Intelligent Early Screening and Warning Method and System for Acute Stroke Based on Deep Learning

Abstract

The present invention provides an early intelligent screening and warning method and system for acute stroke based on deep learning, which relates to the field of medical imaging and artificial intelligence technology, including obtaining multi-phase angiography images of the patient's brain through a pulsed dynamic imaging scheme, extracting hemodynamic characteristics using a dual-stream feature extraction network, and constructing a vascular lesion risk distribution map in combination with a vascular structure analysis model to mark potential embolism risk areas. Further, combining tissue perfusion parameters obtained by multispectral imaging and brain tissue function parameters obtained by electroencephalogram signals, an adaptive threshold segmentation model is used to identify the ischemic penumbra area. Finally, based on the ischemic penumbra area, tissue perfusion parameters, and brain tissue function parameters, a hierarchical decision model with a spatiotemporal attention mechanism is used to calculate the risk level and generate a corresponding clinical intervention plan. The present invention can achieve early and accurate screening and risk warning of acute stroke, and improve the efficiency of clinical intervention.

Description

Technical Field

[0001] The present invention relates to a technology, and in particular, to an early intelligent screening and warning method and system for acute stroke based on deep learning. Background Art

[0002] Acute stroke is a neurological emergency with sudden onset and atypical early symptoms. Clinically, it often relies on imaging means for diagnosis. Existing imaging analysis methods mainly rely on doctors' subjective judgments, lacking systematic and automated screening and warning capabilities, being easily affected by doctors' experience and observation angles, which may lead to inaccurate identification or omission of lesion areas. In addition, traditional imaging analysis techniques are difficult to accurately depict cerebral hemodynamic characteristics, unable to comprehensively evaluate the risk level of vascular lesions, and having limited warning capabilities for potential high-risk patients.

[0003] In recent years, deep learning technology has made remarkable progress in the field of medical image analysis. Its advantages in feature extraction, pattern recognition, and classification provide new possibilities for early stroke screening. However, existing research mainly focuses on static image analysis, lacking in-depth exploration of hemodynamic characteristics and being unable to effectively combine vascular structure, blood flow time series, and functional parameters for comprehensive evaluation. In addition, the current intelligent analysis system has not formed a complete chain from image data acquisition, vascular lesion risk analysis to intervention plan recommendation, and is difficult to meet the clinical needs of rapid screening and early intervention for acute stroke.

[0004] Therefore, there is an urgent need to propose an intelligent analysis method that integrates vascular structure, blood flow characteristics, and functional parameters to achieve automated and high-precision stroke risk assessment and warning. Summary of the Invention

[0005] An embodiment of the present invention provides an early intelligent screening and warning method and system for acute stroke based on deep learning, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiment of the present invention,

[0007] There is provided an early intelligent screening and warning method for acute stroke based on deep learning, including:

[0008] Trigger a scanning sequence to collect multi-phase angiography images within the cardiac cycle through a pulsed dynamic imaging scheme, perform hemodynamic analysis on the multi-phase angiography images to obtain blood flow velocity vector field data, construct a blood flow feature tensor based on the blood flow velocity vector field data, input the blood flow feature tensor into a dual-stream feature extraction network, wherein extract vascular structure features through a spatial feature branch and extract blood flow time series features through a temporal feature branch, and fuse the vascular structure features and the blood flow time series features to generate a hemodynamic feature map;

[0009] Process the hemodynamic feature map through a directional filter to obtain an enhanced vascular image. Use the minimum spanning tree algorithm to reconstruct the vascular topology data. Input the vascular topology data into a recursive neural tensor network to calculate the risk values of vascular branch nodes and generate a vascular lesion risk distribution map. Mark the potential embolism risk area based on the vascular lesion risk distribution map;

[0010] Use a multispectral imaging device to obtain the tissue perfusion parameters of the potential embolism risk area. Collect the electroencephalogram signals of the potential embolism risk area to obtain the brain tissue function parameters. Perform adaptive threshold segmentation on the tissue perfusion parameters and the brain tissue function parameters to obtain the ischemic penumbra area. Calculate the area value of the ischemic penumbra area. Input the area value of the ischemic penumbra area, the tissue perfusion parameters, and the brain tissue function parameters into a hierarchical decision-making model to calculate the risk level value, and generate a clinical intervention plan according to the risk level value.

[0011] In an alternative embodiment,

[0012] Trigger a scan sequence within the cardiac cycle through a pulsed dynamic imaging scheme to acquire multi-phase angiography images. Perform hemodynamic analysis on the multi-phase angiography images to obtain blood flow velocity vector field data. Construct a blood flow feature tensor based on the blood flow velocity vector field data, including:

[0013] Collect the patient's electrocardiogram signal to construct a pulsed dynamic imaging scheme. The pulsed dynamic imaging scheme uses a dual-domain sparse sampling model for undersampling in the frequency domain and the time domain. Based on the patient's electrocardiogram signal, construct multiple trigger windows within the cardiac cycle. Use a spiral angle increasing ray sampling strategy to trigger the scan sequence within the multiple trigger windows to obtain non-uniform grid data of multi-phase angiography;

[0014] Perform iterative reconstruction on the non-uniform grid data, and optimize the data consistency in both the spatial domain and the time domain in each iteration. Finally, obtain multi-phase angiography images with high spatio-temporal resolution;

[0015] Input the multi-phase angiography images into a multi-scale hemodynamic feature extraction network. The multi-scale hemodynamic feature extraction network includes an adaptive deformable convolutional layer, a spatial attention module, and a recurrent feature extraction unit. Among them, the adaptive deformable convolutional layer extracts local blood flow features to obtain vascular feature data. The spatial attention module determines the vascular branch connection relationship based on the vascular feature data to obtain vascular network data. The recurrent feature extraction unit performs temporal analysis on the vascular network data based on the pressure wave propagation constraint, and finally obtains blood flow velocity vector field data;

[0016] The phase enhancement technique is used to calculate the phase difference information representing the magnitude and direction of blood flow velocity between adjacent temporal images. Based on the phase difference information, non-linear correction is performed on the blood flow velocity vector field data. According to the corrected blood flow velocity vector field data, a blood flow feature tensor including blood flow vortex distribution characteristics, blood vessel wall stress distribution characteristics, blood flow complexity characteristics, and pressure wave propagation characteristics is constructed.

[0017] In an alternative embodiment,

[0018] The blood flow feature tensor is input into a two-stream feature extraction network. Among them, the vascular structure features are extracted through the spatial feature branch, and the blood flow temporal features are extracted through the temporal feature branch. The fusion of the vascular structure features and the blood flow temporal features to generate the hemodynamic feature map includes:

[0019] The blood flow feature tensor is input into a two-stream feature extraction network. The two-stream feature extraction network includes a spatial feature branch and a temporal feature branch. The blood flow feature tensor contains blood flow vortex distribution characteristics, blood vessel wall stress distribution characteristics, blood flow complexity characteristics, and pressure wave propagation characteristics;

[0020] The spatial feature branch generates deformable convolution kernel parameters based on the vascular morphology information in the blood flow feature tensor. The deformable convolution kernel groups are respectively subjected to multi-scale convolution operations with the blood flow vortex distribution characteristics, blood vessel wall stress distribution characteristics, blood flow complexity characteristics, and pressure wave propagation characteristics to obtain a local feature map set, and channel dimension weighting is performed on the local feature map set to obtain vascular structure features;

[0021] The temporal feature branch performs phase analysis on the blood flow vortex distribution characteristics, blood vessel wall stress distribution characteristics, blood flow complexity characteristics, and pressure wave propagation characteristics in the blood flow feature tensor respectively. The instantaneous phase and instantaneous amplitude of each feature component are extracted through Hilbert transform, the phase difference between adjacent time windows is calculated to obtain a set of phase change curves, a phase consistency matrix is constructed based on the set of phase change curves, and the instantaneous amplitude is selectively enhanced and reconstructed to obtain blood flow temporal features;

[0022] The mutual information between the vascular structure features and the blood flow temporal features is calculated to obtain a feature correlation matrix. Based on the feature correlation matrix, a bidirectional attention weight is constructed, the vascular structure features and the blood flow temporal features are adaptively enhanced respectively, and the enhanced features are fused through a skip connection structure to obtain a hemodynamic feature map.

[0023] In an alternative embodiment,

[0024] The hemodynamic feature map is processed by a directional filter to obtain an enhanced vascular image. The reconstruction of the vascular topology structure data using the minimum spanning tree algorithm includes:

[0025] Extract the vascular morphological features in the hemodynamic feature map, determine the scale range of the Gaussian second derivative filter based on the vascular morphological features, and rotate the Gaussian second derivative filter within a preset angle range to generate a directional filter;

[0026] Use the directional filter to filter the hemodynamic feature map to obtain multi-scale direction response values, determine the weighting coefficients according to the blood vessel orientation in the hemodynamic feature map, perform weighted processing on the multi-scale direction response values, obtain the weighted maximum response value and the corresponding optimal direction parameter and optimal scale parameter, and generate an enhanced blood vessel image;

[0027] Perform local maximum suppression processing on the enhanced blood vessel image to obtain a set of candidate points for the blood vessel centerline, construct an undirected graph, use the minimum spanning tree algorithm to process the undirected graph, determine the dynamic weight threshold according to the blood vessel density in the enhanced blood vessel image, and remove the node connection relationships greater than the dynamic weight threshold to obtain an initial blood vessel topology structure;

[0028] Extract the blood flow vortex distribution, blood vessel wall stress distribution, blood flow complexity, and pressure wave propagation characteristics on the path corresponding to the connection edges of the initial blood vessel topology structure, calculate the gradient change value to construct a feature consistency index;

[0029] Verify the connection relationships in the initial blood vessel topology structure according to the feature consistency index, identify the connection relationships with the gradient change value greater than the preset gradient threshold as abnormal and correct or delete them to complete the reconstruction of blood vessel connectivity and generate blood vessel topology structure data.

[0030] In an alternative embodiment,

[0031] Input the blood vessel topology structure data into a recursive neural tensor network to calculate the risk values of blood vessel branch nodes and generate a blood vessel lesion risk distribution map. Marking potential embolism risk areas based on the blood vessel lesion risk distribution map includes:

[0032] Extract the feature vectors of blood vessel branch nodes from the blood vessel topology structure data, calculate the topological relationship matrix between nodes, and the topological relationship matrix records the superior-subordinate and peer connection relationships between nodes. Construct a tree-shaped computational graph based on the feature vectors and the topological relationship matrix, and each node in the graph contains the corresponding feature vector and topological relationship data;

[0033] Input the tree-shaped computational graph into a recursive neural tensor network, and the recursive neural tensor network includes a third-order tensor calculation unit. The third-order tensor calculation unit performs tensor product operations on the feature vectors of adjacent nodes to obtain the structural correlation data between nodes and the feature association strength between nodes;

[0034] Construct an attention weight matrix based on the structural correlation data between nodes and the intensity of node feature association. The attention weight matrix includes a feature dimension weight component and a node association weight component. Weight the feature vectors according to the feature dimension weight component and the node association weight component to generate weighted feature vectors that fuse local structural information;

[0035] Use a non-linear activation function to map the weighted feature vectors to obtain initial risk assessment values. Input the initial risk assessment values into a recursive calculation module. The recursive calculation module calculates from the terminal branch nodes in a bottom-up manner, fuses the feature information and risk information of lower-level nodes based on the feature dimension weight component and the node association weight component, and updates the risk assessment values of upper-level nodes until all nodes are traversed to generate the final risk assessment value;

[0036] Normalize the final risk assessment value to generate a risk distribution map. Partition the vascular structure according to the risk value magnitudes in the risk distribution map, and mark the regions with risk values higher than the set risk threshold as potential embolism risk regions.

[0037] In an alternative embodiment,

[0038] Use a multispectral imaging device to obtain tissue perfusion parameters of the potential embolism risk region, collect electroencephalogram signals of the potential embolism risk region to obtain brain tissue function parameters, and perform adaptive threshold segmentation on the tissue perfusion parameters and the brain tissue function parameters to obtain the ischemic penumbra region, including:

[0039] Use a multispectral imaging device to perform continuous wavelength scanning within a preset wavelength range through a liquid crystal tunable filter, and take the average value after accumulating and collecting multiple frames of images at each wavelength point to obtain a spectral data cube;

[0040] Construct an inverse model for tissue optical parameters based on the incident light intensity, transmitted light intensity, and tissue thickness. Input the spectral data cube into the inverse model for tissue optical parameters, and use the least squares fitting method to calculate the concentration of oxyhemoglobin and deoxyhemoglobin. Calculate the tissue blood oxygen saturation as the ratio of the concentration of oxyhemoglobin to the sum of the concentrations of oxyhemoglobin and deoxyhemoglobin, and calculate the tissue reflectance ratio at a preset characteristic wavelength to obtain the tissue perfusion index;

[0041] Collect the electroencephalogram signals of the potential embolism risk region for band-pass filtering processing. Use the wavelet packet decomposition method to decompose the filtered electroencephalogram signals into multiple preset frequency bands, calculate the ratio of the energy of each preset frequency band to the total energy to obtain the relative power spectrum, convert the filtered electroencephalogram signals into a symbol sequence and then statistically obtain the probability distribution of different permutation patterns to obtain the permutation entropy value, and combine the relative power spectrum and the permutation entropy value in a preset order to obtain the brain tissue function parameters;

[0042] Form a feature vector with the tissue perfusion parameters and brain tissue function parameters, calculate the Euclidean distance from each sample point in the feature vector to each cluster center, calculate the membership degree of each sample point to each cluster center based on the Euclidean distance, update the cluster centers according to the membership degree, and repeat the iteration until the difference between the objective function values of two adjacent iterations is less than a preset convergence threshold to obtain the final cluster centers;

[0043] Classify the feature vector based on the final cluster centers to obtain the spatial distributions of the ischemic region, penumbra region, and normal region, extract the set of pixel points in the penumbra region for edge detection to obtain a closed boundary contour, and mark the region within the closed boundary contour as the ischemic penumbra region.

[0044] In an alternative embodiment,

[0045] Calculate the area value of the ischemic penumbra region, input the area value of the ischemic penumbra region, tissue perfusion parameters, and brain tissue function parameters into a hierarchical decision-making model to calculate the risk level value, and generate a clinical intervention plan based on the risk level value, including:

[0046] Perform connected component labeling on the ischemic penumbra region and construct a neuron connection graph, calculate the topological relationship matrix between neurons, count the total number of pixel points in the ischemic penumbra region, and convert the total number of pixel points according to the spatial resolution parameter to obtain the area value of the ischemic penumbra region;

[0047] Construct a hierarchical decision-making model with a spatio-temporal attention mechanism, including a spatial attention module and a temporal attention module; the spatial attention module uses wavelet multi-resolution analysis to decompose the tissue perfusion parameters and the topological relationship matrix into spatial feature matrices at multiple scales, and extracts the dynamic change features of the tissue perfusion parameters through a recurrent neural network to calculate the spatial weight coefficients; the temporal attention module uses a non-linear time series decomposition method to decompose the brain tissue function parameters into feature sequences at different time scales, and extracts the time series change features through an attention autoencoder to calculate the time weight coefficients;

[0048] Construct a multi-modal risk assessment feature set based on the area value of the ischemic penumbra region, weighted tissue perfusion parameters, and weighted brain tissue function parameters, input it together with the neuron connection graph into a graph convolutional neural network to obtain a fused feature vector, and input the fused feature vector into a residual network to obtain the risk level value;

[0049] Predict the risk evolution trend based on the risk level value, select clinical intervention rules by combining historical intervention effect data, generate a clinical intervention decision tree, prioritize the decision tree nodes based on the risk evolution trend, and obtain a prioritized clinical intervention plan with intervention effect and resource consumption as evaluation indicators.

[0050] In the second aspect of the embodiments of the present invention,

[0051] Provide an early intelligent screening and warning system for acute stroke based on deep learning, including:

[0052] The first unit is used to trigger a scanning sequence to collect multi-phase angiography images during the cardiac cycle through a pulsed dynamic imaging scheme, perform hemodynamic analysis on the multi-phase angiography images to obtain blood flow velocity vector field data, construct a blood flow feature tensor based on the blood flow velocity vector field data, and input the blood flow feature tensor into a dual-stream feature extraction network, where the vascular structure features are extracted through the spatial feature branch, the blood flow temporal features are extracted through the temporal feature branch, and the vascular structure features and the blood flow temporal features are fused to generate a hemodynamic feature map;

[0053] The second unit is used to process the hemodynamic feature map through a directional filter to obtain an enhanced vascular image, reconstruct the vascular topology structure data using the minimum spanning tree algorithm, input the vascular topology structure data into a recursive neural tensor network to calculate the risk value of vascular branch nodes to generate a vascular lesion risk distribution map, and mark the potential embolism risk area based on the vascular lesion risk distribution map;

[0054] The third unit is used to obtain tissue perfusion parameters of the potential embolism risk area using a multi-spectral imaging device, collect electroencephalogram signals of the potential embolism risk area to obtain brain tissue function parameters, perform adaptive threshold segmentation on the tissue perfusion parameters and the brain tissue function parameters to obtain the ischemic penumbra area, calculate the area value of the ischemic penumbra area, and input the area value of the ischemic penumbra area, the tissue perfusion parameters, and the brain tissue function parameters into a hierarchical decision model to calculate the risk level value, and generate a clinical intervention plan based on the risk level value.

[0055] In the third aspect of the embodiments of the present invention,

[0056] Provide an electronic device, including:

[0057] A processor;

[0058] A memory for storing instructions executable by the processor;

[0059] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0060] In the fourth aspect of the embodiments of the present invention,

[0061] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0062] In this embodiment, the pulsed dynamic imaging scheme and hemodynamic analysis can be used to more accurately capture the changes in cerebral blood flow, and analyze the vascular structure and timing characteristics, so as to more accurately identify potential embolic risk areas and achieve early warning of stroke. Compared with traditional methods, the present invention can detect signs of stroke earlier and improve the efficiency of early diagnosis. Through the vascular structure analysis model and the recursive neural tensor network, the risk of vascular lesions can be quantitatively evaluated and a vascular lesion risk distribution map can be generated. Combined with the tissue perfusion parameters and brain tissue function parameters obtained by multispectral imaging and electroencephalography technology, the present invention can more comprehensively evaluate the damage of the ischemic penumbra area, so as to more accurately evaluate the risk level of stroke. According to the area value, tissue perfusion parameters, brain tissue function parameters and risk level values ​​of the ischemic penumbra area, a personalized clinical intervention plan can be generated, and the priority of the plan can be adjusted according to the risk level. This personalized intervention strategy can better guide clinical treatment, improve treatment effects, and gain precious rescue time for patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flow chart of an early intelligent screening and warning method for acute stroke based on deep learning according to an embodiment of the present invention;

[0064] Figure 2 This is a comparison chart of the accuracy of blood vessel connectivity reconstruction in an embodiment of the present invention;

[0065] Figure 3 This is a diagram showing the effect of vascular risk distribution according to an embodiment of the present invention;

[0066] Figure 4 A schematic diagram of the structure of an early intelligent screening and warning system for acute stroke based on deep learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] The technical solution of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0069] Figure 1 The following is a schematic flow chart of the method for early intelligent screening and warning of acute stroke based on deep learning according to the embodiments of the present invention, as Figure 1 shown. The method includes:

[0070] Trigger a scanning sequence within the cardiac cycle through a pulsed dynamic imaging scheme to collect multi-phase angiography images, perform hemodynamic analysis on the multi-phase angiography images to obtain blood flow velocity vector field data, construct a blood flow feature tensor based on the blood flow velocity vector field data, and input the blood flow feature tensor into a dual-stream feature extraction network, where vascular structure features are extracted through the spatial feature branch, and blood flow temporal features are extracted through the temporal feature branch, and the vascular structure features and the blood flow temporal features are fused to generate a hemodynamic feature map;

[0071] Process the hemodynamic feature map through a directional filter to obtain an enhanced vascular image, reconstruct vascular topology structure data using the minimum spanning tree algorithm, input the vascular topology structure data into a recursive neural tensor network to calculate the risk value of vascular branch nodes to generate a vascular lesion risk distribution map, and mark potential embolism risk areas based on the vascular lesion risk distribution map;

[0072] Use a multispectral imaging device to obtain tissue perfusion parameters of the potential embolism risk area, collect electroencephalogram signals of the potential embolism risk area to obtain brain tissue function parameters, perform adaptive threshold segmentation on the tissue perfusion parameters and the brain tissue function parameters to obtain the ischemic penumbra area, calculate the area value of the ischemic penumbra area, input the area value of the ischemic penumbra area, the tissue perfusion parameters and the brain tissue function parameters into a hierarchical decision-making model to calculate the risk level value, and generate a clinical intervention plan according to the risk level value.

[0073] In an alternative embodiment, triggering a scanning sequence within the cardiac cycle through a pulsed dynamic imaging scheme to collect multi-phase angiography images, and performing hemodynamic analysis on the multi-phase angiography images to obtain blood flow velocity vector field data, and constructing a blood flow feature tensor based on the blood flow velocity vector field data includes:

[0074] Collect the electrocardiogram signal of the patient to construct a pulsed dynamic imaging scheme. The pulsed dynamic imaging scheme uses a dual-domain sparse sampling model for undersampling in the frequency domain and the time domain, constructs multiple trigger windows within the cardiac cycle based on the electrocardiogram signal of the patient, and uses a spiral angle increasing ray sampling strategy to trigger the scanning sequence within the multiple trigger windows to obtain non-uniform grid data of multi-phase angiography;

[0075] Iteratively reconstruct the non-uniform grid data, optimizing the data consistency in both the spatial domain and the temporal domain simultaneously in each iteration, and finally obtaining multi-phase angiography images with high spatio-temporal resolution;

[0076] Input the multi-phase angiography images into a multi-scale hemodynamic feature extraction network, which includes an adaptive deformable convolutional layer, a spatial attention module, and a recurrent feature extraction unit. Among them, the adaptive deformable convolutional layer extracts local blood flow features to obtain vascular feature data, the spatial attention module determines the vascular branch connection relationship based on the vascular feature data to obtain vascular network data, and the recurrent feature extraction unit performs temporal analysis on the vascular network data based on the pressure wave propagation constraint, and finally obtains blood flow velocity vector field data;

[0077] Adopt a phase enhancement technique to calculate the phase difference information representing the magnitude and direction of blood flow velocity between adjacent phase images, perform non-linear correction on the blood flow velocity vector field data based on the phase difference information, and construct a blood flow feature tensor including blood flow vortex distribution characteristics, vascular wall stress distribution characteristics, blood flow complexity characteristics, and pressure wave propagation characteristics according to the corrected blood flow velocity vector field data.

[0078] Exemplarily, first collect the electrocardiogram (ECG) signal of the patient, which includes a pre-wave signal, a ventricular depolarization wave group signal, and a post-wave signal. Use the ECG signal to construct a pulsed dynamic imaging scheme. This scheme uses a dual-domain sparse sampling model for undersampling in the frequency domain space and the temporal domain, reducing the amount of data acquisition. According to the pre-wave, ventricular depolarization wave group, and post-wave in the ECG signal, set multiple trigger windows within the cardiac cycle. Within each trigger window, use a spiral angle increment ray sampling strategy to trigger the scan sequence to collect data. For example, within one cardiac cycle, the pre-wave is the first trigger window, the ventricular depolarization wave group is the second trigger window, and the post-wave is the third trigger window. In the first trigger window, the spiral ray angle starts from 0 degrees for data collection; in the second trigger window, the spiral ray angle starts from 120 degrees for data collection; in the third trigger window, the spiral ray angle starts from 240 degrees for data collection. This sampling strategy can obtain non-uniform grid data of multi-phase angiography. For example, within one cardiac cycle, non-uniform grid data of three phases are collected, and the data of each phase contains projection data with different spiral angles.

[0079] The obtained non-uniform grid data is input into a directional transform iterative reconstruction model for reconstruction. This model consists of four main modules: a non-uniform Fourier operation module for processing non-uniform sampling data, a directional transform module for extracting blood vessel edge and direction information, a temporal gradient calculation module for constructing temporal continuity constraints, and a matrix decomposition module for reducing image noise. The directional transform adopts, for example, Curvelet transform or Shearlet transform, etc. The temporal gradient calculation module constructs temporal continuity constraints by calculating the pixel value differences between adjacent phase images. The matrix decomposition module adopts methods such as principal component analysis or singular value decomposition, etc. The non-uniform grid data is iteratively reconstructed through this model. In each iteration, the data consistency in both the spatial domain and the temporal domain is optimized simultaneously, and finally, a multi-phase angiography image with high spatio-temporal resolution is obtained. For example, after iterative reconstruction, angiography images of three phases are obtained, with the spatial resolution of each image being 512x512 pixels and the temporal resolution being three different time points within the cardiac cycle.

[0080] The reconstructed multi-phase angiography image is input into a multi-scale hemodynamic feature extraction network. This network includes an adaptive deformable convolutional layer, a spatial attention module, and a recurrent feature extraction unit. The adaptive deformable convolutional layer extracts local blood flow features to obtain blood vessel feature data. The spatial attention module determines the blood vessel branch connection relationship based on the blood vessel feature data to obtain blood vessel network data. The recurrent feature extraction unit performs temporal analysis on the blood vessel network data based on the pressure wave propagation constraint, and finally obtains blood flow velocity vector field data. For example, the adaptive deformable convolutional layer extracts features such as the diameter and curvature of blood vessels, the spatial attention module identifies blood vessel branch points and connection relationships, and the recurrent feature extraction unit analyzes the changes in blood vessel network data at different phases, and finally obtains the magnitude and direction of blood flow velocity at each pixel point.

[0081] The phase enhancement technique is used to calculate the phase difference information representing the magnitude and direction of blood flow velocity between adjacent phase images. Based on the phase difference information, non-linear correction is performed on the blood flow velocity vector field data to eliminate phase aliasing. According to the corrected blood flow velocity vector field data, a blood flow feature tensor including blood flow vortex distribution characteristics, blood vessel wall stress distribution characteristics, blood flow complexity characteristics, and pressure wave propagation characteristics is constructed. For example, the blood flow vortex distribution characteristics are calculated according to the curl of the blood flow velocity vector field, the blood vessel wall stress distribution characteristics are calculated according to the blood flow velocity gradient, the blood flow complexity characteristics are calculated according to the change degree of the blood flow velocity vector field, and the pressure wave propagation characteristics are calculated according to the changes in the blood flow velocity vector field at different phases.

[0082] In this embodiment, through the pulsed dynamic imaging scheme, the scanning sequence can be accurately triggered within the cardiac cycle to obtain multi-phase angiographic images with high spatio-temporal resolution; the blood flow velocity vector field data is extracted through hemodynamic analysis, so as to realize in-depth analysis of blood flow characteristics. The dual-domain sparse sampling and iterative reconstruction model is adopted to optimize data consistency and improve image quality; the multi-scale hemodynamic feature extraction network is used to effectively extract vascular features and blood flow information, and further accurately construct the blood flow feature tensor. The application of the phase enhancement technology eliminates phase aliasing and improves the accuracy of the blood flow velocity vector field, providing a more accurate analysis tool for the diagnosis and treatment of vascular diseases.

[0083] In an alternative embodiment, the blood flow feature tensor is input into a two-stream feature extraction network, wherein the vascular structure features are extracted through the spatial feature branch, and the blood flow temporal features are extracted through the temporal feature branch. The fusion of the vascular structure features and the blood flow temporal features to generate a hemodynamic feature map includes:

[0084] The blood flow feature tensor is input into a two-stream feature extraction network, the two-stream feature extraction network includes a spatial feature branch and a temporal feature branch, and the blood flow feature tensor includes blood flow vortex distribution features, vascular wall stress distribution features, blood flow complexity features, and pressure wave propagation features;

[0085] The spatial feature branch generates deformable convolution kernel parameters based on the vascular morphology information in the blood flow feature tensor, and performs multi-scale convolution operations on the deformable convolution kernel groups respectively with the blood flow vortex distribution features, vascular wall stress distribution features, blood flow complexity features, and pressure wave propagation features to obtain a local feature map set, and performs channel dimension weighting on the local feature map set to obtain vascular structure features;

[0086] The temporal feature branch performs phase analysis on the blood flow vortex distribution features, vascular wall stress distribution features, blood flow complexity features, and pressure wave propagation features in the blood flow feature tensor respectively, extracts the instantaneous phase and instantaneous amplitude of each feature component through Hilbert transform, calculates the phase difference between adjacent time windows to obtain a set of phase change curves, constructs a phase consistency matrix based on the set of phase change curves, selectively enhances the instantaneous amplitude and reconstructs it to obtain blood flow temporal features;

[0087] Calculate the mutual information between the vascular structure features and the blood flow temporal features to obtain a feature correlation matrix, construct a bidirectional attention weight based on the feature correlation matrix, adaptively enhance the vascular structure features and the blood flow temporal features respectively, and fuse the enhanced features through a skip connection structure to obtain a hemodynamic feature map.

[0088] Exemplarily, first, a blood flow feature tensor is obtained. This tensor contains four key dimensions of blood flow information: blood flow vortex distribution characteristics, blood vessel wall stress distribution characteristics, blood flow complexity characteristics, and pressure wave propagation characteristics. For example, computational fluid dynamics simulation technology can be used to simulate a specific blood vessel model to obtain data such as velocity fields and pressure fields at different time points and spatial positions. Then, based on these data, characteristics such as blood flow vortices, blood vessel wall stress, blood flow complexity, and pressure wave propagation are calculated, and these characteristics are organized into a multi-dimensional tensor. Assuming there are 100 time sampling points and 256x256 spatial sampling points, the dimension of the blood flow feature tensor can be 100x256x256x4.

[0089] The blood flow feature tensor is input into a two-stream feature extraction network. This network consists of a spatial feature branch and a temporal feature branch, which are used to extract blood vessel structure features and blood flow temporal features respectively.

[0090] In the spatial feature branch, first, the deformation parameters of the deformable convolution kernel are generated using the blood vessel morphology information in the blood flow feature tensor. Specifically, the sampling point position offsets are calculated based on the blood vessel cross-sectional change rate and local curvature. For example, in a region with a larger blood vessel cross-sectional change rate, the sampling point position offset is also larger, enabling the convolution kernel to pay more attention to the changes in blood vessel morphology. Assuming that within a local region, the blood vessel cross-sectional diameter changes from 2 pixels to 4 pixels, the sampling point position offset in this region can be set to 1 pixel. The calculation method for local curvature is similar, and in a region with a larger curvature, the sampling point position offset is also larger. Then, based on these deformation parameters, the standard convolution kernel is adaptively modulated to obtain a group of deformable convolution kernels. For example, a 3x3 standard convolution kernel, after being modulated by the deformation parameters, the positions of its sampling points may no longer be in a regular grid-like distribution but are adaptively adjusted according to the blood vessel morphology. The obtained group of deformable convolution kernels are respectively subjected to multi-scale convolution operations with the blood flow vortex distribution characteristics, blood vessel wall stress distribution characteristics, blood flow complexity characteristics, and pressure wave propagation characteristics to obtain a local feature map set. Finally, channel dimension weighting is performed on the local feature map set to obtain blood vessel structure features. For example, different weights can be assigned according to the importance of different features. For example, the weight of the blood flow vortex distribution characteristic is 0.4, the weight of the blood vessel wall stress distribution characteristic is 0.3, the weight of the blood flow complexity characteristic is 0.2, and the weight of the pressure wave propagation characteristic is 0.1.

[0091] In the temporal feature branch, phase analysis is respectively performed on the blood flow vortex distribution feature, blood vessel wall stress distribution feature, blood flow complexity feature, and pressure wave propagation feature in the blood flow feature tensor. The instantaneous phase and instantaneous amplitude of each feature component are extracted through Hilbert transform. Then, the phase differences between adjacent time windows are calculated to obtain a group of phase change curves. For example, the phase difference between every two adjacent time points can be calculated to obtain a series of phase change curves. A phase consistency matrix is constructed based on the group of phase change curves, and this matrix is used to characterize the temporal periodicity of each feature component. For example, if the phase change curves of two feature components are highly similar, the corresponding element values in the phase consistency matrix are larger. The instantaneous amplitude is selectively enhanced and reconstructed using the phase consistency matrix to obtain the blood flow temporal features. For example, for feature components with higher phase consistency, their instantaneous amplitudes are enhanced, while for feature components with lower phase consistency, their instantaneous amplitudes are weakened.

[0092] Calculate the mutual information between the blood vessel structure features and the blood flow temporal features to obtain a feature correlation matrix. Based on the feature correlation matrix, a bidirectional attention weight is constructed, and this weight is used to characterize the degree of mutual enhancement between the structure features and the temporal features. For example, if there is a strong correlation between the blood vessel structure features and the blood flow temporal features, the corresponding attention weight is larger.

[0093] The blood vessel structure features and the blood flow temporal features are respectively adaptively enhanced using the bidirectional attention weight. For example, the attention weight is multiplied by the corresponding feature to achieve the adaptive enhancement of the feature. The enhanced features are fused through a skip connection structure to obtain the final hemodynamic feature map.

[0094] In this embodiment, the existing hemodynamic feature extraction methods usually rely on fixed convolution kernels for feature capture, which are difficult to adapt to the influence of complex vascular morphologies on blood flow characteristics, resulting in limited accuracy of spatial feature extraction and restricting the integrity and accuracy of the final feature representation. This application optimizes the spatial information and temporal information of the blood flow feature tensor through a dual-stream feature extraction network. During the spatial feature extraction process, convolution kernel deformation parameters are generated based on vascular morphology information, enabling deformable convolution kernels to adapt to the local structures of different blood flow characteristics, achieving more accurate feature capture, and enhancing key vascular structure features through channel dimension weighting. In terms of temporal feature extraction, the instantaneous phase and instantaneous amplitude of blood flow features are analyzed through Hilbert transform, a phase consistency matrix is constructed to selectively enhance blood flow temporal features, and the expression ability of temporal information is improved. The mutual information between vascular structure features and blood flow temporal features is calculated through a feature correlation matrix, adaptively enhanced using a bidirectional attention mechanism, and features are fused through a skip connection structure to achieve a more discriminative hemodynamic feature representation. The solution of this application introduces deformable convolution to adapt to local changes in different blood flow characteristics compared with the prior art, improving the accuracy of spatial features, while using the instantaneous phase analysis method to enhance the expression ability of temporal features. In addition, the feature correlation matrix and bidirectional attention mechanism are innovatively introduced, enabling closer fusion of spatial and temporal features and enhancing the overall feature expression effect. It can extract hemodynamic features more comprehensively and accurately, providing more reliable data support for hemodynamic analysis.

[0095] In an alternative embodiment, the hemodynamic feature map is processed by a directional filter to obtain an enhanced vascular image, and the reconstruction of the vascular topology structure data using the minimum spanning tree algorithm includes:

[0096] Extract the vascular morphology features in the hemodynamic feature map, determine the scale range of the Gaussian second derivative filter based on the vascular morphology features, and rotate the Gaussian second derivative filter within a preset angle range to generate a directional filter;

[0097] Filter the hemodynamic feature map using the directional filter to obtain multi-scale direction response values, determine the weighting coefficient according to the blood vessel orientation in the hemodynamic feature map, perform weighted processing on the multi-scale direction response values, obtain the weighted maximum response value and the corresponding optimal direction parameter and optimal scale parameter, and generate an enhanced vascular image;

[0098] Perform local maximum suppression processing on the enhanced vascular image to obtain a set of candidate points for the vascular centerline, construct an undirected graph, process the undirected graph using the minimum spanning tree algorithm, determine a dynamic weight threshold according to the vascular density in the enhanced vascular image, and remove the node connection relationships greater than the dynamic weight threshold to obtain an initial vascular topology structure;

[0099] Extract the blood flow vortex distribution, vascular wall stress distribution, blood flow complexity, and pressure wave propagation characteristics on the path corresponding to the connecting edge of the initial vascular topology structure, and calculate the gradient change value to construct a feature consistency index;

[0100] Verify the connection relationships in the initial vascular topology structure according to the feature consistency index, identify the connection relationships with the gradient change value greater than the preset gradient threshold as abnormal and correct or delete them to complete the reconstruction of vascular connectivity and generate vascular topology structure data.

[0101] Exemplarily, first, obtain a hemodynamic feature map. The hemodynamic feature map can be obtained by, for example, computational fluid dynamics (CFD) simulation or an image-based method (such as phase-contrast magnetic resonance imaging). Taking CFD simulation as an example, by establishing a patient-specific vascular geometry model, setting the physical properties of the blood (such as viscosity, density), and defining the boundary conditions (such as inlet velocity, outlet pressure), perform simulation calculations to obtain a hemodynamic feature map containing information such as velocity, pressure, and vorticity. Assume that the resolution of the obtained hemodynamic feature map is 512x512 pixels, and the numerical range is from 0 to 1, representing the magnitude of the blood flow velocity.

[0102] Next, input the hemodynamic feature map into the vascular structure analysis model. The core of this model is to use a directional filter to enhance the vascular image and use the minimum spanning tree algorithm to reconstruct vascular connectivity. The vascular structure analysis model first analyzes the vascular morphological features in the hemodynamic feature map, such as the vascular radius. Assume that the vascular radius change range is analyzed to be 2 to 10 pixels through an image processing algorithm. According to this range, determine that the scale range of the Gaussian second derivative filter is also 2 to 10 pixels to match the size of the blood vessels.

[0103] Then, generate a directional filter. Rotate the Gaussian second derivative filter at a preset angle range (such as 0 to 180 degrees) at a fixed interval (such as 1 degree) to generate multiple directional filters. Each filter corresponds to a specific direction and scale. For example, generate 180 directions, each direction corresponding to a scale of 2 to 10 pixels, for a total of 180*9 = 1620 filters.

[0104] After that, these directional filters are used to filter the hemodynamic feature map. For each pixel, the filter response values at all directions and scales are calculated to obtain the multi-scale directional response values. Suppose that according to the analysis of the blood vessel orientation in the hemodynamic feature map, it is determined that the blood vessels in the vertical direction are more prominent. Therefore, a higher weight is assigned to the filter response values in the vertical direction, for example, the weight coefficient is 1.2, and the weight coefficients in other directions are 1. The multi-scale directional response values of each pixel are weighted to obtain the weighted maximum response value and the corresponding optimal direction parameter and optimal scale parameter. For example, the maximum response value of a certain pixel is 0.8, the corresponding optimal direction is 90 degrees (vertical direction), and the optimal scale is 5 pixels. An enhanced blood vessel image is generated based on the maximum response value.

[0105] Local maximum suppression processing is performed on the enhanced blood vessel image to obtain a set of candidate points for the blood vessel centerline. For example, by comparing the response values of each pixel with its surrounding 8 pixels, if the response value of this pixel is the largest, it is marked as a candidate point for the blood vessel centerline. Suppose 1000 candidate points are obtained.

[0106] An undirected graph is constructed, and the spatial positions of each point in the set of candidate points for the blood vessel centerline are used as the node coordinates in the undirected graph. A weight function for node connection is constructed based on the spatial distance between node coordinates, the difference in direction angles determined by the optimal direction parameters, and the response intensity determined by the maximum response values. For example, the closer the distance between two nodes, the smaller the difference in direction angles, and the larger the response value, the greater the weight between them, indicating a higher possibility of connection.

[0107] The minimum spanning tree algorithm is used to process the undirected graph. The connection priority between nodes is determined according to the values of the weight function. The dynamic weight threshold is determined according to the blood vessel density in the enhanced blood vessel image. For example, the higher the blood vessel density, the smaller the dynamic weight threshold to avoid overconnection. The node connection relationships greater than the dynamic weight threshold are removed to obtain the initial blood vessel topology.

[0108] On the path corresponding to the connection edges of the initial blood vessel topology, the blood flow vortex distribution feature, blood vessel wall stress distribution feature, blood flow complexity feature, and pressure wave propagation feature are extracted from the hemodynamic feature map. For example, on a connection edge, the blood flow velocity values at every 1 pixel are extracted, and the difference between the velocity values of adjacent pixels is calculated as the blood flow complexity feature. And the gradient change values of each feature along the path are calculated. A feature consistency index is constructed according to the gradient change values. For example, if the gradient change value of the blood flow complexity feature on a connection edge is very small, it is considered that the feature consistency of this connection edge is high.

[0109] Verify the connection relationships in the initial vascular topology according to the feature consistency index. Identify the connection relationships with gradient change values greater than the preset gradient threshold as abnormal connections. For example, if the gradient change value of the blood flow complexity feature on a connection edge is greater than the preset threshold of 0.5, then this connection is considered an abnormal connection. Combine the hemodynamic feature map to correct or delete the abnormal connections. For example, if an abnormal connection spans a low-velocity area, then delete it. Complete the vascular connectivity reconstruction to generate vascular topology structure data including node position information and connection relationship information. For example, a vascular topology structure data containing 100 nodes and 90 connection edges.

[0110] Existing vascular topology reconstruction methods usually rely on traditional edge detection or statistical model-based vascular segmentation. These methods are difficult to accurately extract the vascular structure in a complex blood flow environment, especially in areas with drastic changes in hemodynamic features, which easily lead to topological structure breaks or incorrect connections. In addition, when dealing with vascular connectivity, existing methods mostly use fixed thresholds or simple morphological operations and fail to fully optimize by combining blood flow features, resulting in limited accuracy of the topological structure. This application enhances the vascular image through a directional filter and combines the minimum spanning tree algorithm for vascular topology structure reconstruction. The directional filter determines the scale range based on the vascular morphological features and generates filters in different directions through a rotated Gaussian second derivative filter, enabling it to adapt to vascular structures of different scales and directions and improving the contrast and clarity of the vascular image. Subsequently, on the enhanced vascular image, candidate points of the vascular centerline are extracted through the local maximum suppression method, and an undirected graph is constructed. The weight threshold is dynamically adjusted by combining the vascular density information to ensure that the connectivity of the vascular topology structure is more consistent with the real anatomical structure. Further, in the process of optimizing the topological structure, this application introduces hemodynamic features, combines features such as blood flow vortex distribution, vascular wall stress, blood flow complexity, and pressure wave propagation, calculates the gradient change value to construct a feature consistency index, and realizes the accurate identification and correction of abnormal connections. Compared with existing methods, this application not only optimizes the vascular enhancement process based on vascular morphological features but also combines hemodynamic features to verify and correct the topological structure, can more accurately identify abnormal connections, and improves the reliability of vascular connectivity. It significantly improves the accuracy and stability of vascular topology structure reconstruction and provides a more refined vascular model support for hemodynamic analysis.

[0111] In an alternative embodiment, input the vascular topology structure data into a recursive neural tensor network to calculate the vascular branch node risk value and generate a vascular lesion risk distribution map. Marking potential embolism risk areas based on the vascular lesion risk distribution map includes:

[0112] Extract the feature vectors of vascular branch nodes from the vascular topology structure data, calculate the topological relationship matrix between nodes, where the topological relationship matrix records the superior-inferior and peer-to-peer connection relationships of nodes, and construct a tree-shaped computational graph based on the feature vectors and the topological relationship matrix. Each node in the graph contains the corresponding feature vector and topological relationship data;

[0113] Input the tree-shaped computational graph into a recursive neural tensor network, where the recursive neural tensor network includes a third-order tensor calculation unit. The third-order tensor calculation unit performs a tensor product operation on the feature vectors of adjacent nodes to obtain the structural correlation data between nodes and the feature correlation strength of nodes;

[0114] Construct an attention weight matrix based on the structural correlation data between nodes and the feature correlation strength of nodes. The attention weight matrix includes a feature dimension weight component and a node association weight component. Weight the feature vectors according to the feature dimension weight component and the node association weight component to generate a weighted feature vector that fuses local structural information;

[0115] Use a non-linear activation function to map the weighted feature vector to obtain an initial risk assessment value. Input the initial risk assessment value into a recursive calculation module. The recursive calculation module calculates from the terminal branch nodes in a bottom-up manner, fuses the feature information and risk information of lower-level nodes based on the feature dimension weight component and the node association weight component, and updates the risk assessment value of the upper-level nodes until all nodes are traversed to generate the final risk assessment value;

[0116] Normalize the final risk assessment value to generate a risk distribution map. Partition the vascular structure according to the risk value in the risk distribution map, and mark the area with a risk value higher than the set risk threshold as a potential embolism risk area.

[0117] Exemplarily, first, extract the feature vectors of vascular branch nodes from the vascular topology structure data. The topology structure data can be obtained from medical image data (such as Computed Tomography Angiography, Magnetic Resonance Angiography), and is obtained through preprocessing steps such as image segmentation and skeleton extraction. The feature vectors can include geometric features such as the diameter, curvature, length, and bifurcation angle of vascular branch nodes, as well as hemodynamic features such as blood flow velocity and blood flow direction. For example, the feature vector of a certain branch node can be expressed as [diameter: 2.5mm, curvature: 0.1, length: 5mm, bifurcation angle: 45 degrees, blood flow velocity: 10cm / s].

[0118] Next, calculate the topological relationship matrix between vascular branch nodes. This matrix records the superior-inferior connection relationships and the connection relationships between nodes at the same level of vascular branch nodes. For example, if node A is the parent node of node B and node C, then the values in column B and column C of row A in the matrix are 1, indicating a parent-child relationship; if node B and node C are nodes at the same level, then the values in column C of row B and column B of row C in the matrix are 1, indicating a relationship between nodes at the same level. For other unrelated nodes, the matrix value is set to 0.

[0119] Then, construct a tree-shaped computational graph based on the eigenvector and the topological relationship matrix. Each node in the tree-shaped computational graph contains the corresponding eigenvector and topological relationship matrix data. The tree-shaped computational graph reflects the hierarchical structure of vascular branches, where the root node represents the main blood vessel and the leaf nodes represent the terminal vascular branches.

[0120] Input the constructed tree-shaped computational graph into a recursive neural tensor network. This network contains a third-order tensor calculation unit. The third-order tensor calculation unit receives the eigenvectors of adjacent vascular branch nodes in the tree-shaped computational graph, performs a tensor product operation on these eigenvectors, and obtains the structural correlation data between nodes and the node feature correlation strength. For example, calculate the tensor product of node A and its child node B to obtain a third-order tensor, which reflects the correlation strength between A and B in different feature dimensions.

[0121] Construct an attention weight matrix based on the structural correlation data between nodes and the node feature correlation strength. This matrix contains a feature dimension weight component and a node correlation weight component. For example, if the correlation strength of the diameter feature between node A and B is very high, then the weight corresponding to the diameter feature dimension will be larger; if the topological relationship between node A and B is closer, then the node correlation weight will be larger.

[0122] Weight the eigenvector according to the feature dimension weight component and the node correlation weight component to generate a weighted eigenvector that fuses local structural information. For example, multiply the eigenvector of node A by the corresponding weight to obtain a weighted eigenvector, which fuses the structural information of node A and its adjacent nodes.

[0123] Use a non-linear activation function to map the weighted eigenvector to obtain the initial risk assessment value of the vascular branch node. For example, use the Sigmoid function to map the weighted eigenvector to between 0 and 1 to obtain the initial risk value.

[0124] Input the initial risk assessment value into the recursive calculation module. This module calculates starting from the end branch nodes in a bottom-up manner, fuses the feature information and risk information of the lower-level nodes based on the feature dimension weight component and the node association weight component, and updates the risk assessment value of the upper-level nodes until all nodes are traversed to generate the final risk assessment value. For example, the risk values of nodes B and C are 0.8 and 0.6 respectively, and the risk value of the parent node A is calculated and updated according to the weights.

[0125] Normalize the final risk assessment value to generate a risk distribution map representing the probability of vascular lesions. For example, scale the risk values of all nodes to between 0 and 1 to generate the risk distribution map.

[0126] Partition the vascular structure according to the risk value magnitudes in the risk distribution map, and mark the regions with risk values higher than the set risk threshold as potential embolism risk regions. For example, mark the regions with risk values higher than 0.8 as potential embolism risk regions.

[0127] Existing methods for vascular lesion risk assessment usually rely on rule-based statistical models or traditional machine learning methods, and it is difficult to fully utilize the vascular topological structure and its internal hemodynamic characteristics, resulting in inaccurate risk assessment for complex vascular branch regions. In addition, existing methods mostly perform risk assessment based on individual feature indicators and fail to effectively integrate the correlations between multiple features, resulting in limitations in identifying potential embolism risk regions. This application calculates the vascular topological structure data through a recursive neural tensor network to obtain the risk values of vascular branch nodes and generate a vascular lesion risk distribution map. A tree-shaped calculation graph is constructed using the vascular topological structure data, enabling the hierarchical relationship of vascular branches to be clearly expressed, and combining the geometric morphological characteristics and hemodynamic characteristics of blood vessels to more comprehensively characterize the vascular structure information. A third-order tensor calculation unit is used to perform a tensor product operation on the feature vectors of adjacent vascular branch nodes to obtain the structural correlation data between nodes and the node feature association strength. Compared with traditional methods, this method can effectively capture the interaction relationships between different features and improve the accuracy of risk assessment. During the risk calculation process, this application constructs an attention weight matrix to simultaneously consider the weight component of the feature dimension and the node association weight component, enabling the risk assessment result to dynamically adjust the feature importance and enhance the ability to identify key risk factors. In addition, through the recursive calculation module in a bottom-up manner, the feature information and risk assessment value of the lower-level nodes are fused step by step to ensure that the risk assessment of the vascular main trunk fully considers the influence of all branches, thereby improving the rationality of the overall risk assessment. Finally, a risk distribution map is generated through normalization processing, and high-risk regions are accurately marked. Compared with existing methods, this application can more accurately identify potential embolism risk regions, improve the early diagnosis ability of vascular lesions, and provide more reliable support for clinical decision-making.

[0128] In an alternative embodiment, a multispectral imaging device is used to obtain tissue perfusion parameters of a potentially embolized risk area, electroencephalogram signals of the potentially embolized risk area are collected to obtain brain tissue function parameters, and the tissue perfusion parameters and the brain tissue function parameters are subjected to adaptive threshold segmentation to obtain an ischemic penumbra area, including:

[0129] The multispectral imaging device performs continuous wavelength scanning within a preset wavelength range through a liquid crystal tunable filter, and after accumulating and collecting multiple frames of images at each wavelength point and taking the average value, a spectral data cube is obtained;

[0130] An inverse model of tissue optical parameters is constructed according to the incident light intensity, the transmitted light intensity, and the tissue thickness. The spectral data cube is input into the inverse model of tissue optical parameters, and the least squares fitting method is used to calculate the concentration of oxyhemoglobin and the concentration of deoxyhemoglobin. The ratio of the concentration of oxyhemoglobin to the sum of the concentration of oxyhemoglobin and the concentration of deoxyhemoglobin is calculated to obtain the tissue oxygen saturation, and the tissue reflectance ratio is calculated at a preset characteristic wavelength to obtain the tissue perfusion index;

[0131] The electroencephalogram signals of the potentially embolized risk area are collected and subjected to band-pass filtering. The filtered electroencephalogram signals are decomposed into multiple preset frequency bands by using the wavelet packet decomposition method, the ratio of the energy of each preset frequency band to the total energy is calculated to obtain the relative power spectrum, the filtered electroencephalogram signals are converted into a symbol sequence, and the probability distribution of different permutation patterns is statistically obtained to obtain the permutation entropy value. The relative power spectrum and the permutation entropy value are combined in a preset order to obtain the brain tissue function parameters;

[0132] The tissue perfusion parameters and the brain tissue function parameters are combined into a feature vector. The Euclidean distance from each sample point in the feature vector to each cluster center is calculated, the membership degree of each sample point to each cluster center is calculated based on the Euclidean distance, the cluster centers are updated according to the membership degree, and the iteration is repeated until the difference between the objective function values of two adjacent iterations is less than a preset convergence threshold to obtain the final cluster centers;

[0133] Based on the final cluster centers, the feature vector is classified to obtain the spatial distributions of the ischemic area, the penumbra area, and the normal area. The set of pixel points of the penumbra area is extracted and edge detection is performed to obtain a closed boundary contour, and the area within the closed boundary contour is marked as the ischemic penumbra area.

[0134] Exemplarily, first, a multi-spectral imaging device is used to obtain tissue perfusion parameters of the potential embolism risk area. For example, a multi-spectral camera equipped with specific band filters is used to collect multi-spectral images of this area. By analyzing the light intensity information of different bands, parameters reflecting tissue blood perfusion can be obtained, such as hemoglobin concentration, blood oxygen saturation, etc. Suppose images in the red light band and near-infrared band are collected. By calculating the light intensity ratio of these two bands, a parameter reflecting tissue blood oxygen saturation can be obtained.

[0135] Next, a hyperspectral imaging system is used to obtain the reflection spectral data of the potential embolism risk area. The hyperspectral imaging system includes a xenon light source and a liquid crystal tunable filter. The liquid crystal tunable filter performs continuous wavelength scanning at a wavelength interval of 1 nm in the wavelength range of 400 nm to 700 nm. After accumulating and collecting 10 frames of images at each wavelength point and taking the average, a spectral data cube is obtained. For example, at a pixel point, a spectral curve containing 400 to 700 wavelength points can be obtained.

[0136] Then, according to the incident light intensity, transmitted light intensity, and tissue thickness, an inversion model of tissue optical parameters is constructed. The obtained spectral data cube is input into this model, and the least squares fitting method is used to calculate the concentration of oxyhemoglobin and deoxyhemoglobin. For example, assuming the tissue thickness is 1 mm, by measuring the incident light intensity and transmitted light intensity and combining the optical properties of the tissue, the concentrations of oxyhemoglobin and deoxyhemoglobin can be calculated. The ratio of the oxyhemoglobin concentration to the sum of the oxyhemoglobin concentration and deoxyhemoglobin concentration is calculated to obtain the tissue blood oxygen saturation. The tissue perfusion index is obtained by calculating the ratio of tissue reflectance at two characteristic wavelengths of 570 nm and 610 nm. For example, assuming the reflectance at 570 nm is 0.6 and the reflectance at 610 nm is 0.5, then the tissue perfusion index is 1.2.

[0137] At the same time, the electroencephalogram (EEG) signals of the potential embolism risk area are collected. The EEG signals are subjected to band-pass filtering in the frequency range of 1 - 30 Hz to obtain the filtered EEG signals. The wavelet packet decomposition method is used to decompose the filtered EEG signals into four frequency bands: delta (0.5 - 4 Hz), theta (4 - 8 Hz), alpha (8 - 13 Hz), and beta (13 - 30 Hz). The ratio of the energy of each frequency band to the total energy is calculated to obtain the relative power spectrum. For example, the relative power spectrum of the delta frequency band is 0.2, the theta frequency band is 0.3, the alpha frequency band is 0.3, and the beta frequency band is 0.2. After converting the filtered EEG signals into a symbol sequence, the probability distribution of different permutation patterns is statistically obtained to get the permutation entropy value. For example, the calculated permutation entropy value is 0.8. The relative power spectrum and the permutation entropy value are combined in a preset order to obtain the brain tissue function parameters of the potential embolism risk area.

[0138] The tissue perfusion parameters and brain tissue function parameters are composed into a feature vector. An adaptive threshold segmentation model based on fuzzy clustering is constructed. The Euclidean distance from each sample point in the feature vector to each cluster center is calculated. The membership degree of each sample point to each cluster center is calculated based on the Euclidean distance. The cluster centers are updated according to the membership degree, and the iteration is repeated until the difference between the objective functions of two adjacent iterations is less than the convergence threshold of 0.001, and the final cluster centers are obtained.

[0139] Based on the final cluster centers, the feature vector is classified to obtain the spatial distributions of the ischemic region, penumbra region, and normal region. The set of pixel points in the penumbra region is extracted. Edge detection is performed on the set of pixel points to obtain the closed boundary contour of the penumbra region. The region within the closed boundary contour is marked as the ischemic penumbra region. For example, the boundary of the penumbra region is extracted by the Canny edge detection algorithm, and the pixels within the boundary are marked as the ischemic penumbra region.

[0140] In this embodiment, through the comprehensive analysis of multispectral imaging and electroencephalogram signals, tissue perfusion parameters and brain tissue function parameters can be obtained, providing key data for the subsequent identification of ischemic regions and penumbra regions. The hyperspectral imaging system accurately measures the optical parameters of tissues, calculates the blood oxygen saturation and perfusion index of tissues, and ensures a comprehensive assessment of tissue perfusion conditions. The electroencephalogram signal extracts the band energy and permutation entropy through wavelet packet decomposition, can deeply understand the functional state of the brain tissue, and assist in identifying embolism risk regions. The adaptive threshold segmentation model combines the fuzzy clustering method, effectively realizes the fusion of blood perfusion and brain function information, and accurately divides the ischemic region, penumbra region, and normal region. Through edge detection technology, the boundary of the penumbra region is further clarified, the positioning accuracy of the ischemic penumbra region is improved, providing a reliable tool for clinical embolism risk assessment and ischemic focus region identification, and supporting personalized treatment decisions.

[0141] Figure 2 This is a comparison chart of the blood vessel connectivity reconstruction accuracy rate in the embodiment of the present invention, as Figure 2As shown in the figure, it shows the comparison of the reconstruction accuracy of this technical solution with the region growing method and the Dijkstra shortest path method at different branch levels. The horizontal axis represents the blood vessel branch level (levels 1-4), and the vertical axis represents the reconstruction accuracy (60%-100%). This technical solution reaches the highest accuracy of 98.5% at the first-level branch. As the branch level increases, although the accuracy decreases, it still remains at a relatively high level, reaching 91.6% at the fourth-level branch. In contrast, the accuracy of the region growing method is 90% at the first-level branch and drops to 78% at the fourth-level branch; the Dijkstra shortest path method performs the worst, with an accuracy of 85% at the first-level branch and dropping to 72% at the fourth-level branch. The data shows that by introducing hemodynamic characteristics and a feature consistency verification mechanism, this technical solution significantly improves the reconstruction accuracy of multi-level blood vessel structures, especially having obvious advantages in dealing with high-level branches. The specific accuracy values of each level are clearly marked in the figure, highlighting the superiority of this technical solution in blood vessel connectivity reconstruction.

[0142] Figure 3 This is the blood vessel risk distribution effect diagram of the embodiment of the present invention, as Figure 3 shown. This figure shows the analysis results of the blood vessel network risk distribution. At the axial position of the blood vessel trunk, three representative risk regions are clearly identified: the risk value of the Risk-A region on the left reaches 0.95, which is located at 150 mm from the starting section of the blood vessel and shows typical Gaussian distribution characteristics; the risk value of the Risk-B region in the middle is 0.88, which is located at 400 mm, and the risk distribution is more concentrated; the risk value of the Risk-C region on the right is 0.82, which is located at 650 mm and shows a gradual change distribution characteristic. Through 50,000 Monte Carlo sampling analyses, the overall risk distribution is obtained as a lognormal distribution, with a mean of 0.82±0.03. The system detection rate reaches 96.8%, and the false alarm rate is only 2.3%. The calculation time is controlled within 55 ms, meeting the requirements of real-time analysis. There is sufficient spatial separation between the risk regions, ensuring the accuracy and reliability of the risk assessment. The risk distribution curve fully reflects the risk change trend along the blood vessel axis, providing an accurate spatial positioning basis for subsequent clinical interventions. The analysis results fully verify the superiority of this solution in blood vessel network risk assessment.

[0143] Figure 4 This is the structural schematic diagram of an early intelligent screening and warning system for acute stroke based on deep learning in the embodiment of the present invention, as Figure 4 shown. The system includes:

[0144] The first unit is used to trigger the acquisition of multi-phase angiography images during the cardiac cycle through a pulsed dynamic imaging scheme, perform hemodynamic analysis on the multi-phase angiography images to obtain blood flow velocity vector field data, construct a blood flow feature tensor based on the blood flow velocity vector field data, and input the blood flow feature tensor into a dual-stream feature extraction network. Among them, vascular structure features are extracted through the spatial feature branch, and blood flow temporal features are extracted through the temporal feature branch. The vascular structure features and the blood flow temporal features are fused to generate a hemodynamic feature map;

[0145] The second unit is used to process the hemodynamic feature map through a directional filter to obtain an enhanced vascular image, reconstruct vascular topology structure data using the minimum spanning tree algorithm, input the vascular topology structure data into a recursive neural tensor network to calculate the risk values of vascular branch nodes to generate a vascular lesion risk distribution map, and mark potential embolism risk areas based on the vascular lesion risk distribution map;

[0146] The third unit is used to obtain tissue perfusion parameters of the potential embolism risk area using a multi-spectral imaging device, collect electroencephalogram signals of the potential embolism risk area to obtain brain tissue function parameters, perform adaptive threshold segmentation on the tissue perfusion parameters and the brain tissue function parameters to obtain an ischemic penumbra area, calculate the area value of the ischemic penumbra area, and input the area value of the ischemic penumbra area, the tissue perfusion parameters, and the brain tissue function parameters into a hierarchical decision-making model to calculate the risk level value, and generate a clinical intervention plan based on the risk level value.

[0147] In the third aspect of the embodiments of the present invention,

[0148] A kind of electronic device is provided, including:

[0149] A processor;

[0150] A memory for storing instructions executable by the processor;

[0151] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0152] In the fourth aspect of the embodiments of the present invention,

[0153] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0154] The present invention can be a method, device, system, and / or computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An early intelligent screening and warning method for acute stroke based on deep learning, characterized in that: include: A pulsed dynamic imaging scheme is used to trigger a scanning sequence within the cardiac cycle to acquire multi-phase angiographic images, and a hemodynamic analysis is performed on the multi-phase angiographic images to obtain blood flow velocity vector field data. A blood flow feature tensor is constructed based on the blood flow velocity vector field data, and the blood flow feature tensor is input into a dual-flow feature extraction network, wherein the vascular structure features are extracted through a spatial feature branch, and the blood flow time series features are extracted through a time series feature branch, and the vascular structure features and the blood flow time series features are fused to generate a hemodynamic feature map; The hemodynamic characteristic map is processed by a directional filter to obtain an enhanced vascular image, and the vascular topology data is reconstructed using a minimum spanning tree algorithm. The vascular topology data is input into a recursive neural tensor network to calculate the risk value of vascular branch nodes to generate a vascular lesion risk distribution map, and potential embolism risk areas are marked based on the vascular lesion risk distribution map. A multispectral imaging device is used to obtain tissue perfusion parameters in potential embolism risk areas, and electroencephalogram signals in potential embolism risk areas are collected to obtain brain tissue function parameters. The tissue perfusion parameters and brain tissue function parameters are adaptively thresholded to obtain the ischemic penumbra area, and the area value of the ischemic penumbra area is calculated. The area value of the ischemic penumbra area, tissue perfusion parameters, and brain tissue function parameters are input into a hierarchical decision model to calculate the risk level value, and a clinical intervention plan is generated according to the risk level value.

2. The method according to claim 1, characterized in that The pulsed dynamic imaging scheme is used to trigger the scanning sequence to acquire multi-phase angiographic images during the cardiac cycle, and the multi-phase angiographic images are subjected to hemodynamic analysis to obtain blood flow velocity vector field data. The blood flow characteristic tensor is constructed based on the blood flow velocity vector field data, including: The patient's electrocardiogram (ECG) signal is collected to construct a pulsed dynamic imaging scheme, wherein the pulsed dynamic imaging scheme uses a dual-domain sparse sampling model in the frequency domain space and the time domain for undersampling, and multiple trigger windows are constructed within the cardiac cycle based on the patient's ECG signal. A spiral angle incremental ray sampling strategy is used within the multiple trigger windows to trigger a scanning sequence, thereby obtaining non-uniform grid data for multi-phase angiography; Iteratively reconstructing the non-uniform grid data, optimizing the data consistency in the spatial domain and the temporal domain in each iteration, and finally obtaining a multi-phase angiography image with high temporal and spatial resolution; Inputting the multi-phase angiography image into a multi-scale hemodynamic feature extraction network, the multi-scale hemodynamic feature extraction network includes an adaptive deformation convolution layer, a spatial attention module, and a cyclic feature extraction unit, wherein the adaptive deformation convolution layer extracts local blood flow features to obtain vascular feature data, the spatial attention module determines the vascular branch connection relationship based on the vascular feature data to obtain vascular network data, and the cyclic feature extraction unit performs time series analysis on the vascular network data based on pressure wave propagation constraints, and finally obtains blood flow velocity vector field data; Phase enhancement technology is used to calculate the phase difference information between adjacent phase images that characterizes the magnitude and direction of blood flow velocity. Based on the phase difference information, nonlinear correction is performed on the blood flow velocity vector field data. According to the corrected blood flow velocity vector field data, a blood flow characteristic tensor including blood flow vortex distribution characteristics, vascular wall stress distribution characteristics, blood flow complexity characteristics and pressure wave propagation characteristics is constructed.

3. The method according to claim 1, characterized in that The blood flow feature tensor is input into the dual-flow feature extraction network, wherein the vascular structure features are extracted through the spatial feature branch, and the blood flow timing features are extracted through the timing feature branch. The vascular structure features and the blood flow timing features are fused to generate the hemodynamic feature map, including: Inputting the blood flow feature tensor into a dual-flow feature extraction network, the dual-flow feature extraction network includes a spatial feature branch and a temporal feature branch, the blood flow feature tensor includes blood flow vortex distribution features, blood vessel wall stress distribution features, blood flow complexity features and pressure wave propagation features; The spatial feature branch generates a convolution kernel deformation parameter based on the vascular morphology information in the blood flow feature tensor, and adaptively modulates the standard convolution kernel according to the convolution kernel deformation parameter to obtain a deformable convolution kernel group, and performs multi-scale convolution operations on the deformable convolution kernel group with the blood flow vortex distribution characteristics, the vascular wall stress distribution characteristics, the blood flow complexity characteristics and the pressure wave propagation characteristics to obtain a local feature atlas, and performs channel dimension weighting on the local feature atlas to obtain the vascular structure characteristics; The time series feature branch performs phase analysis on the blood flow vortex distribution characteristics, vascular wall stress distribution characteristics, blood flow complexity characteristics and pressure wave propagation characteristics in the blood flow feature tensor, extracts the instantaneous phase and instantaneous amplitude of each characteristic component through Hilbert transform, calculates the phase difference between time adjacent windows to obtain a phase change curve group, constructs a phase consistency matrix based on the phase change curve group, selectively enhances the instantaneous amplitude and reconstructs it to obtain the blood flow time series characteristics; The mutual information between the vascular structure features and the blood flow timing features is calculated to obtain a feature association matrix, a bidirectional attention weight is constructed based on the feature association matrix, the vascular structure features and the blood flow timing features are adaptively enhanced respectively, and the enhanced features are fused through a jump connection structure to obtain a hemodynamic feature map.

4. The method according to claim 1, characterized in that The hemodynamic characteristic map is processed by a directional filter to obtain an enhanced vascular image, and the minimum spanning tree algorithm is used to reconstruct the vascular topology data including: Extracting blood vessel morphological features in the hemodynamic characteristic map, determining the scale range of a Gaussian second-order derivative filter based on the blood vessel morphological features, and rotating the Gaussian second-order derivative filter within a preset angle range to generate a directional filter; Filtering the hemodynamic characteristic map using the directional filter to obtain multi-scale directional response values, determining weighting coefficients according to the direction of blood vessels in the hemodynamic characteristic map, weighting the multi-scale directional response values, obtaining the weighted maximum response value and the corresponding optimal direction parameters and optimal scale parameters, and generating an enhanced blood vessel image; Performing local maximum suppression processing on the enhanced vascular image, obtaining a set of candidate points for the vascular centerline, constructing an undirected graph, processing the undirected graph using a minimum spanning tree algorithm, determining a dynamic weight threshold according to the vascular density in the enhanced vascular image, removing node connections greater than the dynamic weight threshold, and obtaining an initial vascular topology structure; On the path corresponding to the connection edge of the initial vascular topological structure, blood flow vortex distribution, vascular wall stress distribution, blood flow complexity and pressure wave propagation characteristics are extracted, and the gradient change value is calculated to construct a feature consistency index; The connection relationship in the initial vascular topology structure is verified according to the feature consistency index, and the connection relationship with a gradient change value greater than a preset gradient threshold is identified as abnormal and corrected or deleted, so as to complete the vascular connectivity reconstruction and generate vascular topology structure data.

5. The method according to claim 1, characterized in that The vascular topology data is input into the recursive neural tensor network to calculate the risk value of the vascular branch node to generate a vascular lesion risk distribution map. Based on the vascular lesion risk distribution map, the potential embolism risk areas are marked, including: Extracting feature vectors of vascular branch nodes from vascular topological structure data, calculating a topological relationship matrix between nodes, wherein the topological relationship matrix records the superior-subordinate and the same-level connection relationships of nodes, and constructing a tree-shaped computational graph based on the feature vectors and the topological relationship matrix, wherein each node in the graph contains corresponding feature vectors and topological relationship data; Inputting the tree-shaped computation graph into a recursive neural tensor network, wherein the recursive neural tensor network comprises a third-order tensor computation unit, wherein the third-order tensor computation unit performs a tensor product operation on feature vectors of adjacent nodes to obtain structural correlation data between nodes and node feature correlation strength; Based on the inter-node structural correlation data and the node feature association strength, an attention weight matrix is ​​constructed, wherein the attention weight matrix includes a feature dimension weight component and a node association weight component, and the feature vector is weighted according to the feature dimension weight component and the node association weight component to generate a weighted feature vector that integrates local structural information; The weighted feature vector is mapped using a nonlinear activation function to obtain an initial risk assessment value, and the initial risk assessment value is input into a recursive calculation module. The recursive calculation module calculates from the terminal branch node in a bottom-up manner, and fuses the feature information and risk information of the lower-level node based on the feature dimension weight component and the node association weight component, and updates the risk assessment value of the upper-level node until all nodes are traversed to generate a final risk assessment value; The final risk assessment value is normalized to generate a risk distribution map, the vascular structure is partitioned according to the risk value in the risk distribution map, and the area with a risk value higher than a set risk threshold is marked as a potential embolism risk area.

6. The method according to claim 1, characterized in that Multispectral imaging equipment is used to obtain tissue perfusion parameters in potential embolic risk areas, and electroencephalogram signals in potential embolic risk areas are collected to obtain brain tissue function parameters. The tissue perfusion parameters and brain tissue function parameters are segmented by adaptive thresholds to obtain the ischemic penumbra area, including: A multispectral imaging device is used to perform continuous wavelength scanning within a preset wavelength range through a liquid crystal tunable filter, and multiple frames of images are accumulated and collected at each wavelength point, and the average value is taken to obtain a spectral data cube; A tissue optical parameter inversion model is constructed according to the incident light intensity, the transmitted light intensity and the tissue thickness, the spectral data cube is input into the tissue optical parameter inversion model, the oxygenated hemoglobin concentration and the deoxygenated hemoglobin concentration are calculated by the least square fitting method, the tissue blood oxygen saturation is calculated by the ratio of the oxygenated hemoglobin concentration to the sum of the oxygenated hemoglobin concentration and the deoxygenated hemoglobin concentration, and the tissue reflectance ratio is calculated at a preset characteristic wavelength to obtain the tissue perfusion index; The EEG signal of the potential embolism risk area is collected and band-pass filtered, and the filtered EEG signal is decomposed into a plurality of preset frequency bands by using a wavelet packet decomposition method, and the ratio of the energy of each preset frequency band to the total energy is calculated to obtain a relative power spectrum, and the filtered EEG signal is converted into a symbol sequence, and then the probability distribution of different arrangement patterns is statistically obtained to obtain an arrangement entropy value, and the relative power spectrum and the arrangement entropy value are combined in a preset order to obtain brain tissue function parameters; The tissue perfusion parameters and brain tissue function parameters are combined into a feature vector, the Euclidean distance from each sample point in the feature vector to each cluster center is calculated, the membership of each sample point to each cluster center is calculated based on the Euclidean distance, the cluster center is updated according to the membership, and the iteration is repeated until the difference between the objective functions of two adjacent iterations is less than a preset convergence threshold to obtain the final cluster center; The feature vector is classified based on the final cluster center to obtain the spatial distribution of ischemic area, penumbra area and normal area, a set of pixel points in the penumbra area is extracted for edge detection to obtain a closed boundary contour, and the area within the closed boundary contour is marked as the ischemic penumbra area.

7. The method according to claim 1, characterized in that The area value of the ischemic penumbra region is calculated, and the area value of the ischemic penumbra region, tissue perfusion parameters, and brain tissue function parameters are input into the hierarchical decision model to calculate the risk level value. The clinical intervention plan generated according to the risk level value includes: Marking the connected domains of the ischemic penumbra area and constructing a neuron connection map, calculating the topological relationship matrix between neurons, counting the total number of pixels in the ischemic penumbra area, and converting the total number of pixels into an area value of the ischemic penumbra area according to a spatial resolution parameter; A hierarchical decision model of the spatiotemporal attention mechanism is constructed, including a spatial attention module and a temporal attention module; the spatial attention module uses a wavelet multi-resolution analysis method to decompose the tissue perfusion parameters and the topological relationship matrix into spatial feature matrices of multiple scales, and extracts the dynamic change characteristics of the tissue perfusion parameters through a recursive neural network to calculate the spatial weight coefficient; the temporal attention module uses a nonlinear time series decomposition method to decompose the brain tissue function parameters into feature sequences of different time scales, and extracts the time series change characteristics through an attention autoencoder to calculate the time weight coefficient; A multimodal risk assessment feature set is constructed according to the area value of the ischemic penumbra region, the weighted tissue perfusion parameter, and the weighted brain tissue function parameter, and the fused feature vector is input into a graph convolutional neural network together with a neuron connection map to obtain a fused feature vector, and the fused feature vector is input into a residual network to obtain a risk level value; The risk evolution trend is predicted based on the risk level value, and clinical intervention rules are selected in combination with historical intervention effect data to generate a clinical intervention decision tree. The decision tree nodes are prioritized based on the risk evolution trend, and intervention effect and resource consumption are used as evaluation indicators to obtain a clinical intervention plan with priority.

8. An early intelligent screening and warning system for acute stroke based on deep learning, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to trigger a scanning sequence to acquire multi-phase angiography images within a cardiac cycle through a pulsed dynamic imaging scheme, perform hemodynamic analysis on the multi-phase angiography images to obtain blood flow velocity vector field data, construct a blood flow feature tensor according to the blood flow velocity vector field data, input the blood flow feature tensor into a dual-flow feature extraction network, wherein the vascular structure feature is extracted through a spatial feature branch, the blood flow time series feature is extracted through a time series feature branch, and the vascular structure feature is fused with the blood flow time series feature to generate a hemodynamic feature map; The second unit is used to process the hemodynamic characteristic map through a directional filter to obtain an enhanced vascular image, reconstruct the vascular topology data using a minimum spanning tree algorithm, input the vascular topology data into a recursive neural tensor network to calculate the risk value of the vascular branch node to generate a vascular lesion risk distribution map, and mark the potential embolism risk area based on the vascular lesion risk distribution map; The third unit is used to obtain tissue perfusion parameters of potential embolic risk areas using multispectral imaging equipment, collect electroencephalogram signals of potential embolic risk areas to obtain brain tissue function parameters, perform adaptive threshold segmentation on tissue perfusion parameters and brain tissue function parameters to obtain the ischemic penumbra area, calculate the area value of the ischemic penumbra area, input the area value of the ischemic penumbra area, tissue perfusion parameters and brain tissue function parameters into a hierarchical decision model to calculate the risk level value, and generate a clinical intervention plan based on the risk level value.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images