Training prediction model method and device, prediction method and device, electronic device and medium

Abstract

This disclosure presents a training scheme for a hemodynamic parameter prediction model, a hemodynamic parameter prediction scheme, corresponding electronic equipment, and a medium. The training method includes: constructing a vascular topology map representing the geometry and connectivity of blood vessels based on acquired cerebral vascular imaging data to obtain a training sample set; inputting the vascular features from the training samples into the hemodynamic parameter prediction model to be trained to obtain predicted hemodynamic parameters for each blood vessel; calculating a comprehensive loss function value based on the predicted hemodynamic parameters, the comprehensive loss function being a weighted sum of a supervision loss term and a physical consistency loss term; adjusting the parameters of the prediction model based on the comprehensive loss function value until a preset convergence condition is met, resulting in a trained hemodynamic parameter prediction model. According to this disclosure, rapid hemodynamic parameter prediction can be achieved without running a full CFD solution.

Description

Technical Field

[0001] This disclosure generally relates to the field of artificial intelligence technology. More specifically, this disclosure relates to a method and apparatus for training a hemodynamic parameter prediction model, a hemodynamic parameter prediction method and apparatus, and corresponding electronic equipment and media. Background Technology

[0002] In the diagnosis and treatment planning of cerebrovascular diseases such as cerebral aneurysms, arteriovenous malformations, and intracranial arterial stenosis, hemodynamic parameters (such as wall shear stress, oscillatory shear index, and pressure gradient) are crucial for assessing lesion risk, guiding surgical procedure selection, and predicting postoperative outcomes. Currently, clinical assessment primarily relies on computational fluid dynamics (CFD) simulation methods. A typical workflow includes: extracting a three-dimensional vascular model from computed tomography angiography (CTA), magnetic resonance angiography (MRA), or digital subtraction angiography (DSA) images; performing artificial repair and smoothing; mesh generation and boundary condition setting; solving for fluid dynamics using finite volume or finite element methods; and finally deriving hemodynamic parameters such as shear stress, velocity field, and pressure distribution.

[0003] However, existing CFD technologies face multiple challenges in clinical applications, including complex modeling processes with high levels of human intervention, long computation times that make it difficult to meet real-time requirements, sensitivity to boundary conditions leading to limited stability, and difficulty in adapting to batch processing and large-scale research scenarios.

[0004] In recent years, although some studies have attempted to accelerate local blood flow simulation using machine learning or deep learning methods, most methods still use CFD results as input or intermediate processing steps, and a holistic solution capable of directly predicting hemodynamic parameters from medical images has not yet been formed. Therefore, there is an urgent need for a rapid hemodynamic parameter prediction technology that can support real-time clinical assessment and is highly efficient and stable. Summary of the Invention

[0005] In order to at least address one or more of the technical problems mentioned above, this disclosure proposes methods and apparatus for training hemodynamic parameter prediction models, methods and apparatus for predicting hemodynamic parameters, and corresponding electronic equipment and media in several aspects.

[0006] In a first aspect, this disclosure provides a method for training a hemodynamic parameter prediction model, comprising: constructing a vascular topology structure representing vascular geometry and connectivity based on acquired cerebral vascular imaging data to obtain a training sample set; wherein the training samples include vascular features and corresponding hemodynamic label data; inputting the vascular features in the training samples into a hemodynamic parameter prediction model to be trained to obtain predicted hemodynamic parameters for each vascular; calculating a comprehensive loss function value based on the predicted hemodynamic parameters, wherein the comprehensive loss function is obtained by weighted summation of a supervision loss term and a physical consistency loss term; adjusting the parameters of the prediction model based on the comprehensive loss function value until a preset convergence condition is met to obtain a trained hemodynamic parameter prediction model.

[0007] In some embodiments, the hemodynamic tag data includes at least one of wall shear stress, oscillatory shear index, pressure field, and velocity field; the predicted hemodynamic parameters correspond to at least one of predicted wall shear stress, oscillatory shear index, pressure field, and velocity field.

[0008] In some embodiments, the hemodynamic tag data is generated through computational fluid dynamics simulation, specifically including: setting inlet flow velocity boundary conditions and local viscosity parameters, and performing steady-state or quasi-steady-state solutions.

[0009] In some embodiments, the physical consistency loss term includes at least one of the following: a continuity equation constraint term, used to constrain the fluid incompressibility condition; a momentum conservation constraint term, used to constrain the prediction result to satisfy the fluid momentum conservation law; and an energy loss constraint term, used to suppress numerical oscillations of the prediction result in local regions.

[0010] In some embodiments, the preset convergence conditions include: the value of the comprehensive loss function is less than a first threshold, and each item in the physical consistency loss term is less than its corresponding second threshold.

[0011] In some embodiments, the hemodynamic parameter prediction model includes: a feature encoding network for encoding the node features and edge features of the vascular topology structure to obtain node embedding features and edge embedding features; and a graph neural network for aggregating and updating information on the vascular topology based on the node embedding features and the edge embedding features, and outputting the predicted hemodynamic parameters.

[0012] In some embodiments, constructing a vascular topology structure representing the geometry and connectivity of blood vessels based on the acquired cerebral vascular imaging data specifically includes: extracting vascular regions from the cerebral vascular imaging data using a three-dimensional segmentation network or a texture and morphology-based method to obtain a vascular voxel mask; reconstructing a triangular mesh model of the vascular surface based on the vascular voxel mask and extracting a set of vascular centerlines; dividing the blood vessels into multiple vascular segments, each vascular segment corresponding to a tubular structure defined by a centerline; and constructing the vascular topology structure using the vascular segments as nodes and the connectivity between vascular segments as edges.

[0013] In some embodiments, the method further includes: calculating local geometric features for each blood vessel segment, the local geometric features including at least one of radius, rate of change of diameter, curvature, torsion, and relative distance to bifurcation, as feature data of the blood vessel node; and calculating connectivity features for each edge, the connectivity features including at least one of bifurcation angle, transition length, and curvature gradient, as edge features.

[0014] In some embodiments, the hemodynamic tag data includes hemodynamic parameters at multiple sampling points axially distributed along the vessel centerline; the predicted hemodynamic parameters correspond to predicted hemodynamic parameters at multiple sampling points axially distributed along the vessel centerline.

[0015] In some embodiments, the hemodynamic tag data further includes hemodynamic parameters distributed circumferentially in multiple directions within the cross-sectional plane of each sampling point; and the predicted hemodynamic parameters also include predicted hemodynamic parameters distributed circumferentially in multiple directions within the cross-sectional plane of each sampling point. In a second aspect, this disclosure provides a method for predicting hemodynamic parameters, comprising: processing the cerebral vascular imaging data to be tested to construct a topological map structure of the blood vessel to be tested that characterizes the geometry and connectivity of the blood vessels; and inputting the topological map structure of the blood vessel to be tested into a hemodynamic parameter prediction model trained according to the first aspect to obtain the predicted hemodynamic parameters.

[0016] In some embodiments, the method further includes mapping the predicted hemodynamic parameters back to the original three-dimensional vascular space to obtain a parameter field distribution consistent with the three-dimensional vascular geometry.

[0017] In some embodiments, mapping the predicted hemodynamic parameters back to the original three-dimensional vascular space specifically includes: assigning the predicted hemodynamic parameters of the vascular segment to be tested to discrete sampling points on the corresponding centerline based on a pre-established geometric mapping relationship between the vascular segment and the centerline and the three-dimensional vascular region it covers; performing interpolation processing on the sampling points in the boundary region of adjacent segments to achieve a smooth numerical transition; extending the parameters radially along the normal section to the local section region with the centerline sampling point as the center; and for triangular facets in the surface mesh of the vascular segment to be tested, completing parameter backfilling based on its nearest centerline sampling point and the vascular segment to be tested to which it belongs.

[0018] In some embodiments, where the predicted hemodynamic parameters include prediction results in multiple directions distributed circumferentially within the cross-sectional plane of each sampling point, mapping the predicted hemodynamic parameters back to the original three-dimensional vascular space further includes: for a target surface point in the three-dimensional vascular surface grid, determining its circumferential angle within the local cross-sectional plane of the corresponding centerline sampling point; and interpolating the prediction results of adjacent circumferential sampling directions based on the circumferential angle to obtain the hemodynamic parameters of the target surface point.

[0019] In a third aspect, this disclosure provides an apparatus for training a hemodynamic parameter prediction model, comprising: a graph structure construction module, configured to construct a vascular topology graph representing vascular geometry and connectivity based on acquired cerebral vascular imaging data to obtain a training sample set; wherein the training samples include vascular features and corresponding hemodynamic label data; a model prediction module, configured to input the vascular features from the training samples into a hemodynamic parameter prediction model to be trained to obtain predicted hemodynamic parameters for each vascular; a loss calculation module, configured to calculate a comprehensive loss function value based on the predicted hemodynamic parameters, wherein the comprehensive loss function is obtained by weighted summation of a supervision loss term and a physical consistency loss term; and a parameter adjustment module, configured to adjust the parameters of the prediction model based on the comprehensive loss function value until a preset convergence condition is met to obtain a trained hemodynamic parameter prediction model; the apparatus is used to implement the method of the first aspect as described above.

[0020] In a fourth aspect, this disclosure provides a hemodynamic parameter prediction device, comprising: a data processing module for processing cerebral vascular imaging data to be tested and constructing a topological map structure of the blood vessel to be tested that characterizes the geometry and connectivity of the blood vessels; and a model reasoning module for inputting the topological map structure of the blood vessel to be tested into a hemodynamic parameter prediction model trained using the method described in the first aspect to obtain predicted hemodynamic parameters, wherein the device is used to implement the method described in the second aspect.

[0021] In a fifth aspect, this disclosure provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method described in the first aspect, or implements the method described in the second aspect.

[0022] In a sixth aspect, this disclosure provides an electronic device comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of the first aspect, or implement the method of the second aspect.

[0023] The disclosed scheme enables rapid hemodynamic parameter prediction without running a full CFD solution, thus significantly improving computational efficiency while ensuring assessment reliability. Specifically, by leveraging the constraint of the physical consistency loss term, the prediction process effectively suppresses non-physical oscillations, ensuring stable and reliable results and enhancing the model's extrapolation ability in complex lesion scenarios. Furthermore, by constructing a vascular topology graph structure, the continuous and complex geometry of cerebral blood vessels is discretized into a graph structure composed of nodes and edges, effectively reducing the complexity of data processing. Attached Figure Description

[0024] The above and other objects, features, and advantages of exemplary embodiments of this disclosure will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this disclosure are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 An exemplary flowchart of a method for training a hemodynamic parameter prediction model according to some embodiments of this disclosure is shown; Figure 2 An exemplary flowchart illustrating how a vascular topology map representing vascular geometry and connectivity is constructed based on acquired cerebral vascular imaging data, according to some embodiments of this disclosure, is shown. Figure 3 A flowchart of a method for predicting hemodynamic parameters using a trained hemodynamic parameter prediction model according to some embodiments of this disclosure is shown. Figure 4 A block diagram of an apparatus for training a hemodynamic parameter prediction model according to some embodiments of this disclosure is shown; Figure 5 A block diagram of an apparatus for predicting hemodynamic parameters according to some embodiments of this disclosure is shown. Detailed Implementation

[0025] The technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, not all of them. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0026] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0027] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0028] As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0029] The specific embodiments disclosed herein will now be described in detail with reference to the accompanying drawings.

[0030] Figure 1 An exemplary flowchart of a method for training a hemodynamic parameter prediction model according to some embodiments of this disclosure is shown.

[0031] like Figure 1The method 100 includes steps S110, constructing a vascular topology map structure representing the geometry and connectivity of blood vessels based on the acquired cerebral vascular imaging data to obtain a training sample set; wherein the training samples include vascular features and corresponding hemodynamic label data; S120, inputting the vascular features in the training samples into the hemodynamic parameter prediction model to be trained to obtain predicted hemodynamic parameters for each blood vessel; S130, calculating a comprehensive loss function value based on the predicted hemodynamic parameters, wherein the comprehensive loss function is obtained by weighted summation of a supervision loss term and a physical consistency loss term; S140, adjusting the parameters of the prediction model based on the comprehensive loss function value until a preset convergence condition is met to obtain a trained hemodynamic parameter prediction model.

[0032] According to this embodiment, a model capable of rapidly predicting hemodynamic parameters can be obtained without running a full CFD solution, thereby significantly improving computational efficiency while ensuring evaluation reliability. Specifically, by leveraging the constraint of the physical consistency loss term, the model can effectively suppress non-physical oscillations, ensuring stable and reliable results and enhancing the model's extrapolation ability in complex pathological scenarios. Furthermore, by constructing a vascular topology graph structure, the continuous and complex geometry of cerebral blood vessels is discretized into a graph structure composed of nodes and edges, effectively reducing the complexity of data processing.

[0033] Figure 2 An exemplary flowchart illustrating how, based on acquired cerebral vascular image data, a vascular topology structure representing the geometry and connectivity of blood vessels is constructed according to some embodiments of this disclosure. Specifically, step S110 includes sub-steps S111, extracting vascular regions from the cerebral vascular image data using a 3D segmentation network or a texture and morphology-based method to obtain a vascular voxel mask; S112, reconstructing a triangular mesh model of the vascular surface and extracting a set of vascular centerlines based on the vascular voxel mask; S113, dividing the blood vessels into multiple vascular segments, each vascular segment corresponding to a tubular structure defined by a centerline; and S114, constructing the vascular topology structure using the vascular segments as nodes and the connectivity between vascular segments as edges.

[0034] Specifically, in sub-step S111, three-dimensional image data such as CTA, MRA or DSA are read and parsed, and voxel normalization, isotropic resampling and noise filtering preprocessing are performed on them.

[0035] Voxel normalization refers to the process of mapping the grayscale value of each voxel in the original image to a uniform numerical range (usually [0,1] or [-1,1]) through a linear transformation. Since the grayscale range of images acquired by different devices and with different scanning parameters may vary significantly, normalization can eliminate this difference, enabling subsequent segmentation networks or feature extraction algorithms to process input data with a consistent numerical distribution, thereby improving the model's generalization ability and training stability.

[0036] Isotropic resampling is employed because the voxel spacing in the original medical images is often inconsistent across different dimensions (e.g., a planar resolution of 0.5mm × 0.5mm with a slice thickness of 1.0mm), causing voxels to be cuboids rather than cubes. Isotropic resampling uses an interpolation algorithm to resample the original image into a regular grid (e.g., 0.5mm × 0.5mm × 0.5mm) with equal voxel spacing in all three dimensions, ensuring that each voxel represents a cubic region in real space. This process eliminates resolution differences caused by scanning protocols, provides spatially consistent input for subsequent 3D segmentation networks, and facilitates subsequent centerline extraction and geometric measurements.

[0037] Noise filtering is necessary because noise is inevitably introduced during the acquisition and reconstruction of medical images, affecting the clarity of blood vessel boundaries and the accuracy of segmentation. Noise filtering is a process that uses specific algorithms to suppress random noise in images while preserving as much key structural information as possible, such as blood vessel edges and bifurcations. Commonly used methods include anisotropic diffusion filtering (which smooths noise while preserving edges), Gaussian filtering (which smooths noise through neighborhood weighted averaging), and nonlocal mean filtering (which uses image self-similarity for noise reduction).

[0038] The vascular region can be extracted using one of the following two methods: Method 1: Automatic segmentation based on a 3D segmentation network. The preprocessed image data is input into a trained 3D segmentation network (such as 3D U-Net or V-Net). This network extracts multi-scale features through an encoder-decoder structure and outputs a probability map of the same size as the input image. The value of each voxel in the probability map represents the probability that the voxel belongs to a blood vessel region. The probability map is then binarized using threshold segmentation (threshold set to 0.5) to obtain the initial blood vessel voxel mask.

[0039] Method 2: Texture and morphology-based approach. Local texture features (such as gray-level co-occurrence matrix and local binary pattern) are calculated on the preprocessed image data. Combined with the morphological features of blood vessels in the image (such as tubular structure enhancement filtering), the blood vessel region is extracted through methods such as region growing, level set, or adaptive threshold segmentation to obtain the initial blood vessel voxel mask.

[0040] Morphological post-processing is performed on the initial mask, including using three-dimensional connected component analysis to fill the small holes inside the blood vessels, extracting the largest connected component as the final blood vessel region, removing isolated noise and artifacts generated during segmentation, and obtaining the blood vessel voxel mask B.

[0041] In sub-step S112, based on the blood vessel voxel mask B, isosurfaces are extracted using, for example, the traveling cubes algorithm, to generate an initial triangular mesh model of the blood vessel surface. The initial mesh model is then subjected to, for example, Laplacian smoothing, iterated multiple times (e.g., 10 times) to eliminate step artifacts. Simultaneously, a quadratic error metric algorithm is used to simplify the mesh, compressing the number of faces to 30%–50% of the original number. This reduces the computational cost of subsequent processing while maintaining geometric accuracy, resulting in the processed triangular mesh model M of the blood vessel surface.

[0042] Simultaneously, based on the vascular voxel mask B, a topology refinement algorithm is used to extract the vascular centerline. First, the Euclidean distance from each vascular voxel to the vascular boundary is calculated to obtain a distance field map. Starting from the local maxima of the distance field map, the distance field gradient descent direction is traced, and the local maxima are connected to form an initial centerline skeleton. The initial centerline is pruned to remove short branches with a length less than a preset threshold (e.g., 5 mm). The centerline is smoothed using, for example, cubic B-spline fitting to eliminate jagged fluctuations, resulting in a smoothed centerline set C.

[0043] In sub-step S113, based on the set of centerlines C, each centerline is treated as a blood vessel segment. Each centerline corresponds to a complete vascular segment, representing a continuous tubular structural path from start to finish. For each vascular segment... Record the corresponding centerline identifier, start position, end position, length and other attributes, and output the set of blood vessel segments. , where m equals the number of center lines.

[0044] In sub-step S114, the blood vessel segments obtained in sub-step S113 are used. As a graph node Build a node set An edge set E is constructed based on the connectivity of the vessel segments. If two vessel segments share the same bifurcation node, an edge is added between them to indicate a bifurcation connection; if two vessel segments are connected end-to-end on the centerline without any bifurcation point, an edge is added between them to indicate a continuous connection. Each edge... Record the indices of the two connected nodes and the connection type, thereby constructing a graph structure representing the geometric and topological relationships of blood vessels. .

[0045] Simultaneously, a geometric mapping table (Map) can be established for each blood vessel segment. Record the index correspondence between the value and the corresponding centerline, the covered three-dimensional voxel region, and the surface mesh region to form a triplet mapping. ,in This is the centerline corresponding to this segment. This represents the 3D voxel region or surface mesh region covered by the segment. This mapping relationship can be used for subsequent feature extraction and result mapping in the prediction stage. The output is the constructed vascular topology map. And the geometric mapping relationship table Map.

[0046] In some embodiments, the method may further include: calculating local geometric features for each blood vessel segment, the local geometric features including at least one of radius, rate of change of diameter, curvature, torsion, and relative distance to bifurcation, as feature data of the blood vessel node; For each edge, a connectivity feature is calculated, which includes at least one of the following: bifurcation angle, transition length, and curvature gradient, as an edge feature.

[0047] Specifically, based on the completion of blood vessel segmentation and topology graph construction, the corresponding feature data for each node and each edge is calculated and used as input into the hemodynamic parameter prediction model to be trained.

[0048] For each blood vessel segment (Corresponding graph nodes) ), calculate at least one of the following local geometric features: radius The radius reflects the diameter of the blood vessel lumen. Its size directly affects the magnitude of blood flow velocity and wall shear stress, and is a fundamental geometric parameter determining hemodynamic conditions. A smaller radius in a narrow region leads to increased flow velocity and shear stress; conversely, an larger radius in an aneurysm region may result in slowed blood flow and decreased shear stress.

[0049] Pipe diameter change rate This reflects the degree of drastic change in vessel diameter along the axial direction. Dilation areas (positive rate of change) may cause blood flow deceleration and pressure rebound, easily leading to flow separation; constriction areas (negative rate of change) cause blood flow acceleration and pressure drop. Drastic changes in vessel diameter are often associated with lesions, such as dilation areas before and after stenosis.

[0050] curvature This reflects the degree of vascular tortuosity. Vascular tortuosity induces secondary flow, causing uneven circumferential distribution of shear stress on the vessel wall. The outer side of the tortuosity experiences high shear stress, while the inner side experiences low shear stress. Areas with high curvature are common sites for aneurysms and are also high-risk areas for blood flow disturbances.

[0051] Torque This reflects the degree to which the blood vessel deviates from its plane, i.e., the severity of the central axis torsion. Torsion causes complex spiral motions in blood flow, affecting the distribution pattern of wall shear stress and is related to the focal distribution of atherosclerosis.

[0052] Relative distance to the bifurcation This reflects the positional characteristics of vascular segments within the vascular tree. Bifurcation regions exhibit complex blood flow patterns, including flow separation, stagnation points, and secondary flows, making them common sites for aneurysms and atherosclerosis. The closer to the bifurcation, the greater the impact of bifurcation interference on hemodynamics.

[0053] For each edge (Connecting nodes) and ), calculate at least one of the following connectivity features: bifurcation angle The bifurcation angle reflects the size of the angle between the branch vessel and the main vessel. The size of the bifurcation angle directly affects the angle at which blood flows into the branch and the energy loss. Small-angle branches have relatively smooth blood flow, while large-angle branches are prone to flow separation and stagnation zones. The bifurcation angle is closely related to the formation of bifurcation aneurysms.

[0054] Transition length The transition length reflects the abruptness of the geometric transition at the junction of two vessel segments. A longer transition length indicates a smooth geometric change and good blood flow adaptation; a shorter transition length indicates abrupt geometric changes, which can easily lead to flow separation, eddy formation, and drastic changes in wall shear stress. The inlet and outlet transition length of stenotic lesions is an important parameter for assessing hemodynamic effects.

[0055] Curvature gradient This reflects the degree of change in the curvature of blood flow as it passes through the junction. Sudden changes in curvature can disrupt flow stability, induce secondary flow and flow separation, leading to abnormal local wall shear stress, which is closely related to the focal distribution of atherosclerotic plaques.

[0056] In some embodiments, the hemodynamic tag data includes at least one of wall shear stress, oscillatory shear index, pressure field, and velocity field; and therefore, the predicted hemodynamic parameters correspond to at least one of predicted wall shear stress, oscillatory shear index, pressure field, and velocity field.

[0057] Wall shear stress (WSS) is defined as the frictional force acting tangentially on the inner surface of the blood vessel wall. It is determined by both blood viscosity and the near-wall velocity gradient, and is calculated using the following formula: Where μ is blood viscosity, The velocity gradient is perpendicular to the vessel wall. WSS (vessel endothelial septum saturation) is a key indicator for assessing the functional status of vascular endothelial cells. Adequate WSS within the physiological range maintains normal endothelial function and vascular homeostasis; low WSS... This region is prone to endothelial dysfunction, lipid deposition, and atherosclerotic plaque formation; high In certain regions, the distribution pattern of the WSS (wasted surface area) can induce endothelial damage, platelet activation, and aneurysm formation or rupture. In aneurysm assessment, the distribution pattern of the WSS can be used to determine the aneurysm's growth direction and rupture risk.

[0058] The oscillatory shear index (OSI) describes the degree to which the direction of wall shear stress changes during the cardiac cycle, and is defined as follows: The OSI value ranges from 0 to 0.5. OSI=0 indicates that the shear stress direction remains constant, while OSI=0.5 indicates that the alternation time between positive and negative shear stress directions is equal. High OSI regions reflect frequent reversals or violent oscillations in blood flow direction, commonly found on the lateral wall of bifurcations, aneurysm necks, and downstream regions of stenosis. Sustained directional oscillations can interfere with the normal arrangement and function of endothelial cells, promoting inflammatory responses and the progression of atherosclerosis, and are important indicators for assessing the degree of blood flow disturbance and lesion risk.

[0059] The pressure field (P) describes the pressure distribution at various points during blood flow within a blood vessel, including the superposition of static and dynamic pressures. The pressure field directly affects the transmural pressure and wall stress of the vessel wall: high-pressure areas increase the mechanical load on the vessel wall, and long-term effects can lead to vascular remodeling and aneurysm formation; local pressure gradients drive blood flow, and excessively rapid pressure drops indicate significant energy loss. In stenotic lesions, measuring the transstenotic pressure gradient is an important basis for assessing hemodynamic significance; in aneurysm assessment, intraneural pressure distribution is associated with the risk of rupture.

[0060] The velocity field (V) describes the velocity vector distribution of blood particles at various locations, including both magnitude and direction. The velocity field determines convective transport processes and blood flow patterns: high-velocity regions are commonly found in narrow jets and on the outer sides of bends; low-velocity regions are found within aneurysm sacs, bifurcation stagnation zones, and dilated segments. The distribution characteristics of the velocity field (such as jet direction, eddy location, and stagnation zone extent) directly affect the calculation results of WSS and OSI, and are also a direct basis for judging blood flow state (laminar / turbulent), energy dissipation, and thrombosis potential. The three-dimensional distribution of the velocity field fully describes the kinematic behavior of blood.

[0061] In some embodiments, the hemodynamic tag data is generated through computational fluid dynamics simulation, specifically including: setting inlet flow velocity boundary conditions and local viscosity parameters, and performing steady-state or quasi-steady-state solutions.

[0062] Specifically, to construct the labeled data required for supervised learning, a one-time computational fluid dynamics simulation is performed on the cases in the training sample set.

[0063] Based on prior clinical knowledge and specific lesion types, appropriate boundary conditions are set for each vascular model to be simulated.

[0064] Specifically, inlet flow velocity boundary conditions are set. Based on physiological data measured by Doppler ultrasound or phase-contrast MRI, or the normal physiological range reported in the literature, the flow velocity profile at the vessel inlet is set. A fully developed parabolic velocity distribution (laminar flow assumption) or a flat velocity distribution (considering the inlet segment effect) can be used, and the velocity amplitude is adjusted according to the specific vessel location and physiological state (resting / load). For diseased vessels with stenosis or aneurysm, multiple different flow velocity conditions can be set to cover various physiological scenarios.

[0065] Local viscosity parameters are set. Blood is considered either an incompressible Newtonian fluid or a non-Newtonian fluid. For the rapid flow region in large blood vessels, the Newtonian assumption is adopted, and the viscosity is set to a constant value (usually 100%). For low-flow-rate regions or small blood vessels, non-Newtonian models (such as the Carreau or Casson models) can be used to describe shear-thinning behavior to more accurately reflect blood flow characteristics. Blood density can be set to 1060 kg / m³.

[0066] Then, the three-dimensional Navier-Stokes equations are numerically discretized and iteratively solved using the finite volume method or the finite element method.

[0067] Depending on the clinical problem and computational resources, either steady-state or quasi-steady-state solution modes can be selected. Steady-state solutions assume that the flow does not change over time, are computationally efficient, and are suitable for assessing average hemodynamic parameters; quasi-steady-state solutions perform steady-state calculations at multiple time points or use time-averaged boundary conditions across multiple phases within the cardiac cycle, which can approximately reflect the average effect of pulsatile flow.

[0068] Convergence criteria can be set (e.g., residuals decreasing to a certain level). (The following steps) employ appropriate relaxation factors to ensure iterative stability. For complex lesion regions, local mesh refinement can be used to improve computational accuracy.

[0069] After the solution is obtained, the following hemodynamic parameters are extracted from the simulation results as labeled data for supervised learning: wall shear stress field. Oscillatory shear index Pressure field Velocity field The above parameters are combined to form the hemodynamic label data corresponding to each training sample. .

[0070] The simulated hemodynamic label data is associated and stored with the corresponding vascular topology and geometric feature data to construct complete training sample pairs. It is important to emphasize that the above CFD simulation is executed only once during the model training phase to generate supervision signals; in the actual prediction phase after model deployment, there is no need to run any CFD solution process again, thus achieving rapid prediction from medical images to hemodynamic parameters.

[0071] In some embodiments, the hemodynamic parameter prediction model includes: a feature encoding network for encoding the node features and edge features of the vascular topology structure to obtain node embedding features and edge embedding features; and a graph neural network for aggregating and updating information on the vascular topology based on the node embedding features and the edge embedding features, and outputting the predicted hemodynamic parameters.

[0072] Specifically, hemodynamic parameter prediction model It consists of two parts: a feature encoding network and a graph neural network.

[0073] Among them, the feature coding network receives the generated node features. Sum of edge features As input, the network consists of two parallel sub-networks: a node encoder performs an embedding transformation on the geometric features of each blood vessel segment to extract a high-dimensional implicit representation, thus obtaining the node embedding features. The edge encoder performs an embedding transformation on the connection relationship features of each edge to obtain the edge embedding features. The encoding network can be implemented using a multilayer perceptron (MLP) or a one-dimensional convolutional neural network. It maps the original features to the latent space through nonlinear transformation, enabling the subsequent graph neural network to better capture the complex interaction relationships between features.

[0074] Graph neural networks embed features into the nodes output by the feature encoding network. and edge embedding features As input, in the constructed vascular topology structure The graph neural network performs multi-layered information transmission and feature updates. In each layer, each node aggregates the feature information of its neighboring nodes and fuses it with the features of the connecting edges to update its own feature representation. Through multi-layered message passing mechanisms, the model can model the mutual influence of blood flow between vascular branches, long-distance geometric constraints, and complex flow patterns. Graph neural networks can be implemented using variants such as Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), or Graph Isomorphic Networks (GIN).

[0075] training samples (Including the vascular topology and its node and edge features) Input prediction model After feature encoding and graph neural network inference, the model outputs predicted hemodynamic parameters for each vascular node. The predicted hemodynamic parameters include at least one of the following: predicted wall shear stress. This reflects the distribution of tangential frictional force on the vessel wall; the predicted oscillatory shear index. It describes the degree to which the direction of shear stress changes during the cardiac cycle; it predicts the pressure field. It reflects the pressure distribution at various locations within the blood vessel; the predicted velocity field , which describes the velocity vector of blood particles.

[0076] The above prediction results are expressed as .

[0077] To ensure numerical consistency between the model's predictions and the real-label data generated by computational fluid dynamics simulations, a supervision loss term is defined. This loss term measures the numerical similarity between the predicted and labeled values ​​by calculating the L2 norm distance. The specific calculation formula is as follows:

[0078] in, These are hemodynamic label data; L2 norm represents the square root of the sum of squares of the elements in a vector or tensor.

[0079] The purpose of this supervised loss term is to constrain the model's predictions to remain consistent with the CFD simulation results globally, ensuring that the model has good numerical fitting ability on training cases with similar structures.

[0080] In some embodiments, the physical consistency loss term includes at least one of the following: a continuity equation constraint term, used to constrain the fluid incompressibility condition; a momentum conservation constraint term, used to constrain the prediction result to satisfy the fluid momentum conservation law; and an energy loss constraint term, used to suppress numerical oscillations of the prediction result in local regions.

[0081] Specifically, to avoid the model from "rigidly fitting" the CFD simulation results and potentially ignoring underlying physical laws under supervised loss constraints, some embodiments further introduce a physical consistency loss term, so that the prediction results follow the basic laws of fluid mechanics while satisfying the data fitting requirements.

[0082] Among them, the continuity equation constraint terms This constraint is used to ensure the consistency of the fluid incompressibility condition. It measures the degree to which mass conservation is satisfied by calculating the divergence of the predicted velocity field, ensuring that the predicted velocity field satisfies the balance between inflow and outflow at any location.

[0083] The calculation formula is as follows: ,in, For the velocity field predicted by the model, Represents the divergence of the velocity field. The norm is used to represent the constraint term, which approaches zero when the predicted velocity field satisfies the incompressibility condition.

[0084] Momentum conservation constraint This constraint is used to ensure that the predicted results satisfy the law of conservation of fluid momentum. This constraint term is calculated by retrieving the residuals of the Navier-Stokes equations in a weak form, comprehensively considering the coupling relationship between the predicted velocity field and the pressure field, and ensuring that both satisfy the dynamic equilibrium of the fluid motion.

[0085] The calculation formula is as follows:

[0086] in, Blood density, To predict the velocity field, To predict the pressure field, For blood dynamic viscosity, Represents the convection term. Represents the pressure gradient term. This represents the viscous diffusion term. This constraint term approaches zero when the predicted velocity and pressure fields satisfy the Navier-Stokes equations.

[0087] Energy loss constraint This constraint is used to suppress numerical oscillations in the prediction results in local regions. By penalizing the gradient of the predicted wall shear stress field, this constraint reduces non-physical fluctuations that may occur in high-shear regions, thereby improving the spatial smoothness and numerical stability of the prediction results in complex lesion areas.

[0088] in,

[0089] in, The wall shear stress field predicted by the model. This represents the spatial gradient of the wall shear stress field. This constraint term, by penalizing drastic changes in shear stress, suppresses numerical oscillations in high-shear regions, ensuring the physical plausibility of the prediction results.

[0090] Specifically, the supervision loss term and the physical consistency loss term are weighted and summed to obtain the comprehensive loss function for model training:

[0091] in, These are weighting coefficients used to adjust the contribution of different loss terms to the total loss. These weighting coefficients can be adjusted according to the physical characteristics of different disease scenarios (such as aneurysms, stenosis, arteriovenous malformations, etc.). In scenarios with scarce data or high noise, the weight of physical constraint terms can be appropriately increased to strengthen the guiding role of physical laws; in scenarios with sufficient data and high CFD simulation accuracy, the weight of physical constraint terms can be appropriately decreased to prioritize consistency with the labeled data.

[0092] Through the optimization of the aforementioned comprehensive loss function, the model training disclosed herein is not a simple data fitting process, but a learning mechanism driven by both "data supervision and physical consistency." This mechanism enables the model to actively satisfy the fundamental laws of fluid dynamics while fitting the training data, effectively reducing abnormal oscillations in outlier regions and significantly improving the model's extrapolation ability and generalization performance on unseen cases.

[0093] In some embodiments, the preset convergence conditions include: the value of the comprehensive loss function is less than a first threshold, and each item in the physical consistency loss term is less than the corresponding second threshold. These thresholds can be preset based on physical priors or experimental experience.

[0094] Specifically, during model training, the network parameters are updated iteratively by optimizing the comprehensive loss function L. When a preset convergence condition is met, training stops and the current model parameters are saved as the trained hemodynamic parameter prediction model. The preset convergence condition includes the following two aspects: Monitor the change in the value of the comprehensive loss function L, and when it drops to a preset first threshold... The following indicates that the overall prediction error of the model has been reduced to an acceptable range. As mentioned earlier, the first threshold can be preset to a fixed value based on physical priors or experimental experience. Of course, it can also be judged in conjunction with the downward trend of the loss function. If the decrease in the comprehensive loss function is less than the preset rate of change (e.g., 0.1%) within several consecutive iterations (e.g., 50 epochs), it can be regarded as monitoring the convergence of the comprehensive loss function even if the first threshold has not been reached.

[0095] Furthermore, to ensure that the model predictions are not only numerically accurate but also conform to the fundamental laws of fluid mechanics, it is necessary to monitor each physical constraint loss individually. Separate continuity equation constraint terms should be set for each. Momentum conservation constraint Energy loss constraint The corresponding second threshold When all physical losses decrease below their respective thresholds, it indicates that the prediction results have achieved sufficient consistency at the level of physical laws. As mentioned earlier, the second threshold can also be preset to a fixed value based on physical priors or experimental experience. Model training convergence requires both of the above conditions to be met simultaneously: the overall loss function value must be below the first threshold, and each physical consistency loss term must be below its corresponding second threshold. If only the overall loss function meets the threshold while the physical loss term does not, it indicates that the model may have overfitting or numerical oscillation problems, requiring continued training or adjustment of the tradeoff coefficients. If only the physical loss term meets the threshold while the overall loss function does not, it indicates that although the model conforms to physical laws, its fit with the labeled data is insufficient, requiring further optimization of the supervised loss.

[0096] By setting the above multi-level convergence conditions, we can ensure that the trained model meets the requirements in terms of both numerical accuracy and physical consistency, providing a reliable guarantee for subsequent clinical predictions.

[0097] In some embodiments, the hemodynamic tag data includes hemodynamic parameters at multiple sampling points axially distributed along the vessel centerline; the predicted hemodynamic parameters correspond to predicted hemodynamic parameters at multiple sampling points axially distributed along the vessel centerline.

[0098] Specifically, hemodynamic parameters are represented by axial sampling along the blood vessel centerline, aiming to reduce the dimensionality of the model output and improve computational efficiency while ensuring prediction accuracy.

[0099] For each vascular segment, discretization is performed along its centerline at preset sampling intervals to generate a series of axial sampling points. The selection of the sampling interval comprehensively considers factors such as the geometric complexity of the vessel, the spatial rate of change of hemodynamic parameters, and computational resource limitations. In a typical implementation, the sampling interval is set to a fixed value between 0.5 mm and 2.0 mm. For curved areas with large curvature changes, near bifurcations, or critical lesion areas, an adaptive sampling strategy can be adopted, appropriately increasing the sampling points at locations with drastic geometric changes, and appropriately increasing the sampling interval to reduce the amount of data for areas with gentle geometric changes, such as straight segments. Each axial sampling point records its three-dimensional spatial coordinates, its relative position on the centerline, and local coordinate system information.

[0100] When generating computational fluid dynamics simulation label data, the three-dimensional hemodynamic parameter field is mapped onto the aforementioned axial sampling points to form a discretized label representation. For each axial sampling point, the hemodynamic parameter value at that location is extracted.

[0101] Pressure values ​​and velocity vectors can be directly interpolated from the simulation results because the axial sampling points are located on the centerline of the blood vessel, while pressure and velocity are continuous fields defined throughout the volume space.

[0102] Wall shear stress and oscillatory shear index are defined only on the vessel wall and cannot be directly obtained on the vessel centerline. For each cross-section where an axial sampling point is located, the parameters of all circumferential wall positions of that cross-section can be extracted and the average value calculated. This average value can then be used as the representative parameter value of that cross-section and correlated with the axial sampling point on the centerline.

[0103] In the above manner, the set of axial sampling points for each vascular segment corresponds to a set of discretized hemodynamic parameter sequences, which constitute the label data for that segment.

[0104] Accordingly, the hemodynamic parameter prediction model is configured to output predicted values ​​corresponding one-to-one with axial sampling points. After completing graph neural network inference, the model generates a sequence of predicted parameters distributed axially along the centerline for each vessel segment. The model output layer is designed with a dimension matching the number of axial sampling points, and the output at each sampling point includes the predicted wall shear stress, oscillatory shear index, pressure value, and velocity vector. For vessel segments of different lengths, a masking mechanism or a variable-length output strategy can be used to enable the model to handle cases where segment lengths are inconsistent. The number and location distribution of axial sampling points output by the model are consistent with the labeled data during the training phase, ensuring that the two can be compared point-by-point and loss calculated.

[0105] Hemodynamic parameter representation using axial sampling offers several advantages. By adjusting the sampling interval, a balance can be struck between prediction accuracy and computational efficiency, adapting to the needs of different clinical scenarios. More sparse sampling can be used for initial screening, while denser sampling can be used for preoperative planning. The geometry naturally corresponds to the vessel centerline, facilitating subsequent visualization and clinical interpretation; physicians can intuitively view parameter variation curves along the vessel path.

[0106] The discretized representation makes point-by-point comparisons between prediction results and computational fluid dynamics labels more intuitive, and loss calculations more concise and efficient. Simultaneously, it allows analysis of the same blood vessel at different sampling densities, enabling multi-scale observation from global assessment to fine-grained local analysis, providing multi-level information support for clinical decision-making.

[0107] In some embodiments, the hemodynamic tag data further includes hemodynamic parameters distributed circumferentially in multiple directions within the cross-sectional plane of each sampling point; and the predicted hemodynamic parameters also include predicted hemodynamic parameters distributed circumferentially in multiple directions within the cross-sectional plane of each sampling point.

[0108] Specifically, in some other embodiments disclosed herein, in order to further improve the spatial resolution of hemodynamic parameters of the blood vessel wall, a circumferential multi-point sampling mechanism is introduced on the basis of the above-mentioned axial sampling. Hemodynamic parameters are predicted in multiple circumferential directions in the cross-sectional plane of each axial sampling point, thereby forming a two-dimensional parameter field with both axial and circumferential resolution.

[0109] When generating computational fluid dynamics simulation label data, the three-dimensional hemodynamic parameter field is mapped onto the aforementioned circumferential sampling points. For each circumferential sampling direction of each axial sampling point, this direction corresponds to a vessel wall position. The hemodynamic parameter values ​​at this wall position are extracted by interpolation to form a label representation with circumferential resolution.

[0110] Specifically, for each wall position corresponding to a circumferential sampling direction, its coordinates in three-dimensional space are first determined. Then, the mesh nodes or elements adjacent to the position are found on the mesh of the computational fluid dynamics simulation results, and the parameter values ​​at the position are calculated using interpolation methods (such as trilinear interpolation, nearest neighbor interpolation, or interpolation based on finite element shape functions).

[0111] A consistent interpolation method is used for different hemodynamic parameters.

[0112] For wall shear stress, the wall shear stress vector at the wall location is obtained by interpolation on the wall mesh, and its amplitude is calculated as the wall shear stress value in the circumferential direction.

[0113] For a pressure field, the pressure value at the wall location can be obtained by interpolation on the volume mesh or wall mesh.

[0114] For the oscillating shear index, this parameter is calculated based on the change of wall shear stress over time, and its spatial distribution is defined on the wall grid. Therefore, the same interpolation method as the wall shear stress is used to extract the oscillating shear index value at this wall location.

[0115] For the velocity field, since the velocity field is defined on the volume grid, for each circumferential sampling direction, interpolation is performed at the nearest point inside the wall position to obtain the velocity vector of that point, which is used to reflect the blood flow velocity distribution characteristics near the wall.

[0116] In the above manner, each axial sampling point corresponds to a set of wall hemodynamic parameters distributed along the circumference. Combining the circumferential label data of all axial sampling points forms a two-dimensional parameter field label covering the entire blood vessel surface.

[0117] Accordingly, the hemodynamic parameter prediction model is configured to output predicted values ​​corresponding one-to-one with the aforementioned circumferential sampling points. After completing graph neural network inference, the model generates a sequence of predicted parameters distributed circumferentially for each axial sampling point. The model output layer is designed with a two-dimensional output dimension matching the number of axial sampling points and the number of circumferential sampling directions. For each axial sampling point, the corresponding predicted hemodynamic parameters are output in each circumferential direction, including predicted wall shear stress, oscillatory shear index, pressure value, and velocity vector. The number of circumferential sampling points and the angular intervals of the model output are consistent with the labeled data during the training phase, ensuring that the two can be compared point-by-point and loss calculated. For different circumferential sampling densities in different vessel segments or regions, a masking mechanism or adaptive output strategy is used to enable the model to handle inconsistent circumferential sampling.

[0118] Hemodynamic parameters represented using circumferential sampling can precisely characterize complex blood flow patterns such as high-shear regions on the outer side of vessel bends, low-shear regions on the bifurcation side, and localized backflow regions. These patterns are often smoothed or lost in axial homogenization representations. By independently predicting in multiple circumferential directions, the hemodynamic differences at the same axial location in different circumferential directions are preserved, more realistically reflecting the non-uniform mechanical stimulation experienced by the vessel wall. The predilection sites of aneurysms and the focal distribution of atherosclerosis are closely related to the circumferential non-uniformity of wall shear stress. Circumferential sampling representation can provide more accurate quantitative evidence for clinical risk assessment. Furthermore, circumferential sampling representation is based on axial sampling, and both can coexist in the same model. A pure axial sampling mode can be used when rapid assessment is needed, while a circumferential sampling mode can be activated when detailed analysis is required, achieving a flexible trade-off between accuracy and efficiency.

[0119] Figure 3 A flowchart illustrating a method for predicting hemodynamic parameters using a trained hemodynamic parameter prediction model, according to some embodiments of this disclosure, is shown. This prediction method, based on a prediction model built during the training phase, enables rapid prediction of hemodynamic parameters from images of the cerebral blood vessels to be examined.

[0120] like Figure 3 As shown, method 300 includes step S310, which processes the cerebral vascular imaging data to be tested to construct a topological map structure of the blood vessels to be tested, representing the geometry and connectivity of the vessels. This processing procedure is consistent with the data preprocessing workflow in the training phase, and may specifically include the following steps.

[0121] Three-dimensional vessel segmentation is performed on the cerebral vascular imaging data to be tested, and the vessel regions are extracted to obtain the voxel mask of the target vessel. The segmentation method can use the same three-dimensional segmentation network as in the training phase to ensure the consistency of the segmentation results.

[0122] Based on the voxel mask of the vessel to be tested, a triangular mesh model of the vessel surface is reconstructed, and the set of centerlines of the vessel is extracted. Surface reconstruction can be performed using a traveling cubes algorithm to generate the triangular mesh, and centerline extraction uses a topology refinement algorithm to obtain the vascular skeleton.

[0123] The centerline of the vessel to be tested is divided into multiple segments, and each segment corresponds to a tubular structure defined by the centerline.

[0124] Using the segments of the blood vessel to be tested as nodes and the connections between these segments as edges, a topological graph of the blood vessel to be tested is constructed. This graph structure fully represents the geometric shape and topological connections of the blood vessel to be tested.

[0125] Optionally, in step S320, based on the constructed topology map of the vessel to be tested, local geometric features are calculated for each segment of the vessel to be tested, serving as node features of the topology map structure. Local geometric features include at least one of radius, rate of change of diameter, curvature, torsion, and relative distance to bifurcation; these features describe the geometric properties of the vessel segments from different perspectives.

[0126] Simultaneously, connectivity features are calculated for each edge, serving as edge features of the topological structure of the blood vessel under test. These connectivity features include at least one of bifurcation angle, transition length, and curvature gradient, reflecting the spatial connectivity and geometric transition characteristics between blood vessel segments.

[0127] The extraction methods for node features and edge features are exactly the same as those in the training phase, ensuring the consistency of the input data distribution.

[0128] In step S330, the constructed topological graph of the blood vessel to be tested, along with its corresponding node and edge features, is input into the previously trained hemodynamic parameter prediction model. The model first encodes the node and edge features using a feature encoding network to obtain node embedding features and edge embedding features. Subsequently, the graph neural network, based on the blood vessel topological graph structure, combines the node and edge embedding features to perform multi-layer information transmission and feature updates on the graph. Finally, the model outputs the predicted hemodynamic parameters for each node of the blood vessel to be tested.

[0129] In some embodiments disclosed herein, the predicted hemodynamic parameters are represented by predicted values ​​at multiple sampling points axially distributed along the centerline of the vessel to be tested. For each segment of the vessel to be tested, it is discretized into a series of axial sampling points along its centerline at preset sampling intervals, and the model outputs the corresponding hemodynamic parameter value for each sampling point. This axial sampling representation effectively reduces the dimensionality of the model output while ensuring prediction accuracy, improving computational efficiency, and is suitable for rapid clinical assessment scenarios where circumferential resolution requirements are not high. The positional distribution of the axial sampling points remains consistent with that during the training phase.

[0130] In other embodiments disclosed herein, to further enhance the spatial resolution of hemodynamic parameters of the vessel wall, the predicted hemodynamic parameters also include predicted values ​​in multiple directions distributed circumferentially within the cross-sectional plane of each sampling point along the centerline of the vessel under test. Specifically, for each axial sampling point, multiple circumferential sampling directions are generated at preset angular intervals within a local cross-sectional plane perpendicular to the centerline, centered on that point. The model outputs corresponding hemodynamic parameter values ​​for each circumferential direction, forming a two-dimensional parameter field with both axial and circumferential resolution. This circumferential sampling representation can finely characterize complex blood flow patterns such as high-shear regions on the outer side of vessel bends, low-shear regions on the bifurcation side, and local reflux regions, preserving the hemodynamic differences at the same axial position in different circumferential directions, and providing more accurate quantitative evidence for clinical risk assessment.

[0131] To facilitate clinical visualization and quantitative analysis, in some embodiments disclosed herein, the predicted hemodynamic parameters are mapped back to the original three-dimensional vascular space to obtain a parameter field distribution consistent with the three-dimensional vascular geometry. This mapping process can fully utilize the geometric mapping relationships established during the training phase, and through a series of steps, restore the segmented prediction results to a continuous parameter field covering the entire vascular surface.

[0132] First, based on the pre-established geometric mapping relationship between the vascular segment and the centerline and the three-dimensional vascular region it covers, the predicted hemodynamic parameters of the vascular segment to be tested are assigned to the discrete sampling points on the corresponding centerline.

[0133] Subsequently, interpolation processing is performed on the sampling points in the boundary region between adjacent segments to achieve a smooth numerical transition. Different segments exhibit numerical jumps due to the independence of model predictions; linear interpolation or spline interpolation can make the parameter changes in the boundary region continuous and natural.

[0134] Next, using the centerline sampling point as the center, the parameters are radially extended along the normal section to the local section region. For each centerline sampling point, a local section plane is constructed with its tangential vector as the normal, and the parameter values ​​at the centerline point are assigned radially to various positions within the section, forming the parameter distribution on the local section.

[0135] Finally, for the triangular facets in the mesh of the blood vessel surface to be tested, parameter backfilling is completed based on its nearest centerline sampling point and the blood vessel segment to which it belongs. For each surface point, the nearest centerline sampling point is found, and values ​​are assigned according to the parameter value of that sampling point and the blood vessel segment to which that point belongs, ultimately obtaining a three-dimensional parameter field covering the entire blood vessel surface.

[0136] When the predicted hemodynamic parameters are still output using circumferential sampling, the above three-dimensional mapping process needs to be adjusted accordingly to preserve the hemodynamic differences in the circumferential direction. Specifically, for any target surface point in the three-dimensional vascular surface mesh, its circumferential angle in the local cross-sectional plane of the corresponding centerline sampling point is first determined. This angle is obtained by calculating the azimuth angle of the surface point relative to the centerline sampling point.

[0137] Subsequently, the predicted results for adjacent circumferential sampling directions are interpolated based on this circumferential angle to obtain the hemodynamic parameters of the target surface point. Linear interpolation or spline interpolation is used to ensure continuous and smooth parameter changes in the circumferential direction. In this way, surface points at the same axial position in different circumferential directions obtain parameter values ​​that match their azimuth angles, fully preserving the circumferential non-uniformity of parameters such as wall shear stress, and more realistically reflecting the non-uniform mechanical stimulation experienced by the blood vessel wall.

[0138] After completing the above mapping, the final output parameter field distribution is completely consistent with the original three-dimensional vascular geometry. It can be directly used for clinical visualization and quantitative analysis, providing an intuitive and reliable hemodynamic assessment basis for the diagnosis and treatment planning of cerebrovascular diseases such as cerebral aneurysms, arteriovenous malformations, and intracranial arterial stenosis.

[0139] Figure 4 A block diagram of an apparatus for training a hemodynamic parameter prediction model according to some embodiments of this disclosure is shown. The apparatus is used to perform the aforementioned training method, with each module corresponding to a step of the training method.

[0140] The device 400 includes core modules such as a graph structure construction module 410, a model prediction module 420, a loss calculation module 430, and a parameter adjustment module 440.

[0141] The graph structure construction module 410 is used to construct a vascular topology graph structure based on the acquired cerebral vascular image data to obtain a training sample set. The operations performed by this module, such as image segmentation, centerline extraction, vascular segmentation, and topology graph construction, correspond to the data processing steps in the aforementioned training method.

[0142] The model prediction module 420 is used to input vascular features from the training samples into the model to be trained to obtain predicted hemodynamic parameters. The feature encoding, graph neural network inference, and other operations performed by this module correspond to the model forward propagation step in the aforementioned training method.

[0143] The loss calculation module 430 is used to calculate the comprehensive loss function value based on the prediction results. This loss function is obtained by weighted summation of the supervised loss term and the physical consistency loss term. The operations performed by this module, such as supervised loss calculation and physical constraint loss calculation, correspond to the loss function construction steps in the aforementioned training method.

[0144] The parameter tuning module 440 is used to adjust the model parameters according to the loss function value until convergence, thus obtaining a trained prediction model. The backpropagation, gradient update, and convergence determination operations performed by this module correspond to the parameter optimization steps in the aforementioned training method.

[0145] In addition to the core modules mentioned above, the device can also be configured with other auxiliary modules as needed. For example, in some embodiments, an image preprocessing module can be configured to normalize, resample, and filter the original image data; in other embodiments, a feature extraction module can be configured to calculate node and edge features; and in still other embodiments, a label generation module can be configured to generate hemodynamic label data through computational fluid dynamics simulation. These auxiliary modules can be flexibly combined and configured according to the specific implementation of the training method.

[0146] Figure 5 A block diagram of an apparatus for predicting hemodynamic parameters according to some embodiments of this disclosure is shown. The apparatus is used to perform the aforementioned prediction method, with each module corresponding to a step of the prediction method.

[0147] Specifically, the device 500 includes core modules such as a data processing module 510 and a model reasoning module 520.

[0148] The data processing module 510 is used to process the cerebral vascular image data to be tested and construct a topological map structure of the blood vessels to be tested, representing the geometry and connectivity of the blood vessels. The operations performed by this module, such as image segmentation, centerline extraction, blood vessel segmentation, topological map construction, and node and edge feature extraction, correspond to the preprocessing steps in the aforementioned prediction method.

[0149] The model inference module 520, connected to the data processing module, is used to input the constructed topological map of the blood vessel to be tested into the hemodynamic parameter prediction model obtained by the training method described above, and to obtain the predicted hemodynamic parameters. The feature encoding, graph neural network inference, and other operations performed by this module correspond to the model forward propagation step in the aforementioned prediction method.

[0150] Similarly, in addition to the core modules described above, the device can be configured with other auxiliary modules as needed. For example, in some embodiments, a parameter mapping module can be configured to map the predicted hemodynamic parameters back to the original three-dimensional vascular space, obtaining a parameter field distribution consistent with the three-dimensional vascular geometry; in embodiments employing a circumferential sampling prediction mode, this module can also interpolate the prediction results of adjacent circumferential sampling directions based on the circumferential angle of surface points. In other embodiments, a visualization module can be configured to visualize the three-dimensional parameter field in the form of a color cloud map, or generate a risk assessment report containing statistical information of key parameters. In still other embodiments, an output interface module can be configured to export the prediction results to a clinical information system or store them in a local storage device. These auxiliary modules can be flexibly combined and configured according to the specific implementation of the prediction method and clinical needs.

[0151] According to some embodiments of this disclosure, a computer-readable storage medium is also provided. This storage medium stores a computer program that, when executed by a processor, implements the aforementioned method for training a hemodynamic parameter prediction model, or a hemodynamic parameter prediction method.

[0152] The computer-readable storage medium can be any medium capable of storing a computer program, such as a read-only memory, random access memory, disk, or optical disk. When the computer program is executed by a processor, it enables the processor to perform the steps of the training or prediction method described herein, thereby enabling the training of a hemodynamic parameter prediction model from cerebrovascular imaging data, or the use of the trained model for rapid hemodynamic parameter prediction.

[0153] According to some embodiments of this disclosure, an electronic device is also provided. The electronic device includes one or more processors and a memory, wherein the memory is used to store one or more computer programs.

[0154] When the one or more computer programs are executed by the one or more processors, the one or more processors are able to implement the aforementioned method for training a hemodynamic parameter prediction model, or implement the aforementioned hemodynamic parameter prediction method.

[0155] The electronic device can be any device with computing capabilities, such as a server, workstation, personal computer, embedded device, or mobile terminal. By running the computer program, the electronic device can perform the training process from cerebrovascular imaging data to a hemodynamic parameter prediction model, or use the trained model to quickly infer new imaging data to be tested, and output predicted hemodynamic parameters, providing auxiliary decision support for the clinical assessment of cerebrovascular diseases.

[0156] While numerous embodiments of this disclosure have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and intent of this disclosure. It should be understood that various alternatives to the embodiments of this disclosure described herein may be employed in the practice of this disclosure. The appended claims are intended to define the scope of this disclosure and therefore cover equivalents or alternatives within the scope of these claims.

Claims

1. A method for training a hemodynamic parameter prediction model, comprising: Based on the acquired cerebral vascular imaging data, a vascular topology map structure representing the geometry and connectivity of blood vessels is constructed to obtain a training sample set; wherein, the training sample includes vascular features and corresponding hemodynamic label data; The vascular features in the training samples are input into the hemodynamic parameter prediction model to be trained to obtain the predicted hemodynamic parameters for each vascular. Based on the predicted hemodynamic parameters, the comprehensive loss function value is calculated. The comprehensive loss function is obtained by weighted summation of the supervision loss term and the physical consistency loss term. Based on the comprehensive loss function value, the parameters of the prediction model are adjusted until the preset convergence condition is met, thus obtaining a trained hemodynamic parameter prediction model.

2. The method according to claim 1, characterized in that, The hemodynamic tag data includes at least one of wall shear stress, oscillatory shear index, pressure field, and velocity field; the predicted hemodynamic parameters correspond to at least one of predicted wall shear stress, oscillatory shear index, pressure field, and velocity field.

3. The method according to claim 1, characterized in that, The hemodynamic tag data is generated through computational fluid dynamics simulation, specifically including setting inlet flow velocity boundary conditions and local viscosity parameters, and performing steady-state or quasi-steady-state solutions.

4. The method according to claim 1, characterized in that, The physical consistency loss term includes at least one of the following: The continuity equation constraint terms are used to constrain the incompressibility condition of the fluid. The momentum conservation constraint term is used to ensure that the prediction results satisfy the fluid momentum conservation law. The energy loss constraint term is used to suppress numerical oscillations in the prediction results in local regions.

5. The method according to claim 1, characterized in that, The preset convergence conditions include: the value of the comprehensive loss function is less than a first threshold, and each item in the physical consistency loss term is less than its corresponding second threshold.

6. The method according to claim 1, characterized in that, The hemodynamic parameter prediction model includes: A feature encoding network is used to encode the node features and edge features of the blood vessel topology structure to obtain node embedding features and edge embedding features respectively; A graph neural network is used to aggregate and update information on a vascular topology map based on the node embedding features and the edge embedding features, and output the predicted hemodynamic parameters.

7. The method according to claim 1, characterized in that, The step of constructing a vascular topology structure representing the geometry and connectivity of blood vessels based on the acquired cerebral vascular imaging data specifically includes: The vascular regions of the cerebral vascular imaging data are extracted using a three-dimensional segmentation network or a texture and morphology-based method to obtain a vascular voxel mask. Based on the vascular voxel mask, a triangular mesh model of the vascular surface is reconstructed and a set of vascular centerlines is extracted. The blood vessel is divided into multiple blood vessel segments, and each blood vessel segment corresponds to a tubular structure defined by a centerline; The blood vessel topology is constructed using the blood vessel segments as nodes and the connections between the blood vessel segments as edges.

8. The method according to claim 7, further comprising: Local geometric features are calculated for each blood vessel segment. These local geometric features include at least one of radius, rate of change of diameter, curvature, torsion, and relative distance to bifurcation, which serve as feature data for the blood vessel node. For each edge, a connectivity feature is calculated, which includes at least one of the following: bifurcation angle, transition length, and curvature gradient, as an edge feature.

9. The method according to claim 1, characterized in that, The hemodynamic tag data includes hemodynamic parameters at multiple sampling points distributed axially along the vessel centerline; the predicted hemodynamic parameters correspond to predicted hemodynamic parameters at multiple sampling points distributed axially along the vessel centerline.

10. The method according to claim 9, characterized in that, The hemodynamic tag data also includes hemodynamic parameters distributed circumferentially in multiple directions within the cross-sectional plane of each sampling point; and the predicted hemodynamic parameters also include predicted hemodynamic parameters distributed circumferentially in multiple directions within the cross-sectional plane of each sampling point.

11. A method for predicting hemodynamic parameters, comprising: The imaging data of the cerebral blood vessels to be tested are processed to construct a topological map structure of the blood vessels to be tested, representing the geometry and connectivity of the blood vessels. The topology of the blood vessel to be tested is input into the hemodynamic parameter prediction model trained by the method described in any one of claims 1-10 to obtain the predicted hemodynamic parameters.

12. The method according to claim 11, characterized in that, The process of processing the cerebral vascular imaging data to be tested and constructing a topological map structure representing the geometry and connectivity of the blood vessels specifically includes: Three-dimensional vessel segmentation is performed on the cerebral vascular image data to be tested to obtain a voxel mask of the vessel to be tested. Based on the voxel mask of the blood vessel to be tested, the triangular mesh model of the surface of the blood vessel to be tested is reconstructed and the set of center lines of the blood vessel to be tested is extracted. The centerline of the blood vessel to be tested is divided into multiple segments of the blood vessel to be tested, and each segment of the blood vessel to be tested corresponds to a tubular structure defined by the centerline; Using the segments of the blood vessel to be tested as nodes and the connections between the segments as edges, a topological graph of the blood vessel to be tested is constructed.

13. The method according to claim 12, characterized in that, Also includes: For each segment of the blood vessel to be tested, local geometric features are calculated. These local geometric features include at least one of radius, rate of change of diameter, curvature, torsion, and relative distance to bifurcation, which serve as node features of the topological structure of the blood vessel to be tested. as well as For each edge, a connectivity feature is calculated, which includes at least one of bifurcation angle, transition length, and curvature gradient, as an edge feature of the blood vessel topology structure to be tested.

14. The method according to claim 11, characterized in that, The predicted hemodynamic parameters include those from multiple sampling points distributed axially along the centerline of the vessel to be tested.

15. The method according to claim 14, characterized in that, The predicted hemodynamic parameters also include predicted hemodynamic parameters distributed circumferentially in multiple directions within the cross-sectional plane of each sampling point along the centerline of the vessel to be tested.

16. The method according to claim 11, characterized in that, Also includes: The predicted hemodynamic parameters are mapped back to the original three-dimensional vascular space to obtain a parameter field distribution consistent with the three-dimensional vascular geometry.

17. The method according to claim 16, characterized in that, The process of mapping the predicted hemodynamic parameters back to the original three-dimensional vascular space specifically includes: Based on the pre-established geometric mapping relationship between vascular segments and centerlines and the three-dimensional vascular regions they cover, the predicted hemodynamic parameters of the vascular segments to be tested are assigned to discrete sampling points on the corresponding centerlines. Interpolation processing is performed on the sampling points in the boundary region between adjacent segments to achieve a smooth numerical transition; Centered on the centerline sampling point, the parameters are extended radially along the normal section to the local section region; For the triangular facets in the grid on the surface of the blood vessel to be tested, the parameters are backfilled based on the nearest centerline sampling point and the segment of the blood vessel to be tested to which they belong.

18. The method according to claim 17, characterized in that, When the predicted hemodynamic parameters include prediction results distributed circumferentially in multiple directions within the cross-sectional plane of each sampling point, the step of mapping the predicted hemodynamic parameters back to the original three-dimensional vascular space further includes: For a target surface point in a three-dimensional blood vessel surface mesh, determine its circumferential angle in the local cross-sectional plane of the corresponding centerline sampling point; Based on the circumferential angle, the prediction results of adjacent circumferential sampling directions are interpolated to obtain the hemodynamic parameters of the target surface point.

19. An apparatus for training a hemodynamic parameter prediction model, characterized in that, include: The graph structure construction module is used to construct a vascular topology graph structure representing the geometry and connectivity of blood vessels based on the acquired cerebral vascular imaging data, so as to obtain a training sample set; wherein, the training samples include vascular features and corresponding hemodynamic label data; The model prediction module is used to input the vascular features in the training samples into the hemodynamic parameter prediction model to be trained, and obtain the predicted hemodynamic parameters for each vascular. The loss calculation module is used to calculate the comprehensive loss function value based on the predicted hemodynamic parameters. The comprehensive loss function is obtained by weighted summation of the supervision loss term and the physical consistency loss term. The parameter adjustment module is used to adjust the parameters of the prediction model according to the comprehensive loss function value until the preset convergence condition is met, so as to obtain the trained hemodynamic parameter prediction model. The device is used to implement the method according to any one of claims 1-10.

20. A hemodynamic parameter prediction device, characterized in that, include: The data processing module is used to process the cerebral vascular imaging data to be tested and construct a topological map structure of the blood vessels to be tested, representing the geometry and connectivity of the blood vessels. The model inference module is used to input the topological structure of the blood vessel to be tested into the hemodynamic parameter prediction model trained using the method described in any one of claims 1-10, and to obtain the predicted hemodynamic parameters. The device is used to implement the method according to any one of claims 11-18.

21. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method of any one of claims 1-10, or the method of any one of claims 11-18.

22. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-10, or implement the method of any one of claims 11-18.

Images