Preoperative risk prediction method for cerebral aneurysm based on multi-modal deep learning
By combining multimodal deep learning methods with vascular imaging and hemodynamic simulation, the problems of clinical timeliness and physical consistency in preoperative risk assessment of cerebral aneurysms were solved, and rapid and accurate risk prediction and assessment were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JINHUA MUNICIPAL CENT HOSPITAL
- Filing Date
- 2026-02-09
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies for preoperative risk assessment of cerebral aneurysms suffer from insufficient clinical timeliness, physical consistency, and multimodal fusion, making it difficult to provide rapid and accurate individualized hemodynamic assessments and risk predictions.
By employing a multimodal deep learning approach, combining computed tomography angiography and magnetic resonance angiography to reconstruct vascular geometry and tree diagrams, and using Kirchhoff conservation graph networks for cardiac phase and differentiable Wendkesell learning, physical constraint neural operators are invoked to perform hemodynamic simulations. Helmholtz projection and signed distance functions are used for correction to calculate wall shear stress, oscillatory shear exponent, and pressure gradient. Finally, the results are fused with expert product data from imaging, clinical, and textual information to output preoperative risk.
It significantly improves the computational efficiency and accuracy of preoperative risk assessment for cerebral aneurysms, meets clinical timeliness requirements, provides individualized and physically consistent hemodynamic assessment, and enhances the robustness and interpretability of risk prediction.
Smart Images
Figure CN122135967A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image analysis, and more particularly to a method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning. Background Technology
[0002] Risk assessment for rupture of cerebral aneurysms is crucial for interventional or surgical treatment decisions. Existing research indicates that hemodynamic parameters such as wall shear stress, oscillatory shear index, and pressure gradient are associated with rupture. The mainstream approach to obtaining individualized hemodynamics is computational fluid dynamics simulation based on reconstructed geometry from patient images, applying physiological boundary conditions at the inlet and terminal branches, such as the Wendkessel model characterized by proximal resistance, capacitive and distal resistance. On the other hand, four-dimensional fluid magnetic resonance imaging (FMRI) can directly measure the velocity field, but its clinical application is limited by spatial resolution, acquisition time, and equipment accessibility. Meanwhile, radiomics and deep learning methods have been used to extract morphological and textural features from computed tomography (CT) and magnetic resonance angiography (MRI) and to jointly model these features with clinical data and textual reports. In recent years, physically constrained neural networks, neural operators, and graphical models have been used for flow field approximation and network vascular system modeling, showing potential to replace some simulations.
[0003] However, existing technologies still have the following shortcomings:
[0004] 1. Insufficient clinical timeliness and availability of boundary conditions: Traditional simulation relies on fine meshes and solvers, which are complex and computationally time-consuming, making it difficult to meet the preoperative time window. Patient-specific inlet flow and impedance of each terminal usually lack direct measurement and are often set based on experience or simplification, which makes the results sensitive to boundary assumptions.
[0005] 2. Insufficient physical consistency and topological constraints: Data-driven alternative models focus more on regression accuracy and are difficult to strictly satisfy constraints such as incompressibility and solid wall no slip. They lack global consistency constraints based on conservation laws on vascular trees with bifurcations, which can easily lead to branch-level mass conservation deviations, thus affecting the reliability of wall shear stress, oscillatory shear exponent and pressure gradient.
[0006] 3. Insufficient multimodal fusion and uncertainty representation: Existing risk assessments often use empirical weighting or simple early and late fusion, which makes it difficult to robustly integrate imaging features, clinical elements and textual information. They often fail to explicitly include uncertainty and consistency correction for hemodynamic estimation, resulting in insufficient stability and interpretability of risk output.
[0007] Therefore, a method for predicting preoperative risks of cerebral aneurysms that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention
[0008] One objective of this invention is to propose a preoperative risk prediction method for cerebral aneurysms based on multimodal deep learning. Addressing the shortcomings of existing technologies that rely on time-consuming computational fluid dynamics simulations and struggle to obtain individualized boundary conditions, physical consistency, and multimodal fusion, this invention proposes a method to reconstruct vascular geometry and tree diagrams on aligned computed tomography (CT) and magnetic resonance angiography (MRI) images. It combines a Kirchhoff conservation graph network based on cardiac phase with differentiable Wendell-Kessel learning of entry and terminal impedances. By invoking physically constrained neural operators and using Helmholtz projection and signed distance functions to achieve incompressibility and no-slip, the method calculates and corrects wall shear stress, oscillatory shear exponent, and pressure gradient using branch-closed-loop consistency residuals. Finally, it fuses these with image, clinical, and text branches via expert product to output lesion-level and patient-level preoperative risks. This invention offers the technical advantages of rapidly and accurately estimating hemodynamics under physically consistent constraints, providing interpretable risk assessments, and significantly improving clinical timeliness and robustness.
[0009] A method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning according to an embodiment of the present invention includes:
[0010] S1. Obtain the patient's CTA, MRA, blood pressure and heart rate time series, structured clinical data and text reports, and perform time alignment and standardization;
[0011] S2. Perform vessel segmentation and 3D reconstruction on aligned CTA and MRA to obtain a vessel geometric model. Calculate the signed distance function of the vessel wall, vessel centerline, vessel tree diagram, radius distribution, curvature, and bifurcation angle from this model.
[0012] S3. Generate cardiac phase codes based on aligned and denoised blood pressure and heart rate time series, apply Kirchhoff conservation graph neural network on the vascular tree diagram, and attach differentiable Wind-Kessel impedance models to leaf nodes to estimate inlet flow and outlet impedance and Wind-Kessel impedance parameters of each terminal branch.
[0013] S4. Using the inlet flow rate and outlet impedance as boundary conditions and cardiac phase encoding as time conditions, the flow field is approximated by calling the physical constraint neural operator based on the vascular geometry model and combined with the signed distance function to obtain the velocity field and pressure field. Helmholtz projection is performed on the velocity field to satisfy the divergence-free constraint, and boundary embedding is performed based on the signed distance function to satisfy the slip-free constraint.
[0014] S5. Based on the velocity field and pressure field and combined with the signed distance function, calculate the wall shear stress, oscillatory shear index and pressure gradient distribution on the blood vessel wall;
[0015] S6. Based on the vascular tree diagram and centerline, the velocity field is integrated at the cross-section of each branch to obtain the branch flow vector. The branch flow vector is compared with the inlet flow and outlet impedance to obtain the closed-loop consistency residual. Based on this, the wall shear stress, oscillatory shear index and pressure gradient distribution are corrected for consistency.
[0016] S7. Combine the corrected wall shear stress, oscillatory shear index, and pressure gradient distribution with aligned CTA and MRA, standardized structured clinical data, and standardized text reports. Extract risk features through imaging, clinical, and text branches and obtain the risk distribution and confidence level of each branch. Use expert product fusion to obtain the lesion-level and patient-level preoperative risk values and corresponding confidence levels.
[0017] Optionally, step S1 includes:
[0018] For the same patient, computed tomography angiography and magnetic resonance angiography images are time-aligned according to the examination time, and the spatial orientation and voxel spacing of the images are standardized to output time-aligned computed tomography angiography images and time-aligned magnetic resonance angiography images.
[0019] The blood pressure and heart rate time series are resampled at fixed sampling intervals, outlier removal and filtering are performed, and the start and end points are aligned with the time reference of the image to output the time-aligned and denoised blood pressure and heart rate time series.
[0020] The system unifies the execution units, standardizes fields, and normalizes terms for structured clinical data, outputting standardized structured clinical data.
[0021] Perform text cleaning, medical terminology standardization, and time stamping on the text report to output a standardized text report.
[0022] Optionally, step S2 includes:
[0023] Perform vascular segmentation and spatial reconstruction on time-aligned computed tomography angiography and time-aligned magnetic resonance angiography to form a vascular geometry model of the patient and output the vascular geometry model.
[0024] The signed distance function of the blood vessel wall is calculated based on the blood vessel geometry model and used to characterize the position of the blood vessel wall in the subsequent boundary embedding. The signed distance function of the blood vessel wall is output.
[0025] The vessel centerline is extracted based on the vascular geometric model, and a vascular tree diagram is constructed based on the bifurcation relationship of the vessel centerline. This is used to represent the vascular topology in the subsequent graph structure boundary condition learning, and the vascular centerline and vascular tree diagram are output.
[0026] The radius distribution, curvature, and bifurcation angle are calculated on the vascular geometry model to characterize the geometric parameters in subsequent boundary condition estimation and flow field calculation, and the radius distribution, curvature, and bifurcation angle are output.
[0027] Terminology definition:
[0028] The vascular segmentation is a process that distinguishes vascular lumens from non-vascular tissues on time-aligned CTA and MRA and generates binary or probabilistic segmentation results.
[0029] The three-dimensional reconstruction (spatial reconstruction) is the process of forming a three-dimensional geometric representation of the vascular cavity based on the vascular segmentation results, including but not limited to surface mesh, voxel volume or implicit function representation;
[0030] The vascular geometric model is a three-dimensional model that characterizes the shape and topological relationship of the patient's vascular lumen, and is used for subsequent geometric calculations and flow field approximation;
[0031] The vessel wall is the interface surface between the vessel lumen and the surrounding tissue, and is the boundary of the vessel geometry model;
[0032] The signed distance function is a scalar function defined in the spatial domain. Its value is the minimum distance from a spatial point to the vessel wall and it carries a sign that distinguishes between inside and outside the lumen. Preferably, the signs of inside and outside the lumen are opposite.
[0033] The vessel centerline is a continuous curve connecting the inlet to each terminal along the geometric center of the vessel lumen, and can be extracted by methods such as skeletonization, distance field, or minimum cost path.
[0034] The branching relationship is the connection and hierarchy between the parent branch and the child branch recorded at the centerline level;
[0035] The vascular tree diagram is a tree structure with branching points as nodes and vascular segments as edges, used to represent vascular topology, and can be directed or undirected.
[0036] The vascular topology is a collection of the connectivity structure between vascular branches and the hierarchical organization method;
[0037] The radius distribution is a local cavity radius function that varies along the centerline position. The local radius can be obtained through the equivalent circle radius (converted from the local cross-sectional area A to sqrt(A / π)), the inscribed sphere radius, or other geometric estimations.
[0038] The curvature is the geometric curvature of the centerline in three-dimensional space, defined as the norm of the rate of change of the unit tangent vector with respect to the arc length;
[0039] The bifurcation angle is the angle between two branches at the bifurcation point determined by the tangential vector of the centerline, or the angle between the parent branch and the child branch, calculated using three-dimensional spatial vector angles.
[0040] The boundary embedding is a method of expressing and applying geometric boundary conditions in a continuous domain using a signed distance function, which is used to encode conditions such as no-slip in the subsequent physical constraint model.
[0041] Optionally, step S3 includes:
[0042] The time-aligned and denoised blood pressure and heart rate time series are used to calculate and encode the cardiac cycle phase function according to a unified time reference, forming cardiac phase coding.
[0043] Message passing that satisfies Kirchhoff conservation constraints is performed on the vascular tree diagram using a graph neural network, and differentiable Wendkesell impedance models are attached to the leaf nodes of the vascular tree diagram to obtain the proximal resistance parameters, capacitive parameters and distal resistance parameters of each terminal branch.
[0044] The inlet flow rate is estimated based on the cardiac phase coding of the vascular centerline, radius distribution, curvature and bifurcation angle, and the outlet impedance of each terminal branch is obtained by the Wendkessel impedance model.
[0045] Output cardiac phase encoding, inlet flow rate, outlet impedance, and Wind-Kessel impedance parameters.
[0046] Terminology definition:
[0047] The unified time reference is a reference time axis used for synchronizing cross-modal data and physiological signals, serving as the time zero point and scale basis for cardiac phase calculation and conditional modeling;
[0048] The cardiac cycle phase function is a function that maps time to a normalized phase that characterizes the position of the cardiac cycle. Its value range can be [0,1) or [0,2π), and it is estimated from signals such as blood pressure, heart rate or electrocardiogram.
[0049] The cardiac phase encoding is a numerical representation of the cardiac cycle phase function, which serves as a temporal condition variable of the model and can be in the form of a scalar, a pair of sine and cosine functions, or a multi-fundamental frequency vector.
[0050] The graph neural network is a neural network that performs feature propagation and aggregation on nodes and edges of a vascular tree graph, and is used to estimate boundary condition related quantities by combining topological and geometric information.
[0051] The Kirchhoff conservation constraints are constraints that satisfy flow conservation and / or pressure continuity at bifurcation nodes, and are used to reflect the conservation laws in vascular networks.
[0052] The message passing is a feature calculation, aggregation, and update process performed along the graph structure in a graph neural network, used to exchange information between adjacent nodes / edges;
[0053] The leaf node is the terminal node without child branches in the vascular tree diagram, corresponding to the position where the exit boundary condition is applied;
[0054] The terminal branch is a vascular segment that connects to the leaf node and leads to the peripheral circulation; it is the object that defines the outlet impedance and impedance parameters.
[0055] The Wind-Kessel impedance model is an equivalent RCR loop model composed of proximal resistance, capacitive and distal resistance, used to describe the pressure-flow relationship of the terminal vascular bed.
[0056] The differentiable Wendell-Kessel impedance model is a differentiable parameterized implementation of the above RCR model, which enables the parameters to be learned through gradient optimization during end-to-end training.
[0057] The proximal resistance parameter is the resistance element parameter located at the proximal end in the Wendkesel model, used to characterize the magnitude of the proximal resistance.
[0058] The capacitive parameters are the capacitance element parameters in the Wendkesel model, used to characterize the compliance of the vascular bed;
[0059] The distal resistance parameter is the resistance element parameter located at the distal end in the Wendkessel model, used to characterize the magnitude of peripheral resistance.
[0060] The inlet flow rate is the volumetric flow rate boundary condition of the inlet section of the blood vessel model, which varies with the cardiac phase.
[0061] The outlet impedance is the equivalent impedance relationship between pressure and flow rate at the terminal branch, which can be determined by the Wendkesell model.
[0062] The Wind-Kessel impedance parameters are the set of parameters of the Wind-Kessel model, including at least the near-end resistance parameter, capacitive parameter, and far-end resistance parameter.
[0063] Optionally, step S4 includes:
[0064] Based on the inlet flow rate and outlet impedance, and combined with the cardiac phase encoding, boundary conditions that vary with the cardiac phase are constructed at the inlet section and each terminal branch for approximate calculation of the driving flow field;
[0065] Within the flow domain defined by the vascular geometry model, physical constraint neural operators are invoked to generate velocity and pressure fields according to the cardiac phase.
[0066] Perform a Helmholtz projection on the generated velocity field to project the velocity field onto a divergence-free subspace to achieve divergence-free hard constraints;
[0067] Boundary embedding is performed using the signed distance function of the blood vessel wall, so that the velocity field satisfies the no-slip boundary condition at the blood vessel wall;
[0068] Output velocity field and pressure field.
[0069] Terminology definition:
[0070] The inlet section is the section of the blood vessel geometry model used at the inlet boundary to apply inlet boundary conditions and define the inlet flow integral;
[0071] The boundary conditions are a set of conditions that restrict physical quantities such as pressure, flow rate, or velocity at the inlet, outlet, and solid wall, and are used to determine the solution space of the flow problem.
[0072] The flow domain is the fluid-accessible space region defined by the vascular geometry model, serving as the region for defining and calculating the velocity and pressure fields;
[0073] The physical constraint neural operator is a model that explicitly or implicitly encodes fluid physical constraints (including incompressibility and boundary conditions) in a neural operator that maps input conditions to flow field solutions.
[0074] The velocity field is a vector-valued function defined over the flow domain and time / cardiographic phase, used to characterize the fluid velocity at each spatial point;
[0075] The pressure field is a scalar function defined over the flow domain and time / cardiac phase, used to characterize the pressure at each spatial point;
[0076] The Helmholtz projection is an operation that decomposes an arbitrary vector field into a divergence-free part and a gradient field part while retaining the divergence-free part, and is used to make the velocity field incompressible.
[0077] The divergence-free subspace is a set of vector fields with zero divergence, corresponding to the feasible velocity field space of an incompressible fluid.
[0078] The divergence-free hard constraint is a constraint that makes the velocity field divergence strictly zero by means of projection or equation, realizing a soft constraint approximation without loss penalty.
[0079] The no-slip boundary condition is a boundary constraint in which the fluid velocity relative to the wall is zero at the vessel wall, used to characterize the adhesion effect between the solid wall and the fluid.
[0080] Optionally, step S5 includes:
[0081] Based on the velocity and pressure fields, the position of the blood vessel wall is determined by the signed distance function of the blood vessel wall, and the wall normal and tangential directions are extracted for use in calculating hemodynamic parameters on the blood vessel wall.
[0082] By calculating the tangential gradient of the velocity field at the vessel wall and converting it according to a preset fluid viscosity, the wall shear stress distribution is formed;
[0083] The wall shear stress is integrated over time and normalized along the cardiac cycle to form an oscillatory shear index distribution.
[0084] The spatial gradient of the pressure field at the blood vessel wall is calculated and projected along the blood vessel wall to form a pressure gradient distribution.
[0085] Output the wall shear stress distribution, oscillatory shear index distribution, and pressure gradient distribution.
[0086] Terminology definition:
[0087] The wall normal is a unit normal vector at the vessel wall determined by the gradient of a signed distance function, used to distinguish between inside and outside the lumen and to define the tangential plane;
[0088] The wall tangential direction is a unit direction orthogonal to the wall normal, or any direction within the tangential plane spanned by it, used to extract the tangential component and tangential gradient;
[0089] The tangential gradient of the velocity field is the gradient of the tangential component of the velocity field along the normal to the wall at the vessel wall, and is used to characterize the shear rate.
[0090] The preset fluid viscosity is a viscosity parameter used to convert shear rate into shear stress, preferably a constant of blood equivalent viscosity or its model equivalent value;
[0091] The wall shear stress is the shear stress acting on the tangential direction of the blood vessel wall, calculated based on the preset fluid viscosity and the tangential gradient of the velocity field.
[0092] The wall shear stress distribution refers to the spatial distribution of wall shear stress on the blood vessel wall and its variation with cardiac phase or time.
[0093] The oscillatory shear index is a dimensionless index that measures the degree of directional oscillation of wall shear stress during the cardiac cycle, and its value ranges from 0 to 0.5.
[0094] The oscillatory shear index distribution is the spatial distribution of the oscillatory shear index on the blood vessel wall;
[0095] The pressure gradient is the spatial gradient vector of the pressure field, used to characterize the rate of change of pressure with respect to spatial location;
[0096] The projection along the vessel wall is an operation that takes the components of the vector on the tangential plane of the wall, used to obtain quantities related to the tangential direction of the vessel wall;
[0097] The pressure gradient distribution refers to the spatial distribution of the pressure gradient projected along the blood vessel wall at the blood vessel wall and its variation with cardiac phase or time.
[0098] Optionally, step S6 includes:
[0099] Based on the vascular tree diagram, each branch is determined and the branch cross-section is constructed with the tangent of the vascular centerline. The velocity field is integraled with respect to the cardiac phase on the branch cross-section to form the branch flow vector.
[0100] The branch flow vector is compared with the inlet flow for mass conservation, and the reference branch flow based on impedance is calculated by combining the outlet impedance. The mass conservation difference and the reference branch flow difference are combined into a closed-loop consistency residual according to the preset weight.
[0101] Based on the closed-loop consistency residual, the wall shear stress distribution, oscillatory shear index distribution, and pressure gradient distribution are corrected for consistency. The distributions located in branches with significant residuals and their neighborhoods are adjusted proportionally to meet the mass conservation and impedance consistency constraints. The output is the branch flow vector, the closed-loop consistency residual, the corrected wall shear stress distribution, the corrected oscillatory shear index distribution, and the corrected pressure gradient distribution.
[0102] Terminology definition:
[0103] The branch is a vascular segment located between two bifurcation points or between the inlet / terminus and the bifurcation point, based on the vascular tree diagram and the vascular centerline.
[0104] The tangent of the centerline is the unit tangent vector of the blood vessel centerline at a given position, used to determine the normal of the branch cross section;
[0105] The branch cross section is a cross-sectional plane passing through a specified position of the centerline and whose normal direction coincides with the tangent of the centerline at that position, used to define the integration domain of the branch flow.
[0106] The cross-sectional integral is calculated by integrating the normal components of the velocity field over the cross-section of the branch to obtain the volumetric flow rate.
[0107] The branch flow vector is a set of volumetric flow values calculated according to the cardiac phase at each branch cross section, representing the flow in the branch and phase dimensions in vector form;
[0108] The mass conservation comparison involves performing a conservation consistency test and difference measurement on the inlet flow rate and the total flow rate synthesized from the flow rates of each branch.
[0109] The impedance-based reference branch flow is the reference allocation of each branch flow calculated based on the outlet impedance and inlet conditions of each terminal branch.
[0110] The mass conservation difference is a measure of the deviation between the combined inlet flow and branch flow values under the meaning of mass conservation.
[0111] The reference branch flow difference is a measure of the deviation between the actual branch flow and the impedance-based reference branch flow.
[0112] The preset weight is a non-negative weight parameter used to weight the mass conservation difference and the reference branch flow difference;
[0113] The closed-loop consistency residual is a comprehensive residual synthesized according to a preset weight, which is used to characterize the degree of deviation of the current solution from mass conservation and impedance consistency.
[0114] The consistency correction is a quantitative adjustment of the distribution of hemodynamic parameters based on the closed-loop consistency residual, in order to improve the degree of satisfaction of mass conservation and impedance consistency.
[0115] The impedance consistency constraint is a constraint condition that ensures the flow-pressure relationship at the terminal branch is consistent with the outlet impedance model.
[0116] The branch neighborhood is a local region centered on the target branch, determined by the tree diagram adjacency relationship and / or by the geometric distance from the center line, used to define the scope of influence of consistency correction.
[0117] Optionally, step S7 includes:
[0118] In the imaging branch, time-aligned computed tomography angiography images and time-aligned magnetic resonance angiography images are combined, and the corrected wall shear stress distribution, the corrected oscillatory shear index distribution and the corrected pressure gradient distribution are used as auxiliary quantities for the imaging risk characteristics to form the imaging risk characteristics and obtain the risk distribution and confidence of the imaging branch.
[0119] In the clinical branch, clinical risk characteristics are formed based on standardized structured clinical data, and the risk distribution and confidence level of the clinical branch are obtained;
[0120] Within the text branch, text risk characteristics are formed based on standardized text reports, and the risk distribution and confidence level of the text branch are obtained.
[0121] The risk distributions of the imaging branch, clinical branch and text branch are multiplied and normalized by expert product fusion to form a fused risk distribution. The lesion-level preoperative risk value is read from the fused risk distribution, and the patient-level preoperative risk value is generated according to the set of lesions contained in the patient, while the corresponding confidence level is output.
[0122] Terminology definition:
[0123] The image branch is a model subnetwork or processing channel used to extract preoperative risk-related features from time-aligned CTA and MRA and their correlation quantities, and output the risk distribution and confidence of the branch.
[0124] The clinical branch is a model subnetwork or processing channel used to extract preoperative risk-related features from standardized structured clinical data and output the risk distribution and confidence level of the branch.
[0125] The text branch is a model subnetwork or processing channel used to extract features related to preoperative risks from standardized text reports and output the risk distribution and confidence level of the branch.
[0126] The imaging risk features are quantitative representations of preoperative risk estimation extracted from CTA, MRA and their correlation values by imaging branches.
[0127] The clinical risk characteristics are quantitative or categorical representations of preoperative risk estimation extracted from structured clinical data by clinical branches.
[0128] The text risk features are semantic or statistical representations extracted from the text report by the text branch for preoperative risk estimation.
[0129] The auxiliary quantities of the image risk characteristics are quantitative representations of the corrected wall shear stress distribution, oscillatory shear index distribution, and pressure gradient distribution used as additional inputs for image branching.
[0130] The risk distribution is a probability distribution representation of a predefined preoperative risk target (including probability, grade, or category), used to quantify the likelihood of different risk levels;
[0131] The confidence level is a measure of the reliability of the risk distribution, obtained based on the distribution concentration, calibration, or uncertainty estimation, and is used to characterize the degree of certainty of the output result.
[0132] The expert product fusion method is a fusion method that multiplies the risk distributions of the imaging branch, clinical branch, and text branch point by point and then normalizes them to obtain a joint distribution.
[0133] The fusion risk distribution is a joint risk distribution obtained by expert product fusion, which is used to subsequently read the preoperative risk values at the lesion level and the patient level.
[0134] The lesion is a brain aneurysm lesion unit identified in the patient's vascular system, which is the object of lesion-level risk assessment;
[0135] The lesion-level preoperative risk value is a preoperative risk measure read from the fusion risk distribution for a single lesion, and is expressed as a probability value or a risk score converted by a threshold.
[0136] The lesion set refers to the collection of all identified or labeled cerebral aneurysm lesions in the same patient.
[0137] The patient-level preoperative risk value is a patient-level risk measure obtained by using a convergence function (including but not limited to maximum value, weighted average, or logical synthesis) based on the lesion-level preoperative risk value of each lesion in the lesion set.
[0138] The beneficial effects of this invention are:
[0139] 1. Improved computational efficiency and clinical timeliness: Physically constrained neural operators replace traditional computational fluid dynamics solutions. Combined with cardiac phase conditionalization, velocity and pressure fields are rapidly generated under hard constraints of Helmholtz projection and signed distance function boundary embedding. This significantly shortens the time from image to hemodynamic parameters, meeting the time requirements for preoperative assessment.
[0140] 2. Individualized and physically consistent boundary conditions and flow fields: A graph neural network that satisfies Kirchhoff conservation is used on the blood vessel tree diagram, and a differentiable Wendt-Kessel impedance is attached to the leaf nodes. The inlet flow rate and the outlet impedance of each terminal are estimated by combining cardiac phase estimation. With incompressible and no-slip hard constraints, the accuracy and stability of wall shear stress, oscillatory shear index and pressure gradient are improved.
[0141] 3. Improved consistency correction and reliability of risk output: The branch flow vector is formed by integrating the velocity field with the centerline and branch sections. It is compared with the inlet flow and the reference branch flow based on impedance to obtain the closed-loop consistency residual. Based on this, the hemodynamic parameters are corrected at the branch level. The corrected parameters are then fused with the imaging branch, clinical branch and text branch through expert product to output lesion-level and patient-level risks and confidence levels, thereby improving the robustness and interpretability of risk assessment. Attached Figure Description
[0142] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0143] Figure 1 This is a flowchart of a preoperative risk prediction method for cerebral aneurysms based on multimodal deep learning proposed in this invention. Detailed Implementation
[0144] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0145] refer to Figure 1 A method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning, comprising:
[0146] S1. Obtain the patient's CTA, MRA, blood pressure and heart rate time series, structured clinical data and text reports, and perform time alignment and standardization;
[0147] S2. Perform vessel segmentation and 3D reconstruction on aligned CTA and MRA to obtain a vessel geometric model. Calculate the signed distance function of the vessel wall, vessel centerline, vessel tree diagram, radius distribution, curvature, and bifurcation angle from this model.
[0148] S3. Generate cardiac phase codes based on aligned and denoised blood pressure and heart rate time series, apply Kirchhoff conservation graph neural network on the vascular tree diagram, and attach differentiable Wind-Kessel impedance models to leaf nodes to estimate inlet flow and outlet impedance and Wind-Kessel impedance parameters of each terminal branch.
[0149] S4. Using the inlet flow rate and outlet impedance as boundary conditions and cardiac phase encoding as time conditions, the flow field is approximated by calling the physical constraint neural operator based on the vascular geometry model and combined with the signed distance function to obtain the velocity field and pressure field. Helmholtz projection is performed on the velocity field to satisfy the divergence-free constraint, and boundary embedding is performed based on the signed distance function to satisfy the slip-free constraint.
[0150] S5. Based on the velocity field and pressure field and combined with the signed distance function, calculate the wall shear stress, oscillatory shear index and pressure gradient distribution on the blood vessel wall;
[0151] S6. Based on the vascular tree diagram and centerline, the velocity field is integrated at the cross-section of each branch to obtain the branch flow vector. The branch flow vector is compared with the inlet flow and outlet impedance to obtain the closed-loop consistency residual. Based on this, the wall shear stress, oscillatory shear index and pressure gradient distribution are corrected for consistency.
[0152] S7. Combine the corrected wall shear stress, oscillatory shear index, and pressure gradient distribution with aligned CTA and MRA, standardized structured clinical data, and standardized text reports. Extract risk features through imaging, clinical, and text branches and obtain the risk distribution and confidence level of each branch. Use expert product fusion to obtain the lesion-level and patient-level preoperative risk values and corresponding confidence levels.
[0153] In this specific embodiment, S1 includes:
[0154] For the same patient, computed tomography (CT) angiography (CT) images and magnetic resonance angiography (MRI) images are time-aligned according to the examination time, with the start time of the CT angiography examination serving as the zero point of a unified time reference. And establish a unified time benchmark Meanwhile, the spatial orientation and voxel spacing of the two types of images are standardized: the orientation information is read and the volume data is rearranged to the standard coordinate system, and the voxel spacing is resampled to the preset equidistant value using cubic splines or linear interpolation to ensure the consistency of subsequent geometric reconstruction.
[0155] Next, the blood pressure and heart rate time series are resampled at fixed sampling intervals, outlier removal, and filtering denoising based on a unified time reference. Specifically, a fixed sampling step size is set on the unified time reference. An equally spaced sampling grid is formed, and outliers based on the median absolute deviation are removed. Low-pass or band-pass filtering is then applied to suppress high-frequency noise. The start and end points of the time series are then compared with the image's temporal information. and Alignment;
[0156] To facilitate cross-modal synchronization and support subsequent cardiac phase calculations, the raw times of each modality or signal are... Mapped to a unified time base :
[0157] ;
[0158] in, Represents dimensionless time on a unified time base, with a range of values of 1000. The timestamps on the original timeline can be derived from the time of imaging examination or the time of physiological signal sampling. The starting time of the unified time reference is indicated, preferably the start time of the computed tomography angiography image examination; Indicates the alignment window length, used to map times from different sources to... The scale is preferably determined by the end time of the alignment window. relative to the starting time The difference is determined. ;
[0159] After completing time alignment, the structured clinical data is processed to unify units, standardize fields, and normalize terms. Fields involving units such as blood pressure are unified to international units, and the enumerated fields are mapped to a controlled thesaurus to generate standardized structured clinical data.
[0160] Simultaneously, the text report is cleaned, medical terminology is standardized and time-annotated, redundant symbols are removed and synonyms are mapped to a unified medical glossary, and the event times in the report are anchored according to a unified time benchmark.
[0161] Outputs time-aligned computed tomography angiography and time-aligned magnetic resonance angiography, time-aligned and denoised blood pressure and heart rate time series, standardized structured clinical data, and standardized text reports.
[0162] In this specific embodiment, S2 includes:
[0163] Vascular segmentation is performed on time-aligned computed tomography angiography images and time-aligned magnetic resonance angiography images. Preferably, a three-dimensional deep learning segmentation network is used in conjunction with post-processing based on morphology and topological constraints to obtain connected and smooth vascular cavity segmentation results. Based on this, three-dimensional reconstruction is performed to form a vascular geometric model of the patient. Preferably, surface continuity and normal consistency are achieved by using triangular meshes or implicit representation.
[0164] Subsequently, a signed distance function of the vessel wall is calculated based on the vessel geometry model, denoted as . ,in Represents a position vector in three-dimensional space. absolute value representation The minimum distance to the vessel wall, whose sign is used to distinguish between inside and outside the lumen, preferably positive for inside the lumen and negative for outside the lumen, so that it can be directly called in subsequent boundary embedding;
[0165] In terms of geometric center path extraction, let the blood vessel centerline be denoted as... ,in This is the arc length parameter along the centerline, in millimeters. It can be obtained by applying skeletonization, minimum cost path or manifold refinement with distance field constraints to the segmented volume, and calculating the unit tangential vector at each point. Used to define orthogonal sections, and constructing a vascular tree diagram using the bifurcation relationship of the centerline, denoted as... The nodes represent bifurcation points or entrances and terminals, and the edges represent blood vessel segments defined by two adjacent nodes, which are used to characterize blood vessel topology in subsequent graph structure boundary condition learning.
[0166] In terms of geometric parameter calculation, the radius distribution is denoted as... And it is estimated by the equivalent circle radius of the orthogonal section of the centerline, specifically, in the passage And the normal direction and Calculate the cross-sectional area on a uniform cross section Then obtain the following formula :
[0167] ;
[0168] in, Indicates the position at the arc length The equivalent circle radius at that point, Indicates the position at the arc length The cross-sectional area obtained from the orthogonal section, Represents pi;
[0169] Curvature is denoted as Its physical meaning is the degree of curvature of the centerline in three-dimensional space, equivalent to the unit tangential vector. The norm of the rate of change of arc length;
[0170] The bifurcation angle is denoted as The angle between the unit tangential vectors of adjacent branches at the bifurcation point can be obtained and can be calculated separately for the parent branch and the child branch, and for the child branch and the child branch to characterize the local flow geometry.
[0171] Through the above processing, the geometric model of the blood vessel and the signed distance function of the blood vessel wall are output. Central line of blood vessels With blood vessel tree diagram and radius distribution curvature with bifurcation angle .
[0172] In this specific embodiment, S3 includes:
[0173] Based on time-aligned and denoised blood pressure and heart rate time series, according to a unified time reference Calculate the cardiac cycle phase function and form a cardiac phase code. For this purpose, the instantaneous heart rate function obtained from the heart rate time series is denoted as... (Unit: Hertz), with uniform time as the independent variable for integration. Construct the normalized phase and implement it in numerical form. Surrounding, specifically as follows:
[0174] ;
[0175] in, This represents the cardiac phase function, expressed in radians. Represents dimensionless time on a unified time base; This represents the instantaneous heart rate function with uniform time as the independent variable; Represents the integral variable over a unified time base;
[0176] In obtaining Then, numerical encoding was performed using a sine and cosine equivalence method and used as a time-series condition variable input into the subsequent model;
[0177] Subsequently, in the vascular tree diagram A graph neural network is constructed on top of this, and message passing is performed between nodes and edges. Kirchhoff conservation is explicitly applied at each branch node, that is, the sum of the inflow is constrained to be equal to the sum of the outflow and the pressure continuity constraint is satisfied, so that the topological consistency is satisfied during the estimation process.
[0178] The set of leaf nodes in the blood vessel tree diagram is denoted as The position is connected to a differentiable Wendt-Kessel impedance model, and the proximal resistance of each terminal branch is learned in an end-to-end differentiable manner. Capacitive parameters (To avoid aligning with the center line of blood vessels) Confusion, tolerance here (Indicates) and far-end resistance superscript Represents the set of leaf nodes The One terminal branch;
[0179] Combined with the central line of blood vessels Radius distribution curvature with bifurcation angle Isogeometric features, and encoded with cardiac phase As a temporal condition, graph neural networks estimate inlet flow. And in conjunction with the differentiable Wendt-Kessel model, the outlet impedance of each terminal branch is obtained. ;
[0180] Output cardiac phase encoding Entry traffic Impedance of each terminal branch outlet and Windkeisel impedance parameters and .
[0181] In this specific embodiment, S4 includes:
[0182] Based on inbound traffic Output impedance of each terminal branch and encoded by cardiac phase As a temporal condition, in the flow domain defined by the vascular geometry model The boundary conditions that vary with the cardiac phase are constructed by using the signed distance function of the blood vessel wall. Spatial division of entrance boundaries Terminal branch boundary Boundary with blood vessel wall ,exist Apply inlet conditions at which the volumetric flow rate coincides with the cardiac phase, such that the cross-sectional integral equals... ,exist Treat The branch normal flow rate and branch pressure are coupled according to the impedance relationship to form an equivalent impedance boundary. Through based on Boundary embedding achieves slip-free constraints;
[0183] Subsequently in the flow domain Internally calling the physically constrained neural operator, to The velocity field is approximately generated based on temporal conditions and spatial conditions based on geometric and boundary embedding. With pressure field The generated velocity field is then subjected to a Helmholtz projection to achieve an incompressible hard constraint, with the divergence-free condition being:
[0184] ;
[0185] in, This represents the divergence operator with respect to spatial coordinates, used to calculate the volumetric expansion rate of the velocity field. This represents the velocity field vector after Helmholtz projection, defined in the flow domain. cardiac phase on a unified time base superior, A position vector in three-dimensional space is used to locate any point within the flow domain. The dimensionless time variable on a unified time base is used to indicate the cardiac phase position, and 0 represents the zero scalar used to characterize the constraint of zero divergence under incompressible conditions;
[0186] In numerical implementation, the physical constraint neural operator first generates initial velocity-pressure pairs that satisfy the boundary embedding, and then removes the gradient field components through projection to obtain the desired result. The velocity field, and simultaneously the pressure field Apply a reference pressure anchor to eliminate additive uncertainty;
[0187] Output and This is for subsequent calculations of wall shear stress, oscillatory shear index, and pressure gradient distribution on the blood vessel wall.
[0188] In this specific embodiment, S5 includes:
[0189] Based on velocity field With pressure field And combined with the signed distance function of the blood vessel wall Determining the spatial boundaries of the blood vessel wall and its legal information, specifically as The gradient determines the unit normal vector of the wall and in The tangential plane and tangential components of the vessel wall are extracted and used to calculate hemodynamic parameters on the vessel wall.
[0190] First, the gradient of the tangential component of the velocity field along the wall normal is calculated at the wall surface. Then, the shear rate is converted into a wall shear stress vector according to a preset fluid viscosity, forming the wall shear stress distribution. Subsequently, the wall shear stress is integrated and normalized using a unified time reference of the cardiac cycle to obtain the oscillating shear exponential distribution, which is located at a spatial point... The definition is as follows:
[0191] ;
[0192] in, Indicates a point in space The oscillatory shear index at the point is used as a dimensionless index to measure the degree of directional oscillation of wall shear stress during the cardiac cycle. Indicates the boundary of the blood vessel wall spatial point Cardiac phase at a unified time reference The wall shear stress vector; Represents a position vector in three-dimensional space, used to locate the boundary of the blood vessel wall. any point on; A dimensionless time variable representing a uniform time base, used to indicate the position of cardiac phase; Represents the differential component with respect to a unified time base, used for performing time integration; Represents the integral operator over the normalized time interval of the cardiac cycle; The Euclidean norm of a vector is used to measure the magnitude of the shear stress vector.
[0193] Regarding pressure-related indicators, the pressure field at the blood vessel wall... The spatial gradient is calculated and projected along the tangential plane of the vessel wall to form a pressure gradient distribution, which reflects the rate of pressure change along the vessel wall.
[0194] Output the wall shear stress distribution, oscillatory shear index distribution, and pressure gradient distribution for use in subsequent steps.
[0195] In this specific embodiment, S6 includes:
[0196] Based on the vascular tree diagram Identify each branch and use the vessel centerline unit tangential vector Construct the branch cross-section of each branch, and denote the first branch as the first branch. The cross-section of each branch is And its direction is the same as that place Consistency, in Upper velocity field The normal component according to the cardiac phase Perform cross-sectional integration to generate branch volumetric flow rate The flow vectors are then combined into branch flow vectors, and the branch volumetric flow rates are compared with the inlet volumetric flow rates. Perform a mass conservation comparison and base it on the outlet impedance of each terminal branch. The reference branch volumetric flow rate based on impedance is calculated by combining the local pressure-flow relationship. Based on this, the mass conservation difference and the reference branch flow difference are combined according to preset weights to form a closed-loop consistency residual, which has the following scalar form:
[0197] ;
[0198] in, Indicating the phase of the heartbeat The closed-loop consistency residual scalar is used to quantify the degree of deviation of the current solution from mass conservation and impedance consistency; A non-negative weighting parameter representing the difference in mass conservation, used to adjust the contribution of the difference in total inlet and branch flow to the residual; This represents a non-negative weighting parameter based on impedance differences, used to adjust the contribution of the difference between the actual flow rate and the reference flow rate of each branch to the residual. This represents the summation operator for branch indices; Representation of a blood vessel tree diagram The number of branches in the terminal; Indicates the first A branch in the cross section Pressing on the heart phase Volumetric flow rate; The volumetric flow rate boundary condition at the inlet varies with the cardiac phase. Indicates according to the first The output impedance of each terminal branch Reference volumetric flow rate calculated from the local pressure-flow relationship; Indicates the first Equivalent output impedance of each terminal branch; Operator for absolute value of scalars; A dimensionless cardiac phase variable on a unified time base, used to indicate the position of the cardiac cycle;
[0199] After obtaining the closed-loop consistent residuals, the wall shear stress vector is applied to the branches and their neighborhoods based on the branch-level contribution of the residuals and the tree graph adjacency relationship. The oscillatory shear index obtained by integral normalization based on a unified time base In addition, the pressure gradient distribution projected along the vessel wall is corrected for consistency. Specifically, the distribution is proportionally adjusted by a scaling factor that attenuates by geometric distance along the centerline, so that the corrected distribution better meets the mass conservation of the inlet and branches and the consistency of the terminal impedance.
[0200] Output branch flow vector, closed-loop consistency residual, and corrected wall shear stress distribution, corrected oscillatory shear exponent distribution, and corrected pressure gradient distribution.
[0201] In this specific embodiment, S7 includes:
[0202] In the image branch, time-aligned computed tomography angiography images and time-aligned magnetic resonance angiography images are used as the main inputs, and the corrected wall shear stress distribution, the corrected oscillatory shear index distribution and the corrected pressure gradient distribution are used as auxiliary quantities of image risk features. The risk distribution of the image branch is obtained through feature extraction and risk estimation.
[0203] In the clinical branch, standardized structured clinical data is used as input to extract clinical risk features and output the risk distribution of the clinical branch;
[0204] The text branch takes a standardized text report as input, forms text risk features, and outputs the risk distribution of the text branch;
[0205] Each of the three branches simultaneously outputs its corresponding confidence level for subsequent calibration and visualization. During the fusion phase, expert product fusion is used to multiply the risk distributions of the three branches point-by-point and normalize them to form the fused risk distribution. Specifically:
[0206] ;
[0207] in, This represents the fusion risk distribution, serving as a categorical or ordinal variable for predefined preoperative risk objectives. The probability distribution on; The risk distribution of the image branches is derived from modeling time-aligned computed tomography angiography images and time-aligned magnetic resonance angiography images and their auxiliary quantities. The risk distribution of clinical branches is derived from modeling standardized structured clinical data; The risk distribution of text branches is derived from modeling standardized text reports; Category or rank variables representing predefined preoperative risk objectives, used to indicate different risk levels; This represents the normalization constant, used to scale the product result to a valid probability distribution so that the sum of the probabilities of each class is 1;
[0208] After obtaining the fusion risk distribution, at the lesion level, the lesion-level preoperative risk value is read on the fusion risk distribution according to the spatial location of the lesion, and the lesion-level confidence is generated by combining the branch confidence. At the patient level, the lesion-level preoperative risk value is aggregated by the aggregation function according to the set of lesions contained in the patient to obtain the patient-level preoperative risk value and give the corresponding confidence.
[0209] Output the lesion-level and patient-level preoperative risks and their confidence levels for clinical decision-making.
[0210] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
[0211] This invention achieves unified alignment and standardization of multimodal inputs, combined with collaborative modeling of graph structure boundary conditions learning and physically constrained neural operators. It directly generates individualized flow field and hemodynamic indices from patient computed tomography (CT) angiography, magnetic resonance angiography, blood pressure and heart rate time series, and clinical text, reducing reliance on time-consuming computational fluid dynamics simulations. The graph neural network satisfies Kirchhoff conservation on the vascular tree diagram and connects to a differentiable Wendkeser model, outputting inlet flow rate and terminal impedance consistent with the cardiac phase. The neural operator solves for the velocity and pressure fields under hard constraints of Helmholtz projection and signed distance function boundary embedding. At the branch level, it forms a flow vector through cross-sectional integration, which, together with the inlet and impedance-based reference flow rates, forms a closed-loop consistency residual for correction. This improves the accuracy and stability of wall shear stress, oscillatory shear exponent, and pressure gradient while ensuring incompressibility and no-slip. Finally, it integrates with imaging, clinical, and textual risk features to provide confident preoperative risk results.
[0212] In terms of algorithmic structure, this invention addresses the technical problem of rapidly estimating hemodynamic parameters using images to replace time-consuming simulations, proposing several targeted improvements: introducing differentiable Wendkeser projections coupled with graph neural networks to obtain individualized and conserved consistent boundaries; using cardiac phase for conditional driving to characterize time-varying blood flow; employing Helmholtz divergence-free projection and no-slip boundary embedding based on signed distance functions in the neural operators to achieve physical hard constraints; constructing a closed-loop consistency regularization from the graph model to the neural operators and back to the graph to constrain branch mass conservation and impedance consistency; and employing expert product fusion in the risk layer to achieve robust convergence of multiple evidences and uncertainty representation. These structural improvements work synergistically to significantly shorten computation time, reduce sensitivity to empirical boundary settings, and improve physical consistency and generalization ability, thereby better achieving rapid, interpretable, and reliable preoperative risk assessment.
Claims
1. A method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning, characterized in that, include: S1. Obtain the patient's CTA, MRA, blood pressure and heart rate time series, structured clinical data and text reports, and perform time alignment and standardization; S2. Perform vessel segmentation and 3D reconstruction on aligned CTA and MRA to obtain a vessel geometric model. Calculate the signed distance function of the vessel wall, vessel centerline, vessel tree diagram, radius distribution, curvature, and bifurcation angle from this model. S3. Generate cardiac phase codes based on aligned and denoised blood pressure and heart rate time series, apply Kirchhoff conservation graph neural network on the vascular tree diagram, and attach differentiable Wind-Kessel impedance models to leaf nodes to estimate inlet flow and outlet impedance and Wind-Kessel impedance parameters of each terminal branch. S4. Using the inlet flow rate and outlet impedance as boundary conditions and cardiac phase encoding as time conditions, the flow field is approximated by calling the physical constraint neural operator based on the vascular geometry model and combined with the signed distance function to obtain the velocity field and pressure field. Helmholtz projection is performed on the velocity field to satisfy the divergence-free constraint, and boundary embedding is performed based on the signed distance function to satisfy the slip-free constraint. S5. Based on the velocity field and pressure field and combined with the signed distance function, calculate the wall shear stress, oscillatory shear index and pressure gradient distribution on the blood vessel wall; S6. Based on the vascular tree diagram and centerline, the velocity field is integrated at the cross-section of each branch to obtain the branch flow vector. The branch flow vector is compared with the inlet flow and outlet impedance to obtain the closed-loop consistency residual. Based on this, the wall shear stress, oscillatory shear index and pressure gradient distribution are corrected for consistency. S7. Combine the corrected wall shear stress, oscillatory shear index, and pressure gradient distribution with aligned CTA and MRA, standardized structured clinical data, and standardized text reports. Extract risk features through imaging, clinical, and text branches and obtain the risk distribution and confidence level of each branch. Use expert product fusion to obtain the lesion-level and patient-level preoperative risk values and corresponding confidence levels.
2. The method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning according to claim 1, characterized in that, S1 includes: For the same patient, computed tomography angiography and magnetic resonance angiography images are time-aligned according to the examination time, and the spatial orientation and voxel spacing of the images are standardized to output time-aligned computed tomography angiography images and time-aligned magnetic resonance angiography images. The blood pressure and heart rate time series are resampled at fixed sampling intervals, outlier removal and filtering are performed, and the start and end points are aligned with the time reference of the image to output the time-aligned and denoised blood pressure and heart rate time series. The system unifies the execution units, standardizes fields, and normalizes terms for structured clinical data, outputting standardized structured clinical data. Perform text cleaning, medical terminology standardization, and time stamping on the text report to output a standardized text report.
3. The method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning according to claim 1, characterized in that, S2 include: Perform vascular segmentation and spatial reconstruction on time-aligned computed tomography angiography and time-aligned magnetic resonance angiography to form a vascular geometry model of the patient and output the vascular geometry model. The signed distance function of the blood vessel wall is calculated based on the blood vessel geometry model and used to characterize the position of the blood vessel wall in the subsequent boundary embedding. The signed distance function of the blood vessel wall is output. The vessel centerline is extracted based on the vascular geometric model, and a vascular tree diagram is constructed based on the bifurcation relationship of the vessel centerline. This is used to represent the vascular topology in the subsequent graph structure boundary condition learning, and the vascular centerline and vascular tree diagram are output. The radius distribution, curvature, and bifurcation angle are calculated on the vascular geometry model to characterize the geometric parameters in subsequent boundary condition estimation and flow field calculation, and the radius distribution, curvature, and bifurcation angle are output.
4. The method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning according to claim 1, characterized in that, S3 include: The time-aligned and denoised blood pressure and heart rate time series are used to calculate and encode the cardiac cycle phase function according to a unified time reference, forming cardiac phase coding. Message passing that satisfies Kirchhoff conservation constraints is performed on the vascular tree diagram using a graph neural network, and differentiable Wendkesell impedance models are attached to the leaf nodes of the vascular tree diagram to obtain the proximal resistance parameters, capacitive parameters and distal resistance parameters of each terminal branch. The inlet flow rate is estimated based on the cardiac phase coding of the vascular centerline, radius distribution, curvature and bifurcation angle, and the outlet impedance of each terminal branch is obtained by the Wendkessel impedance model. Output cardiac phase encoding, inlet flow rate, outlet impedance, and Wind-Kessel impedance parameters.
5. The method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning according to claim 1, characterized in that, S4 includes: Based on the inlet flow rate and outlet impedance, and combined with the cardiac phase encoding, boundary conditions that vary with the cardiac phase are constructed at the inlet section and each terminal branch for approximate calculation of the driving flow field; Within the flow domain defined by the vascular geometry model, physical constraint neural operators are invoked to generate velocity and pressure fields according to the cardiac phase. Perform a Helmholtz projection on the generated velocity field to project the velocity field onto a divergence-free subspace to achieve divergence-free hard constraints; Boundary embedding is performed using the signed distance function of the blood vessel wall, so that the velocity field satisfies the no-slip boundary condition at the blood vessel wall; Output velocity field and pressure field.
6. The method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning according to claim 1, characterized in that, S5 include: Based on the velocity and pressure fields, the position of the blood vessel wall is determined by the signed distance function of the blood vessel wall, and the wall normal and tangential directions are extracted for use in calculating hemodynamic parameters on the blood vessel wall. By calculating the tangential gradient of the velocity field at the vessel wall and converting it according to a preset fluid viscosity, the wall shear stress distribution is formed; The wall shear stress is integrated over time and normalized along the cardiac cycle to form an oscillatory shear index distribution. The spatial gradient of the pressure field at the blood vessel wall is calculated and projected along the blood vessel wall to form a pressure gradient distribution. Output the wall shear stress distribution, oscillatory shear index distribution, and pressure gradient distribution.
7. The method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning according to claim 1, characterized in that, S6 include: Based on the vascular tree diagram, each branch is determined and the branch cross-section is constructed with the tangent of the vascular centerline. The velocity field is integraled with respect to the cardiac phase on the branch cross-section to form the branch flow vector. The branch flow vector is compared with the inlet flow for mass conservation, and the reference branch flow based on impedance is calculated by combining the outlet impedance. The mass conservation difference and the reference branch flow difference are combined into a closed-loop consistency residual according to the preset weight. Based on the closed-loop consistency residual, the wall shear stress distribution, oscillatory shear index distribution and pressure gradient distribution are corrected for consistency. The above distributions located in branches with significant residuals and their neighborhoods are adjusted proportionally to meet the mass conservation and impedance consistency constraints. The output is the branch flow vector, closed-loop consistency residual, corrected wall shear stress distribution, corrected oscillatory shear index distribution and corrected pressure gradient distribution.
8. The method for preoperative risk prediction of cerebral aneurysms based on multimodal deep learning according to claim 1, characterized in that, S7 includes: In the imaging branch, time-aligned computed tomography angiography images and time-aligned magnetic resonance angiography images are combined, and the corrected wall shear stress distribution, the corrected oscillatory shear index distribution and the corrected pressure gradient distribution are used as auxiliary quantities for the imaging risk characteristics to form the imaging risk characteristics and obtain the risk distribution and confidence of the imaging branch. In the clinical branch, clinical risk characteristics are formed based on standardized structured clinical data, and the risk distribution and confidence level of the clinical branch are obtained; Within the text branch, text risk characteristics are formed based on standardized text reports, and the risk distribution and confidence level of the text branch are obtained. The risk distributions of the imaging branch, clinical branch and text branch are multiplied and normalized by expert product fusion to form a fused risk distribution. The lesion-level preoperative risk value is read from the fused risk distribution, and the patient-level preoperative risk value is generated according to the set of lesions contained in the patient, while the corresponding confidence level is output.