Intelligent detection and positioning method and system for cerebral aneurysms based on multi-feature fusion

Abstract

This invention provides a method and system for intelligent cerebral aneurysm detection and localization based on multi-feature fusion, relating to the field of image processing technology. The method includes receiving and preprocessing head angiography images; utilizing a multi-scale feature extraction network to obtain and fuse feature maps; constructing candidate anatomical regions using density clustering; extracting aneurysm contours using a graph cut energy function; constructing a vascular network topology map based on probabilistic feature mapping, calculating positional feature descriptors; and finally performing classification, discrimination, and annotating and displaying the detection results. By integrating multiple features, the method improves the accuracy and sensitivity of cerebral aneurysm detection and reduces the misdiagnosis rate.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method and system for intelligent detection and positioning of cerebral aneurysms based on multi-feature fusion. Background Art

[0002] Cerebral aneurysms are sac-like structures formed by abnormal localized bulging of intracranial blood vessels. Rupture can lead to subarachnoid hemorrhage, which carries a high rate of disability and mortality. Early detection and localization of cerebral aneurysms are crucial for reducing clinical risk and improving treatment outcomes. Currently, digital subtraction angiography (DSA) remains the gold standard for diagnosing cerebral aneurysms, but manual interpretation based on physician experience is subjective and time-consuming. With the development of computer-aided diagnosis technology, intelligent cerebral aneurysm detection methods based on image processing and artificial intelligence have become a research hotspot.

[0003] Traditional cerebral aneurysm detection methods primarily rely on manually designed features, such as those based on morphological, gradient, or geometric features. In recent years, deep learning technology has been introduced to the field of cerebral aneurysm detection, automatically extracting image features through convolutional neural networks, thereby improving detection accuracy. Existing detection technologies often use a single feature or a combination of simple features for aneurysm identification, and train models through supervised learning methods to achieve detection and localization.

[0004] However, existing intelligent detection and positioning technologies for cerebral aneurysms have many defects and shortcomings. A single feature or a simple combination of features cannot fully express the complex morphology and anatomical structure characteristics of cerebral aneurysms, especially for small-sized or atypical aneurysms, the detection sensitivity is relatively low; existing detection methods are insufficient in distinguishing between vascular branch points and aneurysms, and are prone to false positive results, affecting the reliability of clinical applications; existing technologies lack the ability to globally analyze the topological structure of vascular networks, making it difficult to accurately locate the spatial position relationship of aneurysms relative to surrounding blood vessels, resulting in inaccurate positioning and affecting the formulation of subsequent treatment plans. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for intelligent detection and positioning of cerebral aneurysms based on multi-feature fusion, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for intelligent detection and positioning of cerebral aneurysms based on multi-feature fusion, comprising:

[0007] receiving a head angiography image to be detected and performing preprocessing to obtain a standardized vascular image; extracting features from the standardized vascular image using a multi-scale feature extraction network to obtain multiple sets of feature maps;

[0008] Fuse multiple sets of feature maps to obtain a fused feature map, and construct an initial candidate region based on the fused feature map;

[0009] The density clustering method is used to group the initial candidate regions, extract the main axis characteristics of the blood vessels and establish a global vascular probability model. The anatomical candidate region set is obtained based on the probability score screening.

[0010] A graph cut energy function is constructed to map the set of anatomical candidate regions into a Markov random field. Based on complex-valued features and conditional probabilities, the probabilistic feature map and tumor contour are obtained by minimizing the energy function.

[0011] The vascular network topology map is constructed using probabilistic feature mapping, and the spatial probability distribution is calculated based on the morphological characteristics of the tumor contour to obtain the tumor position feature descriptor.

[0012] Classification and discrimination are performed based on position feature descriptors to obtain the final cerebral aneurysm detection results; the cerebral aneurysm detection results are marked and displayed on the original head angiography image.

[0013] In an optional embodiment,

[0014] The density clustering method is used to group the initial candidate regions, extract the main axis features of the blood vessels and establish a global vascular probability model. The set of anatomical candidate regions obtained based on the probability score screening includes:

[0015] The center coordinates, region radius and region average gray value of the initial candidate region are constructed as a feature vector, and the feature vector is density clustered using a density reachable distance function to obtain a candidate region group;

[0016] Calculating the eigenvalues ​​of the candidate region group based on a multi-scale Hessian matrix, constructing a vascular main axis responsiveness function using the ratio of the eigenvalues, and extracting the vascular main axis of the candidate region group;

[0017] Acquiring cross-sectional grayscale distribution data along the main axis of the blood vessel, establishing a multi-scale probability model for the cross-sectional grayscale distribution data, and constructing a global blood vessel probability model in combination with position constraints and direction constraints;

[0018] The probability integral score of the candidate region group in the global vascular probability model is calculated, an adaptive threshold is constructed based on the statistical characteristics of the probability integral score, and the candidate regions whose probability integral score is higher than the adaptive threshold and whose vascular main axis responsiveness is higher than the preset response threshold are determined as anatomical candidate regions.

[0019] In an optional embodiment,

[0020] Acquiring cross-sectional grayscale distribution data along the main axis of the blood vessel, establishing a multi-scale probability model for the cross-sectional grayscale distribution data, and constructing a global blood vessel probability model in combination with position constraints and direction constraints includes:

[0021] A local coordinate system including the main axis parameters, radial distance and azimuth angle is constructed along the main axis of the blood vessel. According to the coordinate transformation relationship of the local coordinate system, the cross-sectional grayscale distribution data is sampled on the plane orthogonal to the main axis.

[0022] A multi-scale mixed Gaussian probability model is established for the cross-sectional grayscale distribution data, and a scale adaptive mechanism related to radial distance is introduced to dynamically adjust the variance parameters of the mixed Gaussian model;

[0023] constructing a blood vessel position probability model based on the cross-sectional grayscale distribution data, constructing the deviation value of the radial distance from the reference radius and the rate of change of the radial distance to the azimuth angle as shape constraints, and constructing the first-order and second-order rates of change of the radial distance to the main axis parameter as continuity constraints;

[0024] A directional probability distribution model is established based on the local blood vessel orientation, and the directional probability distribution model is fused with the multi-scale mixed Gaussian probability model and the blood vessel position probability model to obtain a joint probability model.

[0025] The spatial distance and tangent vector angle between adjacent main axis positions are calculated, a spatial constraint function is established, and the spatial constraint function is combined with the joint probability model to obtain a global vascular probability model.

[0026] In an optional embodiment,

[0027] A graph cut energy function is constructed to map the set of anatomical candidate regions into a Markov random field. Based on complex-valued features and conditional probabilities, the probabilistic feature map and tumor contour are obtained by minimizing the energy function, including:

[0028] The pixels in the set of candidate anatomical regions are constructed as the vertex set of an undirected graph, and the spatial adjacency between the pixels is constructed as the edge set of the undirected graph. The grayscale value of each pixel in the vertex set is collected to construct a data item, and the grayscale difference between adjacent pixels in the edge set is constructed as a smoothing item. The data item and the smoothing item are combined to generate a graph cut energy function.

[0029] Mapping the undirected graph into a Markov random field, constructing a complex-valued feature containing grayscale value, gradient and phase information in the neighborhood of each vertex, obtaining a local feature vector through Fourier transform, and calculating the conditional probability of the Markov random field;

[0030] Based on the conditional probability, pixel points in the set of anatomical candidate regions are marked, pixel points marked as tumors are used as foreground points, and other pixel points are used as background points, a connection network of foreground points and background points is constructed, network connection weights are set based on the data item and the smoothing item, and an optimal marking field is obtained by minimizing the graph cut energy function;

[0031] The local energy distribution of the optimal labeling field is calculated, a probability feature map is constructed, and the tumor contour is obtained.

[0032] In an optional embodiment,

[0033] The undirected graph is mapped to a Markov random field, a complex-valued feature containing grayscale value, gradient and phase information is constructed in the neighborhood of each vertex, and a local feature vector is obtained by Fourier transform. The conditional probability of the Markov random field is calculated as follows:

[0034] Mapping the undirected graph into a Markov random field, determining the neighborhood range of each vertex in the undirected graph, making the state of each vertex only related to the states of vertices in the corresponding neighborhood, and constructing a conditional dependency relationship that satisfies the Markov property;

[0035] In a neighborhood of each vertex of the Markov random field, calculating a three-dimensional complex-valued feature of the vertex, the three-dimensional complex-valued feature including a first feature set and a second feature set;

[0036] Extracting the grayscale values ​​and gradient vectors of pixels in the neighborhood to construct a first feature set, calculating the phase difference between adjacent pixels in the neighborhood to construct a second feature set; using the first feature set as the real part and the second feature set as the imaginary part to form a complex-valued feature of the vertex;

[0037] Performing Fourier-Mellin transform on the complex-valued features to obtain frequency domain features; activating the frequency domain features using a hyperbolic tangent function to obtain a local feature vector of the vertex;

[0038] The conditional probability of the Markov random field is calculated according to the energy distribution and phase constraint of the local eigenvector.

[0039] In an optional embodiment,

[0040] The vascular network topology map is constructed using probabilistic feature mapping, and the spatial probability distribution is calculated based on the morphological characteristics of the tumor contour. The position feature descriptors of the tumor are obtained, including:

[0041] The probability feature map is processed by non-maximum suppression and adaptive thresholding to obtain the vascular branch structure point set, and B-spline fitting is performed to obtain the topological network backbone.

[0042] Calculating the local curvature value of the topological network trunk, marking points with a curvature value higher than a preset threshold as network nodes, and constructing a topological connection relationship of blood vessels with the network nodes as vertices;

[0043] The vascular topological connection relationship between the network nodes is represented as a weighted graph structure, and the edge weights are calculated based on the spatial distance and the angle between the network nodes to obtain a vascular network topology map;

[0044] Extracting the morphological feature points of the tumor contour, calculating the perimeter, area and eccentricity of the morphological feature points to construct a geometric parameter set, and calculating the mean and variance of the curvature of the morphological feature points to construct a morphological parameter set;

[0045] Determining multiple network nodes closest to the tumor in the vascular network topology map, constructing a three-dimensional coordinate system, and calculating spatial probability parameters of the tumor center;

[0046] Calculating a topological feature distribution of the tumor region based on the vascular network topology map, wherein the topological feature distribution includes the number of adjacent network nodes, connectivity between nodes, and hierarchical distribution of nodes;

[0047] The geometric parameter set, the morphological parameter set, the spatial probability parameter and the topological feature distribution are combined to construct a position feature descriptor of a tumor.

[0048] In an optional embodiment,

[0049] Determining multiple network nodes closest to the tumor in the vascular network topology map, constructing a three-dimensional coordinate system, and calculating the spatial probability parameters of the tumor center include:

[0050] Select the three network nodes closest to the tumor mass center and construct a three-dimensional affine coordinate system based on the spatial positions of the three network nodes;

[0051] Calculating the geodesic distance between the three network nodes, using the geodesic distance as a normalized reference scale, and performing scale normalization on the three-dimensional affine coordinate system;

[0052] Taking the nearest network node as the origin, calculating the relative azimuth angle and polar angle of the tumor mass center based on the three-dimensional affine coordinate system;

[0053] Using a Gaussian kernel function to perform probability density estimation on the distribution of the tumor area in the three-dimensional affine coordinate system to obtain a probability density estimation result;

[0054] The spatial probability parameter of the tumor center is constructed by combining the relative azimuth angle, the polar angle and the probability density estimation result.

[0055] A second aspect of an embodiment of the present invention provides an intelligent cerebral aneurysm detection and positioning system based on multi-feature fusion, comprising:

[0056] The first unit is configured to receive a head angiography image to be detected and perform preprocessing to obtain a standardized vascular image; and extract features from the standardized vascular image using a multi-scale feature extraction network to obtain multiple sets of feature maps.

[0057] The second unit is used to fuse multiple sets of feature maps to obtain a fused feature map, and construct an initial candidate region based on the fused feature map;

[0058] The third unit is used to group the initial candidate regions using a density clustering method, extract the main axis features of the blood vessels and establish a global blood vessel probability model, and obtain a set of anatomical candidate regions based on the probability scores;

[0059] The fourth unit is used to construct a graph cut energy function, map the set of anatomical candidate regions into a Markov random field, and obtain the probabilistic feature map and tumor contour by minimizing the energy function based on complex-valued features and conditional probabilities;

[0060] The fifth unit is used to construct a vascular network topology map using probabilistic feature mapping, calculate the spatial probability distribution based on the morphological characteristics of the tumor contour, and obtain the location feature descriptor of the tumor;

[0061] The sixth unit is used to perform classification and discrimination based on the position feature descriptor to obtain the final cerebral aneurysm detection result; and mark and display the cerebral aneurysm detection result on the original head angiography image.

[0062] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0063] processor;

[0064] a memory for storing processor-executable instructions;

[0065] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0066] According to a fourth aspect of an embodiment of the present invention, 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 method described above is implemented.

[0067] In the embodiment of the present invention, the intelligent detection and precise positioning of cerebral aneurysms are achieved through multi-feature fusion technology, which effectively improves the detection accuracy and efficiency, reduces the misdiagnosis rate and missed diagnosis rate of clinical diagnosis, and provides doctors with reliable auxiliary decision support; it integrates multi-scale feature extraction and vascular topology analysis, overcomes the problem that traditional detection methods have difficulty in identifying tiny and irregular tumors, has a stronger ability to distinguish aneurysms in complex anatomical structures such as vascular bifurcations, and has good generalization performance; through the combination of Markov random fields and graph cut algorithms, the precise segmentation and positioning of tumors are achieved, while ensuring the detection sensitivity, the false positive rate is reduced, the clinical practicality of the detection system is improved, and it can effectively assist neurosurgeons in early diagnosis and treatment planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 Schematic diagram of the process of a method for intelligent detection and positioning of cerebral aneurysms based on multi-feature fusion according to an embodiment of the present invention;

[0069] Figure 2 Comparison of boundary positioning errors of vascular segmentation methods at different levels of vascular branches;

[0070] Figure 3 A radar chart comparing the complex-valued features and the Fourier-Mellin transform effect. DETAILED DESCRIPTION

[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying 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 making creative efforts shall fall within the scope of protection of the present invention.

[0072] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0073] Figure 1 FIG. 1 is a flow chart of a method for intelligent detection and positioning of cerebral aneurysms based on multi-feature fusion according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0074] receiving a head angiography image to be detected and performing preprocessing to obtain a standardized vascular image; extracting features from the standardized vascular image using a multi-scale feature extraction network to obtain multiple sets of feature maps;

[0075] Fuse multiple sets of feature maps to obtain a fused feature map, and construct an initial candidate region based on the fused feature map;

[0076] The density clustering method is used to group the initial candidate regions, extract the main axis characteristics of the blood vessels and establish a global vascular probability model. The anatomical candidate region set is obtained based on the probability score screening.

[0077] A graph cut energy function is constructed to map the set of anatomical candidate regions into a Markov random field. Based on complex-valued features and conditional probabilities, the probabilistic feature map and tumor contour are obtained by minimizing the energy function.

[0078] The vascular network topology map is constructed using probabilistic feature mapping, and the spatial probability distribution is calculated based on the morphological characteristics of the tumor contour to obtain the tumor position feature descriptor.

[0079] Classification and discrimination are performed based on position feature descriptors to obtain the final cerebral aneurysm detection results; the cerebral aneurysm detection results are marked and displayed on the original head angiography image.

[0080] In an optional embodiment, a density clustering method is used to group the initial candidate regions, extract the main axis features of the blood vessels and establish a global blood vessel probability model. The anatomical candidate region set obtained by screening based on the probability score includes:

[0081] The center coordinates, region radius and region average gray value of the initial candidate region are constructed as a feature vector, and the feature vector is density clustered using a density reachable distance function to obtain a candidate region group;

[0082] Calculating the eigenvalues ​​of the candidate region group based on a multi-scale Hessian matrix, constructing a vascular main axis responsiveness function using the ratio of the eigenvalues, and extracting the vascular main axis of the candidate region group;

[0083] Acquiring cross-sectional grayscale distribution data along the main axis of the blood vessel, establishing a multi-scale probability model for the cross-sectional grayscale distribution data, and constructing a global blood vessel probability model in combination with position constraints and direction constraints;

[0084] The probability integral score of the candidate region group in the global vascular probability model is calculated, an adaptive threshold is constructed based on the statistical characteristics of the probability integral score, and the candidate regions whose probability integral score is higher than the adaptive threshold and whose vascular main axis responsiveness is higher than the preset response threshold are determined as anatomical candidate regions.

[0085] In one specific embodiment, an image is preprocessed to obtain initial candidate regions. A CT angiography image is used as input data. High-brightness regions in the image are extracted as initial candidate regions using a threshold segmentation method. The threshold can be set to 1.5 times the image's average grayscale value. For example, for a CT image with a pixel range of 0-255 and an average grayscale value of 120, the threshold is set to 180.

[0086] Construct a feature vector using the center coordinates, radius, and average grayscale value of the initial candidate regions. For each initial candidate region Ri, extract its center coordinates (xi, yi, zi), calculate the equivalent radius ri (the radius when the region's area is equal to the circle's area), and the average grayscale value gi of the pixels within the region. Construct the feature vector Vi = [xi, yi, zi, ri, gi]. For example, the feature vector of a candidate region can be expressed as [128, 96, 45, 3.2, 200].

[0087] The eigenvectors are clustered using a density-reachable distance function, defined as the weighted Euclidean distance between eigenvectors, with a spatial coordinate weight of 1, a radius weight of 2, and a grayscale value weight of 0.5. A density threshold of Eps is set to 10, and a minimum point count of MinPts is set to 4. For any two eigenvectors Vi and Vj, if their weighted distance is less than Eps, they are considered density-reachable. If a sequence of eigenvectors exists such that Vi and Vj are density-reachable, they are assigned to the same cluster group. This step clusters similar candidate regions into a set, resulting in a set of candidate region groups {G1, G2, ..., Gm}.

[0088] For each candidate region group Gk, eigenvalues ​​are calculated based on the multi-scale Hessian matrix to extract the principal axis of the vessel. Specifically, the Gaussian filter scale σ is set within the scale range [1, 5] with a step size of 0.5. For each scale σ, the Hessian matrix at each point p in the candidate region group is calculated, and its eigenvalues ​​λ1, λ2, and λ3 are calculated (sorted by absolute value from smallest to largest). The eigenvalue ratio is used to construct the principal axis responsivity function R(p). This function measures the likelihood that point p belongs to a vascular structure. A higher responsivity indicates that the point is more likely to be located on the vessel centerline. For example, the three eigenvalues ​​of a point are [-0.02, -0.85, -0.87]. Given the characteristic of blood vessels extending in one direction, its responsivity R(p) is high, approximately 0.92.

[0089] Along the detected vessel's main axis, points are sampled at 1mm intervals, and the grayscale distribution of a cross section perpendicular to the main axis is obtained at each point. For each point pi on the main axis, a circular cross section Si centered at that point and perpendicular to the main axis is extracted, with a radius 1.5 times the average radius of the candidate region group. Twenty-four radial rays are sampled at 15° intervals along the radial direction, with each ray sampled at 0.5mm intervals to obtain the cross-sectional grayscale distribution data.

[0090] A multi-scale probability model is constructed for the cross-sectional grayscale distribution data. It is assumed that the grayscale distribution of vascular cross sections approximately conforms to a Gaussian distribution, with high grayscale values ​​at the center and decreasing toward the edges. For each cross section, the radial grayscale distribution is statistically analyzed, and the Gaussian distribution parameters μ (mean) and σ (standard deviation) are fitted. For example, the grayscale distribution parameters for a certain cross section are μ = 210 and σ = 15. A global vascular probability model P(G) is constructed by combining position constraints (the distance from the candidate region to the image center should not exceed 40% of the image size) and orientation constraints (the orientation of adjacent vascular segments should not vary by more than 30°).

[0091] Calculate the probability integral score S(G) for the candidate region group in the global vascular probability model. For each point p in the candidate region group G, calculate its probability P(p) in the global model. Integrate the probability values ​​of all points to obtain the total score S(G). For example, the probability integral score for a candidate region group is 0.86.

[0092] An adaptive threshold Tθ is constructed based on the statistical characteristics of the probability integral score. First, the mean value Savg and standard deviation Sstd of the scores of all candidate region groups are calculated, and then the adaptive threshold Tθ is set to Savg + 0.8 × Sstd. For example, if the mean value of the scores of all candidate region groups is 0.65 and the standard deviation is 0.15, then the adaptive threshold Tθ = 0.65 + 0.8 × 0.15 = 0.77.

[0093] Candidate regions whose probability integral score S(G) exceeds the adaptive threshold Tθ and whose vascular main axis responsivity R(p) exceeds the preset response threshold Tr (set to 0.75) are determined as anatomical candidate regions. This step effectively screens anatomical candidate regions with distinct vascular structures and regular morphology, while excluding non-vascular interference regions.

[0094] In practical applications, for a set of CT angiography images containing 1024×1024×128 voxels, there are approximately 200 initial candidate regions, which are then clustered into approximately 30 candidate region groups. Finally, 8-12 anatomical candidate regions are screened out. These regions are highly consistent with the actual vascular anatomical structure, providing accurate positioning information for subsequent vascular segmentation and lesion detection.

[0095] In this embodiment, feature vectors are constructed using the coordinates of the region center, radius, and grayscale mean, and combined with density clustering to effectively identify potential candidate vascular regions and reduce background interference. Eigenvalues ​​are calculated based on the multi-scale Hessian matrix, and the eigenvalue ratio is used to construct a vascular main axis responsiveness function, which can adapt to vessels of different calibers, enhance linear structures, and improve the accuracy of main axis extraction. A global vascular probability model is constructed by combining a multi-scale probability model of cross-sectional grayscale distribution, position constraints, and direction constraints to achieve overall continuity judgment of vascular structures and avoid local breaks or missed detections. Thresholds are automatically set based on the statistical characteristics of the probability integral score and jointly screened with the responsiveness threshold. This can significantly suppress pseudo-vascular structures while ensuring sensitivity, thereby improving the reliability and robustness of detection. The above multi-stage processing flow takes into account both local detail enhancement and global consistency, can effectively resist noise interference, and achieve high-precision and high-recall vascular anatomical region extraction.

[0096] In an optional embodiment, obtaining cross-sectional grayscale distribution data along the main axis of the blood vessel, establishing a multi-scale probability model for the cross-sectional grayscale distribution data, and constructing a global blood vessel probability model in combination with position constraints and orientation constraints includes:

[0097] A local coordinate system including the main axis parameters, radial distance and azimuth angle is constructed along the main axis of the blood vessel. According to the coordinate transformation relationship of the local coordinate system, the cross-sectional grayscale distribution data is sampled on the plane orthogonal to the main axis.

[0098] A multi-scale mixed Gaussian probability model is established for the cross-sectional grayscale distribution data, and a scale adaptive mechanism related to radial distance is introduced to dynamically adjust the variance parameters of the mixed Gaussian model;

[0099] constructing a blood vessel position probability model based on the cross-sectional grayscale distribution data, constructing the deviation value of the radial distance from the reference radius and the rate of change of the radial distance to the azimuth angle as shape constraints, and constructing the first-order and second-order rates of change of the radial distance to the main axis parameter as continuity constraints;

[0100] A directional probability distribution model is established based on the local blood vessel orientation, and the directional probability distribution model is fused with the multi-scale mixed Gaussian probability model and the blood vessel position probability model to obtain a joint probability model.

[0101] The spatial distance and tangent vector angle between adjacent main axis positions are calculated, a spatial constraint function is established, and the spatial constraint function is combined with the joint probability model to obtain a global vascular probability model.

[0102] In a specific embodiment, a local coordinate system including the main axis parameters, radial distance and azimuth angle is constructed along the main axis of the blood vessel. Specifically, for a given three-dimensional medical image data, starting from a user-specified starting point, the initial direction of the main axis of the blood vessel is determined by calculating the local blood vessel direction vector. On the main axis, a sampling point is set every 0.5 mm, and the position of the sampling point can be represented by the parameter t. At each sampling point, a plane perpendicular to the main axis is established, and the points on the plane can be represented by the radial distance r and the azimuth angle θ. For example, for a sampling point with the main axis parameter t=10.5, in its local coordinate system, the value range of r is 0 to 10 mm, and the value range of θ is 0 to 2π. Through this local coordinate system, any point (x, y, z) in three-dimensional space can be mapped to a point (t, r, θ) in the local coordinate system, and vice versa.

[0103] Based on the established local coordinate system, sampling is performed on the plane orthogonal to the main axis to obtain the grayscale distribution data of the cross section. In specific implementation, the grayscale values ​​of the original image are collected along different azimuth directions (e.g., every 5°) and different radial distances (e.g., starting from 0, with a step size of 0.2 mm, up to 10 mm) on the orthogonal plane where each main axis sampling point is located to form the grayscale distribution data of the cross section. For example, at the sampling point t = 15.5, the grayscale values ​​of 72 azimuth directions (0°, 5°, 10°, ..., 355°) and 50 radial sampling points (r = 0, 0.2, 0.4, ..., 9.8) in each direction can be obtained, for a total of 3600 data points.

[0104] A multiscale mixed Gaussian probability model is established for the acquired cross-sectional grayscale distribution data. This model uses a weighted sum of multiple Gaussian distributions to represent the grayscale value distribution, typically containing 2-4 Gaussian components, corresponding to the vessel interior, vessel wall, and background tissue. Each Gaussian component is determined by a mean, variance, and weight parameter. To address the varying characteristics of vascular diameter, a radial distance-dependent scale adaptation mechanism is introduced to dynamically adjust the variance parameter of the mixed Gaussian model. Specifically, when the radial distance r is small (e.g., r < 2 mm), a smaller variance value (e.g., 0.8) is set to improve the ability to distinguish fine structures. As r increases (e.g., 2 ≤ r < 5 mm), the variance value is increased accordingly (e.g., 1.2). When r is larger (e.g., r ≥ 5 mm), the variance value is further increased (e.g., 1.5). In this way, the model can adapt to the grayscale distribution characteristics of vessels of varying calibers.

[0105] A probabilistic model of vascular location is constructed based on cross-sectional grayscale distribution data. This model considers the shape and continuity characteristics of the blood vessels and includes shape constraints and continuity constraints. The shape constraint consists of two parts: the deviation between the radial distance and the reference radius, and the rate of change of the radial distance with respect to the azimuth angle. The reference radius can be determined by analyzing the grayscale distribution of the local cross section. For example, for the cross section at t = 20.5, the reference radius may be 3.2 mm. If the actual radial distance measured at θ = 45° is 3.5 mm, the deviation is 0.3 mm. Simultaneously, the rate of change of the radial distance with respect to the azimuth angle is obtained by calculating the difference in radial distances at adjacent azimuth angles and dividing it by the angular interval. The continuity constraint includes the first-order and second-order rates of change of the radial distance with respect to the principal axis parameter. These are obtained by calculating the difference in radial distances at corresponding azimuth angles on adjacent cross sections and the difference in the difference between the two values, and dividing them by the principal axis parameter interval. For example, if the radial distances in the θ = 90° direction are 3.6 mm and 3.7 mm at t = 25.5 and t = 26.0, respectively, then the first-order rate of change is 0.2 mm / unit parameter.

[0106] A directional probability distribution model is established based on the local vascular orientation. This model is based on the principle of continuity of vascular orientation and calculates the consistency between the vascular orientation at the current location and the forward predicted orientation. In specific implementation, the direction vectors of the first two main axis points are used to predict the direction of the current point, which is then compared with the actual calculated direction of the current point. The smaller the angle difference, the higher the probability value. For example, if the predicted direction is (0.2, 0.3, 0.9) and the actual direction is (0.25, 0.3, 0.85), the angle difference obtained by calculating the dot product of the two vectors after normalization is approximately 5°, corresponding to a higher probability value of 0.98.

[0107] The direction probability distribution model is fused with the multi-scale mixed Gaussian probability model and the blood vessel position probability model to obtain a joint probability model. In specific implementation, a weighted product method can be used, and the probability values ​​of the three models are respectively given weights (such as 0.4, 0.3, 0.3) and then multiplied. For example, if the multi-scale mixed Gaussian probability of a point is 0.85, the position probability is 0.92, and the direction probability is 0.98, then the joint probability value is 0.85. 0.4 ×0.92 0.3 ×0.98 0.3 ≈0.9.

[0108] The spatial distance and tangent vector angle between adjacent principal axis positions are calculated to establish a spatial constraint function. The spatial distance constraint ensures the continuity of the principal axis, typically setting the spacing between adjacent sampling points to 0.5-1 mm. The tangent vector angle constraint ensures the smoothness of the principal axis, typically limiting the angle between adjacent tangent vectors to no more than 30°. The spatial constraint function is combined with the joint probability model to generate a global vascular probability model, which serves as the final basis for segmentation.

[0109] Existing blood vessel segmentation methods are mainly based on threshold segmentation, region growing or level set methods, which are prone to failure when faced with low contrast, noise interference or changes in blood vessel caliber. Currently widely used technologies such as Frangi filters rely only on the tubular structure characteristics of blood vessels and cannot handle complex lesions and bifurcated structures. This embodiment overcomes the above problems by establishing a global probability framework of local coordinate system, multi-scale mixed Gaussian model and spatial constraints. By introducing a scale adaptation mechanism related to radial distance, the model can adapt to blood vessels of different calibers; combining shape constraints and continuity constraints to construct a comprehensive position probability model; ensuring the continuity of blood vessel direction through a direction probability model; fusing multiple probability models and spatial constraints to achieve global optimal segmentation. This method is superior to existing methods in the segmentation accuracy of major blood vessels and subtle branches, and the segmentation success rate on low-dose CT images is improved and the boundary positioning error is reduced.

[0110] like Figure 2 As shown in the figure, the boundary positioning error of different blood vessel segmentation methods at each level of blood vessel branches is compared. The Hausdorff distance is used as the evaluation index of the boundary positioning error. The smaller the value, the closer the segmentation boundary is to the true boundary. As can be seen from the chart, the present technical solution has achieved the smallest boundary positioning error at all levels of blood vessel branches. Specifically, in the segmentation of main blood vessels, the Hausdorff distance of the present technical solution is only 0.37mm, which is 0.05mm lower than the currently widely used U-Net deep learning method (0.42mm) and 0.31mm lower than the traditional Frangi filter (0.68mm); as the level of blood vessel branches increases, the boundary positioning error of each method generally increases, but the present technical solution always maintains the lowest error level. On the most challenging fourth-level branch, the error of the present technical solution is 1.14mm, which is significantly lower than other methods, especially compared with the 3.05mm of the region growing method, the error is reduced by 62.6%. Notably, this technical solution also demonstrates a clear advantage in terms of error growth trends, with the error increasing by 0.77mm from the main vessel to the fourth-order branches, while other methods such as the Frangi filter and the region growing method have error increases of 1.9mm and 2.2mm, respectively. These data fully demonstrate that this technical solution, by establishing a multi-scale mixed Gaussian model in a local coordinate system and combining shape constraints with continuity constraints, significantly improves the accuracy of vessel boundary positioning, especially for fine vascular branches. This has important clinical significance for accurately identifying small vessel lesions and assessing the degree of vascular stenosis.

[0111] In an optional embodiment, a graph cut energy function is constructed to map the set of anatomical candidate regions into a Markov random field. Based on complex-valued features and conditional probabilities, a probability feature map and a tumor contour are obtained by minimizing the energy function, including:

[0112] The pixels in the set of candidate anatomical regions are constructed as the vertex set of an undirected graph, and the spatial adjacency between the pixels is constructed as the edge set of the undirected graph. The grayscale value of each pixel in the vertex set is collected to construct a data item, and the grayscale difference between adjacent pixels in the edge set is constructed as a smoothing item. The data item and the smoothing item are combined to generate a graph cut energy function.

[0113] Mapping the undirected graph into a Markov random field, constructing a complex-valued feature containing grayscale value, gradient and phase information in the neighborhood of each vertex, obtaining a local feature vector through Fourier transform, and calculating the conditional probability of the Markov random field;

[0114] Based on the conditional probability, pixel points in the set of anatomical candidate regions are marked, pixel points marked as tumors are used as foreground points, and other pixel points are used as background points, a connection network of foreground points and background points is constructed, network connection weights are set based on the data item and the smoothing item, and an optimal marking field is obtained by minimizing the graph cut energy function;

[0115] The local energy distribution of the optimal labeling field is calculated, a probability feature map is constructed, and the tumor contour is obtained.

[0116] In one embodiment, the pixels in the set of candidate anatomical regions are constructed as a vertex set V of an undirected graph, which contains all pixels in the image that may belong to the tumor region. In a real-world example, a candidate anatomical region of 256×256 pixels is extracted from a CT image, with a total of 65,536 vertices. For each pixel p, its grayscale value I(p) is obtained, with grayscale values ​​ranging from 0 to 255. For example, in a lung CT image, the grayscale value of the tumor region is typically between 170 and 220, while the grayscale value of normal lung tissue is typically between 30 and 100.

[0117] The spatial adjacency relationship between pixels is constructed as an edge set E of an undirected graph, using a 4-neighborhood or 8-neighborhood connection mode. In the present embodiment, an 8-neighborhood connection is adopted, that is, each internal pixel is connected to 8 surrounding pixels. For each pixel p in the vertex set V, a data item D(p) is constructed to represent the possibility that the pixel belongs to the tumor, specifically by calculating the difference between the grayscale value I(p) of the pixel p and the predetermined tumor grayscale threshold T_f and the background grayscale threshold T_b. In the present embodiment, T_f=200 and T_b=50 are set. When I(p) is close to T_f, the D(p) value is smaller, indicating that the pixel p is more likely to belong to the tumor area.

[0118] A smoothing term V(p, q) is constructed for the grayscale difference between adjacent pixels p and q in the edge set E. The smoothing term is used to measure the label consistency of adjacent pixels. The smoothing term V(p, q) is inversely proportional to the grayscale difference between adjacent pixels. When the grayscale difference between adjacent pixels is small, the smoothing term is larger, indicating that the two pixels are more likely to belong to the same region. Specifically, for an edge (p, q)∈E, the grayscale difference between adjacent pixels |I(p)-I(q)| is calculated, and a smoothing factor σ=15 is introduced to control the degree of smoothing. When |I(p)-I(q)|<σ, the smoothing term is larger, tending to divide p and q into the same region; when |I(p)-I(q)|≥σ, the smoothing term is smaller, allowing the appearance of a region boundary at this edge.

[0119] The data term and the smoothing term are combined to generate the graph cut energy function E(X), where X represents the set of labels for all pixels and X(p) represents the label of pixel p (1 for tumor and 0 for background). The energy function is a weighted sum of the data term and the smoothing term, with a weight parameter λ = 0.8 used to control the relative importance of the two terms.

[0120] The undirected graph is mapped to a Markov random field, and complex-valued features are constructed within the neighborhood of each vertex. For pixel p, the grayscale value, gradient, and phase information are extracted within its 5×5 neighborhood window to construct the complex-valued feature F(p). The grayscale value is directly collected from the image, and the gradient is calculated using the Sobel operator to obtain the horizontal gradient G_x and vertical gradient G_y. The gradient amplitude is (G_x² + G_y²). 1 / 2 The gradient direction is arctan(G_y / G_x). Phase information is obtained by Gabor filtering a local neighborhood window. Eight directions (0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°) and three scales of Gabor filters are selected to extract texture features.

[0121] The complex-valued features F(p) are subjected to a two-dimensional fast Fourier transform (FFT) to obtain the local feature vector H(p). This vector contains both amplitude and phase spectrum information, effectively characterizing the local texture and structural features surrounding the pixel. In the experiment, the extracted feature vector had 78 dimensions, including grayscale values ​​(1 dimension), gradient features (2 dimensions), and Gabor features (75 dimensions).

[0122] Based on the feature vector H(p), the conditional probability P(X(p)|H(p)) of the Markov random field is calculated. This represents the probability that pixel p belongs to the tumor, given the feature vector H(p). A Gaussian mixture model (GMM) is used to estimate the conditional probability, using three Gaussian components for the tumor and three for the background. During the training phase, feature vectors are extracted from labeled samples and the GMM parameters are trained. During the testing phase, the conditional probabilities of new feature vectors belonging to the tumor and the background are calculated.

[0123] Based on the conditional probability P(X(p)|H(p)), the pixels in the set of anatomical candidate regions are initially labeled. The probability threshold T_p=0.7 is set. When P(X(p)|H(p))>T_p, the pixel p is marked as a tumor foreground point; when P(X(p)|H(p))<(1-T_p), the pixel p is marked as a background point; the remaining pixels are marked as undetermined points. In this embodiment, approximately 20% of the pixels in the initial labeling results are marked as foreground points, 25% are marked as background points, and 55% are undetermined points.

[0124] Construct a network connecting foreground and background points, connecting source point s to all foreground points and all background points to sink point t. Set network connection weights based on the data term D(p) and the smoothing term V(p,q). The connection weight from foreground point p to source point s is D(p), the connection weight from background point q to sink point t is D(q), and the connection weight between adjacent pixels p and q is V(p,q).

[0125] The maximum flow algorithm is applied to solve the minimum cut problem, minimizing the graph cut energy function E(X) to obtain the optimal labeling field X*. In practical implementation, the Boykov-Kolmogorov algorithm is used to calculate the maximum flow, with an algorithm complexity of O(mn²|C|), where m is the number of vertices, n is the number of edges, and |C| is the capacity of the minimum cut. In this example, the image size is 256×256, and the algorithm runs in approximately 0.8 seconds.

[0126] The local energy distribution of the optimal labeling field X* is calculated. For each pixel p, the energy function value E_local(p) is calculated within its neighborhood and normalized to the interval [0, 1] to construct a probabilistic feature map M. In the probabilistic feature map, values ​​closer to 1 indicate a greater likelihood that the pixel belongs to the tumor. The probabilistic feature map M is binarized using a threshold T_m = 0.6 to obtain a binary mask of the tumor region. Morphological operations (opening and closing) are then applied to remove isolated points and fill small holes, ultimately extracting the tumor contour.

[0127] In this embodiment, by constructing pixel points as vertices of the graph, constructing spatial adjacency relationships as edges, and combining data terms and smoothing terms to form a graph cut energy function, accurate distinction between the tumor and the background can be achieved, and the boundary positioning is clearer. ; In the Markov random field framework, not only grayscale values ​​are used, but also gradient and phase information are introduced. The complex-valued feature vector obtained by Fourier transform can more comprehensively characterize local structural features and improve the recognition ability of complex textures or low-contrast areas; Graph cut balances the data fidelity and neighborhood smoothness of each pixel by minimizing the global energy function, that is, it considers the consistency of the overall segmentation while taking into account the fine characterization of local details, avoiding isolated noise points and small block misclassification; Conditional probability is used to mark foreground (tumor) and background pixels, and a special connection network and weight setting are constructed, so that the segmentation process has the ability to adaptively adjust the target shape and grayscale distribution, enhancing the applicability to morphologically diverse lesions; The segmentation framework based on the graph cut algorithm is mature and the computational complexity is controllable. Combined with MRF characteristics and Fourier local feature processing, it can effectively resist noise interference and can balance segmentation speed and accuracy by adjusting the energy function parameters. It is suitable for large-scale clinical image processing.

[0128] In an optional embodiment, the undirected graph is mapped into a Markov random field, a complex-valued feature including grayscale value, gradient, and phase information is constructed in the neighborhood of each vertex, and a local feature vector is obtained by Fourier transform. Calculating the conditional probability of the Markov random field includes:

[0129] Mapping the undirected graph into a Markov random field, determining the neighborhood range of each vertex in the undirected graph, making the state of each vertex only related to the states of vertices in the corresponding neighborhood, and constructing a conditional dependency relationship that satisfies the Markov property;

[0130] In a neighborhood of each vertex of the Markov random field, calculating a three-dimensional complex-valued feature of the vertex, the three-dimensional complex-valued feature including a first feature set and a second feature set;

[0131] Extracting the grayscale values ​​and gradient vectors of pixels in the neighborhood to construct a first feature set, calculating the phase difference between adjacent pixels in the neighborhood to construct a second feature set; using the first feature set as the real part and the second feature set as the imaginary part to form a complex-valued feature of the vertex;

[0132] Performing Fourier-Mellin transform on the complex-valued features to obtain frequency domain features; activating the frequency domain features using a hyperbolic tangent function to obtain a local feature vector of the vertex;

[0133] The conditional probability of the Markov random field is calculated according to the energy distribution and phase constraint of the local eigenvector.

[0134] In a specific embodiment, when an undirected graph is mapped to a Markov random field, an undirected graph structure is constructed for the input image data, in which each pixel is represented as a vertex in the graph, and there are edge connections between adjacent pixels. An 8-neighborhood or 16-neighborhood system is usually used to determine the neighborhood range of each vertex in the undirected graph, that is, for a pixel with a position coordinate of (x, y) in the image, its 8-neighborhood includes pixels with coordinates of (x-1, y-1), (x-1, y), (x-1, y+1), (x, y-1), (x, y+1), (x+1, y-1), (x+1, y), (x+1, y+1). This neighborhood definition method makes the state of each vertex only related to the state of the vertices in its neighborhood, and independent of the state conditions of other vertices, thereby constructing a conditional dependency relationship that satisfies the Markov property. For example, for an image of size 512×512, the pixel at the center position (256, 256) establishes a connection relationship with 8 adjacent pixels under the 8-neighborhood system, forming a local conditional dependency structure.

[0135] Within the neighborhood of each vertex in the Markov random field, the three-dimensional complex-valued features of the vertex are calculated. The three-dimensional complex-valued features include a first feature set and a second feature set, which are the real and imaginary parts of the complex number, respectively. The grayscale values ​​and gradient vectors of the pixels in the neighborhood are extracted to construct the first feature set. The calculation method is as follows: for the pixel at position (x, y), its grayscale value is directly read from the image data, for example, the value is 125; its gradient vector is obtained by calculating the grayscale difference in the x-direction and y-direction, such as the gradient value in the x-direction is 10, the gradient value in the y-direction is -5, the gradient amplitude is 11.18, and the gradient direction angle is -26.57 degrees. For each pixel in the neighborhood, these feature values ​​are calculated to form the first feature set. Taking the 8-neighborhood as an example, the first feature set contains the grayscale value and gradient information of 9 pixels (the center point and its 8 neighbors), a total of 27 real-valued vectors.

[0136] The phase difference between adjacent pixels in the neighborhood is calculated to construct the second feature set. The phase difference calculation is based on the spatial variation trend of the grayscale value of the pixel. In specific implementation, the phase information of each pixel can be obtained by local Fourier transform, and then the phase difference between adjacent pixels is calculated. For example, for the pixel at position (x, y) and the pixel at position (x+1, y) to its right, if the phases of the two points are 1.2 radians and 1.5 radians respectively, the phase difference is 0.3 radians. In the 8-neighborhood system, the phase difference between the center point and the 8 neighbors, as well as the phase difference between the neighbors, is calculated, and a total of 20 phase difference values ​​are obtained to form the second feature set.

[0137] The first feature set is used as the real part, and the second feature set as the imaginary part, to form the complex-valued features of the vertex. To ensure that the real and imaginary parts match, the first feature set can be appropriately reduced in dimension or the second feature set can be padded. For example, the 27-dimensional first feature set can be reduced to 20 dimensions through principal component analysis and matched with the 20-dimensional second feature set. After combination, a 20-dimensional complex-valued feature vector is obtained, such as (125+0.3j, 10-0.2j, -5+0.5j, ...), where each element is a complex number.

[0138] Perform a Fourier-Mellin transform on the complex-valued features to obtain frequency domain features. The Fourier-Mellin transform is a specialized transformation that converts complex-valued vectors into the frequency domain. It effectively captures the energy distribution and phase relationships within complex-valued features. In practice, a 20-dimensional complex-valued feature vector is fast Fourier transformed to obtain 20 complex frequency domain coefficients, each representing the energy and phase information of a specific frequency component. For example, the resulting frequency domain coefficients might be (135+45j, 62-32j, 28+16j, ...).

[0139] The frequency domain features are activated using the hyperbolic tangent function to obtain the local eigenvector of the vertex. The hyperbolic tangent function can perform nonlinear transformations on features, enhancing their expressiveness. For each complex frequency domain coefficient, the hyperbolic tangent function is applied to both its real and imaginary parts. For example, for the frequency domain coefficient 135+45j, applying the hyperbolic tangent function might result in 0.99+0.98j. Applying this operation to all frequency domain coefficients yields a new complex-valued vector, which serves as the local eigenvector for the vertex. This eigenvector captures important information within the pixel neighborhood, such as texture, edges, and phase changes.

[0140] The conditional probability of a Markov random field is calculated based on the energy distribution and phase constraints of the local eigenvectors. The energy distribution can be calculated by squared eigenvector modulo values, while the phase constraints consider the phase consistency of each eigenvector component. In practice, an energy function can be defined that decreases with increasing energy of the eigenvector components and increases with increasing phase inconsistency. For example, components of the local eigenvector with high energy and consistent phases are assigned lower energy values, while components with low energy or inconsistent phases are assigned higher energy values. The energy function is converted to conditional probabilities using a Boltzmann distribution; for example, an energy value of -5 might correspond to a conditional probability of 0.95. These conditional probabilities are used in subsequent inference processes of the Markov random field, such as image segmentation and object detection.

[0141] In the prior art, image processing often uses Markov random field methods based on real-valued features, such as constructing feature vectors based on pixel grayscale and gradient information, and performing image segmentation and recognition through energy minimization. These methods usually only consider real-valued features such as the grayscale and gradient of pixels, ignoring the phase relationship between pixels, resulting in insufficient accuracy when processing complex textures and fine structures. This embodiment introduces complex-valued feature representation, combining phase information as an imaginary part with traditional real-valued features to capture richer image structure information; uses Fourier-Mellin transform to map complex-valued features to the frequency domain, effectively extracting frequency and phase characteristics; uses hyperbolic tangent function to perform nonlinear activation on frequency domain features to enhance feature expression capabilities; and constructs a conditional probability model through energy distribution and phase constraints to achieve more accurate Markov random field inference.

[0142] like Figure 3 The figure shows an analysis of the advantages of complex-valued features and the Fourier-Mellin transform in feature extraction capabilities in this technical solution. The figure uses radar charts to visually compare the performance of four different methods across eight key performance indicators. In particular, in terms of phase sensitivity, this technical solution achieved 92.67%, 14.83 percentage points higher than traditional feature extraction methods. In terms of texture information capture rate, this technical solution achieved 89.43%, significantly superior to the 76.92% of phase features alone and the 72.36% of traditional methods. In terms of edge fidelity, due to the incorporation of phase difference information as the imaginary part, this technical solution achieved a high performance of 95.43%. Low-light enhancement achieved a score of 86.92%, significantly higher than the 68.93% of traditional methods. Particularly noteworthy is that this technical solution strikes a good balance between detail preservation and noise suppression, reaching 88.25% and 82.45% respectively—a balance that traditional methods struggle to achieve. This technical solution also performed well in terms of feature discrimination and anti-interference capabilities, reaching 91.38% and 84.72% respectively. The introduction of complex-valued features enables this technical solution to simultaneously process both amplitude and phase information of the image, thereby achieving more comprehensive feature expression and better performance in various evaluation indicators.

[0143] In an optional embodiment, a vascular network topology map is constructed using a probabilistic feature map, and the spatial probability distribution is calculated in combination with the morphological characteristics of the tumor contour to obtain a position feature descriptor of the tumor, including:

[0144] The probability feature map is processed by non-maximum suppression and adaptive thresholding to obtain the vascular branch structure point set, and B-spline fitting is performed to obtain the topological network backbone.

[0145] Calculating the local curvature value of the topological network trunk, marking points with a curvature value higher than a preset threshold as network nodes, and constructing a topological connection relationship of blood vessels with the network nodes as vertices;

[0146] The vascular topological connection relationship between the network nodes is represented as a weighted graph structure, and the edge weights are calculated based on the spatial distance and the angle between the network nodes to obtain a vascular network topology map;

[0147] Extracting the morphological feature points of the tumor contour, calculating the perimeter, area and eccentricity of the morphological feature points to construct a geometric parameter set, and calculating the mean and variance of the curvature of the morphological feature points to construct a morphological parameter set;

[0148] Determining multiple network nodes closest to the tumor in the vascular network topology map, constructing a three-dimensional coordinate system, and calculating spatial probability parameters of the tumor center;

[0149] Calculating a topological feature distribution of the tumor region based on the vascular network topology map, wherein the topological feature distribution includes the number of adjacent network nodes, connectivity between nodes, and hierarchical distribution of nodes;

[0150] The geometric parameter set, the morphological parameter set, the spatial probability parameter and the topological feature distribution are combined to construct a position feature descriptor of a tumor.

[0151] In one embodiment, a probabilistic feature map is processed to construct a vascular network topology map. A non-maximum suppression algorithm is applied to the probabilistic feature map of blood vessels extracted from a 3D medical image, retaining local maxima and suppressing non-maximum points. Specifically, for each point in the map, the probability values ​​within its 26-neighborhood are examined. A point is retained only if its probability value is greater than that of all neighboring points; otherwise, its probability value is set to zero.

[0152] Adaptive thresholding is performed on the feature maps processed with non-maximum suppression. The adaptive threshold is determined using the Otsu algorithm, which maximizes the inter-class variance between foreground and background to determine the optimal threshold. In practice, for normal liver vascular data, a typical threshold range is between 0.35 and 0.65, dynamically adjusted based on image contrast and noise level. After thresholding, the resulting binary point set represents the points of the vascular branching structure.

[0153] A B-spline fitting algorithm is applied to the binary point set to reconstruct the vascular trunks. A third-order B-spline curve is used, with control point spacing set to 3 to 5 voxels and a smoothing factor of 0.85 to smooth out subtle noise while maintaining accurate vascular orientation. In a real-world case, a liver vascular network containing approximately 5,000 structural points can generate approximately 200 to 300 smooth curve segments through B-spline fitting, forming the vascular trunk network.

[0154] Local curvature values ​​were calculated for the fitted vascular trunks. At each point, curvature was calculated using the five-point differential method, which approximates the curvature by differencing the coordinates of five adjacent points. Experience has shown that for liver vascular networks, an appropriate curvature threshold ranges from 0.12 to 0.25, with a value of 0.18 achieving optimal results on a test dataset. Points above this threshold were labeled as network nodes, typically corresponding to vascular bifurcations or sharp bends.

[0155] The vascular topology connection relationship is constructed based on the identified network nodes. The specific implementation method is to detect the connectivity between all adjacent network nodes. If there is a continuous vascular pathway between two nodes, a connection relationship is established. In actual cases, a typical liver vascular network can contain 60 to 100 network nodes, forming 80 to 150 topological connections between these nodes.

[0156] The topological connectivity of blood vessels is represented as a weighted graph structure, with edge weights taking into account both spatial distance and strike angle. In the weight calculation formula, spatial distance is calculated using Euclidean distance, and strike angle is determined by the angle between the line connecting two nodes and their local tangent vectors. In practice, a weight ratio of 7:3 between distance and angle factors is used, which has shown optimal performance in precise location tasks.

[0157] Morphological features of the tumor contour were extracted and analyzed. Morphological feature points of the tumor contour were extracted using an edge detection algorithm. The Canny algorithm was used, with the low threshold set at 40% of the high threshold. The high threshold was determined based on the image histogram and was typically 75% of the peak value.

[0158] The geometric parameters of the tumor's morphological features are calculated, including perimeter, area, and eccentricity. The perimeter is calculated by summing the distances between adjacent feature points, the area is calculated from the contour coordinates using Green's formula, and the eccentricity is determined by fitting an ellipse and calculating the ratio of the major and minor axes. In a real-world case, a liver tumor with a diameter of 32 mm had a perimeter of approximately 100.5 mm, an area of ​​804.2 square millimeters, and an eccentricity of 0.72.

[0159] The mean and variance of the curvature of morphological feature points were calculated to construct a morphological parameter set. Local curvature was calculated for 100 points uniformly sampled from the contour. In a typical liver tumor case, with a smoothing window size of 5% of the contour perimeter, the mean curvature was 0.068 and the variance was 0.0023. These parameters can effectively distinguish between benign and malignant tumors.

[0160] In the vascular network topology, multiple network nodes closest to the tumor are identified. Typically, three to five nodes are selected, though four is preferred in practice to balance computational complexity and localization accuracy. A three-dimensional coordinate system is constructed using these nodes as reference points. The relative position of the tumor center relative to this coordinate system is calculated to derive spatial probability parameters.

[0161] The topological distribution of the tumor region is calculated based on the vascular network topology map. Specifically, this involves counting the number of network nodes within twice the tumor radius, typically 3 to 8; analyzing the connectivity between these nodes and calculating the connectivity degree, with an average value typically between 1.5 and 2.8; and assessing the hierarchical distribution of nodes, categorizing them into levels 1 to 3 based on their distance from the main vessels, and calculating the proportion of nodes at each level.

[0162] The researchers combined geometric parameters (perimeter, area, and eccentricity), morphological parameters (curvature mean and variance), spatial probability parameters (relative position coordinates), and topological feature distributions (number of nodes, connectivity, and hierarchical proportions) into a high-dimensional vector to form a tumor location feature descriptor. This descriptor typically has 15 to 20 dimensions, and normalization is performed to map each dimension to a range of 0 to 1, making it easier for subsequent machine learning algorithms to process.

[0163] In this embodiment, by simultaneously extracting the topological trunk of vascular branches, the geometric shape of the tumor contour, the spatial probability distribution and the network topological features, a multi-dimensional characterization of the tumor location is achieved, the amount of descriptor information is enhanced, and the discrimination ability of subsequent analysis or classification is improved; based on the topological network nodes and weighted graph structure, the spatial distance and orientation relationship between the blood vessels and the tumor is accurately calculated, and the spatial probability parameters of the tumor center are quantified in the three-dimensional coordinate system, which can effectively capture the spatial coupling characteristics of the tumor and the surrounding blood vessel distribution; the network nodes are automatically marked by the curvature threshold, and combined with B Spline fitting generates a smooth trunk, which not only ensures the sensitive capture of key bifurcation points of blood vessels, but also avoids interference from noise or small pseudo-structures, and realizes robust topological network reconstruction; the statistical features of contour perimeter, area, eccentricity and curvature are extracted, providing comprehensive parameters for describing the diversity of tumor morphology, and improving the adaptability and generalization ability of the model to lesions of different shapes; the node set and tumor center are embedded in the three-dimensional coordinate system, and combined with the spatial probability model, it supports the quantitative analysis of the true three-dimensional position and distribution of the tumor in the anatomical structure, which is helpful for clinical positioning and surgical planning; the geometric, morphological, spatial probability and topological features are integrated into a unified position feature descriptor, which is convenient for direct use by subsequent machine learning or statistical models, and realizes end-to-end automated tumor positioning and risk assessment.

[0164] In an optional embodiment, determining a plurality of network nodes closest to the tumor in the vascular network topology map, constructing a three-dimensional coordinate system, and calculating the spatial probability parameters of the tumor center includes:

[0165] Select the three network nodes closest to the tumor mass center and construct a three-dimensional affine coordinate system based on the spatial positions of the three network nodes;

[0166] Calculating the geodesic distance between the three network nodes, using the geodesic distance as a normalized reference scale, and performing scale normalization on the three-dimensional affine coordinate system;

[0167] Taking the nearest network node as the origin, calculating the relative azimuth angle and polar angle of the tumor mass center based on the three-dimensional affine coordinate system;

[0168] Using a Gaussian kernel function to perform probability density estimation on the distribution of the tumor area in the three-dimensional affine coordinate system to obtain a probability density estimation result;

[0169] The spatial probability parameter of the tumor center is constructed by combining the relative azimuth angle, the polar angle and the probability density estimation result.

[0170] In one embodiment, for a detected cerebral aneurysm, the system calculates the Euclidean distance from all network nodes in the cerebral vascular network to the tumor center of mass. A network node refers to a point on the centerline of a vascular vessel obtained after cerebral vascular segmentation and skeletonization. The tumor center of mass is obtained by averaging the positions of all voxels in the tumor region. For example, assuming that the tumor region contains 250 voxels, each with spatial coordinates (x, y, z), the coordinates of the tumor center of mass are the arithmetic mean of all voxel coordinates. After the calculation is completed, the system sorts all network nodes in ascending order according to their distance to the tumor center of mass and selects the three network nodes closest to them, denoted as nodes A, B, and C. Assume that in a certain case, the coordinates of node A are (125.3, 87.6, 105.2) mm, the coordinates of node B are (124.1, 89.2, 106.5) mm, and the coordinates of node C are (126.7, 88.4, 104.8) mm.

[0171] Based on the three selected network nodes, the system constructs a three-dimensional affine coordinate system. Node A, closest to the tumor's centroid, is set as the coordinate origin. Vector AB is used as the x-axis, the projection of vector AC on a plane perpendicular to vector AB is used as the y-axis, and the cross product of the x- and y-axes is used as the z-axis. This establishes a local three-dimensional affine coordinate system. Using the coordinates of the three nodes in the example above, the x-axis vector is calculated to be (-1.2, 1.6, 1.3) mm. After normalization, the unit vectors are (-0.51, 0.68, 0.55). Similarly, the unit vectors for the y and z axes are calculated.

[0172] To eliminate scale differences between cases, the system calculates the geodesic distance between the three network nodes and uses it as a normalized reference scale. Geodesic distance refers to the shortest path distance along the centerline of the blood vessel and is implemented using the Dijkstra algorithm. For example, the geodesic distance from node A to node B is 5.8mm, the geodesic distance from node A to node C is 6.2mm, and the geodesic distance from node B to node C is 4.5mm. The system takes the average of these three distances, 5.5mm, as the normalized reference scale and normalizes the scale of the three-dimensional affine coordinate system, that is, dividing all coordinate values ​​by this reference scale.

[0173] In a standardized three-dimensional affine coordinate system, the system calculates the azimuth and polar angles of the tumor's center of mass relative to the origin (node ​​A). The azimuth angle is the angle between the projection of the tumor's center of mass on the xy plane and the positive x-axis, with a range of [0, 360] degrees. The polar angle is the angle between the tumor's center of mass and the positive z-axis, with a range of [0, 180] degrees. Assume the calculated azimuth angle is 75.3 degrees and the polar angle is 42.8 degrees. These two angles together describe the spatial orientation of the tumor's center of mass in the local coordinate system.

[0174] For the distribution characteristics of the tumor area in the three-dimensional affine coordinate system, the system uses the Gaussian kernel function to estimate the probability density. Each voxel in the tumor area is mapped to a standardized three-dimensional affine coordinate system to form point cloud data. Select an appropriate bandwidth parameter h (h = 0.8 in this embodiment) and assign a Gaussian kernel function to each point. For any point p in space, its probability density estimate is the sum of the Gaussian kernel function values ​​of all points divided by the total number of points. For example, at the center of the tumor, the calculated probability density estimate is 0.85, indicating that the probability of this position belonging to the tumor is high.

[0175] The system divides the tumor region into multiple subregions and calculates a probability density estimate in each subregion, forming a probability density distribution map. For example, if the tumor region is divided into 27 subregions (a 3×3×3 grid), and the probability density estimate is calculated at the center point of each subregion, a vector containing 27 values ​​is obtained, such as [0.23, 0.56, 0.42, ..., 0.31].

[0176] The system combines the relative azimuth, polar angle, and probability density estimates to construct spatial probability parameters for the tumor center. These parameters include an azimuth value of 75.3 degrees, a polar angle value of 42.8 degrees, and a probability density distribution vector. The system then combines these parameters into a feature vector, which characterizes the tumor's precise location and distribution within the cerebral vascular network. This feature vector can be expressed as [75.3, 42.8, 0.23, 0.56, 0.42, ..., 0.31].

[0177] In practical applications, the system can calculate the spatial probability parameters of newly detected suspected cerebral aneurysms and compare them with a database of known cerebral aneurysms to assist doctors in making diagnostic and treatment decisions. For example, cerebral aneurysms with certain locations and distribution characteristics may have a higher risk of rupture and require timely intervention and treatment.

[0178] Traditional cerebral aneurysm localization methods primarily rely on global coordinate systems or anatomical landmarks, making them difficult to adapt to inter-individual differences in cerebrovascular anatomy. For example, commonly used localization methods include absolute coordinate positioning based on a standard head coordinate system or qualitative descriptions based on specific vascular names (such as the anterior communicating artery and posterior communicating artery). These methods struggle to accurately locate the aneurysm within the complex and ever-changing cerebrovascular network, especially for aneurysms located in atypical locations.

[0179] The technical solution of this embodiment overcomes the above-mentioned shortcomings. By selecting the vascular network node closest to the tumor as a reference point, a local relative coordinate system is established. This system is independent of global coordinates or standard templates and can better adapt to individual differences. At the same time, the geodesic distance normalization process solves the problem of comparing cases of different scales. The use of the Gaussian kernel function for probability density estimation not only describes the central position of the tumor, but also characterizes the spatial distribution characteristics of the tumor, providing more comprehensive information for subsequent risk assessment.

[0180] ;

[0181] The table above compares the key performance indicators of three cerebral aneurysm localization methods. It clearly contrasts the performance differences between the traditional global coordinate method, the anatomical landmark method, and this technical solution across multiple key indicators, and quantifies the improvements achieved by this technical solution.

[0182] In terms of positioning accuracy, the traditional global coordinate method achieved 4.87±1.26mm, the anatomical landmark method achieved 3.65±0.98mm, and this technical solution significantly improved it to 1.73±0.42mm, a 52.6% improvement over the best traditional method. In terms of individual adaptability score (0-10 point scale), the traditional global coordinate method scored 5.3 points, the anatomical landmark method scored 6.8 points, and this technical solution reached 8.9 points, an improvement of 30.9%. In terms of complex location recognition rate, the traditional global coordinate method scored 62.4%, the anatomical landmark method scored 75.8%, and this technical solution reached 94.2%, an improvement of 24.3%.

[0183] While the algorithm execution time of this solution (1.65 seconds) is slightly longer than the traditional global coordinate method (1.23 seconds), it is 42.5% shorter than the anatomical landmark method (2.87 seconds), maintaining high computational efficiency. In terms of spatial description completeness, the traditional global coordinate method achieved 68.5%, the anatomical landmark method achieved 72.3%, and the solution achieved 96.7%, a 33.7% improvement.

[0184] The data of this technical solution in the table intuitively demonstrates its advantages in various indicators, especially the significant improvement in positioning accuracy, complex location recognition rate and spatial description integrity, which fully demonstrates the technical superiority of the aneurysm localization method based on the local relative coordinate system.

[0185] The intelligent cerebral aneurysm detection and positioning system based on multi-feature fusion according to an embodiment of the present invention includes:

[0186] The first unit is configured to receive a head angiography image to be detected and perform preprocessing to obtain a standardized vascular image; and extract features from the standardized vascular image using a multi-scale feature extraction network to obtain multiple sets of feature maps.

[0187] The second unit is used to fuse multiple sets of feature maps to obtain a fused feature map, and construct an initial candidate region based on the fused feature map;

[0188] The third unit is used to group the initial candidate regions using a density clustering method, extract the main axis features of the blood vessels and establish a global blood vessel probability model, and obtain a set of anatomical candidate regions based on the probability scores;

[0189] The fourth unit is used to construct a graph cut energy function, map the set of anatomical candidate regions into a Markov random field, and obtain the probabilistic feature map and tumor contour by minimizing the energy function based on complex-valued features and conditional probabilities;

[0190] The fifth unit is used to construct a vascular network topology map using probabilistic feature mapping, calculate the spatial probability distribution based on the morphological characteristics of the tumor contour, and obtain the location feature descriptor of the tumor;

[0191] The sixth unit is used to perform classification and discrimination based on the position feature descriptor to obtain the final cerebral aneurysm detection result; and mark and display the cerebral aneurysm detection result on the original head angiography image.

[0192] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0193] processor;

[0194] a memory for storing processor-executable instructions;

[0195] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0196] According to a fourth aspect of an embodiment of the present invention, 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 method described above is implemented.

[0197] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0198] 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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, 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 embodiments of the present invention.

Claims

1. An intelligent cerebral aneurysm detection and positioning method based on multi-feature fusion, characterized in that: include: receiving a head angiography image to be detected and performing preprocessing to obtain a standardized vascular image; extracting features from the standardized vascular image using a multi-scale feature extraction network to obtain multiple sets of feature maps; Fuse multiple sets of feature maps to obtain a fused feature map, and construct an initial candidate region based on the fused feature map; The density clustering method is used to group the initial candidate regions, extract the main axis characteristics of the blood vessels and establish a global vascular probability model. The anatomical candidate region set is obtained based on the probability score screening. A graph cut energy function is constructed to map the set of anatomical candidate regions into a Markov random field. Based on complex-valued features and conditional probabilities, the probabilistic feature map and tumor contour are obtained by minimizing the energy function. The vascular network topology map is constructed using probabilistic feature mapping, and the spatial probability distribution is calculated based on the morphological characteristics of the tumor contour to obtain the tumor position feature descriptor. Classification and discrimination are performed based on position feature descriptors to obtain the final cerebral aneurysm detection results; the cerebral aneurysm detection results are marked and displayed on the original head angiography image.

2. The method according to claim 1, characterized in that The density clustering method is used to group the initial candidate regions, extract the main axis features of the blood vessels and establish a global vascular probability model. The set of anatomical candidate regions obtained based on the probability score screening includes: The center coordinates, region radius and region average gray value of the initial candidate region are constructed as a feature vector, and the feature vector is density clustered using a density reachable distance function to obtain a candidate region group; Calculating the eigenvalues ​​of the candidate region group based on a multi-scale Hessian matrix, constructing a vascular main axis responsiveness function using the ratio of the eigenvalues, and extracting the vascular main axis of the candidate region group; Acquiring cross-sectional grayscale distribution data along the main axis of the blood vessel, establishing a multi-scale probability model for the cross-sectional grayscale distribution data, and constructing a global blood vessel probability model in combination with position constraints and direction constraints; The probability integral score of the candidate region group in the global vascular probability model is calculated, an adaptive threshold is constructed based on the statistical characteristics of the probability integral score, and the candidate regions whose probability integral score is higher than the adaptive threshold and whose vascular main axis responsiveness is higher than the preset response threshold are determined as anatomical candidate regions.

3. The method according to claim 2, characterized in that Acquiring cross-sectional grayscale distribution data along the main axis of the blood vessel, establishing a multi-scale probability model for the cross-sectional grayscale distribution data, and constructing a global blood vessel probability model in combination with position constraints and direction constraints includes: A local coordinate system including the main axis parameters, radial distance and azimuth angle is constructed along the main axis of the blood vessel. According to the coordinate transformation relationship of the local coordinate system, the cross-sectional grayscale distribution data is sampled on the plane orthogonal to the main axis. A multi-scale mixed Gaussian probability model is established for the cross-sectional grayscale distribution data, and a scale adaptive mechanism related to radial distance is introduced to dynamically adjust the variance parameters of the mixed Gaussian model; constructing a blood vessel position probability model based on the cross-sectional grayscale distribution data, constructing the deviation value of the radial distance from the reference radius and the rate of change of the radial distance to the azimuth angle as shape constraints, and constructing the first-order and second-order rates of change of the radial distance to the main axis parameter as continuity constraints; A directional probability distribution model is established based on the local blood vessel orientation, and the directional probability distribution model is fused with the multi-scale mixed Gaussian probability model and the blood vessel position probability model to obtain a joint probability model. The spatial distance and tangent vector angle between adjacent main axis positions are calculated, a spatial constraint function is established, and the spatial constraint function is combined with the joint probability model to obtain a global vascular probability model.

4. The method according to claim 1, wherein A graph cut energy function is constructed to map the set of anatomical candidate regions into a Markov random field. Based on complex-valued features and conditional probabilities, the probabilistic feature map and tumor contour are obtained by minimizing the energy function, including: The pixels in the set of candidate anatomical regions are constructed as the vertex set of an undirected graph, and the spatial adjacency between the pixels is constructed as the edge set of the undirected graph. The grayscale value of each pixel in the vertex set is collected to construct a data item, and the grayscale difference between adjacent pixels in the edge set is constructed as a smoothing item. The data item and the smoothing item are combined to generate a graph cut energy function. Mapping the undirected graph into a Markov random field, constructing a complex-valued feature containing grayscale value, gradient and phase information in the neighborhood of each vertex, obtaining a local feature vector through Fourier transform, and calculating the conditional probability of the Markov random field; Based on the conditional probability, pixel points in the set of anatomical candidate regions are marked, pixel points marked as tumors are used as foreground points, and other pixel points are used as background points, a connection network of foreground points and background points is constructed, network connection weights are set based on the data item and the smoothing item, and an optimal marking field is obtained by minimizing the graph cut energy function; The local energy distribution of the optimal labeling field is calculated, a probability feature map is constructed, and the tumor contour is obtained.

5. The method according to claim 4, characterized in that The undirected graph is mapped to a Markov random field, a complex-valued feature containing grayscale value, gradient and phase information is constructed in the neighborhood of each vertex, and a local feature vector is obtained by Fourier transform. The conditional probability of the Markov random field is calculated as follows: Mapping the undirected graph into a Markov random field, determining the neighborhood range of each vertex in the undirected graph, making the state of each vertex only related to the states of vertices in the corresponding neighborhood, and constructing a conditional dependency relationship that satisfies the Markov property; In a neighborhood of each vertex of the Markov random field, calculating a three-dimensional complex-valued feature of the vertex, the three-dimensional complex-valued feature including a first feature set and a second feature set; Extracting the grayscale values ​​and gradient vectors of pixels in the neighborhood to construct a first feature set, calculating the phase difference between adjacent pixels in the neighborhood to construct a second feature set; using the first feature set as the real part and the second feature set as the imaginary part to form a complex-valued feature of the vertex; Performing Fourier-Mellin transform on the complex-valued features to obtain frequency domain features; activating the frequency domain features using a hyperbolic tangent function to obtain a local feature vector of the vertex; The conditional probability of the Markov random field is calculated according to the energy distribution and phase constraint of the local eigenvector.

6. The method according to claim 1, wherein The vascular network topology map is constructed using probabilistic feature mapping, and the spatial probability distribution is calculated based on the morphological characteristics of the tumor contour. The position feature descriptors of the tumor are obtained, including: The probability feature map is processed by non-maximum suppression and adaptive thresholding to obtain the vascular branch structure point set, and B-spline fitting is performed to obtain the topological network backbone. Calculating the local curvature value of the topological network trunk, marking points with a curvature value higher than a preset threshold as network nodes, and constructing a topological connection relationship of blood vessels with the network nodes as vertices; The vascular topological connection relationship between the network nodes is represented as a weighted graph structure, and the edge weights are calculated based on the spatial distance and the angle between the network nodes to obtain a vascular network topology map; Extracting the morphological feature points of the tumor contour, calculating the perimeter, area and eccentricity of the morphological feature points to construct a geometric parameter set, and calculating the mean and variance of the curvature of the morphological feature points to construct a morphological parameter set; Determining multiple network nodes closest to the tumor in the vascular network topology map, constructing a three-dimensional coordinate system, and calculating the spatial probability parameters of the tumor center; Calculating a topological feature distribution of the tumor region based on the vascular network topology map, wherein the topological feature distribution includes the number of adjacent network nodes, connectivity between nodes, and hierarchical distribution of nodes; The geometric parameter set, the morphological parameter set, the spatial probability parameter and the topological feature distribution are combined to construct a position feature descriptor of a tumor.

7. The method according to claim 6, characterized in that Determining multiple network nodes closest to the tumor in the vascular network topology map, constructing a three-dimensional coordinate system, and calculating the spatial probability parameters of the tumor center include: Select the three network nodes closest to the tumor mass center and construct a three-dimensional affine coordinate system based on the spatial positions of the three network nodes; Calculating the geodesic distance between the three network nodes, using the geodesic distance as a normalized reference scale, and performing scale normalization on the three-dimensional affine coordinate system; Taking the nearest network node as the origin, calculating the relative azimuth angle and polar angle of the tumor mass center based on the three-dimensional affine coordinate system; Using a Gaussian kernel function to perform probability density estimation on the distribution of the tumor area in the three-dimensional affine coordinate system to obtain a probability density estimation result; The spatial probability parameter of the tumor center is constructed by combining the relative azimuth angle, the polar angle and the probability density estimation result.

8. An intelligent cerebral aneurysm detection and positioning system based on multi-feature fusion, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is configured to receive a head angiography image to be detected and perform preprocessing to obtain a standardized vascular image; and extract features from the standardized vascular image using a multi-scale feature extraction network to obtain multiple sets of feature maps. The second unit is used to fuse multiple sets of feature maps to obtain a fused feature map, and construct an initial candidate region based on the fused feature map; The third unit is used to group the initial candidate regions using a density clustering method, extract the main axis features of the blood vessels and establish a global blood vessel probability model, and obtain a set of anatomical candidate regions based on the probability scores; The fourth unit is used to construct a graph cut energy function, map the set of anatomical candidate regions into a Markov random field, and obtain the probabilistic feature map and tumor contour by minimizing the energy function based on complex-valued features and conditional probabilities; The fifth unit is used to construct a vascular network topology map using probabilistic feature mapping, calculate the spatial probability distribution based on the morphological characteristics of the tumor contour, and obtain the location feature descriptor of the tumor; The sixth unit is used to perform classification and discrimination based on the position feature descriptor to obtain the final cerebral aneurysm detection result; and mark and display the cerebral aneurysm detection result on the original head angiography image.

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 according to 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