Early intelligent screening and early warning method and system for acute cerebral stroke based on deep learning

Abstract

The application provides an acute cerebral stroke early intelligent screening early warning method and system based on deep learning, relates to the technical field of medical image processing, and comprises the following steps: acquiring patient brain CT images, historical medical records and cerebral vascular data, extracting density abnormality features and symmetry change features to judge lesions, constructing a multi-scale feature pyramid and recursively refining lesion boundaries, combining three-dimensional lesion feature maps and illness change time sequence features, adopting a self-attention mechanism and a variational autoencoder to analyze cerebral vascular tree structures, and finally generating a cerebral vascular function impairment probability early warning map, so that early accurate screening and early warning of cerebral stroke can be realized, and the timeliness of treatment is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and in particular to an acute stroke early intelligent screening and early warning method and system based on deep learning. BACKGROUND

[0002] Acute stroke is a fatal disease with high disability rate and high mortality rate, and is a major public health problem worldwide. According to epidemiological surveys, stroke has become the second leading cause of death and the third leading cause of disability worldwide. Acute stroke is mainly divided into two categories: ischemic stroke and hemorrhagic stroke, of which ischemic stroke accounts for about 80%. Because about 1.9 million neurons in the brain tissue die every minute after the onset of acute stroke, time is brain tissue, so early and rapid and accurate diagnosis and treatment are crucial for patient prognosis.

[0003] The existing method mainly analyzes the CT image in a single dimension, ignores the time sequence feature information contained in the patient's historical medical record data, and is difficult to comprehensively evaluate the patient's disease development trajectory and potential risk factors, resulting in limited early warning accuracy. The evaluation of cerebral vascular structure is relatively rough, and lacks fine analysis capability for different caliber blood vessels, and cannot accurately quantify the influence of vascular stenosis or occlusion on brain tissue perfusion, thereby affecting the evaluation of stroke degree and prognosis. The existing method has insufficient accuracy in lesion region segmentation and quantization, especially for the recognition ability of low-density contrast features such as early ischemic changes, and these early imaging signs are crucial for the development of acute treatment strategies. With the development of deep learning technology, the combination of multi-source data fusion analysis and advanced image processing technology is expected to break through the existing technical bottlenecks and provide more accurate and timely acute stroke early warning solutions. SUMMARY

[0004] The embodiment of the present application provides an acute stroke early intelligent screening and early warning method and system based on deep learning, which can solve the problems in the prior art.

[0005] In a first aspect, the embodiment of the present application provides an acute stroke early intelligent screening and early warning method based on deep learning, comprising:

[0006] Obtaining brain CT image data, historical medical record data and cerebral vascular data of a patient;

[0007] Extracting density abnormality features and symmetry change features in the brain CT image data, and determining whether there is an acute stroke lesion; if there is the acute stroke lesion, constructing a multi-scale feature pyramid based on the brain CT image data to capture a lesion boundary of the acute stroke lesion, and using a recursive refinement strategy to iteratively optimize the lesion boundary to determine a lesion region segmentation result and a lesion volume data;

[0008] constructing a three-dimensional lesion feature map based on the lesion region segmentation result and the lesion volume data, performing time sequence feature extraction on the historical medical record data to obtain a time sequence feature of disease condition change;

[0009] based on the cerebral vascular data, performing hierarchical feature extraction on cerebral blood vessels of different diameters using a self-attention mechanism, introducing morphological priori knowledge for constraint optimization, generating multi-level features of the cerebral vascular tree structure based on a variational autoencoder, combining the three-dimensional lesion feature map and the time sequence feature of disease condition change, establishing a spatial dependency relationship between cerebral vascular segments through a conditional random field mechanism, and outputting a hierarchical cerebral vascular function impairment probability early warning map;

[0010] calculating an acute stroke risk prediction score and generating early warning information according to the cerebral vascular function impairment probability early warning map.

[0011] extracting density abnormality features and symmetry change features in the brain CT image data, and determining whether there is an acute stroke lesion, comprising:

[0012] performing hierarchical processing on the brain CT image data, extracting density features of local regions at different levels, performing constraint screening on the density features based on a preset Hounsfield unit threshold interval, eliminating the density features that are not within the Hounsfield unit threshold interval, and constructing a density abnormality feature map;

[0013] determining a reference midline position of the brain CT image data, taking the reference midline as a symmetry axis, calculating a local symmetry difference degree and an overall symmetry coefficient of corresponding regions of brain tissue on both sides of the symmetry axis and constructing a symmetry feature vector, and mapping the symmetry feature vector to the same feature space as the density abnormality feature map to obtain a symmetry change feature map;

[0014] adaptively fusing the density abnormality feature map and the symmetry change feature map to generate a fusion feature map;

[0015] based on the feature distribution of each target region in the fusion feature map and the feature association of its adjacent regions, constructing a hierarchical region abnormality scoring matrix, aggregating the region abnormality scoring matrix using a weighted voting mechanism to obtain an acute stroke lesion probability value of the target region, and determining that the target region has the acute stroke lesion when the acute stroke lesion probability value exceeds a preset lesion determination threshold.

[0016] based on the brain CT image data, constructing a multi-scale feature pyramid to capture a lesion boundary of the acute stroke lesion, and using a recursive refinement strategy to iteratively optimize the lesion boundary to determine a lesion region segmentation result and lesion volume data, comprising:

[0017] performing multi-level down-sampling processing on the brain CT image data and setting a local feature extraction window, extracting local response features by using the local feature extraction window to obtain a multi-scale feature response graph sequence; calculating a feature gradient for each feature response graph in the multi-scale feature response graph sequence, performing maximum and minimum value mapping on the feature gradient to obtain a boundary saliency value, and fusing the boundary saliency values of different scales to obtain a segmented straight line boundary graph;

[0018] calculating a continuity measure value between adjacent boundary points in the segmented straight line boundary graph, constructing a boundary smoothing constraint, extracting a gray scale statistical feature of an image block inside and outside the segmented straight line boundary graph, and constructing a feature consistency constraint based on the gray scale statistical feature; performing iterative optimization on the segmented straight line boundary graph, updating a boundary position based on the boundary smoothing constraint and the feature consistency constraint in each iteration, and obtaining a smooth curve boundary graph when an iterative update amount of the boundary position is less than a preset convergence threshold.

[0019] determining a contour of a lesion region according to the smooth curve boundary graph, segmenting the lesion region along the contour to obtain a lesion region segmentation result, and calculating the lesion volume data in the contour.

[0020] Based on the brain blood vessel data, self-attention mechanism is used to extract hierarchical features of brain blood vessels of different diameters, and morphological prior knowledge is introduced for constraint optimization. The multi-level features of the brain blood vessel tree structure are generated based on a variational autoencoder, including:

[0021] The brain blood vessel data is divided into multiple brain blood vessel levels according to the size of the brain blood vessel diameter. In each brain blood vessel level, the topological connection relationship between adjacent brain blood vessel nodes is calculated. Based on the topological connection relationship, a query vector and a key vector are extracted in the brain blood vessel node. The attention score matrix between the query vector and the key vector is calculated. The brain blood vessel node features are calculated according to the attention score matrix.

[0022] Based on the morphological prior knowledge, the continuity constraint of the feature difference between adjacent brain blood vessel nodes, the angle constraint of the brain blood vessel bifurcation point angle, and the curvature smoothness constraint of the brain blood vessel direction are constructed as multiple constraint conditions.

[0023] The brain blood vessel node features are input into a variational autoencoder to generate a mean vector and a log variance vector of the hidden space feature distribution. The hidden space features are obtained by random sampling based on the mean vector and the log variance vector, and the hidden space features are reconstructed by a decoder. Under the multiple constraint conditions, the multi-level features of the brain blood vessel tree structure are obtained by minimizing the reconstruction loss and the distribution loss.

[0024] inputting the brain blood vessel node feature into a variational autoencoder to generate a mean vector and a log variance vector of a latent space feature distribution, performing random sampling based on the mean vector and the log variance vector to obtain a latent space feature, and reconstructing the latent space feature through a decoder includes:

[0025] transforming the brain blood vessel node feature through a multi-layer perceptron structure in the variational autoencoder to generate the mean vector and the log variance vector representing the latent space feature distribution; performing reparameterization processing on the mean vector and the log variance vector, sampling a random noise vector from a standard normal distribution, multiplying the random noise vector and an exponential transformation result of the log variance vector element by element, and then superimposing the result onto the mean vector to obtain the latent space feature conforming to a target probability distribution;

[0026] sequentially performing non-linear mapping of the latent space feature through a multi-layer fully connected network of the decoder, performing non-linear transformation of an output of each layer of the fully connected network through a rectified linear unit, and mapping to an initial feature space at a last layer of the decoder to obtain a reconstructed feature;

[0027] calculating an Euclidean distance between the reconstructed feature and the brain blood vessel node feature as a reconstruction loss, and calculating a Kullback-Leibler divergence between the latent space feature distribution and a standard normal distribution as a distribution loss.

[0028] and combining the three-dimensional lesion feature map and the disease change time series feature to establish a spatial dependency relationship between brain blood vessel segments through a conditional random field mechanism to output a hierarchical brain blood vessel function impairment probability warning map including:

[0029] aligning the multi-level feature, the three-dimensional lesion feature map, and the disease change time series feature in a spatial dimension, extracting lesion features and time series change features of each brain blood vessel node position, and fusing the features with the multi-level feature of the corresponding position to generate a multi-modal fusion feature;

[0030] constructing a conditional random field mechanism based on the multi-modal fusion feature, the conditional random field mechanism including a node potential and an edge potential; calculating an Euclidean distance between nodes based on three-dimensional spatial coordinates of brain blood vessel nodes, calculating a feature similarity between nodes based on the multi-modal fusion feature, generating a spatial dependency matrix between brain blood vessel nodes based on the Euclidean distance and the feature similarity; performing non-linear feature mapping on the node potential to obtain a node response value, and constraining the edge potential based on the spatial dependency matrix; substituting the node response value and the constrained edge potential into the conditional random field mechanism, and calculating a function state probability distribution of each brain blood vessel node based on a Gibbs distribution;

[0031] Marginal probabilities of different diameters of cerebral vessels are calculated respectively, and based on the marginal probabilities, the functional state probability distribution of the cerebral vessel nodes in the same level is combined to obtain the hierarchical cerebral vessel function impairment probability early warning map.

[0032] In a second aspect of the embodiment of the present application, an early intelligent screening and early warning system for acute cerebral stroke based on deep learning is provided, comprising:

[0033] A first unit is configured to acquire brain CT image data, historical medical record data and cerebral vessel data of a patient.

[0034] A second unit is configured to extract density abnormality features and symmetry change features in the brain CT image data, and determine whether there is an acute cerebral stroke lesion.

[0035] A third unit is configured to construct a three-dimensional lesion feature map based on the lesion region segmentation result and the lesion volume data, extract time sequence features of the historical medical record data to obtain time sequence features of disease changes.

[0036] A fourth unit is configured to extract hierarchical features of cerebral vessels of different diameters based on the cerebral vessel data using a self-attention mechanism, and simultaneously introduce morphological prior knowledge for constraint optimization.

[0037] A fifth unit is configured to calculate an acute cerebral stroke risk prediction score and generate early warning information according to the cerebral vessel function impairment probability early warning map.

[0038] In a third aspect of the embodiment of the present application,

[0039] An electronic device is provided, comprising:

[0040] A processor;

[0041] A memory for storing processor-executable instructions;

[0042] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0043] In a fourth aspect of the embodiment of the present application,

[0044] A computer readable storage medium is provided, and computer program instructions are stored on the computer readable storage medium, and the computer program instructions are executed by a processor to implement the method.

[0045] The beneficial effects of the present application are as follows:

[0046] By extracting the density abnormality features and symmetry change features in the brain CT image data, combining the multi-scale feature pyramid and the recursive refinement strategy, accurate positioning and segmentation of acute cerebral stroke lesions are realized, the accuracy of lesion recognition and the precision of boundary positioning are improved, and reliable imaging basis is provided for early diagnosis.

[0047] The self-attention mechanism is used to extract hierarchical features of blood vessels of different diameters, and morphological prior knowledge and variational autoencoder are combined to construct multi-level feature expression of cerebral vascular tree structure, which overcomes the problem of insufficient analysis ability of complex vascular structure in traditional methods, and can more comprehensively evaluate the functional state of blood vessels.

[0048] By integrating three-dimensional lesion feature maps, disease progression time series features and cerebral vascular structure features, a conditional random field model is established to capture the spatial dependence relationship between cerebral vascular segments, and a hierarchical cerebral vascular function impairment probability warning map is generated, early and accurate prediction of acute cerebral stroke risk is realized, and the timeliness and accuracy of clinical warning are improved, providing decision support for early intervention. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The flowchart of the method for early intelligent screening and early warning of acute cerebral stroke based on deep learning of the embodiments of the present application is shown.

[0050] Figure 2 The flowchart of the multi-scale feature pyramid lesion boundary detection and segmentation process is shown. DETAILED DESCRIPTION

[0051] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0052] The technical scheme of the present application will be described in detail in specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0053] Figure 1A flowchart of an acute stroke early intelligent screening and early warning method based on deep learning for an embodiment of the present application is shown in FIG. 1, which comprises the following steps: Figure 1

[0054] Obtaining brain CT image data, historical medical record data and cerebral vascular data of a patient;

[0055] Extracting density abnormality features and symmetry change features in the brain CT image data, and determining whether there is an acute stroke lesion; if there is an acute stroke lesion, constructing a multi-scale feature pyramid based on the brain CT image data to capture a lesion boundary of the acute stroke lesion, and iteratively optimizing the lesion boundary using a recursive refinement strategy to determine a lesion region segmentation result and a lesion volume data;

[0056] Based on the lesion region segmentation result and the lesion volume data, constructing a three-dimensional lesion feature map, and extracting time sequence features of the historical medical record data to obtain a disease condition change time sequence feature;

[0057] Based on the cerebral vascular data, using a self-attention mechanism to extract hierarchical features of cerebral blood vessels of different diameters, while introducing morphological priori knowledge for constraint optimization, generating multi-level features of cerebral vascular tree structure based on a variational autoencoder, and combining the three-dimensional lesion feature map and the disease condition change time sequence feature, establishing a spatial dependency relationship between cerebral vascular segments through a conditional random field mechanism, and outputting a hierarchical cerebral vascular function impairment probability warning map;

[0058] According to the cerebral vascular function impairment probability warning map, calculating an acute stroke risk prediction score and generating early warning information.

[0059] In an optional implementation, extracting the density abnormality features and the symmetry change features in the brain CT image data includes:

[0060] Performing hierarchical processing on the brain CT image data, extracting density features of local regions at different levels, constraining and screening the density features based on a preset Hounsfield unit threshold interval, eliminating the density features not within the Hounsfield unit threshold interval, and constructing a density abnormality feature map;

[0061] Determining a reference midline position of the brain CT image data, taking the reference midline as a symmetry axis, calculating a local symmetry difference degree and an overall symmetry coefficient of corresponding regions of brain tissues on both sides of the symmetry axis and constructing a symmetry feature vector, mapping the symmetry feature vector to the same feature space as the density abnormality feature map, and obtaining a symmetry change feature map;

[0062] ​Adaptively fuse the density anomaly feature map and the symmetry change feature map to generate a fused feature map;

[0063] Based on the feature distribution of each target region in the fused feature map and the feature correlation of its adjacent region, a hierarchical region anomaly score matrix is constructed, and a weighted voting mechanism is used to aggregate the region anomaly score matrix to obtain the acute cerebral stroke lesion probability value of the target region. When the acute cerebral stroke lesion probability value exceeds a preset lesion determination threshold, it is determined that the target region has the acute cerebral stroke lesion.

[0064] The brain CT image data is processed in layers. The layering processing is mainly based on the multi-layer slice characteristics of the CT image. The entire brain CT slice sequence is divided into multiple levels by setting the interlayer distance to 5mm, and each level contains 1-3 adjacent slices. A local area of 20x20 pixels is selected as a basic unit in each layer, and the overlap rate between adjacent units is set to 30% to ensure the continuity of the features. For each local area, the average density value, the maximum density value, the minimum density value, and the standard deviation are calculated to form the density feature four-tuple of the region: Dmean, Dmax, Dmin, Dstd. For the features of acute cerebral stroke, the Hounsfield unit threshold interval is set to [18, 38] HU, which corresponds to the density change range of early ischemic stroke. The extracted density features are constrained and selected, and only the feature points with the Dmean value falling within the preset interval are retained. For example, if the average density value of a region is 25 HU, which is within the threshold interval, the feature is retained; if it is 15 HU, it is excluded. The selected feature points are marked as 1 at the corresponding position of the original CT image, and the rest are marked as 0 to construct a binary density anomaly feature map Mdensity.

[0065] The image processing technology is used to locate the edge points of the inner side of the skull, the reference midline is obtained by fitting the points through the least square method, and the local symmetry difference degree and the overall symmetry coefficient of the corresponding regions of the brain tissue on both sides of the symmetry axis are calculated with the reference midline as the symmetry axis. The local symmetry difference degree calculation method is that the CT image is divided into left and right parts with the reference midine as the axis, for each local region (i, j), the corresponding region (i', j') is found at the symmetric position, and the density difference value Ddiff(i, j) of the two regions is calculated, wherein D represents the average density of the region. The symmetry difference threshold is set to 10HU, and when Ddiff exceeds the threshold, the region is marked as having symmetry abnormality. The overall symmetry coefficient calculation involves the symmetry degree of the ventricle and the middle cerebral artery region. In the specific implementation, the brain tissue is divided into three functional regions, i.e., front, middle and back, and the symmetry similarity of each region is calculated. The symmetry similarity weight proportion of the three regions is 3:4:3, and the overall symmetry coefficient Soverall is obtained. The symmetry feature vector is mapped to the same feature space as the density abnormality feature map, and the symmetry change feature map Msymmetry is constructed.

[0066] The density abnormality feature map and the symmetry change feature map are adaptively fused, and the adaptive weight distribution strategy is used in the fusion process. Different weights are given according to the anatomical characteristics of different brain regions. For the front and back regions of the brain, the density feature weight wd is set to 0.65, and the symmetry feature weight ws is 0.35; for the middle region, especially the basal ganglia region, wd=0.55 and ws=0.45, because these regions are more susceptible to symmetry changes. The fused feature map Mfusion(x, y) = wd×Mdensity(x, y) + ws×Msymmetry(x, y), wherein (x, y) represents the pixel coordinates.

[0067] The CT image is divided into 8x8 grids, and each grid unit represents a target region. For each target region, the abnormal feature proportion, the similarity with the surrounding region, and the density gradient change in the region are calculated. The abnormal feature proportion calculation method is the ratio of the number of abnormal points in the region to the total number of pixels; the similarity between adjacent regions is obtained by calculating the cosine similarity of the feature vectors of the target region and the surrounding 8 regions; and the density gradient change is calculated by the Sobel operator. The three indicators are combined according to the weight proportion of 6:3:1 to obtain the preliminary score of each target region. On the basis of the preliminary score, the clinical anatomical knowledge is further considered, such as increasing the weight of the target region corresponding to the middle cerebral artery blood supply area by 20%. Finally, the regional abnormality score matrix S is constructed, and the matrix size is consistent with the grid division. Each element represents the abnormal score of the corresponding region.

[0068] The weighted voting mechanism is used to aggregate the regional anomaly scoring matrix, the influence radius of the adjacent region is set to 2 grid units, for the abnormal region with a score exceeding 0.7, the score distribution of the surrounding region is investigated, if there are multiple high-score regions (score > 0.6) around the center region, the final probability value of the center region is increased, and the final acute cerebral apoplexy lesion probability value P is calculated as the weighted sum of the regional score and the regional importance weight. When the P value exceeds the preset lesion determination threshold 0.75, it is determined that the target region has an acute cerebral apoplexy lesion. In actual application, if the P value of a certain region is 0.82, which exceeds the preset threshold, the region is marked as an acute cerebral apoplexy lesion.

[0069] In an optional implementation, based on the brain CT image data, a multi-scale feature pyramid is constructed to capture the lesion boundary of the acute cerebral apoplexy lesion, a recursive refinement strategy is used to iteratively optimize the lesion boundary, and the lesion region segmentation result and the lesion volume data are determined, including:

[0070] The brain CT image data is subjected to multi-level down-sampling processing and a local feature extraction window is set, the local response features are extracted by using the local feature extraction window, and a multi-scale feature response graph sequence is obtained. The feature gradient is calculated for each feature response graph in the multi-scale feature response graph sequence, the feature gradient is subjected to maximum and minimum value mapping to obtain a boundary saliency value, the boundary saliency values of different scales are fused to obtain a piecewise straight line boundary graph.

[0071] The continuity measure value between adjacent boundary points in the piecewise straight line boundary graph is calculated, a boundary smoothing constraint is constructed, the gray scale statistical features of the image blocks inside and outside the piecewise straight line boundary graph are extracted, and a feature consistency constraint is constructed based on the gray scale statistical features. The piecewise straight line boundary graph is iteratively optimized, and in each iteration, the boundary position is updated based on the boundary smoothing constraint and the feature consistency constraint. When the iterative update amount of the boundary position is less than a preset convergence threshold, a smooth curve boundary graph is obtained.

[0072] The outline of the lesion region is determined according to the smooth curve boundary graph, the lesion region is segmented along the outline to obtain the lesion region segmentation result, and the lesion volume data in the outline is calculated.

[0073] As shown in Figure 2 the method comprises:

[0074] The CT image data of the brain is subjected to multi-level down-sampling processing, and the original CT image with a resolution of 512x512 pixels is sequentially down-sampled into 256x256, 128x128 and 64x64 pixels to form a multi-level image pyramid. A local feature extraction window is set on each level, and the window size can be set to 7x7 pixels. The local response features are extracted by calculating the local statistical features of the pixel gray value in each window, including mean, standard deviation, maximum value, minimum value, etc. For the 512x512 resolution level, the local features with a mean of 32.5 and a standard deviation of 8.3 can be extracted; for the 256x256 resolution level, the local features with a mean of 31.8 and a standard deviation of 7.9 can be extracted; for the 128x128 resolution level, the local features with a mean of 30.2 and a standard deviation of 7.6 can be extracted. In this way, a multi-scale feature response graph sequence is obtained.

[0075] For each feature response graph in the multi-scale feature response graph sequence, the Sobel operator is used to calculate the horizontal and vertical gradients. For example, for a feature response graph with a resolution of 256x256, the horizontal gradient value at coordinate (120, 135) is calculated to be 15.2, and the vertical gradient value is calculated to be 9.7. The calculated feature gradient is mapped to the maximum and minimum values, and the gradient value range is mapped from the original [-25, 25] to the [0, 1] interval to obtain the boundary saliency value. For the above example, the gradient value mapping of the boundary saliency value at coordinate (120, 135) is 0.78. The boundary saliency values of different scales are fused, and a weighted average method can be used, in which the weight of the high-resolution level is 0.5, the weight of the medium-resolution level is 0.3, and the weight of the low-resolution level is 0.2. After fusion, a piecewise straight line boundary graph is obtained, which contains multiple boundary points, each boundary point has a clear coordinate position.

[0076] A continuity measure value between adjacent boundary points in the piecewise straight line boundary map is calculated: traverse the sequence of boundary points, calculate the straight line distance for each boundary point and its adjacent points, analyze the degree of turning angle change between adjacent boundary segments, the closer the distance and the smaller the angle change, the better the boundary continuity, by comprehensively considering the distance and angle two factors, a continuity score is assigned to each pair of adjacent boundary points, the score range is between zero and one, the higher the value, the smoother the boundary. For the two adjacent boundary points with coordinates (120, 135) and (121, 137), the continuity measure value can be obtained by calculating the Euclidean distance, which is 2.24. A boundary smoothing constraint is constructed based on the calculated continuity measure value: set the continuity threshold to 0.8, when the continuity measure value of a certain boundary is lower than this threshold, a smoothing adjustment force is applied to the boundary point, the size of the adjustment force is inversely proportional to the continuity measure value, the worse the continuity, the greater the adjustment force received by the boundary point, the adjustment direction points to the middle position of the line connecting the adjacent two boundary points, the purpose is to reduce the sharp turning and irregular fluctuations of the boundary, the smoothing constraint strength parameter is set to 0.3, to ensure that the boundary accuracy is maintained while achieving moderate smoothing. Extract the gray scale statistical features of the image blocks inside and outside the piecewise straight line boundary map, the average gray value of the inside region (i.e. the lesion region) is 18.5, and the standard deviation is 5.2; the average gray value of the outside region (i.e. the normal tissue region) is 42.8, and the standard deviation is 6.7. A feature consistency constraint is constructed based on these gray scale statistical features: small image regions are extracted around each boundary point on the inside and outside, the average gray value and the degree of change of these regions are calculated respectively, the rationality of the current boundary position is evaluated by comparing the gray difference between the inside and outside regions of the boundary, the greater the gray difference, the higher the credibility of the boundary positioning, an adjustment force is constructed based on this difference, which drives the boundary to move to the position with more obvious gray contrast, the consistency constraint strength is set to 0.4, the adjustment direction is along the direction of the most severe local gray change, to ensure that the final boundary can accurately separate the lesion region and the normal tissue region.

[0077] The piecewise straight line boundary map is iteratively optimized, in the first iteration, based on the boundary smoothing constraint (0.85) and the feature consistency constraint (0.92), the new position of the boundary point is calculated, for example, the boundary point with coordinates (120, 135) is updated to (119.8, 134.7) after the first iteration, in the second iteration, the position of this point is further updated to (119.6, 134.5). The iteration process continues until the update amount of the boundary position is less than the preset convergence threshold, for example 0.01 pixel. In this example, after 12 iterations, the update amount is 0.008 pixels, which is lower than the convergence threshold, the iteration stops, and a smooth curve boundary map is obtained.

[0078] According to the smooth curve boundary map, the contour of the lesion region is determined, which is composed of a series of consecutive boundary points, such as {(110.2, 130.5), (111.5, 132.3),..., (142.3, 165.8)}, and the lesion region is segmented along the contour to obtain the lesion region segmentation result, that is, all the pixel points within the contour constitute the lesion region. The lesion volume data within the contour is calculated, which can be achieved by counting the number of pixels within the lesion region and combining the voxel size of the CT image, for example, if the voxel size of the CT image is 0.5*0.5*1.0 mm, and the lesion region contains 2500 pixels, then the lesion volume is 2500*0.5*0.5*1.0=625 cubic millimeters.

[0079] Through the above embodiments, the brain CT image of an acute stroke patient can be accurately segmented, and objective and quantitative lesion volume data can be provided for clinical diagnosis to assist doctors in developing treatment plans and evaluating treatment effects. The method is suitable for different types of acute stroke cases, including ischemic and hemorrhagic stroke, and has high universality and robustness.

[0080] In an optional embodiment, based on the brain blood vessel data, self-attention mechanism is used for hierarchical feature extraction of brain blood vessels of different diameters, and morphological prior knowledge is introduced for constraint optimization, and the multi-level features of the brain blood vessel tree structure are generated based on a variational autoencoder, including:

[0081] The brain blood vessel data is divided into multiple brain blood vessel levels according to the diameter of the brain blood vessels, and in each brain blood vessel level, the topological connection relationship between adjacent brain blood vessel nodes is calculated, based on which, query vectors and key vectors are extracted from the brain blood vessel nodes, an attention score matrix between the query vectors and the key vectors is calculated, and brain blood vessel node features are calculated according to the attention score matrix;

[0082] Based on the morphological prior knowledge, the continuity constraint of the difference between adjacent brain blood vessel node features, the angle constraint of the angle of the brain blood vessel bifurcation point, and the curvature smoothness constraint of the direction of the brain blood vessel are constructed as multiple constraint conditions;

[0083] The brain blood vessel node features are input into a variational autoencoder to generate a mean vector and a log variance vector of the hidden space feature distribution, the hidden space features are obtained by random sampling based on the mean vector and the log variance vector, and the hidden space features are reconstructed by a decoder; under the multiple constraint conditions, the multi-level features of the brain blood vessel tree structure are obtained by minimizing the reconstruction loss and the distribution loss.

[0084] The cerebral vascular data containing three-dimensional coordinates and diameter information of the cerebral vessels are acquired, and the cerebral vascular data are divided into multiple cerebral vessel levels according to the diameter of the cerebral vessels. In this embodiment, the cerebral vessels are divided into three levels according to the diameter: a large vessel level (diameter greater than 2.5 mm), a medium vessel level (diameter between 1.0 and 2.5 mm), and a microvessel level (diameter less than 1.0 mm). The cerebral vascular data of each level are processed respectively to capture the feature information of vessels of different scales.

[0085] In each cerebral vessel level, the sampling points on the centerline of the cerebral vessels are taken as the cerebral vessel nodes, the topological connection relationship is determined by calculating the Euclidean distance between the nodes, the distance threshold is set to 0.5 mm, and the connection relationship is represented as an adjacency matrix A, wherein A[i][j]=1 represents that node i is connected with node j, otherwise A[i][j]=0.

[0086] For each cerebral vessel node, its three-dimensional coordinates, diameter, and local geometric features (such as curvature and tortuosity) are combined to form a feature vector with a dimension of 16, and a query vector Q and a key vector K are generated by linear mapping, and the dimensions of the mapping matrices WQ and WK are 16x32, that is, the original features are mapped to a 32-dimensional space. In specific implementation, the query vector is calculated as Q=XWQ, and the key vector is calculated as K=XWK, wherein X is the original feature matrix of the node.

[0087] For node i and node j, the attention score is calculated as the dot product of the query vector and the key vector, and then divided by the scaling factor 8. If node i and node j are not connected, the corresponding attention score is set to a very small negative number to ensure that the attention mechanism only focuses on the topologically connected nodes. The attention score is normalized by applying a softmax function to obtain the final attention weight matrix.

[0088] The original node features are mapped to a value vector V=XWV by linear mapping, and the mapping matrix WV has a dimension of 16x32. The new features of the nodes are calculated by the weighted sum of the attention weight matrix and the value vector. In actual operation, a multi-head attention mechanism is adopted, 8 attention heads are set, and the output dimension of each head is 4. After merging, a 32-dimensional feature vector is obtained. The feature representation is further optimized through residual connection and layer normalization.

[0089] The continuity constraint of the feature difference between adjacent cerebral vessel nodes is realized by calculating the L2 norm of the feature vectors of adjacent nodes, and the constraint threshold is set to 0.2. The angle constraint of the cerebral vessel bifurcation point is based on physiological knowledge, which requires the bifurcation angle of the subvessel to be between 20 degrees and 120 degrees. The bifurcation points deviating from this range will be applied with a penalty term. The curvature smoothness constraint of the cerebral vessel course is realized by calculating the local curvature change rate, and the curvature calculation is based on a circle formed by three adjacent nodes. The curvature change rate threshold is set to 0.15. Multiple constraint conditions are obtained.

[0090] inputting the brain vessel node features into the variational autoencoder to generate a mean vector and a log-variance vector of a latent space feature distribution, randomly sampling the mean vector and the log-variance vector to obtain a latent space feature, and reconstructing the latent space feature through a decoder.

[0091] Under multiple constraint conditions, multi-level features of the brain vessel tree structure are obtained by minimizing reconstruction loss and distribution loss. The reconstruction loss calculates the Euclidean distance between the original features and the reconstructed features, and the distribution loss adopts KL divergence to measure the difference between the latent space distribution and the standard normal distribution. The constraint conditions are added to the total loss function in the form of regularization terms, and the weight coefficients are respectively: continuity constraint 0.3, angle constraint 0.2, and curvature constraint 0.5. The model parameters are optimized through a gradient descent algorithm (learning rate is set to 0.001) for 200 iterations, and data with a batch size of 64 is used for training in each iteration.

[0092] The final multi-level features contain structural and morphological information of blood vessels at different scales, which can be used for subsequent tasks such as brain vessel disease diagnosis and blood flow dynamics analysis.

[0093] In an alternative embodiment, inputting the brain vessel node features into the variational autoencoder to generate a mean vector and a log-variance vector of a latent space feature distribution, randomly sampling the mean vector and the log-variance vector to obtain a latent space feature, and reconstructing the latent space feature through a decoder includes:

[0094] transforming the brain vessel node features through a multi-layer perceptron structure in the variational autoencoder to generate the mean vector and the log-variance vector representing the latent space feature distribution; performing reparameterization processing on the mean vector and the log-variance vector, sampling a random noise vector from a standard normal distribution, multiplying the random noise vector and an exponential transformation result of the log-variance vector element by element, and then adding the result to the mean vector to obtain the latent space feature subject to a target probability distribution;

[0095] sequentially mapping the latent space feature through a multi-layer fully connected network of the decoder for non-linear mapping, and performing non-linear transformation on the output of each layer of the fully connected network through a rectified linear unit, and mapping to an initial feature space at the last layer of the decoder to obtain a reconstructed feature;

[0096] calculating the Euclidean distance between the reconstructed feature and the brain vessel node features as the reconstruction loss, and calculating the Kullback-Leibler divergence between the latent space feature distribution and the standard normal distribution as the distribution loss.

[0097] In the feature encoding stage, the input cerebral vessel node features are received, which can be represented as a multi-dimensional vector, for example, a 96-dimensional vector containing spatial position, diameter, bifurcation angle and other attribute information of the node. These cerebral vessel node features are processed by the encoder part in the variational autoencoder. The encoder adopts a multi-layer perceptron structure, including three fully connected networks. The first layer maps the 96-dimensional input features to a 64-dimensional hidden layer, the second layer further maps the 64-dimensional hidden layer to a 32-dimensional hidden layer, and the last layer maps the 32-dimensional hidden layer to two 16-dimensional vectors respectively, representing the mean vector μ and the logarithmic variance vector logσ of the latent space feature distribution 2 . A rectified linear unit (ReLU) is applied as an activation function after each fully connected network to enhance the non-linear expression ability of the network, which is implemented as f(x) = max(0, x). For example, when the input feature vector is [0.23, 0.45,..., -0.12], after transformation by the first layer weight matrix W1 (size 96x64) and the bias vector b1, the intermediate result is [0.31, -0.08,..., 0.56], and after ReLU activation function processing, it becomes [0.31, 0,..., 0.56].

[0098] In the latent space sampling stage, the mean vector μ and the logarithmic variance vector logσ 2 output by the encoder are reparameterized to realize a differentiable random sampling process. A random noise vector ε is sampled from the standard normal distribution N(0, I), which has the same size as the mean vector, for example, dimension 16. For the logarithmic variance vector logσ 2 , the exponential transformation result σ is calculated, that is, the exponential operation is performed on each element in logσ 2 . The random noise vector ε is multiplied element by element with the standard deviation vector σ (the square root of σ 2 ), to obtain a random offset, which is added to the mean vector μ to obtain the latent space feature z conforming to the target probability distribution N(μ, σ 2 ). For example, when μ = [0.5, -0.2,..., 0.7], logσ 2 = [-1.0, -0.5,..., -0.8], and ε = [0.1, 0.3,..., -0.2] is sampled, the standard deviation σ = [0.607, 0.707,..., 0.670] is calculated, and the final latent space feature z = [0.561, 0.012,..., 0.566] is obtained.

[0099] In the feature decoding stage, the hidden space feature z generated in the previous stage is input into the decoder, and is decoded through a multi-layer fully connected network to map the feature back to the original feature space. The network structure of the decoder is generally symmetrical to that of the encoder and also contains three layers of fully connected networks. The first layer maps the 16-dimensional hidden space feature to a 32-dimensional hidden layer, the second layer maps the 32-dimensional hidden layer to a 64-dimensional hidden layer, and the last layer maps the 64-dimensional hidden layer back to the 96-dimensional original feature space to obtain the reconstructed feature x'. Similarly, ReLU activation function is used for non-linear transformation after each layer of fully connected network except the last layer, and linear activation function is used for the last layer so that the reconstructed value can match the range of the original feature. For example, when the hidden space feature z = [0.561, 0.012,..., 0.566] is input into the decoder, the reconstructed feature x' = [0.25, 0.42,..., -0.15] is finally obtained through multi-layer conversion of weight matrix and activation function.

[0100] In the training process, the loss function is calculated to optimize the parameters of the variational autoencoder. The loss function consists of two parts: reconstruction loss and distribution loss. The reconstruction loss measures the difference between the reconstructed feature and the original brain vessel node feature, and is calculated using the Euclidean distance, that is, the square sum of the difference between the corresponding elements of the two vectors is calculated, and then the square root is taken. For example, when the original feature x = [0.23, 0.45,..., -0.12] and the reconstructed feature x' = [0.25, 0.42,..., -0.15], the Euclidean distance is 0.087; the distribution loss adopts the Kullback-Leibler divergence, which calculates the difference between the distribution of the hidden space feature and the standard normal distribution. It can be calculated by the mean vector μ and the logarithmic variance vector logσ 2 directly, the result is the sum of the square of the mean vector and the sum of the variance, minus the sum of the logarithmic variance and the vector dimension. For example, when μ = [0.5, -0.2,..., 0.7] and logσ 2 = [-1.0, -0.5,..., -0.8], the calculated KL divergence is about 2.36.

[0101] In an optional implementation, the spatial dependency relationship between the brain vessel segments is established through a conditional random field mechanism in combination with the three-dimensional lesion feature map and the disease change time series feature, and a hierarchical brain vessel function impairment probability early warning map is output, including:

[0102] The multi-level features, the three-dimensional lesion feature map and the disease change time series feature are aligned in the spatial dimension, the lesion feature and the time series change feature of each brain vessel node position are extracted, and are fused with the multi-level features of the corresponding position to generate multi-modal fusion features;

[0103] A conditional random field mechanism is constructed based on the multi-modal fusion features, and the conditional random field mechanism includes node potentials and edge potentials; a Euclidean distance between nodes is calculated using three-dimensional spatial coordinates of the brain blood vessel nodes, a feature similarity between nodes is calculated in combination with the multi-modal fusion features, and a spatial dependence matrix between the brain blood vessel nodes is generated based on the Euclidean distance and the feature similarity; a node response value is obtained by performing nonlinear feature mapping on the node potentials, and the edge potentials are constrained based on the spatial dependence matrix; the node response value and the constrained edge potentials are substituted into the conditional random field mechanism, and a functional state probability distribution of each brain blood vessel node is calculated based on Gibbs distribution.

[0104] Marginal probabilities are calculated for brain blood vessel levels of different diameters, and the functional state probability distributions of the brain blood vessel nodes in the same level are combined based on the marginal probabilities to obtain a hierarchical brain blood vessel function impairment probability early warning map.

[0105] Multi-level features, three-dimensional lesion feature maps, and disease condition change time series features are accurately aligned in the spatial dimension through three-dimensional spatial registration technology, ensuring that all features are in the same spatial coordinate system. For example, for a brain blood vessel image data with a size of 256x256x128, a spatial interpolation algorithm is used to map features of different resolutions to a unified spatial grid. At each brain blood vessel node position (x, y, z), the lesion feature vector Fl(x, y, z) and the time series change feature vector Ft(x, y, z) corresponding to the position are extracted, and the feature vector Fm(x, y, z) of the corresponding position is obtained from the multi-level features. Through feature connection operation, the three feature vectors are combined into a unified multi-modal fusion feature vector F(x, y, z) = [Fm(x, y, z), Fl(x, y, z), Ft(x, y, z)], and the dimension is the sum of the original feature dimensions.

[0106] Based on the obtained multi-modal fusion features, a conditional random field mechanism is constructed, which is composed of node potentials and edge potentials, for modeling the interdependent relationship between brain blood vessel nodes. For each node i, the node potential is represented as Φ(yi, F(i)), where yi represents the functional state of node i, and F(i) represents the multi-modal fusion features at node i. To calculate the spatial dependence between nodes, the Euclidean distance between nodes is calculated using the three-dimensional spatial coordinates of the brain blood vessel nodes. Specifically, for node i and node j, the Euclidean distance is d(i, j) = ||p(i)-p(j)||2, where p(i) and p(j) represent the three-dimensional spatial coordinates of node i and node j, respectively.

[0107] Meanwhile, the feature similarity s(i,j) between nodes i and j is calculated by combining the multi-modal fusion features and computing the cosine similarity between their feature vectors F(i) and F(j). For example, when the feature vectors of two nodes are [0.7, 0.2, 0.5, 0.8] and [0.6, 0.3, 0.4, 0.7] respectively, their cosine similarity is about 0.985, indicating that the features of the two nodes are very similar.

[0108] Based on the Euclidean distance and the feature similarity, the spatial dependency matrix W between the nodes of the brain blood vessels is generated, specifically, W(i,j) = s(i,j) × exp(-d(i,j) 2 / σ 2 ), where σ is a spatial decay parameter for controlling the influence of distance on the dependency strength. For example, when σ = 5.0, d(i,j) = 3.0, and s(i,j) = 0.985, W(i,j) is about 0.672, indicating that there is a strong spatial dependency relationship between the two nodes.

[0109] When processing the node potential, the multi-modal fusion feature F(i) is mapped to the node response value r(i) using a nonlinear feature mapping function. The mapping function can be a multilayer perceptron containing two hidden layers with 64 and 32 neurons respectively, using a ReLU activation function. For example, when the input feature vector is [0.7, 0.2, 0.5, 0.8], the node response value obtained after nonlinear mapping is [0.65, 0.35], indicating that the probability of the node being in a normal state is 0.65 and the probability of being in a damaged state is 0.35.

[0110] Based on the spatial dependency matrix W, the edge potential is constrained, represented as ψ(yi, yj, W(i,j)), which is used to model the mutual dependency between the states of nodes i and j. Specifically, when W(i,j) is large, nodes i and j tend to have similar functional states; when W(i,j) is small, their functional states are different. For example, when W(i,j) = 0.672, if the state of node i is normal, node j also has a high probability of being in a normal state.

[0111] The node response value and the constrained edge potential are substituted into the conditional random field mechanism, and specifically, the probability that the node i is in the state yi is P(yi|F) ∝ exp(Φ(yi,F(i)) + ∑j∈N(i)ψ(yi,yj,W(i,j))), wherein N(i) represents a neighbor set of the node i. The probability distribution is solved by an iterative optimization algorithm, for example, after 20 iterations, the probabilities that a certain node is in the normal, slightly damaged, moderately damaged and severely damaged states are [0.75, 0.15, 0.08, 0.02] respectively.

[0112] According to the diameter of the blood vessels, the cerebral blood vessels are divided into three levels of main arteries (diameter ≥ 2.5 mm), medium arteries (1.0 mm ≤ diameter < 2.5 mm) and microvessels (diameter < 1.0 mm). For each level, the marginal probability of the functional state of all nodes in the level is calculated, for example, for the main artery level, the functional damage probability is 0.12; for the medium artery level, the functional damage probability is 0.23; for the microvessel level, the functional damage probability is 0.45. The functional state probability distribution of the cerebral blood vessel nodes in the same level is combined to generate a hierarchical cerebral blood vessel functional damage probability early warning map. The early warning map is presented in the form of a heat map, which can intuitively identify the functional damage risk of blood vessels at different levels and develop targeted treatment plans.

[0113] The early warning system for intelligent screening of acute cerebral stroke based on deep learning comprises:

[0114] The first unit is configured to acquire brain CT image data, historical medical record data and cerebral blood vessel data of a patient.

[0115] The second unit is configured to extract density abnormality features and symmetry change features in the brain CT image data, and determine whether there is an acute cerebral stroke lesion. If the acute cerebral stroke lesion exists, a multi-scale feature pyramid is constructed based on the brain CT image data to capture a lesion boundary of the acute cerebral stroke lesion, a recursive refinement strategy is used to iteratively optimize the lesion boundary, and a lesion region segmentation result and lesion volume data are determined.

[0116] The third unit is configured to construct a three-dimensional lesion feature map based on the lesion region segmentation result and the lesion volume data, and extract time sequence features of the historical medical record data to obtain disease condition change time sequence features.

[0117] The fourth unit is configured to perform hierarchical feature extraction on brain blood vessels with different diameters based on the brain blood vessel data by using a self-attention mechanism, constrain and optimize by introducing morphological prior knowledge, generate multi-level features of the brain blood vessel tree structure based on a variational autoencoder, combine the three-dimensional lesion feature map and the disease change time series feature, establish a spatial dependency relationship between brain blood vessel segments by a conditional random field mechanism, and output a hierarchical brain blood vessel function impairment probability warning map.

[0118] The fifth unit is configured to calculate an acute stroke risk prediction score and generate warning information according to the brain blood vessel function impairment probability warning map.

[0119] In a third aspect, an electronic device is provided, including:

[0120] a processor;

[0121] a memory for storing processor-executable instructions;

[0122] The processor is configured to invoke the instructions stored in the memory to perform the method described above.

[0123] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0124] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which are used to perform various aspects of the present application.

[0125] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A deep learning-based intelligent screening and early warning method for acute stroke, characterized in that, include: Acquire the patient's brain CT imaging data, historical medical records, and cerebrovascular data; Density abnormality features and symmetry change features are extracted from the brain CT image data to determine whether there is an acute stroke lesion. If the acute stroke lesion exists, a multi-scale feature pyramid is constructed based on the brain CT image data to capture the lesion boundary. A recursive refinement strategy is used to iteratively optimize the lesion boundary to determine the lesion region segmentation result and lesion volume data. Based on the lesion region segmentation results and the lesion volume data, a three-dimensional lesion feature map is constructed, and the historical medical record data is subjected to temporal feature extraction to obtain the temporal features of disease changes. Based on the cerebral vascular data, a self-attention mechanism is used to extract hierarchical features of cerebral blood vessels of different diameters. At the same time, morphological prior knowledge is introduced for constraint optimization. Based on the variational autoencoder, a multi-level feature of the cerebral vascular tree structure is generated, including: dividing the cerebral vascular data into multiple cerebral vascular levels according to the diameter of the cerebral blood vessels; calculating the topological connection relationship between adjacent cerebral vascular nodes in each cerebral vascular level; extracting query vectors and key vectors from the cerebral vascular nodes based on the topological connection relationship; calculating the attention score matrix between the query vector and the key vector; and calculating the cerebral vascular node features based on the attention score matrix. Based on the aforementioned morphological prior knowledge, multiple constraints are constructed, including the continuity constraint of the differences in features between adjacent cerebral vascular nodes, the angular constraint of the angle of the bifurcation point of cerebral vessels, and the curvature smoothness constraint of the cerebral vessel orientation. The features of the cerebral vascular nodes are input into a variational autoencoder to generate the mean vector and logarithmic variance vector of the latent space feature distribution. The latent space features are obtained by random sampling based on the mean vector and the logarithmic variance vector, and the latent space features are reconstructed by a decoder. Under the multiple constraints, the multi-level features of the cerebral vascular tree structure are obtained by minimizing the reconstruction loss and the distribution loss. Combining the three-dimensional lesion feature map and the temporal features of the disease progression, the spatial dependency between cerebral vascular segments is established through a conditional random field mechanism, and a hierarchical early warning map of the probability of cerebral vascular function impairment is output. Based on the cerebrovascular function impairment probability early warning map, an acute stroke risk prediction score is calculated and early warning information is generated.

2. The method according to claim 1, characterized in that, Extracting density abnormalities and symmetry changes from the brain CT imaging data to determine the presence of acute stroke lesions includes: The brain CT image data is processed in layers, and density features of local regions are extracted at different layers. The density features are constrained and filtered based on a preset Hounsfield unit threshold interval, and the density features that are not within the Hounsfield unit threshold interval are removed to construct a density abnormality feature map. The baseline midline position of the brain CT image data is determined. Using the baseline midline as the axis of symmetry, the local symmetry difference and overall symmetry coefficient of the corresponding regions of brain tissue on both sides of the axis of symmetry are calculated and constructed as a symmetry feature vector. The symmetry feature vector is mapped to the same feature space as the density abnormality feature map to obtain the symmetry change feature map. The density anomaly feature map and the symmetry change feature map are adaptively fused to generate a fused feature map; Based on the feature distribution of each target region in the fused feature map and the feature association of its adjacent regions, a hierarchical regional anomaly scoring matrix is ​​constructed. The regional anomaly scoring matrix is ​​aggregated using a weighted voting mechanism to obtain the probability value of acute stroke lesions in the target region. When the probability value of acute stroke lesions exceeds a preset lesion determination threshold, it is determined that the target region has acute stroke lesions.

3. The method according to claim 1, characterized in that, Based on the brain CT image data, a multi-scale feature pyramid is constructed to capture the lesion boundary of the acute stroke lesion. A recursive refinement strategy is used to iteratively optimize the lesion boundary, and the lesion region segmentation result and lesion volume data are determined, including: The brain CT image data is subjected to multi-level downsampling processing and a local feature extraction window is set. Local response features are extracted using the local feature extraction window to obtain a multi-scale feature response map sequence. The feature gradient is calculated for each feature response map in the multi-scale feature response map sequence. The feature gradient is mapped to the maximum and minimum values ​​to obtain the boundary saliency value. The boundary saliency values ​​at different scales are fused to obtain a piecewise straight line boundary map. Calculate the continuity metric between adjacent boundary points in the segmented straight line boundary map, construct boundary smoothing constraints, extract grayscale statistical features of the inner and outer image blocks of the segmented straight line boundary map, and construct feature consistency constraints based on the grayscale statistical features; perform iterative optimization on the segmented straight line boundary map, updating the boundary position based on the boundary smoothing constraints and the feature consistency constraints in each iteration, and obtaining a smooth curve boundary map when the iterative update amount of the boundary position is less than a preset convergence threshold; The contour of the lesion region is determined based on the smooth curve boundary map, the lesion region is segmented along the contour to obtain the lesion region segmentation result, and the volume data of the lesion within the contour is calculated.

4. The method according to claim 1, characterized in that, The cerebral vascular node features are input into a variational autoencoder to generate a mean vector and a log-variance vector of the latent space feature distribution. Latent space features are obtained by random sampling based on the mean vector and the log-variance vector, and the latent space features are reconstructed by a decoder, including: The cerebral vascular node features are transformed using the multilayer perceptron structure in the variational autoencoder to generate the mean vector and the log-variance vector representing the latent space feature distribution. The mean vector and the log-variance vector are then reparameterized, and random noise vectors are sampled from the standard normal distribution. The random noise vector and the exponential transformation result of the log-variance vector are multiplied element-wise and then superimposed onto the mean vector to obtain the latent space features that follow the target probability distribution. The latent space features are sequentially passed through the multi-layer fully connected network of the decoder for nonlinear mapping. The output of each layer of the fully connected network is nonlinearly transformed by a rectified linear unit and then mapped to the initial feature space in the last layer of the decoder to obtain the reconstructed features. The Euclidean distance between the reconstructed features and the cerebral vascular node features is calculated as the reconstruction loss, and the Kohlbek-Leibler divergence between the latent space feature distribution and the standard normal distribution is calculated as the distribution loss.

5. The method according to claim 1, characterized in that, Combining the three-dimensional lesion feature map and the temporal characteristics of disease progression, a spatial dependency relationship between cerebral vascular segments is established through a conditional random field mechanism, outputting a hierarchical early warning map of cerebral vascular function impairment probability, including: Align the multi-level features, the three-dimensional lesion feature map, and the temporal features of disease progression in the spatial dimension, extract the lesion features and temporal change features at each cerebrovascular node location, and fuse them with the multi-level features at the corresponding location to generate multimodal fusion features; A conditional random field (CRF) mechanism is constructed based on the multimodal fusion features. The CRF mechanism includes node potential and edge potential. The Euclidean distance between nodes is calculated using the three-dimensional spatial coordinates of the cerebrovascular nodes. The feature similarity between nodes is calculated by combining the multimodal fusion features. A spatial dependency matrix between cerebrovascular nodes is generated based on the Euclidean distance and the feature similarity. The node potential is nonlinearly mapped to obtain node response values. The edge potential is constrained based on the spatial dependency matrix. The node response values ​​and the constrained edge potential are substituted into the CRF mechanism, and the functional state probability distribution of each cerebrovascular node is calculated based on the Gibbs distribution. Marginal probabilities are calculated for different diameter cerebral blood vessel levels. Based on these marginal probabilities, the functional state probability distributions of cerebral blood vessel nodes within the same level are combined to obtain a hierarchical early warning map of the probability of cerebral blood vessel functional impairment.

6. A deep learning-based intelligent screening and early warning system for acute stroke, used to implement the method as described in any one of claims 1-5, characterized in that, include: The first unit is used to acquire the patient's brain CT imaging data, historical medical records, and cerebrovascular data. The second unit is used to extract density abnormality features and symmetry change features from the brain CT image data to determine whether there is an acute stroke lesion. If the acute stroke lesion exists, a multi-scale feature pyramid is constructed based on the brain CT image data to capture the lesion boundary. A recursive refinement strategy is used to iteratively optimize the lesion boundary to determine the lesion region segmentation result and lesion volume data. The third unit is used to construct a three-dimensional lesion feature map based on the lesion region segmentation results and the lesion volume data, and to extract temporal features from the historical medical record data to obtain temporal features of disease changes. The fourth unit is used to extract hierarchical features of cerebral blood vessels of different diameters based on the cerebral blood vessel data using a self-attention mechanism. At the same time, morphological prior knowledge is introduced for constraint optimization. Based on the variational autoencoder, a multi-level feature of the cerebral blood vessel tree structure is generated. Combined with the three-dimensional lesion feature map and the temporal features of the disease change, the spatial dependency relationship between cerebral blood vessel segments is established through the conditional random field mechanism, and a hierarchical cerebral blood vessel function damage probability warning map is output. The fifth unit is used to calculate the acute stroke risk prediction score based on the cerebrovascular function impairment probability early warning map and generate early warning information.

7. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 5.

Images