A three-dimensional spatio-temporal hemodynamic reconstruction method for lower extremity arterial blood perfusion

Abstract

The application discloses a kind of lower limb arterial blood perfusion three-dimensional space-time hemodynamics reconstruction methods, comprising the following steps: S1: the computer tomography angiography CTA three-dimensional volume data of patient lower limb and the time sequence digital subtraction angiography DSA image of same part are collected, and pretreatment is carried out;S2: space-time registration is carried out to the DSA sequence after pretreatment with CTA three-dimensional blood vessel model, and the spatial mapping relationship of two-dimensional dynamic information and three-dimensional anatomical structure is established;The application not only can accurately locate stenosis, assess collateral circulation function, identify ischemic risk area, but also can provide hemodynamic basis for preoperative planning and postoperative prediction of interventional therapy (such as stent implantation), so as to promote the diagnosis of lower limb arterial disease from traditional morphological evaluation to a new stage of "anatomy-function-mechanics" multi-modal fusion, significantly improve the accuracy, foresight and personalized level of diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the field of arterial blood perfusion technology, specifically a method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion. Background Technology

[0002] Lower extremity arterial disease (PAD) is a major cause of lower extremity ischemia, intermittent claudication, and even gangrene. Accurate blood perfusion assessment is crucial for treatment planning. Currently, commonly used perfusion assessment methods in clinical practice include: Digital subtraction angiography (DSA) is considered the "gold standard" for blood flow assessment. It can provide high temporal resolution of dynamic vascular imaging, but it is an invasive procedure and only provides two-dimensional projection information, and cannot directly obtain three-dimensional flow velocity distribution.

[0003] Computed tomography angiography (CTA): can acquire three-dimensional vascular structures, but has low temporal resolution and is usually used for anatomical assessment rather than dynamic blood flow analysis.

[0004] Ultrasonic Doppler: non-invasive and repeatable, but highly dependent on the operator and difficult to obtain the full-field flow velocity in complex geometric regions.

[0005] In recent years, progress has been made in image-based hemodynamic calculations. For example, optical flow has been used to estimate two-dimensional blood flow velocities from DSA sequences. However, optical flow alone lacks physical constraints, has large errors in low-contrast or overlapping vessels, and cannot be directly extended to three dimensions.

[0006] Physical Information Neural Networks (PINNs) provide a new framework for fusing sparse measurements with physical laws, but their application in medical imaging still requires solving the problems of multimodal data fusion and flow velocity initialization.

[0007] Currently, there is no systematic solution that combines DSA-CTA multimodal registration, optical flow method initial velocity estimation, and PINN inverse problem solving for three-dimensional spatiotemporal blood flow reconstruction and perfusion assessment of lower limb arteries. Summary of the Invention

[0008] The purpose of this invention is to provide a method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion, so as to solve the problems mentioned in the background art.

[0009] A method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion includes the following steps: S1: Acquire three-dimensional volume data of computed tomography angiography (CTA) of the patient's lower extremities and time-series digital subtraction angiography (DSA) images of the same site, and perform preprocessing. S2: Spatiotemporal registration of the preprocessed DSA sequence with the CTA three-dimensional vascular model is performed to establish the spatial mapping relationship between two-dimensional dynamic information and three-dimensional anatomical structure. S3: Based on the optical flow method, the two-dimensional blood flow velocity field is estimated from the registered DSA sequence, and it is mapped to the three-dimensional blood vessel through the inverse projection geometry transformation to obtain the blood flow velocity prior; SA: The coupled system of solving the three-dimensional Navier-Stokes equations and Windkessel outlet boundary conditions using the volume method generates high-resolution velocity and pressure fields on the geometry of the lower limb arteries as reference solutions. Data is sampled in the flow field to form a dataset for training PINNs. S4: Construct a physical information neural network PINN with spatiotemporal coordinates (x,y,z,t) as input and three-dimensional flow velocity (u,v,w) and pressure (p) as output, and embed the incompressible Navier-Stokes equation as a physical constraint into the network; S5: Integrate the flow velocity prior and physical constraints, train the PINN by minimizing the total loss function that includes data fitting terms and physical residual terms, and invert hemodynamic parameters during the training process; S6: Utilize the trained PINN to perform inference in the three-dimensional spatiotemporal domain, reconstruct a continuous four-dimensional blood flow field, and calculate clinical blood perfusion assessment indicators.

[0010] Furthermore, in S1, the preprocessing includes segmenting the CTA data to extract a three-dimensional vascular surface model and generating a computational domain mesh; and denoising, enhancing contrast, and extracting vascular structures from the DSA sequence. The blood vessel segmentation is achieved using a deep learning-based convolutional neural network or a traditional image segmentation algorithm, supplemented by blood vessel centerline extraction to optimize the model's geometric accuracy.

[0011] Furthermore, in S2, the spatiotemporal registration adopts a non-rigid registration algorithm based on mutual information and elastic deformation model, and uses the blood vessel centerline as the registration guidance feature to achieve accurate spatial alignment between each frame of DSA image and the three-dimensional CTA model; The registration algorithm further incorporates multi-resolution strategies and / or deep learning-based registration networks such as VoxelMorph to improve the robustness and efficiency of registration.

[0012] Furthermore, in S3, the optical flow method employs an end-to-end optical flow estimation model based on deep learning, such as FlowNet, PWC-Net, or variants thereof, and introduces an attention mechanism into the network to enhance the estimation robustness in areas with overlapping blood vessels, low contrast, and noise. The optical flow estimation model incorporates the fluid continuity equation as a physical regularization term during training or inference to ensure the physical rationality of the estimated velocity field.

[0013] Mapping two-dimensional optical flow velocity to three-dimensional blood vessels includes: extracting the three-dimensional blood vessel centerline, combining DSA projection geometric parameters, calculating the three-dimensional lumen position corresponding to the two-dimensional pixel point through back projection, and decomposing the two-dimensional velocity vector into axial components to obtain the estimated axial velocity values ​​at each position of the three-dimensional lumen.

[0014] Furthermore, in S4, PINN adopts a fully connected neural network architecture, including an input layer, multiple hidden layers and an output layer, and the activation function adopts tanh, Swish or adaptive activation function; a Fourier feature embedding layer can be introduced before the network input layer to improve the network's ability to fit high-frequency spatiotemporal signals; The incompressible Navier-Stokes equations define the physical residual terms: , , , , Where ρ is blood density and ν is dynamic viscosity, which can be used as network input parameters; The physical constraints are implemented using automatic differentiation technology, and the derivative of the network output with respect to the input coordinates is used to construct the equation residuals.

[0015] Furthermore, in S5, the total loss function is defined as follows: , in: The fitting error of the three-dimensional cavity velocity from DSA optical flow inversion is used to constrain the consistency between the network output and the velocity prior obtained from S3. The residuals of the Navier-Stokes equations at collocation points within the computational domain are evaluated at collocation points within the computational domain. There are no conditions for slippage in the vessel wall and no inlet / outlet boundary constraints. : Regularization of flow rate range based on clinical priors, such as peak systolic velocity constraint; : Parameter constraints of the 3D model export Windkessel model; The weights of each item in the loss function , , and An adaptive weighting strategy is employed for adjustment, including uncertainty-based weighting, gradient normalization, or soft attention mechanisms, to automatically balance the contributions of different loss terms during training.

[0016] Furthermore, in S5, the training employs the Adam optimizer for pre-training, followed by fine-tuning using the L-BFGS optimizer to accelerate convergence and improve solution accuracy, utilizing automatic differential calculation. u、 p and 2 u, ensuring the precise embedding of physical constraints.

[0017] Furthermore, in S6, the blood perfusion assessment indicators include three-dimensional wall shear force distribution, oscillatory shear index, perfusion intensity index, pulse wave conduction velocity, and ischemic risk area marking. The trained PINN is inferred in the three-dimensional spatiotemporal domain to obtain continuous blood flow velocity u(x,y,z,t) and pressure p(x,y,z,t). Calculate key hemodynamic parameters: Wall shear stress (WSS); Oscillatory shear index (OSI); Perfusion Intensity Index (PII): This index assesses tissue-level perfusion by combining flow velocity and vascular density. It is calculated by fusing blood flow velocity with vascular density information extracted based on CTA or DSA and is used to evaluate the blood perfusion status at the tissue level.

[0018] Furthermore, S6 also includes outputting blood flow animation, parameter distribution map and ischemia risk assessment report through a three-dimensional visualization platform, combining the reconstructed three-dimensional blood flow field with a risk assessment model based on machine learning or deep learning to predict the progression of lower extremity arterial disease or the improvement of hemodynamics after interventional treatment.

[0019] Furthermore, in step S4, within the PINNs framework, the Windkessel model parameters for each exit of the 3D blood vessel model are first initialized. The loss function constraint term L... WK The equations include pressure equations and differential equations, where the flow rate is obtained by integrating the velocity output by the network at the outlet surface. Training employs a two-stage strategy: first, the Windkessel parameters are fixed, and only the neural network weights are optimized to fit the flow field; then, both the network weights and parameters are optimized simultaneously to minimize the total loss. Through backpropagation, the parameters are gradually updated, eventually converging to an estimate that minimizes the coupled physical residual.

[0020] Furthermore, in the SA, commercial or open-source software such as OpenFoam can be used to simulate batch 3D models. The inlet flow velocity boundary conditions are assigned based on the flow velocity at the inlet position of the model obtained in S3, and the outlet Windkessel parameters can be inverted based on the flow velocity information to obtain the impedance parameters.

[0021] By employing the aforementioned technical solutions, and deeply integrating the high temporal resolution dynamic information of digital subtraction angiography (DSA) with the high spatial resolution three-dimensional anatomical structure of computed tomography angiography (CTA), an innovative integrated technical framework of "multimodal image registration - deep learning optical flow estimation - physical information neural network inversion" was constructed. This successfully elevates traditional blood flow observation based on two-dimensional projection to three-dimensional spatiotemporal blood flow dynamics reconstruction under physical constraints. The technical effects are reflected in multiple dimensions: at the data fusion level, a non-rigid registration algorithm accurately aligns DSA sequences and CTA vascular models, solving the spatial mapping problem between two-dimensional dynamic information and three-dimensional static anatomy, and maximizing the utilization of image information; at the flow velocity estimation level, a non-rigid registration algorithm is used to precisely align DSA sequences and CTA vascular models, solving the spatial mapping problem between two-dimensional dynamic information and three-dimensional static anatomy, and maximizing the utilization of image information; at the flow velocity estimation level, a non-rigid registration algorithm is used to maximize the spatial mapping between two-dimensional dynamic information and three-dimensional static anatomy, maximizing the spatial mapping between two-dimensional dynamic information and three-dimensional static anatomy, and maximizing the utilization of image information. Improved optical flow networks based on attention mechanisms (such as PWC-Net variants) significantly enhance the robustness of DSA sequence estimation in low-contrast and overlapping vascular regions. Furthermore, they obtain the tangential velocity distribution of the 3D vascular surface through projective geometric inversion, providing high-quality prior data for subsequent modeling. At the physical modeling level, the core innovation lies in constructing a deep neural network embedding incompressible Navier-Stokes equations. Through automatic differentiation, momentum and mass conservation equations are directly incorporated as physical regularization terms into the loss function. This allows the network to simultaneously fit optical flow observation data and satisfy fundamental fluid dynamics laws during training, fundamentally overcoming the lack of physical consistency in traditional purely data-driven methods and ensuring the accuracy of velocity measurements. The framework demonstrates the biomechanical rationality of the flow field and pressure field in three-dimensional space. At the inverse problem-solving level, it starts from sparse surface velocity measurements (derived from optical flow estimation), jointly optimizing neural network parameters and physical parameters (such as blood viscosity) to inversely derive the complete three-dimensional spatiotemporal blood flow distribution of the entire vascular bed. This represents a breakthrough from "partial observation" to "full-field reconstruction," and possesses the ability to infer clinical inverse problems such as boundary conditions (terminal branch vascular bed impedance) and vascular wall characteristics. At the output level, the system can generate continuous and smooth four-dimensional (spatial three-dimensional + temporal) blood flow velocity and pressure fields, thereby accurately calculating spatiotemporal distributions of wall shear force, oscillatory shear index, particle residence time, and other parameters closely related to the occurrence and development of atherosclerosis, which are unavailable through traditional methods. This invention provides a non-invasive, quantitative, and repeatable three-dimensional blood perfusion assessment scheme. The output results are presented intuitively in the form of dynamic three-dimensional visualization, parameter cloud maps, and structured reports. It can not only accurately locate stenosis, assess collateral circulation function, and identify ischemic risk areas, but also provide hemodynamic evidence for preoperative planning and postoperative prediction of interventional treatments (such as stent placement). This advances the diagnosis of lower extremity arterial diseases from traditional morphological assessment to a new stage of "anatomical-functional-mechanical" multimodal fusion, significantly improving the accuracy, foresight, and personalization of diagnosis and treatment.

[0022] Compared with the prior art, the beneficial effects of the present invention are: the three-dimensional spatiotemporal hemodynamic reconstruction method for lower limb arterial blood perfusion, Multimodal information fusion: Fully utilize the high temporal resolution of DSA and the high spatial resolution of CTA to achieve complementary advantages of anatomical and functional data.

[0023] Velocity optimization under physical constraints: Optical flow provides high-resolution initial estimates, while PINN performs global optimization and three-dimensional expansion through physical equations, overcoming the limitation of traditional optical flow methods lacking physical consistency.

[0024] Full three-dimensional spatiotemporal blood flow field reconstruction: obtaining continuous and physically reliable three-dimensional velocity and pressure distributions without invasive measurement conditions, breaking through the projection limitations of two-dimensional DSA.

[0025] High clinical applicability: Comprehensive output parameters support hemodynamic mechanism analysis and precise localization of ischemic areas, providing quantitative basis for interventional treatment planning.

[0026] The system is highly adaptive: through trainable physical parameters, it can adapt to individual differences in blood viscosity, vascular elasticity, etc., among different patients. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall steps of the present invention; Figure 2 This is a flowchart of the image acquisition and preprocessing process of the present invention; Figure 3 This is a block diagram of multimodal spatiotemporal registration and a flowchart of two-dimensional blood flow velocity estimation using optical flow method according to the present invention. Figure 4 This is a flowchart illustrating the mapping of two-dimensional flow velocity to three-dimensional blood vessel surface according to the present invention; Figure 5 The diagram shows the PINN network structure and the PINN multi-loss training flowchart of this invention. Figure 6 This invention relates to the four-dimensional blood flow field reconstruction and index calculation. Figure 7 This is a flowchart of the clinical visualization and assessment output of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Please see Figure 1-3 This invention provides a technical solution: a three-dimensional spatiotemporal hemodynamic reconstruction method for lower limb arterial blood perfusion, which involves acquiring lower limb CTA and DSA images of the patient; segmenting CTA vessels using open-source tools (such as 3D Slicer) to generate a three-dimensional mesh, or using custom code; using the Elastix toolbox to complete non-rigid registration of DSA-CTA, or using custom code; processing the DSA sequence using the PWC-Net optical flow model to obtain two-dimensional flow velocity for each frame; mapping to a three-dimensional surface to obtain approximately 5000 surface flow velocity data points; constructing a 10-layer PINN with 96 neurons per layer, using the tanh activation function; training on an NVIDIA V100 GPU for approximately 2 hours with loss convergence; outputting a three-dimensional blood flow animation, WSS cloud map, and markers of post-stenotic hypoperfusion areas; comparing the results with ultrasound Doppler measurements (or 4D flow comparison), the velocity profile shows good agreement (relative error <8%), and successfully identifying collateral circulation blood flow not visualized by CTA.

[0030] By constructing a multi-layered fusion system of "multimodal registration-optical flow prior-physical constraint neural network," the shortcomings of traditional two-dimensional optical flow methods are systematically overcome. The entire process begins with the non-rigid registration of the DSA dynamic sequence and the CTA three-dimensional model, mapping the two-dimensional temporal projection information to the three-dimensional anatomical space, thus solving the dimensional limitation problem. Next, a deep learning optical flow network based on the attention mechanism (such as a variant of PWC-Net) is used to extract a high spatiotemporal resolution two-dimensional velocity field from the registered DSA sequence, and inverts it into a three-dimensional tangential velocity distribution on the blood vessel surface through projection onto the vessel centerline, serving as a strong data prior for the subsequent physical model. A physical information neural network (PINN) is constructed, with spatiotemporal coordinates as input and three-dimensional velocity and pressure as output. The network structure adopts residual connections and adaptive activation functions, and its unique feature is that it uses incompressible N... The Navier-Stokes equations are directly embedded in the loss function as physical regularization terms—specifically, the spatiotemporal gradients of velocity and pressure are calculated through automatic differentiation, and momentum and mass conservation are enforced at numerous collocation points within the computational domain. This ensures that the generated blood flow field strictly follows the laws of fluid mechanics, fundamentally correcting the shortcomings of traditional optical flow methods in lacking physical consistency. Simultaneously, the system weightedly integrates data fitting terms from optical flow estimation, Navier-Stokes residual constraints, no-slip boundary conditions of the vessel wall, and parameter regularization terms based on clinical priors. This design enables the network to start from sparse, noisy surface velocity data (provided by optical flow), driven by physical equations, and inversely derive a complete and smooth three-dimensional spatiotemporal blood flow field, achieving a paradigm shift from "two-dimensional forward modeling" to "three-dimensional inverse problem solving." Ultimately, as... Figure 4 and Figure 5As shown, the system outputs biomechanical parameters such as full three-dimensional blood flow velocity vector field, pressure distribution, wall shear force cloud map, and oscillatory shear index. It can also couple perfusion assessment at the vascular and tissue levels through perfusion intensity index to form a complete diagnostic report. This process not only restores the three-dimensional information and physical authenticity missing in traditional methods, but also achieves deep fusion of multimodal data and robust solution of clinical inverse problems through an end-to-end trainable framework.

[0031] Example 1

[0032] A 65-year-old male patient was admitted to the hospital due to intermittent claudication in the right lower extremity, with an ankle-brachial index of 0.5. CTA showed severe stenosis of about 80% in the middle segment of the right superficial femoral artery, with a stenosis length of about 3 cm. Acquired lower extremity CTA and DSA images of the patient; used 3D Slicer software to segment blood vessels from the CTA data, extracted 3D surface models of stenotic segments and adjacent blood vessels, and generated tetrahedral meshes using its built-in mesh generation module; used the Elastix toolbox to complete non-rigid registration of DSA-CTA, using the vessel centerline as a guiding feature to achieve spatial alignment between the 2D sequence and the 3D model; used the PWC-Net optical flow model to process the registered DSA sequence, obtaining the 2D velocity field of each frame image, and mapped the velocity onto the 3D blood vessel surface through inverse projection geometry transformation, obtaining approximately 5000 surface velocity data points; constructed a 10-layer fully connected physical information neural network based on TensorFlow, with 96 neurons per layer, using tanh as the activation function, with spatiotemporal coordinates as input and 3D velocity and pressure as output; The residuals of the Navier-Stokes equations were incorporated as physical constraints into the loss function, and blood viscosity was set as a trainable parameter. After training on an NVIDIA V100 GPU for approximately 2 hours, the loss function converged. The trained network was used to infer the three-dimensional blood flow field throughout the cardiac cycle, outputting a three-dimensional blood flow animation, wall shear force contour map, and markers of the post-stenotic hypoperfusion area. The results were compared with the downstream velocity profile of the stenosis measured by preoperative Doppler ultrasound, and the waveforms of the two methods matched well, with a relative error of less than 8% in peak velocity. Furthermore, this method successfully identified collateral circulation blood flow that was not shown in the original CTA images, providing additional evidence for clinical interventional decision-making.

[0033] Example 2

[0034] A 70-year-old female diabetic patient presented with a persistent plantar ulcer on her left foot. CTA revealed multi-segmental stenosis of the posterior tibial artery and peroneal artery below the knee, with the narrowest point being approximately 70%, and significant calcification in the vessel walls. Acquired lower extremity CTA and DSA images from the patient; used a self-developed Python script to segment CTA vessels based on a 3D U-Net model, performed morphological completion on calcified areas, generated a continuous 3D vessel surface model, and meshed it; employed the deep learning-based registration network VoxelMorph to complete non-rigid registration of DSA-CTA, and used a pre-trained model to quickly align each frame of images; used an improved version of PWC-Net to process DSA sequences, introducing an attention mechanism in optical flow estimation to enhance robustness in small vessel regions, and obtained 2D flow velocities for each frame; back-projected the 2D flow velocities onto the 3D vessel surface using DSA system calibration parameters, obtaining approximately 3000 high-quality surface flow velocity data points; constructed a 10-layer physical information neural network with 96 neurons per layer, using the Swish activation function, and incorporated Fourier feature embedding into the input layer to improve the fitting ability for high-frequency signals from small vessels below the knee; embedded the Navier-Stokes equation as a physical constraint, and used blood viscosity as a trainable parameter; The model was trained on an NVIDIA V100 GPU for approximately 2 hours, and the loss converged. It output a 3D blood flow animation, wall shear force distribution, and foot perfusion intensity index map. The reconstructed posterior tibial artery velocity waveform was compared with the ultrasound Doppler measurement results. The relative error of the peak systolic velocity was 7.2%, and the waveform morphology was consistent. The model successfully located the inadequate perfusion area corresponding to the plantar ulcer region and quantified the oscillatory shear index downstream of the stenosis, providing a hemodynamic basis for the selection of blood revascularization target points.

Claims

1. A method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion, characterized in that, Includes the following steps: S1: Acquire three-dimensional volume data of computed tomography angiography (CTA) of the patient's lower extremities and time-series digital subtraction angiography (DSA) images of the same site, and perform preprocessing. S2: Spatiotemporal registration of the preprocessed DSA sequence with the CTA three-dimensional vascular model is performed to establish the spatial mapping relationship between two-dimensional dynamic information and three-dimensional anatomical structure. S3: Based on the optical flow method, the two-dimensional blood flow velocity field is estimated from the registered DSA sequence, and it is mapped to the three-dimensional blood vessel through the inverse projection geometry transformation to obtain the blood flow velocity prior; SA: The coupled system of solving the three-dimensional Navier-Stokes equations and Windkessel outlet boundary conditions using the volume method generates high-resolution velocity and pressure fields on the geometry of the lower limb arteries as reference solutions. Data is sampled in the flow field to form a dataset for training PINNs. S4: Construct a physical information neural network PINN with spatiotemporal coordinates (x,y,z,t) as input and three-dimensional flow velocity (u,v,w) and pressure (p) as output, and embed the incompressible Navier-Stokes equation as a physical constraint into the network; S5: Integrate the flow velocity prior and physical constraints, train the PINN by minimizing the total loss function that includes data fitting terms and physical residual terms, and invert hemodynamic parameters during the training process; S6: Utilize the trained PINN to perform inference in the three-dimensional spatiotemporal domain, reconstruct a continuous four-dimensional blood flow field, and calculate clinical blood perfusion assessment indicators.

2. The method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion according to claim 1, characterized in that, In step S1, the preprocessing includes segmenting CTA data to extract a three-dimensional vascular surface model and generating a computational domain mesh; and denoising, contrast enhancement, and vascular structure extraction of DSA sequences. The blood vessel segmentation is achieved using a deep learning-based convolutional neural network or a traditional image segmentation algorithm, supplemented by blood vessel centerline extraction to optimize the model's geometric accuracy.

3. The method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion according to claim 1, characterized in that, In S2, the spatiotemporal registration adopts a non-rigid registration algorithm based on mutual information and elastic deformation model, and uses the blood vessel centerline as the registration guidance feature to achieve accurate spatial alignment between each frame of DSA image and the three-dimensional CTA model. The registration algorithm further incorporates multi-resolution strategies and / or deep learning-based registration networks such as VoxelMorph to improve the robustness and efficiency of registration.

4. The method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion according to claim 1, characterized in that, In S3, the optical flow method employs an end-to-end optical flow estimation model based on deep learning, such as FlowNet, PWC-Net, or variants thereof, and introduces an attention mechanism into the network to enhance the estimation robustness in areas with overlapping blood vessels, low contrast, and noisy regions. The optical flow estimation model incorporates the fluid continuity equation as a physical regularization term during training or inference to ensure the physical rationality of the estimated velocity field. Mapping two-dimensional optical flow velocity to three-dimensional blood vessels includes: extracting the three-dimensional blood vessel centerline, combining DSA projection geometric parameters, calculating the three-dimensional lumen position corresponding to the two-dimensional pixel point through back projection, and decomposing the two-dimensional velocity vector into axial components to obtain the estimated axial velocity values ​​at each position of the three-dimensional lumen.

5. The method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion according to claim 1, characterized in that, In S4, PINN adopts a fully connected neural network architecture, which includes an input layer, multiple hidden layers and an output layer. The activation function is tanh, Swish or adaptive activation function. A Fourier feature embedding layer can be introduced before the network input layer to improve the network's ability to fit high-frequency spatiotemporal signals. The incompressible Navier-Stokes equations define the physical residual terms: ， ， ， ， Where ρ is blood density and ν is dynamic viscosity, which can be used as network input parameters; The physical constraints are implemented using automatic differentiation technology, and the derivative of the network output with respect to the input coordinates is used to construct the equation residuals.

6. The method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion according to claim 5, characterized in that, In S5, the total loss function is defined as follows: ， in: The fitting error of the three-dimensional cavity velocity from DSA optical flow inversion is used to constrain the consistency between the network output and the velocity prior obtained from S3. The residuals of the Navier-Stokes equations at collocation points within the computational domain are evaluated at collocation points within the computational domain. There are no conditions for slippage in the vessel wall and no inlet / outlet boundary constraints. : Regularization of flow rate range based on clinical priors, such as peak systolic velocity constraint; : Parameter constraints of the 3D model export Windkessel model; The weights of each item in the loss function , , and An adaptive weighting strategy is employed for adjustment, including uncertainty-based weighting, gradient normalization, or soft attention mechanisms, to automatically balance the contributions of different loss terms during training.

7. The method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion according to claim 1, characterized in that, In step S5, the Adam optimizer is used for pre-training, followed by fine-tuning using the L-BFGS optimizer to accelerate convergence and improve solution accuracy, utilizing automatic differential calculation. u、 p and 2 u, ensuring the precise embedding of physical constraints.

8. The method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion according to claim 1, characterized in that, In S6, the blood perfusion assessment indicators include three-dimensional wall shear force distribution, oscillatory shear index, perfusion intensity index, pulse wave conduction velocity, and ischemic risk area marking. The trained PINN is inferred in the three-dimensional spatiotemporal domain to obtain continuous blood flow velocity u(x,y,z,t) and pressure p(x,y,z,t). Calculate key hemodynamic parameters: Wall shear stress (WSS); Oscillatory shear index (OSI); Perfusion Intensity Index (PII): This index assesses tissue-level perfusion by combining flow velocity and vascular density. It is calculated by fusing blood flow velocity with vascular density information extracted based on CTA or DSA and is used to evaluate the blood perfusion status at the tissue level.

9. The method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion according to claim 1, characterized in that, The S6 also includes outputting blood flow animation, parameter distribution map and ischemia risk assessment report through a three-dimensional visualization platform, combining the reconstructed three-dimensional blood flow field with a risk assessment model based on machine learning or deep learning to predict the progression of lower extremity arterial disease or the improvement of hemodynamics after interventional treatment.

10. A method for three-dimensional spatiotemporal hemodynamic reconstruction of lower limb arterial blood perfusion according to claim 1, characterized in that, In step S4, within the PINNs framework, the Windkessel model parameters for each exit of the 3D blood vessel model are first initialized. The loss function constraint term L... WK The algorithm includes pressure equations and differential equations, where the flow rate is obtained by integrating the velocity output by the network at the outlet. Training employs a two-stage strategy: first, the Windkessel parameters are fixed, and only the neural network weights are optimized to fit the flow field; then, both network weights and parameters are optimized simultaneously to minimize the total loss. Through backpropagation, the parameters are gradually updated, eventually converging to an estimate that minimizes the coupling physical residual. In the SA (Automatic Stress Analysis), commercial or open-source software such as OpenFoam can be used to simulate batches of 3D models. The inlet velocity boundary conditions are assigned based on the inlet velocity obtained from S3, and the outlet Windkessel parameters can be derived from the impedance parameters based on the velocity information.

Images