Cerebral stroke early warning method based on cerebrovascular image representation group

Abstract

The invention relates to the technical field of medical image early warning, and discloses a cerebral apoplexy early warning method based on a cerebrovascular image representation group. The method comprises the following steps: acquiring a cerebrovascular image sequence, and constructing multi-modal information containing a vascular structure and hemodynamics; based on this, a multi-level representation network is constructed, and a dynamic time-varying risk representation graph is generated through node interaction and space-time modulation. Then performing multi-scale grid segmentation on the risk map to obtain a plurality of local evaluation units, and synchronously screening out candidate units with abnormal evolution according to a preset risk dynamics model; and aggregating the candidate unit features to generate a global risk evolution path, backtracking and positioning to a high-risk frame set in the original image, and outputting an early warning signal after feature decoupling and recombination. According to the method, quantitative tracking and early-stage accurate positioning of the dynamic evolution process of the stroke risk in the space-time space are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image early warning, in particular to a stroke early warning method based on a cerebrovascular image representation group. BACKGROUND

[0002] Existing stroke image early warning techniques mainly rely on independent analysis of blood vessel morphology or blood flow parameters. Conventional methods measure isolated indicators such as blood vessel stenosis degree and assess blood flow velocity, and make risk judgments based on empirical thresholds or statistical models. Some studies attempt to integrate multi-modal image data, but usually use simple feature stacking or late decision fusion, and fail to establish deep interaction calculation models between different physiological parameters.

[0003] The existing technology has deficiencies. Static parameter measurement cannot depict the dynamic evolution law of the cerebrovascular system in the space-time dimension, especially ignoring the real-time interaction between blood vessel geometry and local hemodynamics. At the same time, individual historical brain event information is often evaluated independently and cannot be effectively integrated with the current physiological state in time sequence. This makes it difficult for the early warning model to reflect the dynamic accumulation process of the risk in specific areas of the brain, and unable to identify early and local abnormal evolution patterns that meet the pathophysiological development rules of stroke, thereby limiting the timeliness and spatial specificity of early warning.

[0004] The present application constructs a multi-level representation network that can integrate blood vessel morphology, blood flow dynamics and historical event information, and drives the interaction calculation between nodes to generate a time-sequenced atlas reflecting the coupling and evolution of risk factors. This method performs multi-scale analysis on the atlas and selects local abnormal areas based on the biophysical model of risk evolution, ultimately positioning and warning high-risk brain regions with specific evolution patterns from dynamic images. SUMMARY

[0005] The present application aims to provide a stroke early warning method based on a cerebrovascular image representation group to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a stroke early warning method based on a cerebrovascular image representation group, which comprises:

[0007] Obtaining a cerebrovascular image sequence of a target individual;

[0008] Constructing multi-modal information of the cerebrovascular image sequence, the multi-modal information including blood vessel structure information and hemodynamic information;

[0009] Based on the multi-modal information, a multi-level representation network for representing stroke risk is constructed, the initial network nodes of the multi-level representation network including blood vessel morphology nodes, blood flow dynamics nodes and historical brain event nodes;

[0010] activating network nodes of the multi-level representation network, driving the vascular morphology nodes to interact with the blood flow dynamics nodes to obtain a spatial coupling relationship;

[0011] applying spatiotemporal modulation on the spatial coupling relationship according to the historical brain event node to generate a time-varying risk representation map;

[0012] performing multi-scale grid segmentation on the time-varying risk representation map to obtain a plurality of local risk assessment units;

[0013] synchronously evaluating evolution trends of the plurality of local risk assessment units to screen out candidate early warning units conforming to a preset risk dynamics model;

[0014] aggregating features of all the candidate early warning units to generate a global risk evolution path;

[0015] backtracking to the cerebral vascular image sequence according to the global risk evolution path to demarcate a high-risk image frame set;

[0016] performing feature decoupling and recombination on the high-risk image frame set to output a final stroke risk early warning signal.

[0017] Preferably, the multi-modal information of the cerebral vascular image sequence is constructed by the following steps:

[0018] performing density profile analysis along a blood vessel centerline on the cerebral vascular image sequence to extract continuous blood vessel wall morphology profiles to form blood vessel structure information;

[0019] calculating a pixel intensity gradient field of the cerebral vascular image sequence at consecutive time points, and calculating local blood flow velocity and direction by tracking motion trajectories of gradient field maximum points to form hemodynamic information;

[0020] aligning and fusing the blood vessel structure information and the hemodynamic information according to acquisition time stamps and spatial coordinates to generate the multi-modal information.

[0021] Preferably, the multi-level representation network representing stroke risk is constructed based on the multi-modal information by the following steps:

[0022] taking blood vessel bifurcation points and curvature extreme points as network anchor points, establishing a vascular morphology node based on the blood vessel structure information, the vascular morphology node storing local blood vessel diameter, wall roughness and orientation information;

[0023] in the hemodynamic information, taking the spatial position corresponding to the network anchor point as a reference, establishing a blood flow dynamics node, the blood flow dynamics node storing instantaneous blood flow velocity, pulsatility index and shear force information at the anchor point;

[0024] parsing a historical brain event record from an image description field of the brain blood vessel image sequence, mapping the historical brain event record to a corresponding spatial region to form a historical brain event node, the historical brain event node storing an event type, an occurrence time and an influence range;

[0025] taking the blood vessel morphology node, the blood flow dynamics node and the historical brain event node as initial network nodes of the multi-level representation network.

[0026] Preferably, the driving the blood vessel morphology node and the blood flow dynamics node to interact and operate to obtain a spatial coupling relationship comprises the following steps:

[0027] For each network anchor point, calculating an attribute mean of all blood vessel morphology nodes within a preset radius spatial range centered on the network anchor point to obtain a local blood vessel morphology summary;

[0028] calculating an attribute mean of all blood flow dynamics nodes within the same spatial range to obtain a local blood flow dynamics summary;

[0029] establishing a mutual information matrix of the local blood vessel morphology summary and the local blood flow dynamics summary, solving a maximum mutual information value of the mutual information matrix through an iterative optimization algorithm, and taking the maximum mutual information value and a corresponding feature mapping relationship as a spatial coupling relationship at the network anchor point;

[0030] integrating spatial coupling relationships at all network anchor points to form a complete spatial coupling relationship.

[0031] Preferably, the imposing a space-time modulation on the spatial coupling relationship according to the historical brain event node to generate a time-varying risk representation map comprises the following steps:

[0032] in the time dimension, defining a time decay window with an occurrence time recorded in the historical brain event node as a time origin;

[0033] in the spatial dimension, taking an influence range recorded in the historical brain event node as a spatial scope;

[0034] within a space-time range defined by the time decay window and the spatial scope, introducing a modulation weight to the spatial coupling relationship, the modulation weight decaying with time away from the time origin and decaying with spatial position away from a center of the spatial scope;

[0035] projecting the modulated spatial coupling relationship to a two-dimensional plane according to its spatial coordinates, and assigning a time stamp to each coordinate point to generate the time-varying risk representation map.

[0036] Preferably, the time-varying risk characterization map is subjected to multi-scale grid segmentation to obtain a plurality of local risk assessment units, including the following steps:

[0037] The time-varying risk characterization map is spatially divided using a hierarchical grid division strategy from coarse to fine;

[0038] At the coarse-scale grid level, the time-varying risk characterization map is divided into a plurality of primary grids, and the mean and variance of the risk characteristics of all points in each primary grid are calculated;

[0039] At the fine-scale grid level, the primary grids with risk characteristic variance exceeding a dynamic threshold are further subdivided to form secondary grids;

[0040] Each final grid that cannot be further subdivided is defined as a local risk assessment unit, which includes its spatial boundary, internal risk characteristic time series, and grid level information.

[0041] Preferably, the candidate early warning units that meet the preset risk dynamics model are screened, including the following steps:

[0042] For each local risk assessment unit, its risk characteristic time series is extracted;

[0043] The differential entropy, sample entropy, and trend slope of the risk characteristic time series are calculated;

[0044] The preset risk dynamics model includes an entropy change threshold and a slope threshold;

[0045] The differential entropy and the sample entropy are compared with the entropy change threshold, respectively, and the trend slope is compared with the slope threshold;

[0046] When the differential entropy and / or the sample entropy of the local risk assessment unit exceeds the corresponding entropy change threshold, and the absolute value of the trend slope exceeds the slope threshold, it is determined that the local risk assessment unit meets the preset risk dynamics model, and it is marked as a candidate early warning unit;

[0047] The construction of the preset risk dynamics model includes the following steps:

[0048] Cerebrovascular image data and corresponding stroke event records of historical stroke cases are collected to form a training data set;

[0049] From the training data set, the risk characteristic time series of each case is extracted, which is generated based on cerebrovascular morphological changes and hemodynamic changes indicators;

[0050] The differential entropy, sample entropy, and trend slope of each risk characteristic time series are calculated;

[0051] According to the stroke event record, the cases are divided into a high-risk group and a low-risk group;

[0052] By statistical analysis method, the differential entropy, sample entropy and trend slope distribution of the high-risk group and the low-risk group are compared, and the entropy threshold and the slope threshold are determined respectively;

[0053] The entropy threshold and the slope threshold are integrated into a risk dynamics model.

[0054] Preferably, the features of all the candidate early warning units are aggregated to generate a global risk evolution path, including the following steps:

[0055] All the candidate early warning units are sorted according to the starting time of their risk feature time series;

[0056] The distance of the spatial center point and the correlation of the risk features in time between adjacent candidate early warning units are calculated;

[0057] Based on the distance and the correlation, a risk propagation link describing the spatio-temporal relationship between the candidate early warning units is constructed;

[0058] All the risk propagation links are connected in chronological order to form a global risk evolution path that runs through the entire time series, which records the origin, propagation direction and intensity evolution of the risk.

[0059] Preferably, the global risk evolution path is traced back to the cerebral vascular image sequence to demarcate a high-risk image frame set, including the following steps:

[0060] The starting time point and the ending time point of all the risk propagation links in the global risk evolution path are extracted;

[0061] All the time points are mapped back to the acquisition timeline of the cerebral vascular image sequence;

[0062] In the cerebral vascular image sequence, the image frame corresponding to each mapped time point is located;

[0063] The image frame and all the adjacent image frames within a preset time window before and after the image frame are included to form the high-risk image frame set.

[0064] Preferably, the high-risk image frame set is subjected to feature decoupling and recombination to output a final stroke risk early warning signal, including the following steps:

[0065] For each image frame in the high-risk image frame set, its structural features related to the vascular morphological nodes and its dynamic features related to the blood flow dynamic nodes are extracted respectively;

[0066] Perform principal component analysis on the structural features and the dynamic features to realize feature decoupling and obtain feature components independent of each other;

[0067] Only the feature components with a variance contribution rate exceeding a contribution rate threshold are screened out;

[0068] The screened feature components are weighted and reorganized according to the time sequence in which they appear in the global risk evolution path to generate a comprehensive risk indicator sequence;

[0069] A peak detection algorithm is applied to the comprehensive risk indicator sequence, and the time, spatial position and peak intensity corresponding to the peak points exceeding a risk peak threshold are integrated into the final stroke risk early warning signal.

[0070] Compared with the prior art, the present application has the following beneficial effects:

[0071] A multi-level representation network containing blood vessel morphology, blood flow dynamics and historical brain event nodes is constructed, and spatial coupling relationships are obtained through driving node interaction operation, and then time and space modulation is performed through the historical event nodes. This technology realizes deep spatiotemporal fusion of anatomical structure, real-time function and individual history in a unified computing framework. The time-varying risk representation map generated is no longer a superposition of static parameters, but a dynamic field depicting the interaction of risk factors in the brain and their evolution over time, which can reveal the early risk dynamic evolution process driven by structural-functional coupling imbalance that cannot be captured by traditional methods, thereby advancing the early warning time to an earlier physiological and pathological change stage.

[0072] The time-varying risk representation map is subjected to multi-scale grid segmentation to form local evaluation units, and its evolution trend is simultaneously screened according to a preset risk dynamics model. This technology decomposes the global risk field into discrete dynamic units, and through a model-driven screening mechanism, it can accurately identify micro-regions whose change patterns conform to the typical pathophysiological rules of stroke occurrence. Aggregating the global risk evolution paths generated by these candidate units realizes quantitative backtracking and calibration from micro-abnormal dynamic patterns to macro-risk development trajectories, enabling the early warning signal to specifically point to the precise anatomical location and time window where the risk is substantially evolving, thereby improving the spatial directionality and temporal accuracy of the early warning. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 A working principle diagram of the stroke early warning method based on the brain vascular image representation group according to the present application;

[0074] Figure 2 A flowchart for constructing multi-modal information of brain vascular image sequences;

[0075] Figure 3A flow chart for driving the blood vessel morphology node and the blood flow dynamic node to interact and obtain the spatial coupling relationship;

[0076] Figure 4 A spatiotemporal distribution thermograph of the spatiotemporal modulation weight for stroke early warning;

[0077] Figure 5 A risk intensity time-varying correlation graph of the brain blood vessel image sequence. DETAILED DESCRIPTION

[0078] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0079] Please refer to Figure 1 The present application provides a stroke early warning method based on a brain blood vessel image feature set, which comprises the following steps: first, acquiring a brain blood vessel image sequence of a target individual, which is usually a continuous medical image frame. Then, constructing multi-modal information of the brain blood vessel image sequence, which integrates blood vessel structure information and blood flow dynamics information. Based on the constructed multi-modal information, a multi-level feature network for representing stroke risk is established, and the initial network nodes of the network include a blood vessel morphology node, a blood flow dynamic node and a historical brain event node. Then, the multi-level feature network is activated to drive the blood vessel morphology node and the blood flow dynamic node to interact and operate, so as to obtain a spatial coupling relationship reflecting the correlation between the two. Further, according to the information carried by the historical brain event node, the spatial coupling relationship is subjected to spatiotemporal modulation to generate a time-varying risk representation graph containing a time dimension. The time-varying risk representation graph is subjected to multi-scale grid segmentation to divide it into a plurality of local risk assessment units. The evolution trend of the local risk assessment units is synchronously evaluated, and candidate early warning units that meet a preset risk dynamics model are selected from the local risk assessment units. The features of all candidate early warning units are aggregated to generate a global risk evolution path describing the risk evolution process. The global risk evolution path is traced back to the initial brain blood vessel image sequence to demarcate a high-risk image frame set. Finally, the high-risk image frame set is subjected to feature decoupling and recombination operations to output a final stroke risk early warning signal.

[0080] Embodiment 1: Please refer to Figure 2After obtaining the brain vessel image sequence of the target individual, multi-modal information thereof is constructed. Density profile analysis along the centerline of the brain vessel image sequence is performed to extract continuous vessel wall profile, which forms the vessel structure information. The pixel intensity gradient field of the brain vessel image sequence at consecutive time points is calculated, and the local blood flow velocity and direction are calculated by tracking the motion trajectory of the gradient field maximum points, which forms the hemodynamic information. The vessel structure information and the hemodynamic information are aligned and fused according to their acquisition time stamps and spatial coordinates to generate unified multi-modal information.

[0081] Based on the generated multi-modal information, a multi-level representation network representing the risk of stroke is constructed. The vessel bifurcation points and the curvature extreme points are taken as the network anchors, and the vessel morphology nodes are established at these anchors based on the vessel structure information. The vessel morphology nodes store the local vessel diameter, wall roughness and orientation information. In the hemodynamic information, the blood flow dynamic nodes are established based on the spatial positions corresponding to the same network anchors, and the blood flow dynamic nodes store the instantaneous blood flow velocity, pulsatility index and shear force information at the anchors. The historical brain event records are parsed from the image description field of the brain vessel image sequence, and the historical brain event records are mapped to the corresponding spatial regions in the image sequence to form the historical brain event nodes, which store the event type, occurrence time and impact range. The vessel morphology nodes, the blood flow dynamic nodes and the historical brain event nodes are taken as the initial network nodes of the multi-level representation network.

[0082] In specific implementation, constructing the multi-modal information of the brain vessel image sequence involves two parallel processing procedures. Density profile analysis along the centerline of the brain vessel image sequence is performed, and the operation object of the density profile analysis is the segmented three-dimensional vessel structure in the brain vessel image sequence. The vessel centerline is extracted, and the gray distribution profile perpendicular to the centerline direction is calculated for each point on the centerline to obtain continuous vessel wall profile, which contains the contour information of the inner wall and the outer wall, forming the vessel structure information representing the vessel geometry and texture. The pixel intensity gradient field of the brain vessel image sequence at consecutive time points is calculated, and the quantitative data of the local blood flow velocity and direction are calculated by tracking the motion trajectory of the gradient field maximum points based on the optical flow method or the particle image velocimetry principle, forming the hemodynamic information representing the blood flow state. The vessel structure information and the hemodynamic information are derived from the same set of brain vessel image sequences but reflect different physical properties. The vessel structure information and the hemodynamic information are aligned and fused according to the acquisition time stamps and spatial coordinates. The alignment operation ensures that the coordinates of the same anatomical position in different modal information are consistent, and the fusion operation generates a multi-modal information dataset integrating the structural attributes and dynamic attributes.

[0083] In some embodiments, the extraction of the vessel wall shape profile can be quantified by an edge strength function. The edge strength function is defined as the integral of the length of the gradient modulus of the gray scale along the normal profile of the vessel centerline:

[0084]

[0085] wherein: denotes the edge strength of a point on the vessel wall, the integral path is along the direction of a straight line passing through the point and perpendicular to the vessel centerline, and are the end points of the integral path, is the pixel intensity value at the position on the path, is the gradient operator. This calculation helps to accurately determine the inner and outer boundaries of the vessel wall, thereby obtaining the vessel diameter and wall roughness.

[0086] A multi-level representation network representing the risk of stroke is constructed based on multi-modal information. The construction of the network starts from defining network anchor points. The blood vessel bifurcation points and the curvature extreme points are taken as the network anchor points. The blood vessel bifurcation points are the positions where branches appear in the vascular tree structure, and the curvature extreme points are the positions where the bending degree of the vessel centerline is the largest. The blood vessel shape nodes are established at the network anchor points based on the blood vessel structure information. The blood vessel shape nodes store local blood vessel diameter, wall roughness and orientation information. The local blood vessel diameter is directly measured from the blood vessel wall shape profile, the wall roughness is obtained by analyzing the periodic fluctuations of the blood vessel wall profile, and the orientation information is represented by the tangent vector of the blood vessel centerline at the anchor point.

[0087] In some embodiments, the blood flow dynamic nodes are established in the blood hemodynamic information with the same spatial position of the network anchor point as the reference, which ensures that the blood vessel morphology nodes and the blood flow dynamic nodes strictly correspond in space. The blood flow dynamic nodes store the instantaneous blood flow velocity, pulsatility index and shear force information at the anchor point, the instantaneous blood flow velocity is directly read from the result of tracking the motion trajectory of the gradient field maximum point, the pulsatility index is obtained by calculating the difference between the maximum value and the minimum value of the blood flow velocity in a cardiac cycle divided by the average value, and the shear force information is calculated according to the blood flow velocity profile and the blood vessel diameter through the Newtonian fluid model. In specific implementation, the specific implementation of the Newtonian fluid model is to calculate based on the two key parameters of the blood flow velocity profile and the blood vessel diameter; the blood flow velocity profile is the distribution data of the instantaneous blood flow velocity in the radial direction of the blood vessel obtained by tracking the motion trajectory of the gradient field maximum point from the cerebral vascular image sequence, and the blood vessel diameter is the local blood vessel geometric size extracted from the blood vessel structure information. The historical brain event record is parsed from the image description field of the cerebral vascular image sequence, and the image description field contains the text information in the digital imaging and communication standard format file. The historical brain event record is mapped to the corresponding spatial region in the cerebral vascular image sequence, and the mapping process is registered according to the anatomical position description in the record and the image coordinate system to form the historical brain event node. The historical brain event node stores the event type, occurrence time and influence range, the event type includes past ischemic lesions or microhemorrhage, the occurrence time is the date of the event record, and the influence range is the descriptive position or the lesion area segmented from the past image.

[0088] It can be understood that the one-to-one correspondence between the blood vessel morphology nodes and the blood flow dynamic nodes in space is the premise for effective interactive operation of the multi-level representation network. This correspondence ensures that in the subsequent steps, the attributes of the blood vessel morphology nodes can be directly associated and calculated with the attributes of the blood flow dynamic nodes at the same position.

[0089] Optionally, the tracking of the gradient field maximum points can adopt an algorithm based on local image block matching. The algorithm defines an image block centered at the maximum point in the image of the continuous frame, and finds the most similar image block position in the search area of the next frame, calculates the displacement of the image block position to calculate the motion trajectory, and thus obtains the blood flow velocity vector. This method can adapt to a certain degree of image noise and artifacts. In the specific implementation, after the multi-level representation network is activated, the core is to drive the blood vessel morphology nodes and the blood flow dynamic nodes to interact to obtain the spatial coupling relationship. For each network anchor point, the attribute mean of all blood vessel morphology nodes within a preset radius of the network anchor point is calculated, and the attribute mean includes the local blood vessel diameter mean, the wall roughness mean, and the average direction of the heading vector. The calculation result is defined as the local blood vessel morphology summary. The attribute mean of all blood flow dynamic nodes within the same spatial range is calculated, and the attribute mean includes the instantaneous blood flow velocity mean, the pulsatility index mean, and the shear force mean. The calculation result is defined as the local blood flow dynamic summary. The local blood vessel morphology summary and the local blood flow dynamic summary respectively represent the statistical characteristics of the blood vessel structure and the blood flow state in the same local region.

[0090] An mutual information matrix of the local blood vessel morphology summary and the local blood flow dynamic summary is established, and the mutual information matrix is used to quantify the nonlinear dependence between the two sets of statistical characteristics. The maximum mutual information value of the mutual information matrix is solved by an iterative optimization algorithm. The iterative optimization algorithm can adopt a gradient ascent method, which adjusts the feature mapping relationship in each iteration to maximize the mutual information value. The solved maximum mutual information value and the feature mapping relationship corresponding to the maximum value are jointly used as the spatial coupling relationship at the current network anchor point. The feature mapping relationship describes the functional relationship between how the local blood vessel morphology characteristics affect the local blood flow dynamic characteristics. The spatial coupling relationships at all network anchor points are integrated, and the integration operation is to collect the spatial coupling relationship data at all anchor points according to their spatial coordinates to form a complete spatial coupling relationship atlas covering the entire cerebral vascular network.

[0091] Embodiment 2: see Figure 3 After the multi-level representation network is activated, the blood vessel morphology nodes and the blood flow dynamic nodes are driven to interact to obtain the spatial coupling relationship. For each network anchor point, the attribute mean of all blood vessel morphology nodes within a preset radius of the network anchor point is calculated to obtain the local blood vessel morphology summary. The attribute mean of all blood flow dynamic nodes within the same spatial range is calculated to obtain the local blood flow dynamic summary. An mutual information matrix of the local blood vessel morphology summary and the local blood flow dynamic summary is established, and the maximum mutual information value of the mutual information matrix is solved by an iterative optimization algorithm. The maximum mutual information value and the corresponding feature mapping relationship are used as the spatial coupling relationship at the network anchor point. The spatial coupling relationships at all network anchor points are integrated to form a complete spatial coupling relationship.

[0092] In a specific implementation, for each network anchor in the multi-level representation network, a preset radius is set centered on the three-dimensional spatial coordinates of the network anchor, which defines a local spherical spatial range for interaction operation. The attribute mean values of all vessel morphology nodes located within this preset radius spatial range are calculated, including the local vessel diameter mean value, the wall roughness mean value, and the mean value of the direction information, wherein the mean value of the direction information is obtained by averaging the average direction of the direction vectors of all vessel morphology nodes in the range, and the above calculation results are uniformly defined as the local vessel morphology summary. Within the same preset radius spatial range, the attribute mean values of all blood flow dynamic nodes are calculated, including the instantaneous blood flow velocity mean value, the pulsatility index mean value, and the shear force mean value, and the above calculation results are uniformly defined as the local blood flow dynamic summary. The local vessel morphology summary and the local blood flow dynamic summary respectively depict the comprehensive state of the local cerebral vascular region centered on the network anchor from the structural domain and the dynamic domain.

[0093] In some embodiments, a mutual information matrix of the local vessel morphology summary and the local blood flow dynamic summary is established, which is a mathematical construction whose rows correspond to each feature interval in the local vessel morphology summary after discretization processing, and whose columns correspond to each feature interval in the local blood flow dynamic summary after discretization processing, and each element value in the matrix represents the correlation degree of the joint probability and the marginal probability product between the row features and the column features. Solving the maximum mutual information value of the mutual information matrix requires an iterative optimization algorithm, which adjusts the parameters of a nonlinear mapping function such that the mutual information measure between the local vessel morphology summary and the local blood flow dynamic summary after transformation is maximized. This process can be expressed as finding the optimal parameters to maximize the objective function:

[0094]

[0095] wherein: represents the local vessel morphology summary, represents the local blood flow dynamic summary, is a nonlinear feature mapping function with parameters , and is a mutual information function that measures the dependence between two variables. The parameters are iteratively updated by the gradient ascent method, and the maximum value obtained when the final convergence is reached is the maximum mutual information value, and the feature mapping function ​​​is the corresponding feature mapping relationship. The maximum mutual information value and the feature mapping relationship are collectively taken as the spatial coupling relationship at the current network anchor point.

[0096] In some embodiments, integrating the spatial coupling relationships at all network anchor points is to collect the outputs calculated by all network anchor points, and the output of each network anchor point includes its three-dimensional spatial coordinates, the maximum mutual information value, and the parameter set of the feature mapping function. By associating and storing the spatial coordinates of each network anchor point with its corresponding spatial coupling relationship data, a complete spatial coupling relationship atlas covering the entire anatomical region corresponding to the cerebral vascular image sequence is formed.

[0097] It can be understood that the setting of the preset radius directly affects the range of the local summary calculation. If the preset radius is too small, the local information will not be statistically representative. If the preset radius is too large, the local information will be too smooth and lose detailed features. The value of the preset radius needs to be adaptively selected according to the scale of the vascular network. Optionally, when calculating the shear force mean value of the local blood flow dynamic summary, if the shear force data of an individual blood flow dynamic node has an abnormal outlier value due to the original image quality problem, a robust statistical method based on the median can be used for preprocessing, that is, the median of all shear force values in the local range is first calculated, then the outlier points with an absolute deviation from the median exceeding a preset threshold are excluded, and then the average value of the remaining data is calculated. When updating the parameters of the iterative optimization algorithm, an adaptive moment estimation algorithm can be used to accelerate convergence and improve stability. The adaptive moment estimation algorithm dynamically adjusts the learning rate of each parameter by calculating the first moment estimation and the second moment estimation of the gradient.

[0098] Embodiment 3: According to the historical brain event node, the spatial coupling relationship is subjected to spatiotemporal modulation, and a time-varying risk representation map is generated. In the time dimension, the event occurrence time recorded in the historical brain event node is taken as the time origin, and a time decay window is defined. In the spatial dimension, the influence range recorded in the historical brain event node is taken as the spatial scope. Within the spatiotemporal range defined by the time decay window and the spatial scope, a modulation weight is introduced to the spatial coupling relationship. The modulation weight decays with the time away from the time origin and decays with the spatial position away from the center of the spatial scope. The modulated spatial coupling relationship is projected onto a two-dimensional plane according to its spatial coordinates, and a timestamp is assigned to each coordinate point, thereby generating a time-varying risk representation map.

[0099] The time-varying risk characterization map is divided into multiple local risk assessment units through multi-scale gridding. A hierarchical gridding strategy from coarse to fine is used to spatially partition the time-varying risk characterization map. At the coarse-scale grid level, the time-varying risk characterization map is divided into several primary grids, and the mean and variance of the risk characteristics of all points within each primary grid are calculated. At the fine-scale grid level, primary grids with risk characteristic variance exceeding the dynamic square threshold are further subdivided to form secondary grids. Each final grid that cannot be further subdivided is defined as a local risk assessment unit, which includes its spatial boundary, the time series of its internal risk characteristics, and the grid hierarchy information to which it belongs.

[0100] In practical implementation, in the temporal dimension, the event occurrence time recorded in the historical brain event nodes is used as the time origin, and a time decay window is defined. The time decay window is a time interval extending into the future from the time origin. In the spatial dimension, the influence range recorded in the historical brain event nodes is used as the spatial domain, which is a specific region in the three-dimensional image space. Within the spatiotemporal range jointly defined by the time decay window and the spatial domain, a modulation weight is introduced into the spatial coupling relationship. The modulation weight is a coefficient between 0 and 1. The modulation weight decays as time moves away from the time origin, and as spatial location moves away from the center of the spatial domain. The modulated spatial coupling relationship is projected onto a two-dimensional plane according to the spatial coordinates recorded by the spatial coupling relationship itself, and each coordinate point is assigned a timestamp corresponding to the calculation time. Through this series of assignment and mapping operations, a time-varying risk characterization map is generated. The time-varying risk characterization map is a three-dimensional data field containing spatial two-dimensional coordinates, timestamps, and modulated coupling strength values.

[0101] In some embodiments, the specific calculation of applying a temporal control mechanism to spatial coupling relationships based on historical brain event nodes can be expressed as follows:

[0102]

[0103] in: Indicates time spatial point The overall modulation weight at the point, This represents the time of occurrence of events recorded in the historical brain event nodes, i.e., the time origin. This represents the coordinates of the center point of the spatial domain recorded in the historical brain event node. This represents the absolute time difference between the current time and the origin. This represents the Euclidean distance between the current spatial point and the center point of the spatial domain. The time decay coefficient is used to control the rate at which the weight decays over time. The spatial decay coefficient is used to control the rate at which the weight decays with spatial distance. It is a balancing factor for temporal and spatial influences, used to adjust the relative importance of temporal and spatial attenuation effects. The modulated spatial coupling value is equal to the original spatial coupling value multiplied by the comprehensive modulation weight at the corresponding spatiotemporal point. .

[0104] In practical implementation, the time-varying risk characterization map is segmented into multi-scale grids. This multi-scale grid segmentation transforms the two-dimensional planar risk map containing spatiotemporal information into a series of regularized local units. A hierarchical grid division strategy, from coarse to fine, is adopted to spatially divide the time-varying risk characterization map. From coarse to fine means first using a larger grid size to cover the entire area, and then subdividing specific areas using smaller grid sizes. At the coarse-scale grid level, the time-varying risk characterization map is divided into several primary grids, and the mean and variance of the risk characteristics for all coordinate points within each primary grid are calculated. At the fine-scale grid level, primary grids with risk characteristic variance exceeding the dynamic square threshold are further subdivided. This subdivision involves dividing the primary grid into smaller sub-grids in both the row and column directions to form secondary grids. Each final grid that cannot be further subdivided is defined as a local risk assessment unit. The inability to be further subdivided means that the risk characteristic variance of the grid has fallen below the dynamic square threshold or has reached the preset minimum grid size. The data structure of the local risk assessment unit includes its spatial boundary coordinates, the time series of risk characteristic values ​​of all points inside, and the grid level information to which the unit belongs.

[0105] In some embodiments, the time decay window can be set to 6 months extending into the future from the origin of time. The spatial domain can be defined as a spherical region with the center point of the affected area record as the center and a certain distance as the radius. The dynamic square threshold is a preset value used to determine whether the fluctuation of the risk characteristics within the grid is large enough to warrant further subdivision. Optionally, the hierarchical grid partitioning strategy can be implemented using a quadtree structure. Initially, the entire time-varying risk characterization map is used as a root node. If the variance of the risk characteristics calculated for a certain node exceeds the dynamic square threshold, the node is split into four child nodes of equal size. This process is recursively executed until the variance of all nodes is below the threshold or the maximum split depth is reached.

[0106] It is understandable that the values ​​of the time decay coefficient α and the spatial decay coefficient β determine the spatiotemporal range of the influence of historical brain events. Larger coefficient values ​​lead to rapid decay of modulation weights with distance or time, making the influence range more concentrated, while smaller coefficient values ​​produce a wider and smoother spatiotemporal modulation effect. When projecting spatial coupling relationships onto a two-dimensional plane, for coordinate points in a three-dimensional vascular network, orthogonal projection can be performed along a preset projection direction, assigning the spatial coupling relationship value of the three-dimensional point to its corresponding pixel position on the two-dimensional plane. If multiple three-dimensional points are projected onto the same pixel, the average or maximum value of these points is taken as the final value of that pixel. The setting of the dynamic square threshold is the key to controlling the fineness of grid segmentation. A higher dynamic square threshold means that only areas with very drastic fluctuations in risk characteristics will be subdivided, resulting in fewer and larger local risk assessment units; a lower dynamic square threshold means that more areas will be subdivided, resulting in more and smaller local risk assessment units, which can capture more subtle risk changes.

[0107] See Figure 4 This is a heatmap showing the spatiotemporal distribution of modulation weights in stroke early warning, primarily illustrating the impact of "time difference" and "spatial distance" on the overall modulation weight. The overall modulation weight is the influence coefficient of historical brain events on current risk assessment; a higher value indicates a stronger influence of the historical event. The smaller the time difference (closer to the time of the historical event) and the closer the spatial distance (closer to the area affected by the event), the higher the modulation weight (the upper left area of ​​the map, with a weight close to 1.0). As the time difference increases and the spatial distance increases, the weight decreases exponentially. Spatiotemporal combinations on the same contour line correspond to the same modulation weight, reflecting the complementary influence of "time" and "space." This map is used for risk characterization correction in stroke early warning: by quantifying the spatiotemporal influence range of historical brain events, it provides dynamic weights for the risk assessment of cerebrovascular imaging, ensuring that risk analysis simultaneously considers the "timeliness of historical events" and "spatial correlation," thereby improving the accuracy of early warning.

[0108] Example 4: Simultaneously evaluate the evolution trend of multiple local risk assessment units and screen out candidate early warning units that conform to the preset risk dynamics model. Extract the risk characteristic time series for each local risk assessment unit. Calculate the differential entropy, sample entropy, and trend slope of the risk characteristic time series. The preset risk dynamics model includes an entropy change threshold and a slope threshold. Compare the differential entropy and sample entropy of the local risk assessment unit with the entropy change threshold, and compare its trend slope with the slope threshold. When the differential entropy and / or sample entropy of the local risk assessment unit exceeds the corresponding entropy change threshold, and the absolute value of the trend slope exceeds the slope threshold, the local risk assessment unit is determined to conform to the preset risk dynamics model and is marked as a candidate early warning unit.

[0109] The construction process of the pre-defined risk dynamics model is as follows: Cerebrovascular imaging data and corresponding stroke event records of historical stroke cases are collected to form a training dataset. Risk feature time series for each case are extracted from the training dataset; these time series are generated based on cerebral vascular morphological changes and hemodynamic changes. The differential entropy, sample entropy, and trend slope of each risk feature time series are calculated. Based on stroke event records, cases are divided into high-risk and low-risk groups. The distributions of differential entropy, sample entropy, and trend slope are compared between the high-risk and low-risk groups using statistical analysis methods to determine the entropy change threshold and slope threshold, respectively. The determined entropy change threshold and slope threshold are then integrated into the risk dynamics model.

[0110] Features from all candidate early warning units are aggregated to generate a global risk evolution path. All candidate early warning units are sorted according to the start time of their risk characteristic time series. The distance between the spatial center points of adjacent candidate early warning units and the temporal correlation of their risk characteristics are calculated. Based on these distances and correlations, risk propagation links describing the spatiotemporal connections between candidate early warning units are constructed. All risk propagation links are connected chronologically to form a global risk evolution path that runs through the entire time series. This global risk evolution path records the origin, propagation direction, and intensity evolution of the risk.

[0111] In practical implementation, the risk feature time series stored within each local risk assessment unit is extracted. This risk feature time series is composed of statistical features calculated from the modulated spatial coupling relationship values ​​of all points within the unit at each timestamp, arranged chronologically. The differential entropy, sample entropy, and trend slope of the risk feature time series are calculated. Differential entropy measures the randomness and uncertainty of the time series, sample entropy quantifies its complexity and regularity, and the trend slope is obtained by linearly fitting the time series values ​​over time. The preset risk dynamics model includes entropy change thresholds and slope thresholds. The entropy change thresholds include both differential entropy and sample entropy thresholds. The differential entropy and sample entropy of the local risk assessment unit are compared with the corresponding entropy change thresholds in the preset risk dynamics model, and the trend slope of the local risk assessment unit is compared with the slope threshold in the preset risk dynamics model. When the differential entropy of a local risk assessment unit exceeds its corresponding differential entropy threshold, and / or the sample entropy of the local risk assessment unit exceeds its corresponding sample entropy threshold, and the absolute value of the trend slope of the local risk assessment unit exceeds the slope threshold, the local risk assessment unit is determined to conform to the preset risk dynamics model and is marked as a candidate early warning unit.

[0112] In some embodiments, the trend slope can be calculated using a linear regression model. The risk characteristic time series is viewed as a set of observed data points. ,in Representing the At a certain point in time, This represents the risk characteristic value corresponding to that time point. A linear equation is fitted using the least squares method. The obtained slope This is the trend slope, calculated using the following formula:

[0113]

[0114] in: The total number of data points in the time series representing risk characteristics. It is the first The value at each point in time. It is the first Risk characteristic values ​​corresponding to each time point, It is the calculated trend slope. It is the intercept of the fitted line. The slope... The sign and magnitude of the value directly reflect the upward or downward trend and rate of change of risk characteristics over time.

[0115] In some embodiments, the construction of a pre-defined risk dynamics model includes a series of steps. First, collect cerebrovascular imaging data and corresponding stroke event records for historical stroke cases. The stroke event records must clearly indicate the time and type of the event, forming a training dataset. Second, extract the risk characteristic time series for each case from the training dataset. The risk characteristic time series is generated based on cerebrovascular morphological changes and hemodynamic changes calculated from historical cerebrovascular imaging data. Third, calculate the differential entropy, sample entropy, and trend slope of the risk characteristic time series for each case. Fourth, based on the stroke event records, divide cases into high-risk and low-risk groups. The high-risk group includes cases that experienced stroke during the follow-up period, while the low-risk group includes cases that did not experience stroke. Fifth, compare the distributions of differential entropy, sample entropy, and trend slope between the high-risk and low-risk groups using statistical analysis methods. The statistical analysis method can use ROC curve analysis to determine the optimal discrimination threshold, and determine the differential entropy threshold, sample entropy threshold, and trend slope threshold used to distinguish between the two groups of cases. Sixth, integrate the determined differential entropy threshold, sample entropy threshold, and trend slope threshold into a pre-defined risk dynamics model, which uses these thresholds as the judgment criteria. Refer to Table 1, which shows a simplified threshold determination table for a risk dynamics model:

[0116] Table 1: Threshold Determination Table for Risk Dynamics Model

[0117] Indicator High risk group mean ± standard deviation Low risk group mean ± standard deviation Determined threshold Differential entropy 2.15±0.43 1.62±0.38 >1.95 Sample entropy 0.78±0.22 0.51±0.18 >0.65 Trend slope absolute value 0.12±0.05 0.04±0.02 >0.08

[0118] In practice, features from all candidate early warning units are aggregated to generate a global risk evolution path. All candidate early warning units are sorted in ascending order according to the start time of their risk feature time series, where the start time is the timestamp corresponding to the first data point in the unit's time series. Between adjacent candidate early warning units after sorting, the Euclidean distance between their spatial center points is calculated, and the temporal correlation of their risk features is also calculated. Temporal correlation can be obtained by calculating the Pearson correlation coefficient of the risk feature sequences of two units within overlapping time periods. Based on the calculated spatial center point distance and temporal correlation of risk features, a risk propagation link describing the spatiotemporal connection between candidate early warning units is constructed. The weight of the link can be determined by both distance and correlation coefficient. All risk propagation links are connected in chronological order to form a global risk evolution path that runs through the entire observation time series. This global risk evolution path records the origin and time of the risk, its propagation direction in the cerebrovascular network, and the evolution of risk intensity during propagation.

[0119] It is understandable that the differential entropy and sample entropy of risk feature time series describe the dynamic characteristics of the series from different perspectives. Differential entropy focuses more on randomness, while sample entropy focuses more on complexity. Combining the two can improve the robustness of identifying abnormal evolution patterns. When calculating sample entropy, two parameters need to be set: embedding dimension m and tolerance r. The embedding dimension m is usually 1 or 2, and the tolerance r is usually 0.1 to 0.25 times the standard deviation of the time series to evaluate the self-similarity of the time series under fixed parameters. Optionally, when constructing risk propagation links, if the spatial center point distance between two adjacent candidate early warning units exceeds a preset maximum spatial propagation distance threshold, even if the temporal correlation is high, a direct link between them is not constructed. This is consistent with the reasonable assumption that risk propagation has spatial limitations in physiological terms.

[0120] Example 5: Based on the global risk evolution path, backtrack to the cerebrovascular imaging sequence to identify a high-risk image frame set. Extract the start and end times of all risk propagation links in the global risk evolution path. Map all time points back to the acquisition timeline of the cerebrovascular imaging sequence. In the cerebrovascular imaging sequence, locate the image frame corresponding to each mapped time point. Include the located image frame and all adjacent image frames within a preset time window before and after it to form a high-risk image frame set.

[0121] Feature decoupling and recombination are performed on a high-risk image frame set to output a final stroke risk warning signal. For each image frame in the high-risk image frame set, structural features related to vascular morphology nodes and dynamic features related to blood flow dynamic nodes are extracted. Principal component analysis is performed on the structural and dynamic features to achieve feature decoupling and obtain mutually independent feature components. Only feature components with a variance contribution rate exceeding a contribution rate threshold are selected. The selected feature components are weighted and recombinated according to their temporal order of appearance in the global risk evolution path to generate a comprehensive risk index sequence. A peak detection algorithm is applied to this comprehensive risk index sequence to integrate the time, spatial location, and peak intensity corresponding to the peak points exceeding the risk peak threshold into the final stroke risk warning signal.

[0122] In practice, the start and end times of all risk propagation links in the global risk evolution path are extracted. These risk propagation links are the spatiotemporal relationships connecting adjacent candidate early warning units, defined when constructing the global risk evolution path. All extracted start and end times are mapped back to the acquisition timeline of the cerebrovascular imaging sequence. This mapping operation involves finding the closest corresponding timestamp to the extracted time point on the timeline based on the precise acquisition timestamp recorded for each image frame in the cerebrovascular imaging sequence. Within the cerebrovascular imaging sequence, the image frame corresponding to each mapped time point is located by matching timestamps. The located image frame and all adjacent image frames within a preset time window before and after it are included to form a high-risk image frame set. The preset time window defines the time range of additional image frames to be included before and after the located time point.

[0123] In some embodiments, for each image frame in the high-risk image frame set, structural features related to vascular morphology nodes are extracted, including local vessel diameter, wall roughness, and vessel curvature, as well as dynamic features related to blood flow dynamic nodes, including instantaneous blood flow velocity, pulsatility index, and vessel wall shear force. Principal component analysis is performed on the extracted structural and dynamic features. Principal component analysis transforms the original multiple related feature variables into a set of linearly uncorrelated variables, i.e., principal components, through orthogonal transformation, achieving feature decoupling and obtaining mutually independent feature components. Only feature components with a variance contribution rate exceeding a contribution rate threshold are selected. The variance contribution rate measures the proportion of each feature component that explains the total variance of the original data, and the contribution rate threshold is a preset percentage. The selected feature components are weighted and recombined according to their chronological order of appearance in the global risk evolution path to generate a comprehensive risk index sequence. The weighted recombination assigns a weight to each feature component, which is related to the temporal position of its corresponding image frame in the global risk evolution path.

[0124] In practical implementation, the process of generating a comprehensive risk indicator sequence through weighted reorganization can be described as the following calculation. Assuming that after screening, the following results are obtained... The feature component, for the in the high-risk image frame set Each image frame, its corresponding The eigencomponent values ​​constitute a vector. The temporal order weight of this image frame in the global risk evolution path is: Then the first Comprehensive risk index corresponding to each image frame The calculation is as follows:

[0125]

[0126] in: Indicates the first The first image frame The values ​​of each feature component It is the first The eigenvalues ​​corresponding to each eigencomponent in principal component analysis Assigning the first based on time order The weight of each image frame is generally determined by the image frame being more recent in time. The larger the value, It is the calculated number of The comprehensive risk index of each image frame. The above calculation is performed on all image frames in the high-risk image frame set in chronological order, resulting in a chronologically arranged sequence. This refers to the comprehensive risk indicator sequence.

[0127] A peak detection algorithm is applied to the comprehensive risk index sequence to identify local maxima. Peaks exceeding a predetermined risk threshold are identified as significant risk peaks. The time corresponding to each significant risk peak, the spatial location reflected in the associated image frame, and the peak intensity are integrated into a final stroke risk warning signal. This final stroke risk warning signal is a structured output containing information such as the time of risk occurrence, anatomical location, and risk intensity level.

[0128] It's understandable that setting a preset time window is intended to capture the dynamic changes before and after a risk event. A time window that's too long introduces too much irrelevant information, while a window that's too short misses crucial evolution frames. The preset time window needs to be set based on the image acquisition frequency and the temporal characteristics of the physiological process. Optionally, before performing principal component analysis, the extracted structural and dynamic features can be standardized to eliminate the influence of differences in feature dimensions and numerical ranges on the analysis results. Standardization typically uses the Z-score method. It's understandable that the contribution rate threshold determines how much data variation information is retained. A higher contribution rate threshold retains fewer feature components and achieves higher information compression, but loses detail; a lower contribution rate threshold retains more feature components and provides more complete information, but requires more computational reconstructive analysis.

[0129] See Figure 5 This is a graph showing the correlation between the risk intensity of a cerebrovascular imaging sequence and time, primarily illustrating the temporal correspondence between the imaging sequence and the local risk intensity. The local risk intensity fluctuates periodically, exhibiting high risk in the 0-20 minute and 80-100 minute timeframes, and decreasing to a low risk zone in the 40-60 minute timeframe. The time points of the red sampled image frames correspond to the peaks / troughs of the risk intensity curve, allowing direct location of the risk level at the corresponding time through the image frames. This graph is used for image-risk temporal correlation analysis in stroke early warning: by aligning the image sequence with the risk intensity curve, high-risk corresponding image frames can be quickly located, assisting doctors in recalling cerebrovascular imaging characteristics during high-risk periods, and improving the interpretability and accuracy of risk warnings.

[0130] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0131] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A stroke early warning method based on cerebrovascular imaging characterization, characterized in that, Includes the following steps: Obtain cerebral vascular imaging sequences of the target individual; Construct multimodal information of the cerebral vascular imaging sequence, wherein the multimodal information includes vascular structure information and hemodynamic information; Based on the multimodal information, a multi-level representation network for stroke risk is constructed. The initial network nodes of the multi-level representation network include vascular morphology nodes, blood flow dynamic nodes, and historical brain event nodes. Activate the network nodes of the multi-level representation network to drive the blood vessel morphology node and the blood flow dynamic node to perform interactive operations and obtain the spatial coupling relationship. Based on the historical brain event nodes, a temporal control is applied to the spatial coupling relationship to generate a time-varying risk characterization map; The time-varying risk characterization map is divided into multiple local risk assessment units by multi-scale grid segmentation. Simultaneously evaluate the evolution trend of multiple local risk assessment units and select candidate early warning units that conform to the preset risk dynamics model; By aggregating the features of all the candidate early warning units, a global risk evolution path is generated; Based on the global risk evolution path, the cerebrovascular imaging sequence is traced back to identify the set of high-risk image frames; The high-risk image frame set is subjected to feature decoupling and recombination to output the final stroke risk warning signal.

2. The stroke early warning method based on cerebrovascular imaging characterization group according to claim 1, characterized in that, The construction of the multimodal information of the cerebral vascular imaging sequence includes the following steps: Density profile analysis along the blood vessel centerline is performed on the cerebral vascular image sequence to extract continuous blood vessel wall morphology profiles and form blood vessel structure information; The pixel intensity gradient field of the cerebral vascular imaging sequence at continuous time points is calculated. By tracking the motion trajectory of the maximum point of the gradient field, the local blood flow velocity and direction are inferred to form hemodynamic information. The vascular structure information and the hemodynamic information are aligned and fused according to the acquisition timestamp and spatial coordinates to generate the multimodal information.

3. The stroke early warning method based on cerebrovascular imaging characterization group according to claim 2, characterized in that, The construction of a multi-level representation network for stroke risk based on the multimodal information includes the following steps: Using the bifurcation points and curvature extreme points of blood vessels as network anchor points, blood vessel morphology nodes are established based on the blood vessel structure information. The blood vessel morphology nodes store local blood vessel diameter, wall roughness and orientation information. In the hemodynamic information, a blood flow dynamic node is established based on the spatial position corresponding to the network anchor point. The blood flow dynamic node stores the instantaneous blood flow velocity, pulsatility index and shear force information at the anchor point. Historical brain event records are parsed from the image description field of the cerebrovascular imaging sequence, and the historical brain event records are mapped to the corresponding spatial regions to form historical brain event nodes. The historical brain event nodes store the event type, occurrence time and scope of influence. The vascular morphology node, the blood flow dynamic node, and the historical brain event node are used as the initial network nodes of the multi-level representation network.

4. The stroke early warning method based on cerebrovascular imaging characterization group according to claim 1, characterized in that, The process of driving the blood vessel morphology node and the blood flow dynamic node to perform interactive calculations to obtain spatial coupling relationships includes the following steps: For each network anchor point, calculate the average attribute value of all vascular morphology nodes within a preset radius space centered on the network anchor point to obtain a local vascular morphology summary; Calculate the mean attribute values ​​of all blood flow dynamic nodes within the same spatial range to obtain a local blood flow dynamic summary; A mutual information matrix is ​​established between the local vascular morphology summary and the local blood flow dynamics summary. The maximum mutual information value of the mutual information matrix is ​​solved by an iterative optimization algorithm. The maximum mutual information value and the corresponding feature mapping relationship are used as the spatial coupling relationship at the network anchor point. Integrate the spatial coupling relationships at all the network anchor points to form a complete spatial coupling relationship.

5. The stroke early warning method based on cerebrovascular imaging characterization group according to claim 1, characterized in that, The process of applying a temporal constraint to the spatial coupling relationship based on the historical brain event nodes to generate a time-varying risk characterization map includes the following steps: In the time dimension, the time of occurrence of events recorded in the historical brain event nodes is taken as the time origin, and a time decay window is defined. In terms of spatial dimension, the scope of influence recorded in the historical brain event nodes is the spatial domain. Within the spatiotemporal range jointly defined by the time decay window and the spatial domain, a modulation weight is introduced into the spatial coupling relationship. The modulation weight decays as time moves away from the time origin and as the spatial location moves away from the center of the spatial domain. The modulated spatial coupling relationship is projected onto a two-dimensional plane according to its spatial coordinates, and a timestamp is assigned to each coordinate point to generate the time-varying risk characterization map.

6. The stroke early warning method based on cerebrovascular imaging characterization group according to claim 5, characterized in that, The process of performing multi-scale gridded segmentation on the time-varying risk characterization map to obtain multiple local risk assessment units includes the following steps: The time-varying risk characterization map is spatially divided using a hierarchical grid partitioning strategy from coarse to fine. At the coarse-scale grid level, the time-varying risk characterization map is divided into several primary grids, and the mean and variance of the risk characteristics of all points in each primary grid are calculated. At the fine-scale grid level, the primary grids with risk characteristic variances exceeding the dynamic square threshold are further subdivided to form secondary grids. Each final grid that cannot be further subdivided is defined as a local risk assessment unit, which includes its spatial boundary, time series of internal risk characteristics, and grid level information.

7. The stroke early warning method based on cerebrovascular imaging characterization group according to claim 1, characterized in that, The process of selecting candidate early warning units that conform to the preset risk dynamics model includes the following steps: Extract the time series of risk characteristics for each local risk assessment unit; Calculate the differential entropy, sample entropy, and trend slope of the time series of the risk characteristics; The preset risk dynamics model includes entropy change threshold and slope threshold; The differential entropy and the sample entropy are compared with the entropy change threshold, and the trend slope is compared with the slope threshold. When the differential entropy and / or the sample entropy of the local risk assessment unit exceed the corresponding entropy change threshold, and the absolute value of the trend slope exceeds the slope threshold, the local risk assessment unit is determined to conform to the preset risk dynamics model and is marked as a candidate early warning unit. The construction of the preset risk dynamics model includes the following steps: Collect cerebrovascular imaging data and corresponding stroke event records from historical stroke cases to form a training dataset; Risk feature time series for each case were extracted from the training dataset, and the risk feature time series were generated based on indicators of cerebral vascular morphological changes and hemodynamic changes. Calculate the differential entropy, sample entropy, and trend slope for each of the aforementioned risk feature time series; Based on the stroke event records, the cases were divided into high-risk and low-risk groups; By comparing the differential entropy, sample entropy, and trend slope distribution of the high-risk group and the low-risk group using statistical analysis methods, the entropy change threshold and slope threshold were determined respectively. The entropy change threshold and slope threshold are integrated into a risk dynamics model.

8. The stroke early warning method based on cerebrovascular imaging characterization group according to claim 1, characterized in that, The process of aggregating the features of all candidate early warning units to generate a global risk evolution path includes the following steps: All candidate early warning units are sorted according to the start time of their risk characteristic time series; Between adjacent candidate early warning units, calculate the distance between their spatial center points and the temporal correlation of their risk characteristics; Based on the distance and the correlation, a risk propagation link describing the spatiotemporal relationship between the candidate early warning units is constructed; Connect all the aforementioned risk propagation links in chronological order to form a global risk evolution path that runs through the entire time series. This global risk evolution path records the origin, propagation direction, and intensity evolution of the risk.

9. The stroke early warning method based on cerebrovascular imaging characterization group according to claim 8, characterized in that, The step of tracing back to the cerebrovascular imaging sequence based on the global risk evolution path and identifying the set of high-risk image frames includes the following steps: Extract the start and end times of all risk propagation links in the global risk evolution path; Map all time points back to the acquisition timeline of the cerebral vascular imaging sequence; In the cerebral vascular imaging sequence, the image frame corresponding to each mapped time point is located; The image frame and all adjacent image frames within the preset time window before and after it are included together to form the high-risk image frame set.

10. The stroke early warning method based on cerebrovascular imaging characterization group according to claim 1, characterized in that, The process of performing feature decoupling and reconstruction on the high-risk image frame set to output the final stroke risk warning signal includes the following steps: For each image frame in the high-risk image frame set, extract the structural features related to the blood vessel morphology node and the dynamic features related to the blood flow dynamic node. Principal component analysis is performed on the structural features and the dynamic features to achieve feature decoupling and obtain mutually independent feature components. Only feature components whose variance contribution rate exceeds the contribution rate threshold are selected. The selected feature components are weighted and recombined according to their chronological order of appearance in the global risk evolution path to generate a comprehensive risk index sequence. A peak detection algorithm is applied to the comprehensive risk index sequence to integrate the time, spatial location, and peak intensity corresponding to the peak points that exceed the risk peak threshold into the final stroke risk warning signal.

Images