of the 4D Flow MRI synthetic data of this application on the model is clearly demonstrated. Figure 9 As shown, the model's reconstruction performance is significantly insufficient when training data is insufficient. With the addition of the 4DFlow MRI synthetic data of this application, its reconstruction performance has been improved to a certain extent.

[0182] Furthermore, embodiments of this application also provide a four-dimensional flow field magnetic resonance data synthesis system based on CT images, which may include:

[0183] The acquisition module is used to acquire CTA images;

[0184] The first processing module is used to perform image segmentation and three-dimensional reconstruction processing on CTA images to obtain the target three-dimensional model of the thoracic aortic system.

[0185] The second processing module is used to perform fluid dynamics simulation processing based on the target three-dimensional model to obtain the blood flow field simulation parameters of the thoracic aortic system.

[0186] The third processing module is used to perform four-dimensional flow field magnetic resonance data synthesis based on blood flow field simulation parameters to obtain magnetic resonance synthesis data corresponding to CTA images.

[0187] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0188] In summary, this application proposes a personalized 4D Flow MRI data synthesis technique. Based on clinical CTA images that cannot be independently analyzed for hemodynamics, it generates individualized pathological vascular models through vessel segmentation and geometric reconstruction. After optimization, CFD simulation is performed using dynamic boundary conditions with personalized parameters to solve for spatiotemporally resolved blood flow velocity and pressure fields. Finally, high-fidelity 4D Flow MRI synthetic data is generated through phase-encoded mapping, k-space undersampling, and frequency domain conversion techniques with multi-coil noise injection. The embodiments of this application can yield high-quality 4D Flow MRI datasets containing pathological models, covering various typical cardiovascular lesions (including Marfan syndrome, aortic coarctation, and coronary artery stenosis). Each dataset includes an individualized CFD simulation original flow field, high-quality 4D Flow MRI synthetic data, and corresponding segmentation labels (obtainable through labeling operations).

[0189] Specifically, as mentioned above, the 4DflowNet model suffers from drawbacks including simplification assumptions in CFD models, deficiencies in the source of 3D models, and insufficient validation data. The SynRealSeg-4Dflow model, on the other hand, suffers from drawbacks including a lack of pathological models and the inability to perform individualized modeling. In contrast, the embodiments of this application demonstrate the advantages of personalized modeling, namely, generating corresponding 4D Flow MRI synthetic data for specific CTA images. In addition, it also possesses at least one of the following technical effects:

[0190] (1) Regarding the problems existing in the current technology, such as the lack of pathological models, the deficiency in the source of three-dimensional models, and the lack of validation data:

[0191] On the one hand, this application generates a three-dimensional model of the thoracic aortic system by combining CTA images with image segmentation and three-dimensional reconstruction techniques, which enables three-dimensional modeling to no longer rely on 4D Flow MRI data, i.e., it is not limited by the scarcity of 4D Flow MRI data, thereby solving the problem of the lack of source of three-dimensional models.

[0192] On the other hand, the CTA images synthesized in this application include both healthy and diseased cases. The synthesized data generated thereby comes not only from healthy thoracic aortic models but also from diseased thoracic aortic models. This allows pathological and healthy models to coexist, which helps to improve the performance of models trained based on these data and thus solves the problem of lacking pathological models.

[0193] On the other hand, the entire synthesis process of this application does not rely on existing 4D Flow MRI data, and CFD simulation can fully capture the complex noise patterns, motion artifacts and pathology-related blood flow heterogeneity in real 4D Flow MRI data, thereby ensuring the accuracy of 4D Flow MRI data generation.

[0194] (2) Regarding the problem of deficiencies in the source of 3D models in existing technologies:

[0195] This application accurately segments the thoracic aorta region based on CTA images, effectively improving the segmentation accuracy of the thoracic aorta region. Subsequently, this application performs 3D reconstruction and optimization based on the thoracic aorta region. In the optimization process, preprocessing the 3D model effectively eliminates noise and unreasonable distortions in the initial 3D model and improves the surface smoothness of the initial 3D model. Based on the preprocessed 3D model, corresponding surface patches and grids are constructed, achieving surface morphology optimization and curvature distribution balance of the 3D model of the thoracic aorta system. Transforming the topologically continuous high-quality parametric grid into a finite element mesh dominated by structured quadrilaterals enables surface fitting of the 3D model of the thoracic aorta system, thereby effectively improving the modeling quality of the 3D model of the thoracic aorta system. While preserving complete vascular anatomical features, it ensures that the 3D model of the thoracic aorta system, such as element orthogonality and aspect ratio, meets the requirements of subsequent CFD simulation for mesh convergence and computational stability, thus improving the accuracy of CFD simulation.

[0196] (3) Regarding the problem of simplifying assumptions in existing CFD models:

[0197] In this application, the velocity inlet boundary conditions of the thoracic aorta region are set according to the heartbeat characteristics in CFD simulation, the pressure outlet boundary conditions of each aortic branch are set according to the three-element Windkessel model, and the hemodynamic control equations of the thoracic aorta region are set in combination with the laws of fluid physics. Based on this, CFD simulation is performed to obtain accurate CFD simulation data. This ensures the accuracy of downstream flow in the thoracic aorta system, makes the flow mode of CFD simulation close to the real physiological state, and thus improves the accuracy of CFD simulation.

[0198] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0199] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for synthesizing four-dimensional flow field magnetic resonance data based on CT images, characterized in that, Includes the following steps: Acquire CTA images; The CTA images are segmented and reconstructed in three dimensions to obtain a target three-dimensional model of the thoracic aortic system. Based on the target three-dimensional model, fluid dynamics simulation processing is performed to obtain the blood flow field simulation parameters of the thoracic aortic system; Based on the blood flow field simulation parameters, four-dimensional flow field magnetic resonance data synthesis processing is performed to obtain magnetic resonance synthesized data corresponding to the CTA image; The blood flow field simulation parameters include velocity field parameters; the four-dimensional flow field magnetic resonance data synthesis processing based on the blood flow field simulation parameters to obtain magnetic resonance synthesized data corresponding to the CTA image includes: The velocity field parameters are mapped to based on the velocity coding parameter threshold. Phase interval, a phase map generated by VENC mapping of the velocity field parameters is obtained. The amplitude of the CTA image is multiplied element-wise by a preset fluid domain binary mask to obtain a non-zero amplitude image of the fluid region. The phase map and the non-zero amplitude image of the fluid region are superimposed to generate a phase amplitude composite image of the thoracic aortic system. In this case, the blood region of the fluid domain binary mask is set to 1 and the tissue region is set to 0. The phase amplitude composite image is converted to the k-space domain by fast Fourier transform, and then downsampled by truncating high-frequency components to obtain the downsampled phase amplitude composite image. Zero-mean Gaussian noise is superimposed on the downsampled phase amplitude composite image to obtain the denoised phase amplitude composite image. An inverse Fourier transform is performed on the denoised phase amplitude composite image to obtain magnetic resonance synthetic data corresponding to the CTA image.

2. The method according to claim 1, characterized in that, The process of image segmentation and 3D reconstruction of the CTA images to obtain a target 3D model of the thoracic aortic system includes: The CTA images were segmented to obtain the target tissue region associated with the thoracic aorta; Based on the target tissue region, a three-dimensional reconstruction process is performed to obtain an initial three-dimensional model of the thoracic aortic system; The initial 3D model is optimized to obtain the target 3D model.

3. The method according to claim 2, characterized in that, The step of performing image segmentation processing on the CTA image to obtain the target tissue region associated with the thoracic aorta includes: The CTA image was segmented using a threshold segmentation method to obtain the initial tissue region associated with the thoracic aorta; The initial tissue region is refined using the region growing method to obtain the target tissue region.

4. The method according to claim 2, characterized in that, The optimization process of the initial 3D model to obtain the target 3D model includes: The initial 3D model is preprocessed to obtain the preprocessed initial 3D model; The preprocessed initial 3D model is decomposed to obtain multiple surface patches; Multiple grid nodes are obtained by constructing multiple surface patches; The target 3D model is generated based on multiple grid nodes.

5. The method according to claim 1, characterized in that, The fluid dynamics simulation based on the target 3D model yields the blood flow field simulation parameters of the thoracic aortic system, including: Obtain the hemodynamic control equations and boundary conditions of the thoracic aortic system; The target 3D model is meshed to obtain a finite element model; wherein the finite element model includes multiple discrete elements, and each discrete element is constrained by the boundary conditions; The finite element model is solved based on the hemodynamic control equations to obtain the blood flow field simulation parameters.

6. The method according to claim 5, characterized in that, The process of obtaining the hemodynamic control equations and boundary conditions of the thoracic aortic system includes: Obtain the mass conservation equation, momentum conservation equation, energy conservation equation, state equation, and heat conduction equation of the target three-dimensional model; The mass conservation equation and the momentum conservation equation are simplified to obtain the simplified mass conservation equation and the simplified momentum conservation equation. The energy conservation equation, the state equation, the heat conduction equation, the simplified mass conservation equation, and the simplified momentum conservation equation are determined as the hemodynamic control equations.

7. The method according to claim 5, characterized in that, The finite element model includes a velocity inlet; the boundary conditions include velocity inlet boundary conditions, which are used to limit the velocity inlet; obtaining the hemodynamic control equations and boundary conditions of the thoracic aortic system includes: The velocity time-varying functions of the systolic cycle, isovolumetric relaxation cycle, and diastolic cycle in the thoracic aortic system are determined as the velocity inlet boundary conditions; wherein, the velocity time-varying function of the systolic cycle adopts a positive exponential-sine composite function, the velocity time-varying function of the isovolumetric relaxation cycle adopts a negative decay function, and the value of the velocity time-varying function of the diastolic cycle is always zero.

8. The method according to claim 5, characterized in that, The finite element model includes four pressure outlets; the boundary conditions include pressure outlet boundary conditions, which are used to limit the pressure outlets. The process of obtaining the hemodynamic control equations and boundary conditions of the thoracic aortic system includes: Based on the attribute data of the thoracic aortic system, the first resistance, second resistance, and capacitance of each aortic branch in the thoracic aortic system are set; Based on the first resistance, second resistance, and capacitance of each aortic branch, each aortic branch is equivalent to a corresponding hemodynamic circuit model; wherein, the hemodynamic circuit model includes a first resistance, a second resistance, and a capacitor, one end of the first resistance is connected to one end of the capacitor and one end of the second resistance, the other end of the capacitor and the other end of the second resistance are both grounded, and the capacitor is connected in parallel with the second resistance; Based on the hemodynamic circuit model of each aortic branch, the mapping relationship between the coupled blood pressure and volumetric flow rate of each aortic branch is obtained as the pressure outlet boundary condition.

9. A four-dimensional flow field magnetic resonance data synthesis system based on CT images, characterized in that, include: The acquisition module is used to acquire CTA images; The first processing module is used to perform image segmentation and three-dimensional reconstruction processing on the CTA image to obtain a target three-dimensional model of the thoracic aortic system. The second processing module is used to perform fluid dynamics simulation processing based on the target three-dimensional model to obtain the blood flow field simulation parameters of the thoracic aortic system. The third processing module is used to perform four-dimensional flow field magnetic resonance data synthesis processing based on the blood flow field simulation parameters to obtain magnetic resonance synthesized data corresponding to the CTA image. The blood flow field simulation parameters include velocity field parameters; the four-dimensional flow field magnetic resonance data synthesis processing based on the blood flow field simulation parameters to obtain magnetic resonance synthesized data corresponding to the CTA image includes: The velocity field parameters are mapped to based on the velocity coding parameter threshold. Phase interval, a phase map generated by VENC mapping of the velocity field parameters is obtained. The amplitude of the CTA image is multiplied element-wise by a preset fluid domain binary mask to obtain a non-zero amplitude image of the fluid region. The phase map and the non-zero amplitude image of the fluid region are superimposed to generate a phase amplitude composite image of the thoracic aortic system. In this case, the blood region of the fluid domain binary mask is set to 1 and the tissue region is set to 0. The phase amplitude composite image is converted to the k-space domain by fast Fourier transform, and then downsampled by truncating high-frequency components to obtain the downsampled phase amplitude composite image. Zero-mean Gaussian noise is superimposed on the downsampled phase amplitude composite image to obtain the denoised phase amplitude composite image. An inverse Fourier transform is performed on the denoised phase amplitude composite image to obtain magnetic resonance synthetic data corresponding to the CTA image.

Images