Pulmonary embolism lesion segmentation method and system based on image recognition and storage medium

Abstract

The application relates to the technical field of image processing, and discloses a pulmonary embolism lesion segmentation method and system based on image recognition and a storage medium. The method comprises the following steps: extracting a multi-level blood vessel topological structure of a CTPA image through blood vessel diameter gradient analysis; modeling blood vessel density distribution by using a Weibull mixed model to obtain an embolism characteristic parameter; estimating a pixel embolism probability by using variational Bayesian inference to generate a lesion distribution map; performing multi-scale feature fusion on the lesion probability map to obtain a segmentation boundary; and optimizing the boundary based on blood vessel connectivity constraints to obtain a final embolism lesion segmentation result. The application solves the problems that blood vessel level differentiation processing cannot be realized, accurate probability modeling is lacked, and anatomic constraint verification is lacked. The application improves the precision of pulmonary embolism lesion segmentation.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method, system and storage medium for segmenting pulmonary embolism lesions based on image recognition. Background Technology

[0002] Current image recognition-based methods for pulmonary embolism lesion segmentation primarily rely on traditional thresholding and region growing techniques. These methods segment CTPA images by setting contrast agent density thresholds or manually setting seed points. Monitoring systems typically employ a uniform segmentation algorithm for all vessel levels, lacking the ability to differentiate between different levels of vessels, such as the main pulmonary artery, segmental branches, and subsegmental branches. Quality assessment is mainly based on simple density differences and morphological features, lacking quantitative analysis of the probability distribution and uncertainty of embolic lesions.

[0003] However, existing technologies have significant shortcomings. First, they lack the ability to process blood vessel layers effectively. Traditional methods cannot effectively distinguish the morphological differences and hemodynamic characteristics of blood vessels at different layers, resulting in low detection accuracy and a high false positive rate for microemboli in subsegmental branches. Second, they lack accurate probabilistic modeling capabilities. Existing methods have failed to establish accurate probability distribution models for different tissue components within blood vessels, making it impossible to accurately distinguish the density distribution characteristics of contrast agents, thrombi, and vessel walls. Furthermore, the segmentation results lack anatomical constraint validation, relying mainly on image features for segmentation and lacking post-processing optimization based on vascular connectivity and hemodynamic principles.

[0004] The inability to accurately identify and classify vascular layers makes it impossible to establish specialized density distribution models for different vascular layers, thus hindering accurate estimation of embolism probability and quantification of uncertainty. Furthermore, the lack of precise probability distribution modeling capabilities prevents the construction of an effective multi-scale feature fusion mechanism, hindering unified detection and accurate segmentation of embolic lesions of different sizes. Finally, the lack of multi-scale feature fusion support makes it impossible to establish a complete vascular connectivity constraint and hemodynamic verification system, thereby hindering the verification of the anatomical rationality of the segmentation results and boundary optimization. Summary of the Invention

[0005] This application provides a method, system, and storage medium for segmenting pulmonary embolism lesions based on image recognition, addressing the problems of inability to achieve differential processing at the vascular level, lack of accurate probabilistic modeling, and anatomical constraint verification. This application improves the accuracy of pulmonary embolism lesion segmentation.

[0006] In a first aspect, this application provides a method for segmenting pulmonary embolism lesions based on image recognition. The method includes: extracting vascular tree hierarchically from CTPA three-dimensional image data using vascular diameter gradient analysis to obtain multi-level vascular topology data; modeling the intravascular density distribution using a Weibull mixture model based on the multi-level vascular topology data to obtain embolism density feature parameters; estimating the embolism probability of vascular pixels using variational Bayesian inference based on the embolism density feature parameters to obtain a lesion probability distribution map; performing multi-scale feature fusion processing on the lesion probability distribution map to obtain lesion segmentation boundaries; and optimizing vascular connectivity constraints based on the lesion segmentation boundaries to obtain the final embolism lesion segmentation result.

[0007] Secondly, this application provides an image recognition-based pulmonary embolism lesion segmentation system, the image recognition-based pulmonary embolism lesion segmentation system comprising:

[0008] The extraction module is used to perform hierarchical extraction of vascular tree data from CTPA three-dimensional image data through vascular diameter gradient analysis to obtain multi-level vascular topology data.

[0009] The modeling module is used to model the intravascular density distribution based on the multi-level vascular topology data using a Weibull mixture model to obtain embolism density characteristic parameters.

[0010] The estimation module is used to perform embolism probability estimation on blood vessel pixels based on the embolism density feature parameters through variational Bayesian inference to obtain a lesion probability distribution map.

[0011] The fusion module is used to perform multi-scale feature fusion processing on the lesion probability distribution map to obtain the lesion segmentation boundary;

[0012] The optimization module is used to optimize the vascular connectivity constraints based on the lesion segmentation boundary to obtain the final embolized lesion segmentation result.

[0013] Thirdly, an image recognition-based pulmonary embolism lesion segmentation device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the image recognition-based pulmonary embolism lesion segmentation device to execute the above-described image recognition-based pulmonary embolism lesion segmentation method.

[0014] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the above-described image recognition-based pulmonary embolism lesion segmentation method.

[0015] The technical solution provided in this application utilizes the technique of extracting vascular tree levels from CTPA three-dimensional image data through vascular diameter gradient analysis. This technique enables precise hierarchical identification of the main pulmonary artery, segmental branches, and subsegmental branches, overcoming the limitations of traditional methods that uniformly process all vascular levels. It significantly improves the adaptability and detection accuracy for vessels of different sizes. The application of the Weibull mixture model in intravascular density distribution modeling effectively describes the complex density distribution characteristics of contrast agents, thrombi, and vascular wall tissue components. In particular, it reveals a significant difference between the bimodal or multimodal distribution pattern of embolic regions and the unimodal distribution of normal vessels, providing a mathematical basis for accurately distinguishing different tissue components. The application of variational Bayesian inference technology in embolism probability estimation not only provides pixel-level embolism probability values ​​but also quantifies the uncertainty of the prediction results. This solves the problem of the lack of confidence assessment in traditional binary segmentation methods, enabling clinicians to make more accurate diagnostic decisions based on the confidence level of the probability distribution.

[0016] The implementation of multi-scale feature fusion processing technology, through the synergistic effect of large-scale morphological features, medium-scale texture features, and small-scale boundary features, achieves full-coverage detection from major trunk embolisms to subsegmental microembolisms. The adaptive weight allocation of the attention mechanism ensures that embolic lesions of different sizes can obtain the most suitable feature combination. The application of vascular connectivity constraint optimization processing technology integrates prior knowledge of vascular anatomy and hemodynamic principles into the post-segmentation processing, effectively eliminating erroneous segmentation that violates physiological laws and ensuring that the segmentation results conform to both image features and medical rationality. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of an embodiment of the pulmonary embolism lesion segmentation method based on image recognition in this application.

[0019] Figure 2 This is a schematic diagram of an embodiment of the pulmonary embolism lesion segmentation system based on image recognition in this application.

[0020] Figure 3 This is a schematic block diagram of the pulmonary embolism lesion segmentation device based on image recognition in an embodiment of the present invention. Detailed Implementation

[0021] This application provides a method, system, and storage medium for segmenting pulmonary embolism lesions based on image recognition. The terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0022] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the pulmonary embolism lesion segmentation method based on image recognition in this application includes:

[0023] Step S101: Perform vascular tree hierarchical extraction processing on CTPA three-dimensional image data through vascular diameter gradient analysis to obtain multi-level vascular topology data.

[0024] Step S102: Based on the multi-level vascular topology data, the density distribution within the blood vessel is modeled using the Weibull mixture model to obtain embolism density characteristic parameters;

[0025] Step S103: Based on the embolism density feature parameters, perform embolism probability estimation on the vascular pixel points through variational Bayes inference to obtain the lesion probability distribution map;

[0026] Step S104: Perform multi-scale feature fusion processing on the lesion probability distribution map to obtain the lesion segmentation boundary;

[0027] Step S105: Optimize the vascular connectivity constraints based on the lesion segmentation boundary to obtain the final embolized lesion segmentation result.

[0028] It is understood that the executing entity of this application can be an image recognition-based pulmonary embolism lesion segmentation system, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0029] Specifically, CTPA 3D image data is processed through vessel diameter gradient analysis. First, anisotropic diffusion filtering is applied to the CTPA 3D image data. Anisotropic diffusion filtering is an edge-preserving denoising technique that removes noise caused by uneven contrast agent distribution by strong diffusion in smooth regions and suppressing diffusion in edge regions. The filtering process adaptively adjusts the diffusion coefficient according to the magnitude of the image gradient. Regions with large gradients have small diffusion coefficients at the vessel edges to maintain edge clarity, while regions with small gradients have large diffusion coefficients inside the vessels to smooth noise, resulting in denoised contrast agent distribution data. Next, Hessian matrix eigenvalue decomposition is performed. The Hessian matrix is ​​a second-order partial derivative matrix that describes the local curvature information of the image at each pixel. Vessel structures exhibit a tubular morphology, and the eigenvalues ​​of their Hessian matrix have a specific distribution pattern. The largest eigenvalue corresponds to the curvature in the vessel cross-section direction, medium eigenvalues ​​are close to zero, and the smallest eigenvalue corresponds to the curvature along the vessel axis. By analyzing the magnitude and sign of the eigenvalues, the position of the vessel centerline can be accurately identified, forming the coordinates of the vessel centerline skeleton. Then, the local diameter change rate and curvature features are calculated. The diameter change rate is obtained by analyzing the local cross-sectional area change at each point on the vessel centerline, and the curvature features are obtained by calculating the first and second derivatives of the centerline. These geometric parameters reflect the morphological characteristics of the vessels. Vessels at different levels have different diameter ranges and degrees of curvature, forming a set of vessel geometric parameters. Subsequently, the set of vessel geometric parameters is input into a hierarchical discriminant function. The hierarchical discriminant function establishes classification rules based on vessel diameter thresholds and bifurcation angle features. The main pulmonary artery diameter is usually greater than a certain threshold and has a large bifurcation angle, the segmental branch artery diameter is in the medium range, and the subsegmental branch artery has the smallest diameter and a small bifurcation angle. The discriminant function automatically classifies the vessels into three levels of labels. Finally, topological relationships are constructed. Based on the connectivity and bifurcation structure of the vessel centerline, a hierarchical relationship of the vessel tree is established, recording the parent-child relationship, spatial location, and hierarchical identifier of each vessel branch, forming a multi-level vessel topology data containing complete topological information.

[0030] Based on multi-level vascular topology data, a Weibull mixture model was used to model the intravascular density distribution. First, the density values ​​of the cross-sections of vascular branches were extracted. The spatial location and geometry of each vascular branch were determined according to the vascular topology data. Cross-sectional images were extracted on a plane perpendicular to the vessel centerline, and the CT density value of each pixel within the cross-section was calculated. In normal vessels, contrast agents exhibit a uniform high-density distribution, while embolic areas show a low-density or mixed-density distribution due to blood flow obstruction. Statistical analysis yielded layered vascular density sampling data. Next, the density sampling data was input into the Weibull mixture model for component fitting. The Weibull distribution is a flexible probability distribution; its shape parameter controls the shape characteristics of the distribution, and its scale parameter controls the broadening degree of the distribution. The mixture model assumes that the intravascular density distribution is a linear combination of multiple Weibull distributions with certain weights. Each component corresponds to different tissue components such as contrast agents, thrombi, and vessel walls. The fitting process used maximum likelihood estimation to determine the parameters of each Weibull distribution, obtaining the initial values ​​of the shape parameter, scale parameter, and mixture weights. Then, the expectation-maximization algorithm is used for iterative optimization. Expectation-maximization is an iterative algorithm for handling latent variable problems. In the expectation step, the posterior probability of each data point belonging to each component is estimated based on the current parameters. In the maximization step, the parameters of the Weibull distribution are updated based on the posterior probability. The iteration process continues until the parameters converge or the maximum number of iterations is reached, resulting in a converged Weibull parameter combination. Subsequently, the probability density functions of the three components—contrast agent, thrombus, and vessel wall—are calculated. Based on the converged Weibull parameters and mixed weights, the probability density of each component at different density values ​​is calculated. The contrast agent component exhibits a high-density peak distribution, the thrombus component exhibits a medium-to-low density distribution, and the vessel wall component exhibits a low-density distribution. The weighted sum of the three probability density functions constitutes the density distribution model. Finally, a mapping relationship between vessel hierarchy and Weibull parameters is established. Due to differences in diameter and blood flow conditions, the Weibull parameters of vessels at different hierarchy levels also differ. Through statistical analysis, a correspondence between hierarchy identifiers and Weibull parameters is established, forming embolism density characteristic parameters.

[0031] Based on embolism density feature parameters, probability estimation of vascular pixels is performed using variational Bayesian inference. First, the embolism density feature parameters are input into a variational posterior distribution to construct a Bayesian classification model. Variational Bayesian inference is an approximate inference method that simplifies computational complexity by approximating the true posterior distribution through a variational distribution. During the construction process, a prior distribution of the embolism state of each pixel is defined, and the prior probability of embolism occurrence is set based on medical statistics and clinical experience. This prior distribution of embolism probability is established by combining the parameter information of the Weibull mixture model. Next, the likelihood function is calculated for the density values ​​of the vascular pixels. The likelihood function describes the probability of observing a specific density value under a given embolism state. Based on the probability density function of the Weibull mixture model, the likelihood of each pixel under embolism and non-embolism states is calculated. The density values ​​of embolic pixels better conform to the Weibull distribution of thrombus components, while the density values ​​of non-embolic pixels better conform to the Weibull distribution of contrast agent components, yielding pixel-level embolism likelihood values. Then, the parameters are updated using a variational lower bound optimization algorithm. The variational lower bound is the lower bound of the log-marginal likelihood of the true posterior distribution. The parameters of the variational distribution are optimized by maximizing the variational lower bound. The optimization process uses the coordinate ascent method to alternately update the variational parameters until the variational lower bound converges, resulting in a converged set of variational parameters. Subsequently, the variational parameters are input into the posterior probability calculation module for embolism probability inference. Based on Bayes' theorem combined with the prior distribution, likelihood function, and variational parameters, the posterior probability of embolism for each pixel is calculated. The posterior probability integrates prior information and observational information, reflecting the confidence that the pixel belongs to the embolism region, thus obtaining the pixel embolism probability value. Finally, spatial mapping and uncertainty quantification are performed. The embolism probability values ​​of the pixels are mapped onto the original CTPA image according to spatial coordinates to form a probability distribution map. At the same time, the uncertainty of the probability estimate is calculated. The uncertainty is quantified by the variance or entropy of the posterior distribution. High uncertainty regions indicate that the classification result is unreliable and requires further investigation, resulting in a lesion probability distribution map containing probability values ​​and uncertainty information.

[0032] Multi-scale analysis of the lesion probability distribution map is performed to obtain segmentation boundaries. First, multi-scale convolutional kernel feature extraction is performed based on the lesion probability distribution map. Different sized convolutional kernels are used to extract local and global features of the image. Large-scale convolutional kernels extract the overall morphological features and long-distance spatial relationships of blood vessels, suitable for detecting large-area embolic lesions. Medium-scale convolutional kernels extract the texture features and medium-range spatial patterns of blood vessels, suitable for identifying medium-sized embolic regions. Small-scale convolutional kernels extract the boundary features and detailed information of blood vessels, suitable for accurately locating the boundary of small emboli. Features at different scales describe the performance of embolic lesions at different spatial resolutions, resulting in large-scale morphological features, medium-scale texture features, and small-scale boundary features. Next, the multi-scale features are input into an attention mechanism for weight allocation. The attention mechanism adaptively assigns importance weights to different features. The attention weight for each scale feature is calculated based on the local environment and global context information of the current pixel. Important features receive higher weights, and less important features receive lower weights. The calculation of attention weights is based on the similarity and complementarity between features; similar features have lower weights to avoid redundancy, while complementary features have higher weights to enhance expressive power, resulting in a scale-adaptive feature weight vector. Then, multi-scale features are weighted and fused according to the weight vector. Weighted fusion integrates features from different scales into a unified feature representation through linear combination. The fusion process preserves important features while suppressing noise and interference. The fused feature integrates information from multiple scales, exhibiting stronger expressive power and robustness, resulting in a fused feature response map. Subsequently, spatial consistency constraints are applied using Conditional Random Fields (CRFs). CRFs are probabilistic graphical models that can establish spatial dependencies between pixels and label consistency constraints. Constraints include label similarity between adjacent pixels, label consistency within regions, and boundary smoothness. The optimal pixel label allocation is solved by optimizing the energy function, eliminating isolated misclassified points and discontinuous segmentation regions, resulting in spatially continuous segmentation results. Finally, distance constraints from the vessel centerline are validated. The distance from each pixel to the centerline is calculated based on the vessel centerline's location information. Embolized areas must be located inside the vessel; pixels too far from the centerline are considered incorrectly segmented and need to be removed. Distance constraints ensure that the segmentation results conform to the vessel's geometry and anatomical structure, yielding the lesion segmentation boundary.

[0033] The final optimization is performed based on the lesion segmentation boundaries to ensure the medical rationality of the segmentation results. First, a vascular branch connectivity graph is constructed based on the lesion segmentation boundaries. A connectivity graph is a graph data structure where nodes represent vascular branches and edges represent the connectivity relationships between branches. The connectivity graph is constructed based on vascular topology data and segmentation boundary information. Each node contains information on the location, size, and embolism status of the branch, and each edge contains information on connectivity relationships and blood flow direction. Graph theory algorithms are used to analyze the connectivity of the vascular tree and the embolism propagation path to obtain embolism propagation path constraints. Next, the constraints are input into a graph theory algorithm for abnormal segmentation region identification. Graph theory algorithms include depth-first search, breadth-first search, and shortest path algorithms. By traversing the connectivity graph, segmentation regions that violate anatomical rules are identified, such as spatially discontinuous isolated embolic regions, embolism propagation patterns against the blood flow direction, and embolism distributions that do not conform to the vascular bifurcation rules. These abnormal regions are usually errors in the segmentation algorithm and need to be corrected, resulting in isolated embolism markers that violate anatomical rules. Then, anomaly regions are removed using a connectivity analysis algorithm. Based on the concept of connected components, connectivity analysis retains spatially continuous embolic regions that conform to anatomical patterns, while deleting isolated or unreasonable embolic regions. The removal process considers the size, shape, and location of the embolic regions; excessively small isolated regions are identified as noise, regions with overly irregular shapes are identified as mis-segmented, and regions with unreasonable locations are identified as artifacts, resulting in a connectivity-corrected segmentation result. Next, hemodynamic constraint verification is performed. Hemodynamic constraints are based on the physical laws of blood flow; complete embolism of an upstream vessel will lead to interruption of blood flow in a downstream vessel, and the downstream vessel should not be filled with contrast agent. The verification process checks whether the embolism distribution conforms to the blood flow direction and pressure gradient. Segmentation results that do not conform to hemodynamic laws need to be corrected, resulting in a consistency check result for upstream and downstream vessel embolism. Finally, iterative boundary optimization is performed. Boundary optimization uses an iterative algorithm to gradually adjust the segmentation boundary to make it more accurate and smooth. The optimization process combines image gradient information, regional statistical information, and shape prior knowledge, solving for the optimal boundary position by minimizing the energy function. The iterative process continues until the boundary converges or reaches the preset accuracy requirement, obtaining the final embolic lesion segmentation result.

[0034] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0035] Anisotropic diffusion filtering was performed on the CTPA three-dimensional image data to obtain the noise-reducing contrast agent distribution data;

[0036] Based on the noise-reducing contrast agent distribution data, Hessian matrix eigenvalue decomposition was performed to obtain the coordinates of the vascular centerline skeleton.

[0037] Based on the coordinates of the vascular centerline skeleton, the local diameter change rate and curvature characteristics are calculated to obtain a set of vascular geometric parameters;

[0038] The set of vascular geometric parameters is input into a hierarchical discriminant function for branch classification processing to obtain three-level classification labels for the main pulmonary artery, segmental branches, and subsegmental branches.

[0039] Topological relationship construction was performed on the three-level classification labels to obtain multi-level vascular topology data.

[0040] Specifically, anisotropic diffusion filtering performs edge-preserving denoising on CTPA 3D image data to eliminate artifacts caused by uneven contrast agent distribution. Anisotropic diffusion filtering is an image denoising technique based on partial differential equations. Its core principle is to perform strong diffusion in smooth areas of the image while suppressing diffusion in edge areas. During the filtering process, the gradient magnitude of each pixel is first calculated. The gradient magnitude reflects the intensity of local gray-level changes; the gradient is larger at the vessel edges and embolism boundaries, and smaller inside the vessel and in the background. The diffusion coefficient function is adaptively adjusted according to the gradient magnitude. When the gradient magnitude is small, the diffusion coefficient is large, allowing for strong smoothing to remove noise; when the gradient magnitude is large, the diffusion coefficient approaches zero, suppressing diffusion to maintain edge sharpness. The diffusion equation is solved using the finite difference method, iteratively updating the gray-level value of each pixel. During the iteration process, high-frequency noise from the contrast agent distribution is gradually smoothed, while the sharp features of the vessel edges are preserved. The filtering parameters include diffusion time, gradient threshold, and number of iterations. These parameters need to be optimized according to the characteristics of CTPA images to form denoised contrast agent distribution data that removes contrast agent distribution noise while preserving vascular structure.

[0041] The Hessian matrix eigenvalue decomposition process extracts the vessel centerline based on denoised contrast agent distribution data. The Hessian matrix, composed of the second-order partial derivatives of the image, describes the local curvature information at each pixel. For 3D CTPA images, the Hessian matrix is ​​a 3x3 symmetric matrix containing six independent second-order partial derivative components. Vascular structures appear as tubular shapes in images, exhibiting specific curvature characteristics: smaller curvature along the vessel axis and larger curvature along the radial direction. Eigenvalue decomposition decomposes the Hessian matrix into three eigenvalues ​​and corresponding eigenvectors. The magnitude and sign of the eigenvalues ​​reflect the geometric properties of the local structure. For an ideal tubular structure, the largest eigenvalue corresponds to the principal curvature along the vessel's cross-section (large and negative), the medium eigenvalue corresponds to the curvature along another cross-section (medium and negative), and the smallest eigenvalue corresponds to the curvature along the vessel axis (close to zero). By analyzing the magnitude and sign pattern of the three eigenvalues, the vessel response function calculates the probability that each pixel belongs to a vascular structure; pixels with high response values ​​constitute candidate regions for the vessel. The centerline extraction algorithm finds local maxima within the candidate blood vessel region. These maxima correspond to the center positions of the blood vessels. By connecting adjacent center points, a continuous blood vessel centerline is formed, resulting in the coordinates of the blood vessel centerline skeleton, which includes spatial coordinates and connectivity relationships.

[0042] The calculation process for vascular geometric parameters involves calculating geometric parameters describing the morphological characteristics of the blood vessel based on the coordinates of the vessel's centerline skeleton. The calculation of the local diameter change rate is based on cross-sectional analysis. For each point on the vessel's centerline, a cross-sectional image is extracted on a plane perpendicular to the centerline. The vessel's boundary contour is determined using thresholding or region growing algorithms, the cross-sectional area is calculated, and the equivalent diameter is estimated based on the circular assumption. The diameter change rate is obtained by dividing the diameter difference between adjacent centerline points by the distance interval; a positive value indicates vessel dilation, a negative value indicates vessel constriction, and a zero value indicates a stable diameter. The curvature feature is calculated based on the geometric properties of the centerline. The first derivative represents the tangent direction of the centerline, and the second derivative represents the rate of change of the tangent direction, i.e., curvature. Curvature calculation in three-dimensional space uses the curvature formula, calculating the curvature value of each point using the first and second derivatives of the centerline coordinates. The curvature value reflects the degree of tortuosity of the vessel; straight segments have zero curvature, while curved segments have greater curvature. The torsional feature describes the spirality of the blood vessel in three-dimensional space, calculated using the cross product of the third derivative and the aforementioned derivatives. The bifurcation angle is calculated at the bifurcation point of the blood vessel by analyzing the angle between the direction vectors of the centerlines before and after the bifurcation. These geometric parameters comprehensively describe the morphological characteristics of the blood vessel, including diameter, diameter variation, degree of tortuosity, torsion characteristics, and bifurcation angle, forming a comprehensive set of vascular geometric parameters.

[0043] The hierarchical discriminant function (HDI) branch classification process inputs a set of vascular geometric parameters into the classification algorithm for automatic vascular hierarchical identification. The HDI is a classification model built upon vascular anatomy and statistical principles, comprehensively considering multiple features such as vessel diameter, location, bifurcation angle, and geometric morphology for classification decisions. The main pulmonary artery is categorized based on its large diameter, proximity to the heart, and large bifurcation angle. The diameter threshold is typically set to be greater than a specific value. Location is measured by spatial distance from the hilum, and the bifurcation angle reflects the typical anatomical structure of the main pulmonary artery dividing into left and right pulmonary arteries. Segmental branches are categorized based on their medium diameter, moderate bifurcation angle, and specific distribution pattern. Their diameter falls between that of the main pulmonary artery and subsegmental branches, their bifurcation pattern is relatively regular, and their spatial distribution follows the anatomical division of the lung segments. Subsegmental branches are categorized based on their small diameter, small bifurcation angle, and dense distribution. These vessels are located within lung segments, their diameter is usually smaller than a set threshold, their bifurcation angle is relatively sharp, and their distribution exhibits a dendritic diffusion pattern. The classification algorithm employs machine learning methods such as decision trees or support vector machines to learn the characteristic patterns of blood vessels at each level through training data, establishing a mapping relationship from geometric parameters to level labels. During the classification process, the probability or confidence score of each blood vessel branch belonging to each level is calculated, and the level with the highest probability is selected as the final classification result, resulting in a three-level classification label identifying the level to which each blood vessel branch belongs: the main pulmonary artery, segmental branches, and subsegmental branches.

[0044] The topology relationship construction process performs spatial relationship analysis on the three-level classification labels to establish the hierarchical structure of the vascular tree. Topology relationship construction is based on the connectivity and bifurcation structure of the vascular centerline, establishing parent-child hierarchical relationships by analyzing the spatial adjacency relationships between vascular branches and the blood flow direction. The construction process first identifies vascular bifurcation points, which are the convergence points of multiple vascular branches. The number and identification of vascular branches connected to each bifurcation point are determined by analyzing the connectivity of the centerline. The parent-child relationship is determined based on the vascular hierarchy and blood flow direction; higher-level vascular branches are designated as parent nodes, and lower-level vascular branches as child nodes, with blood flow from parent nodes to child nodes. The main pulmonary artery is the root node at the top level of the vascular tree, segmental arteries are intermediate nodes connecting the main pulmonary artery and subsegmental arteries, and subsegmental arteries are leaf nodes at the bottom level. The tree structure is constructed using a depth-first search or breadth-first search algorithm, traversing all connected vascular branches starting from the root node and recursively establishing the hierarchical relationships. Each node records its hierarchical identifier, spatial coordinates, geometric parameters, parent and child node information, while edges record connectivity, blood flow direction, and bifurcation angle information. The topological relationships also include the naming and coding of vascular branches; each vascular branch is assigned a unique identifier according to anatomical standards to facilitate subsequent localization and tracking. This ultimately forms a multi-level vascular topology data structure containing complete hierarchical relationships, spatial locations, geometric features, and connectivity information. This data structure not only records the static morphological information of blood vessels but also reflects the dynamic functional relationships and blood flow propagation paths of the vascular system.

[0045] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0046] Based on multi-level vascular topology data, the density values ​​of vascular branch cross sections are extracted and processed to obtain layered vascular density sampling data;

[0047] The layered vascular density sampling data is input into the Weibull mixture model for component fitting to obtain the shape parameters, scale parameters and initial values ​​of the mixture weights.

[0048] The shape parameter, scale parameter, and initial value of the mixed weights are iteratively optimized using the expectation-maximization algorithm to obtain a convergent Weibull parameter combination.

[0049] The density distribution model is obtained by processing the probability density functions of the three components of contrast agent, thrombus and vessel wall based on the convergent Weibull parameter combination;

[0050] Based on the density distribution model, a vascular hierarchy-Weibull parameter mapping relationship was established to obtain embolism density characteristic parameters.

[0051] Specifically, density values ​​are acquired based on multi-level vascular topology data. During extraction, the centerline position and spatial orientation of each vascular branch are first determined based on the vascular topology data. The vascular centerline provides the geometric axis and orientation information of the vessel. Cross-sectional extraction is performed on a plane perpendicular to the vascular centerline. By calculating the tangent direction of the centerline at each sampling point, a cross-sectional plane perpendicular to the tangent is constructed. Three-dimensional interpolation techniques are used to extract the cross-sectional image. The original CTPA three-dimensional data is resampled along the cross-sectional plane to obtain a two-dimensional vascular cross-sectional image. Density value extraction requires determining the vascular boundary. The location of the vascular wall is identified using threshold segmentation or gradient detection methods, dividing the cross-section into an internal and external region. The internal region contains different components such as contrast agent, thrombus, and residual blood. Each component exhibits different density values ​​in the CT image: contrast agent has a high density, typically within a specific range; thrombus has a medium density; and blood has a low density. Statistical analysis of the density values ​​includes the calculation of statistics such as mean, variance, skewness, and kurtosis. These statistics describe the central tendency, dispersion, and shape characteristics of the density distribution. Density data were grouped according to vascular hierarchical identifiers. The density distribution characteristics of the main pulmonary artery, segmental branches, and subsegmental branches differed, requiring separate processing to improve modeling accuracy. During sampling, sampling points were placed at equal intervals along the centerline of each vascular branch. The sampling interval was determined based on the vessel diameter and the required analytical precision; smaller vessels required denser sampling to capture local variations. The final result was hierarchical vascular density sampling data grouped by vascular hierarchy, with each data point containing spatial location, hierarchy identifier, and density value information.

[0052] Layered vascular density sampling data are input into a Weibull mixture model for probability distribution modeling. The Weibull distribution is a flexible continuous probability distribution whose probability density function is controlled by shape and scale parameters. The shape parameter determines the shape of the distribution curve, while the scale parameter determines the broadening and location of the distribution. The mixture model assumes that the observed density distribution is a linear combination of multiple Weibull distributions, with each component corresponding to different tissue components within the blood vessel. The component fitting process begins with parameter initialization, estimating initial parameter values ​​using the moment estimation method or a simplified version of maximum likelihood estimation. The initial value of the shape parameter is estimated by analyzing the skewness and kurtosis characteristics of the density distribution. Skewness reflects the degree of asymmetry in the distribution, and kurtosis reflects the sharpness of the distribution; these characteristics have a mathematical relationship with the shape parameter of the Weibull distribution. The initial value of the scale parameter is estimated by analyzing the mean and variance of the density distribution. Both the mean and variance of the Weibull distribution are functions of the shape and scale parameters; the initial parameter estimate can be obtained by solving a system of equations. The initial values ​​of the mixture weights represent the proportion of each component in the overall distribution, estimated through cluster analysis or peak detection methods. The k-means clustering algorithm divides the density data into several clusters, each corresponding to a Weibull component. The cluster size determines the initial values ​​of the mixing weights. Peak detection methods search for local maxima in the histogram of the density distribution; each peak corresponds to a potential component, and the height and width of the peak provide information for parameter initialization. The fitting process also needs to determine the number of components. Information criteria, such as the Akaike information criterion or the Bayesian information criterion, are used to compare the model performance with different numbers of components, and the optimal number of components is selected. Finally, an initial parameter set containing the shape parameters, scale parameters, and mixing weights for each Weibull component is obtained.

[0053] The Expectation-Maximization (EM) algorithm iteratively optimizes the parameters of a Weibull mixture model to obtain the optimal fit. EEM is a classic optimization method for probabilistic models with latent variables, particularly suitable for parameter estimation in mixture models. The core idea is to alternately execute the expectation step and the maximization step until the parameters converge to a local optimum. In the expectation step, the posterior probability of each observed data point belonging to each Weibull component is calculated; these probabilities are called the membership degree or responsibility degree. The posterior probability is calculated based on Bayes' theorem, combining the current parameter estimates, prior probabilities, and likelihood function to calculate the probability density of each data point under each component. Normalization is then performed to ensure that the sum of the posterior probabilities of all components is one. In the maximization step, the model parameters, including shape parameters, scale parameters, and mixture weights, are updated based on the posterior probabilities calculated in the expectation step. The update formula for the mixture weights is the average of the posterior probabilities of all data points for that component, reflecting the relative importance of that component in the population. Updating the parameters of the Weibull distribution requires solving the maximum likelihood estimation equation. Since the likelihood function of the Weibull distribution is quite complex, numerical optimization methods such as the Newton-Raphson method or gradient descent are typically used. Updating the shape parameters involves solving transcendental equations, requiring iterative numerical methods. Updating the scale parameter is relatively simple, having an analytical solution. The convergence of the algorithm is determined by monitoring changes in the likelihood function; convergence is considered achieved when the increment of the likelihood function between consecutive iterations is less than a preset threshold. To avoid getting trapped in local optima, the algorithm typically employs a multiple random initialization strategy, selecting the result with the largest likelihood function value as the final solution. The iterative process also needs to address numerical stability issues; excessively small weights of a component can lead to numerical instability, requiring regularization techniques or parameter constraints to ensure algorithm stability. After multiple iterations of optimization, a convergent combination of Weibull parameters is obtained, which best fits the observed blood vessel density distribution data.

[0054] The probability density function calculation process calculates the probability distribution characteristics of each tissue component based on a convergent Weibull parameter combination. The probability density function of the contrast agent component is calculated based on its corresponding Weibull distribution parameters. Contrast agents typically exhibit high density characteristics in CTPA images, and their Weibull distribution scale parameter is relatively large, while the shape parameter determines the sharpness of the distribution. The probability density function of the thrombus component reflects the density distribution pattern of the embolic tissue. Thrombus density is usually lower than that of contrast agents but higher than that of blood, and its Weibull distribution parameters are between those of contrast agents and the vessel wall. Variations in the shape parameter reflect the complexity and heterogeneity of the thrombus component. The probability density function of the vessel wall component describes the density characteristics of the vessel wall tissue. The vessel wall density is relatively stable and low, and its Weibull distribution typically has a small scale parameter and specific shape parameters. The probability density function is calculated using the standard formula for the Weibull distribution, and the function value represents the probability density of the component at a specific density value. The probability density functions of the three components are linearly combined according to mixed weights to form a density distribution model, which can describe the mixed distribution characteristics of various tissue components within the blood vessel. Model validation was conducted by comparing the degree of agreement between the theoretical distribution and actual observational data, using statistical methods such as the chi-square test and the Kolmogorov-Smirnov test to evaluate the quality of fit. The distribution model also provides the ability to separate components; by comparing the probability densities of each component at a specific density value, it is possible to determine which tissue component that density value most likely corresponds to, providing a probabilistic basis for subsequent embolism detection.

[0055] The process of establishing the mapping relationship between vascular levels and Weibull parameters is based on a density distribution model to establish the correspondence between different vascular levels and Weibull parameters. The mapping relationship is established based on vascular anatomy and hemodynamic principles. Due to differences in diameter, blood flow velocity, and contrast agent concentration, the density distribution characteristics of vessels at different levels exhibit systematic differences. The main pulmonary artery, as a large vessel, has high blood flow velocity and high contrast agent concentration; its Weibull parameters typically show a large combination of scale parameters and specific shape parameters, with the contrast agent component dominating in the mixed weights. The Weibull parameters of segmental arteries fall between those of the main pulmonary artery and subsegmental arteries, with moderate blood flow velocity and moderate contrast agent concentration; their parameter combinations reflect transitional distribution characteristics. Subsegmental arteries, as small vessels, have slow blood flow velocity and relatively low contrast agent concentration, making them prone to incomplete filling; their Weibull parameters typically show smaller scale parameters, with a larger range of variation in shape parameters. The mapping relationship is established using statistical learning methods. By analyzing the association patterns between vascular levels and Weibull parameters in a large amount of sample data, a mapping function from level identifiers to parameter space is established. The mapping function can take various forms, such as linear functions, polynomial functions, or neural networks. The selection criterion is a balance between mapping accuracy and generalization ability. Dimensionality reduction of the parameter space is achieved through methods such as principal component analysis or linear discriminant analysis, projecting the high-dimensional Weibull parameter space into a low-dimensional space to facilitate visualization analysis and pattern recognition. The mapping relationship also considers the influence of individual differences and pathological states. Different patients have different vascular characteristics, and pathological states such as arteriosclerosis and inflammation can affect the density distribution of blood vessels. The mapping model needs to have a certain degree of robustness to adapt to these changes. The final established mapping relationship organically combines vascular hierarchical information with Weibull parameter features to form embolism density feature parameters. These parameters not only include the statistical characteristics of density distribution but also incorporate prior knowledge of vascular anatomy and pathology.

[0056] In one specific embodiment, the process of inputting the layered vascular density sampling data into the Weibull mixture model for component fitting may specifically include the following steps:

[0057] Statistical analysis of density values ​​was performed on the layered vascular density sampling data to obtain the vascular density distribution characteristics.

[0058] Based on the characteristics of blood vessel density distribution, Weibull probability density function fitting is performed to obtain the initial Weibull model parameters;

[0059] The initial Weibull model parameters are input into the maximum likelihood estimation algorithm for parameter optimization to obtain optimized Weibull parameter values;

[0060] Based on the optimized Weibull parameter values, parameter analysis is performed to obtain the shape and scale parameters of the Weibull distribution;

[0061] The shape parameters and scale parameters are processed by component weight calculation to obtain the initial values ​​of shape parameters, scale parameters and mixed weights.

[0062] Specifically, the statistical analysis process first calculates the central tendency indicators of the density data, including the arithmetic mean, median, and mode. The arithmetic mean reflects the overall level of the density data; the median indicates the central location of the density distribution and is insensitive to outliers; and the mode shows the most frequently occurring value in the density data. Measures of dispersion include the calculation of variance, standard deviation, interquartile range, and range. Variance and standard deviation quantify the degree of dispersion of density values ​​around the mean; the interquartile range describes the distribution range of the middle 50% of the data; and the range reflects the overall magnitude of variation in density values. The analysis of the distribution shape is conducted using skewness and kurtosis. Skewness measures the degree of asymmetry of the density distribution relative to a normal distribution; positive skewness indicates a longer right tail, and negative skewness indicates a longer left tail. Kurtosis describes the sharpness of the distribution; kurtosis indicates a concentrated and sharp distribution, while hypokurtosis indicates a flat and dispersed distribution. Histogram analysis, by dividing the density values ​​into several intervals and counting the frequency of each interval, reveals the overall shape and multimodal characteristics of the density distribution. A multimodal distribution suggests the presence of multiple distinct organizational components. The cumulative distribution function is constructed by sorting density values ​​and calculating cumulative probabilities, providing a complete description of the density distribution and quantile information. Outlier detection uses box plots or z-scores to identify density values ​​that deviate from the normal range; these outliers may correspond to image artifacts, contrast agent extravasation, or special pathological structures. Stratified analysis calculates the statistical characteristics of the main pulmonary artery, segmental branches, and subsegmental branches based on the hierarchical identifiers in the vascular topology data. The density distribution characteristics of vessels at different levels exhibit systematic differences, requiring separate modeling to improve accuracy. Finally, a comprehensive vascular density distribution feature is obtained, incorporating information such as mean, variance, skewness, kurtosis, quantiles, and distribution shape.

[0063] The Weibull distribution is a widely used continuous probability distribution. Its probability density function contains two key parameters: shape and scale. The shape parameter controls the shape of the distribution curve, while the scale parameter controls the location and breadth of the distribution. The fitting process first uses the method of moments (MoM) to obtain initial estimates of the parameters. MoM is based on the principle that sample moments equal population moments, establishing a system of equations to solve for the parameters. The first and second moments of the Weibull distribution are functions of both the shape and scale parameters. A system of equations containing the two unknown parameters can be established using the sample mean and sample variance. MoM estimation of the shape parameter requires solving transcendental equations, typically using numerical iterative methods such as the bisection method or Newton's method. Estimating the scale parameter is relatively simple and can be directly calculated using analytical formulas. The probability paper method is another parameter estimation technique. It estimates parameters by converting the Weibull distribution into a linear relationship. Specifically, the density data is sorted, empirical cumulative probabilities are calculated, a logarithmic transformation is performed, and then the data points are plotted on Weibull probability paper. The Weibull parameters are determined by the slope and intercept of the fitted line using linear regression. Goodness-of-fit evaluation employed statistical indicators such as the coefficient of determination, mean squared error, and chi-square test. The coefficient of determination measures the degree to which the fitted line explains data variation, the mean squared error quantifies the difference between predicted and observed values, and the chi-square test assesses the consistency between the theoretical and empirical distributions. Preliminary identification of multi-component fitting was performed by analyzing the multi-peak characteristics of the density distribution. When the histogram shows multiple peaks, it indicates the presence of multiple Weibull components, each peak corresponding to a potential component. The position of the peaks provides information on the scale parameter, and the sharpness of the peaks provides information on the shape parameter. Finally, initial Weibull model parameters capable of describing the basic characteristics of vascular density distribution were obtained.

[0064] The parameter optimization process of the maximum likelihood estimation algorithm involves inputting the initial Weibull model parameters into the algorithm for parameter optimization. Maximum likelihood estimation is an important parameter estimation method; its basic idea is to find the parameter values ​​that maximize the probability of the observed data. The likelihood function is constructed based on the Weibull probability density function and the assumption of independent and identically distributed (ICD). The joint probability density function, i.e., the likelihood function, is obtained by multiplying the probability density functions of all observed data points. The log-likelihood function is obtained by taking the logarithm of the likelihood function. The logarithmic transformation converts the product operation into an addition operation, simplifying numerical computation and improving numerical stability. The maximum likelihood estimation solution is established by taking the partial derivative of the log-likelihood function with respect to the parameters and setting it to zero. The likelihood equations for the Weibull distribution often cannot obtain analytical solutions and require numerical optimization methods. The Newton-Raphson method is a commonly used numerical optimization algorithm that iteratively solves the problem by calculating the first and second derivatives of the objective function, moving one step towards the optimal solution in each iteration until convergence to a local optimum. Gradient descent is another optimization algorithm that determines the search direction by calculating the gradient of the objective function. The step size is chosen using a line search or adaptive strategy to ensure convergence. Constrained optimization considers the physical meaning and range of the Weibull parameters; both shape and scale parameters must be positive, and non-negativity constraints are applied during optimization to ensure parameter rationality. A multi-starting-point optimization strategy involves running the algorithm multiple times with different initial values, comparing the results from different starting points to select the global optimum and avoid getting trapped in local optima. Convergence is judged based on indices such as parameter changes, objective function changes, or gradient norms; convergence is considered achieved when these indices are below preset thresholds. Numerical stability measures include avoiding computational overflow, handling singular matrices, and controlling rounding errors. Finally, Weibull parameter values ​​optimized through maximum likelihood estimation are obtained. These parameters exhibit good statistical properties such as consistency, asymptotic normality, and asymptotic efficiency. The parameter analysis process involves detailed parameter analysis and interpretation of the physical meaning based on the optimized Weibull parameter values. The parameter analysis first extracts two fundamental parameters of the Weibull distribution: the shape parameter, which controls the shape characteristics of the distribution; when the shape parameter is less than 1, the distribution exhibits a decreasing exponential shape; when it is equal to 1, it degenerates into an exponential distribution; and when it is greater than 1, it exhibits a bell-shaped but right-skewed distribution. The larger the parameter value, the closer the distribution is to a normal distribution. The scale parameter determines the location and broadening of the distribution; as the scale parameter increases, the distribution shifts to the right and broadens; as the scale parameter decreases, the distribution shifts to the left and contracts. This parameter is directly related to the characteristic lifetime or eigenvalue of the distribution. The confidence interval estimation of the parameters is based on the asymptotic normality theory. The standard error of the parameter estimation is calculated using the Fisher information matrix, and the confidence interval is constructed to assess the estimation accuracy. The correlation analysis of the parameters assesses the independence of the two parameter estimates by calculating the correlation coefficient between the estimated shape and scale parameters. High correlation indicates the existence of collinearity in the parameter estimates, requiring the use of regularization techniques.The calculation of distribution characteristics includes statistics such as mean, variance, skewness, and kurtosis. These characteristics are all functions of Weibull parameters, and various statistical properties of the distribution can be directly calculated from the parameter values. The quantile function is calculated using the inverse cumulative distribution function of the Weibull distribution. Quantiles provide location information and tail characteristics of the distribution, which are of significant diagnostic value in medical image analysis. Parameter sensitivity analysis assesses the impact of parameter estimation errors on subsequent analysis results by calculating the degree of influence of parameter changes on distribution characteristics; highly sensitive parameters require more precise estimation. The model's extrapolation capability is evaluated by testing the model's predictive performance outside the parameter estimation range, ensuring the model's reliability in practical applications. Finally, the shape and scale parameters of the Weibull distribution, containing both statistical and physical interpretations, are obtained.

[0065] The component weight calculation process performs weight allocation calculations for shape and scale parameters using a mixture model. Component weight calculation is based on mixture model theory, assuming that the observed blood vessel density distribution is a weighted combination of multiple Weibull distributions, with each component corresponding to different tissue components within the blood vessel. The weight calculation first determines the number of components by comparing model performance with different numbers of components using information criteria such as the Akaike information criterion or Bayesian information criterion, selecting the optimal number of components that balances model complexity and fitting accuracy. Initial weight allocation is based on clustering analysis results. The k-means clustering algorithm divides the density data into several clusters, each corresponding to a Weibull component. The size of the cluster determines the initial weight of the corresponding component. Weight constraints include non-negativity and normalization constraints. All component weights must be non-negative, and the sum of the weights must equal one. These constraints ensure the probabilistic interpretation of the weights. Weight optimization uses the Lagrange multiplier method or projected gradient method to handle the constrained optimization problem, maximizing the model's fitting effect while satisfying the constraints. Weight stability analysis observes the sensitivity to weight changes by perturbation parameters. Stable weight allocation indicates good model robustness. The physical meaning of the weights is interpreted based on tissue composition analysis of medical images. Larger weights correspond to major tissue components such as contrast agents, while smaller weights correspond to minor components such as thrombi or vessel walls. Spatiotemporal variation analysis of the weights considers the differences in weight allocation between different vascular levels and patients, establishing a correlation between weights and vascular features. Cross-validation evaluates the generalization performance of the weight calculation, testing the stability of the weight allocation on independent datasets using leave-one-out or k-fold cross-validation. Finally, optimized and validated shape parameters, scale parameters, and initial values ​​for the mixed weights are obtained, which constitute the Weibull mixture model.

[0066] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0067] The embolism density feature parameters are input into the variational posterior distribution and processed by Bayesian classification model construction to obtain the embolism probability prior distribution.

[0068] Based on the prior distribution of embolism probability, the likelihood function is calculated on the density values ​​of blood vessel pixels to obtain pixel-level embolism likelihood values.

[0069] Based on the pixel-level embolism likelihood values, the parameters are updated using a variational lower bound optimization algorithm to obtain a convergent variational parameter set.

[0070] The convergent variational parameter set is input into the posterior probability calculation module for embolism probability inference processing to obtain the pixel embolism probability value.

[0071] Spatial mapping and uncertainty quantification are performed on the pixel embolism probability values ​​to obtain the lesion probability distribution map.

[0072] Specifically, the variational posterior distribution Bayesian classification model construction process inputs embolism density feature parameters into a variational inference framework to establish a probabilistic classification model. The variational posterior distribution construction is based on variational Bayesian theory. This method introduces a variational distribution to approximate the complex true posterior distribution, transforming the cumbersome integral calculation into an optimization problem. The model construction process first defines the latent variable of embolism state. The embolism state of each vascular pixel is represented as a binary random variable, taking the values ​​of embolism or non-embolism. The prior distribution is set based on medical statistics and clinical experience. The prior probability of embolism is determined according to epidemiological data of different vascular levels and patient groups; the prior probability of embolism in the main pulmonary artery is usually higher than that in subsegmental branches. Embolism density feature parameters are input into the model as observed variables. These parameters include the shape parameters, scale parameters, and mixture weights of the Weibull mixture model, reflecting the distribution characteristics of different tissue components within the vessel. The variational distribution is chosen using the mean-field approximation assumption, assuming that the embolism states of different pixels are independent, and that the variational distribution of each pixel is an independent Bernoulli distribution. The variational parameters are initialized through random sampling or heuristic methods based on prior information. The choice of initial parameters affects the convergence speed and final result of the optimization algorithm. The hierarchical structure of the model considers vascular topology information; the embolism state of higher-level vessels influences lower-level vessels, and the dependencies between levels are established through conditional probabilities. The setting of hyperparameters includes parameters of the prior distribution and parameters of the variational distribution. These hyperparameters control the complexity and fitting ability of the model and need to be determined through cross-validation or empirical Bayesian methods. Finally, an embolism probability prior distribution containing prior information on embolism state and variational distribution parameters is obtained. This distribution provides a probabilistic framework for subsequent likelihood calculations and probabilistic inference.

[0073] The likelihood function calculation process performs probabilistic modeling and likelihood calculation on the density values ​​of vessel pixels based on the prior distribution of embolism probability. The likelihood function describes the probability of observing a specific density value under a given embolism state and is a core component of Bayesian inference. The calculation process first extracts the CT density value of each vessel pixel, which reflects the tissue characteristics and contrast agent distribution in the corresponding region. The likelihood function under embolism is calculated based on the Weibull distribution of thrombus components. The density value of the thrombus region is typically lower than that of normal contrast agent but higher than that of background tissue, and its distribution characteristics are described by the corresponding Weibull parameters. The likelihood function under non-embolism state is calculated based on the Weibull distribution of contrast agent components. Normal vessels are filled with contrast agent, resulting in higher density values ​​and a relatively uniform distribution. The handling of mixed components considers partial volume effects and incomplete embolism. Pixels may simultaneously contain both contrast agent and thrombus components; the likelihood function uses a weighted average to combine the contributions of different components. Likelihood values ​​are calculated using point estimation of the probability density function. The pixel density value is substituted into the Weibull probability density function for the corresponding state to obtain the probability density of that density value. Numerical stability is addressed by avoiding zero or extremely low probability densities, and by adding a smoothing term or using logarithmic probability to prevent underflow. Likelihood function normalization ensures comparability of likelihoods between different pixels by summing the likelihoods across all possible states. Spatial correlation is considered by introducing neighborhood information to enhance the robustness of the likelihood function. The density values ​​of adjacent pixels and the embolism state have a certain correlation, which can be modeled using methods such as Markov random fields. Uncertainty is quantified by calculating the confidence interval of the likelihood or estimating the posterior distribution of the parameters using Bayesian methods to assess the reliability of the likelihood estimation. Finally, pixel-level embolism likelihood values ​​for each blood vessel pixel in both embolism and non-embolism states are obtained. These values ​​quantify the degree of agreement between the observed density values ​​and different embolism states.

[0074] The variational lower bound optimization algorithm iteratively updates model parameters based on pixel-level embolism likelihood values ​​using a variational inference optimization algorithm. The variational lower bound is a core concept in variational Bayesian inference, representing a lower bound on the true marginal log-likelihood. Maximizing the variational lower bound approximates the optimal posterior distribution. The lower bound function is constructed based on Jensen's inequality, transforming the integral form of the marginal likelihood into an expectation form expression for the lower bound. The lower bound consists of a likelihood term and a regularization term. The likelihood term encourages the model to fit the observed data, while the regularization term prevents overfitting and maintains consistency with the prior distribution. The optimization algorithm uses the coordinate ascent method for parameter updates. This method alternately optimizes different sets of variational parameters, fixing other parameters each time while optimizing the current parameter set. The iteration process continues until convergence. The update formula for the variational parameters is obtained by taking the partial derivative of the variational lower bound and setting it to zero. For the Bernoulli variational distribution, the update formula has an analytical form, allowing direct calculation of the new parameter values. Likelihood values ​​are fused by integrating likelihood information from different pixels through weighted averaging or multiplication, with weights allocated considering the reliability and importance of each pixel. Gradient calculation employs automatic or numerical differentiation techniques to ensure accuracy and efficiency; gradient calculation for complex models may involve the chain rule and backpropagation. An adaptive strategy, such as the Adam optimizer or AdaGrad algorithm, is used to automatically adjust the step size based on historical gradient information. Convergence is determined based on metrics such as the change in variational lower bound, parameter change, or gradient norm; the algorithm is considered convergent when these metrics remain below a preset threshold for several consecutive iterations. An early stopping strategy prevents overfitting by monitoring performance on the validation set; training is terminated early when validation performance no longer improves. Multi-starting point initialization uses different random seeds to run the algorithm multiple times, selecting the result with the largest lower bound as the final solution to increase the probability of finding the global optimum. Finally, a convergent set of variational parameters optimized by the variational lower bound is obtained; these parameters represent the optimal variational posterior distribution.

[0075] The posterior probability calculation module's embolism probability inference process inputs the converged variational parameter set into the Bayesian inference framework for accurate calculation of the embolism probability. Posterior probability inference, based on Bayes' theorem, combines prior probability, likelihood function, and variational parameters to calculate the posterior probability of embolism for each pixel. The inference process first reads the converged variational parameters, which contain the variational distribution parameters of the embolism state for each pixel, reflecting the updated embolism probability information based on observation data. The application of Bayes' formula multiplies the prior probability by the likelihood function to obtain the non-normalized posterior probability, which is then normalized to obtain the true posterior probability distribution. The variational approximation correction considers the difference between the variational distribution and the true posterior distribution, improving the approximation accuracy through importance sampling or higher-order variational methods. The embolism probability value is calculated using the mean of the variational distribution as a point estimate, representing the expected probability that a pixel belongs to the embolism state. The uncertainty is estimated using the variance or entropy of the variational distribution; high variance or high entropy indicates greater uncertainty in the embolism state, requiring further examination or verification. Probabilistic calibration adjusts probability values ​​using methods such as Platt scaling or ordinal-preserving regression to better reflect the true likelihood of embolism. Threshold selection determines the optimal classification threshold through ROC curve analysis or precision-recall curves, balancing sensitivity and specificity requirements. Multi-scale inference considers embolism probabilities at different resolutions, capturing features of emboli of different sizes using image pyramids or multi-scale convolutional networks. Temporal information fusion improves the accuracy of embolism detection for dynamically enhanced CT images by combining information from multiple time points through time-series analysis. Ensemble learning enhances the stability and accuracy of embolism probability estimation by combining the predictions of multiple models; commonly used ensemble methods include voting, averaging, and stacking. Finally, the embolism probability value for each blood vessel pixel is obtained, quantitatively describing the confidence that the pixel belongs to an embolic region.

[0076] Spatial mapping and uncertainty quantification processes perform spatial reconstruction and uncertainty analysis on pixel embolism probability values ​​to generate a complete probability distribution map. The spatial mapping process reorganizes the one-dimensional pixel probability sequence into a three-dimensional spatial structure corresponding to the original CTPA image, restoring the spatial adjacency relationships and anatomical location information between pixels. Coordinate transformation processing considers geometric transformations during image acquisition, including rotation, translation, scaling, and projection transformations, and correctly maps the probability values ​​to the anatomical coordinate system through inverse transformation. Interpolation algorithms handle the resolution difference between the original sampling grid and the target display grid, using methods such as bilinear interpolation, cubic spline interpolation, or sine interpolation to achieve a smooth transition of probability values. Spatial filtering smooths the probability distribution using methods such as Gaussian filtering or bilateral filtering, removing isolated high-probability points and noise in the probability distribution, maintaining the continuity of the embolism region. Uncertainty quantification employs multiple methods to evaluate the reliability of probability estimates, including Bayesian confidence intervals, bootstrap confidence intervals, and model-based uncertainty estimation. Understanding uncertainty reflects the uncertainty of model parameter estimation; the variability of probability predictions is calculated through the posterior distribution of the parameters. Random uncertainty reflects the randomness and noise inherent in the data, and is quantified through the variance of the observation model. Uncertainty is visualized through color coding, contour plots, or 3D surface maps, intuitively displaying the spatial distribution of uncertainty. Probability thresholds are set according to clinical needs and diagnostic requirements, generating corresponding binary segmentation results. Multi-level display considers vascular hierarchical information, showing the embolism probability distribution of the main pulmonary artery, segmental branches, and subsegmental branches separately. An interactive query function allows users to click on any location in the image to view the embolism probability value and uncertainty information for that point. Quality control indicators include evaluation criteria such as the smoothness, continuity, and anatomical rationality of the probability distribution. Finally, a complete lesion probability distribution map containing embolism probability values, uncertainty information, and spatial location information is obtained. This distribution map provides clinicians with intuitive, quantitative, and reliable embolism detection results, supporting diagnostic decisions and treatment planning.

[0077] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0078] Multi-scale convolutional kernel feature extraction is performed based on the lesion probability distribution map to obtain large-scale morphological features, medium-scale texture features and small-scale boundary features.

[0079] Large-scale morphological features, medium-scale texture features, and small-scale boundary features are input into an attention mechanism for weight allocation, resulting in a scale-adaptive feature weight vector.

[0080] The multi-scale features are weighted and fused according to the scale-adaptive feature weight vector to obtain the fused feature response map.

[0081] Based on the fused feature response map, spatial consistency constraint processing is performed using a conditional random field to obtain spatial continuity segmentation results;

[0082] The spatial continuity segmentation results are validated by the distance constraint of the blood vessel centerline to obtain the lesion segmentation boundary.

[0083] Specifically, the multi-scale convolutional kernel feature extraction process analyzes and extracts features at different spatial scales based on the lesion probability distribution map. The feature extraction process uses multiple convolutional kernels of different sizes to simultaneously process the lesion probability distribution map. Each scale of convolutional kernel can capture image features and patterns within a specific spatial range. Large-scale convolutional kernel feature extraction uses a larger convolutional window to perform convolution operations on the probability distribution map. Large-scale convolutional kernels can capture the overall morphological information and long-distance spatial relationships of embolic lesions, including the overall shape, size, and spatial distribution pattern of the embolic region. Large-scale morphological features mainly reflect the macroscopic geometric characteristics of embolic lesions, such as shape descriptors like ellipticity, compactness, elongation, and directionality. These features are particularly effective for identifying large-area main vessel embolism. Medium-scale convolutional kernels use a medium-sized convolutional window to extract local texture information from the probability distribution. Medium-scale features can capture density variation patterns and texture structures within the embolic region. Mesoscale texture features include gray-level co-occurrence matrix features, local binary pattern features, and Gabor filter responses. These features describe the internal heterogeneity and histological characteristics of embolic tissue, helping to distinguish different types and stages of thrombi. Small-scale convolutional kernels use smaller convolutional windows to accurately extract detailed information about embolic boundaries. Small-scale features mainly focus on gradient changes in probability distributions and edge responses. Small-scale boundary features include geometric descriptors such as edge intensity, edge direction, corner detection, and contour curvature. These features are crucial for accurately locating embolic boundaries and distinguishing the interface between emboli and normal tissue. Convolution operations employ different kernel functions, including Gaussian kernels, Laplacian kernels, Sobel kernels, and custom kernels. Different kernel functions have different sensitivities to specific types of features. Feature mapping is generated through convolution operations and nonlinear activation functions. Activation functions such as ReLU or sigmoid enhance the nonlinear expressive power of the features. Feature normalization eliminates numerical differences between features at different scales. Batch normalization or layer normalization techniques ensure the numerical stability of the features. Ultimately, we obtained large-scale morphological features, medium-scale texture features, and small-scale boundary features that describe the characteristics of embolic lesions at different spatial scales.

[0084] The attention mechanism's weight allocation process inputs multi-scale features into the attention calculation module for adaptive weight allocation. The attention mechanism is a computational framework capable of automatically learning feature importance, assigning corresponding weights by calculating the correlation and importance between different features. The weight allocation process first calculates the similarity matrix between features, measuring the degree of association between features at different scales through dot product, cosine similarity, or a learned similarity function. Correlation analysis between large-scale morphological features and mesoscale texture features reveals the relationship between the overall morphology and internal structure of the embolism; high correlation indicates that morphological and texture features describe different aspects of the same embolic region. The correlation between mesoscale texture features and small-scale boundary features reflects the connection between the heterogeneity within the embolism and boundary clarity; clear boundaries typically correspond to uniform internal texture. The calculation of attention weights uses a softmax function to ensure the probabilistic nature of the weights; the sum of the weights for all scale features equals one, and the weight value reflects the contribution of the corresponding feature to the final segmentation result. The self-attention mechanism enhances the expression of important features by calculating the correlation between features and themselves, suppressing the influence of unimportant or noisy features. The cross-attention mechanism calculates the interaction between features at different scales, capturing the synergistic effects and complementary information of multi-scale features. The introduction of positional encoding considers the spatial location information of features; the same feature may have different importance at different locations. Positional encoding is implemented through a sine function or a learned embedding vector. The multi-head attention mechanism employs multiple independent attention computation heads, each focusing on different aspects of the feature. Parallel computation and result fusion enhance the expressive power of the attention mechanism. Weight regularization uses L1 or L2 regularization to prevent excessive weight concentration on a few features, maintaining the diversity and robustness of feature selection. Dynamic weight adjustment adaptively adjusts the weight allocation strategy based on the characteristics of the input data; different embolism types and sizes may require different combinations of feature weights. Finally, a scale-adaptive feature weight vector reflecting the importance of features at different scales is obtained, which guides the subsequent feature fusion process.

[0085] The weighted fusion process organically integrates and unifies multi-scale features based on scale-adaptive feature weight vectors. The weighted fusion process uses a linear combination method to weight and average features at different scales according to their weight vectors; the fusion formula is the dot product of the weights and features. Feature alignment ensures the spatial and semantic correspondence of features at different scales, unifying feature mapping to the same spatial resolution through upsampling or downsampling techniques. When large-scale morphological features have a larger weight, the fusion result reflects the overall shape and spatial distribution of the embolism, suitable for detecting large-area main vessel embolisms. When medium-scale texture features have a larger weight, the fusion result focuses more on the internal structure and density distribution of the embolic region, helping to distinguish mixed regions of different tissue components. When small-scale boundary features have a larger weight, the fusion result emphasizes the precise location and detailed description of the embolic boundary, improving the accuracy of boundary segmentation. Nonlinear extensions of feature fusion enhance the expressive power of the fusion by introducing nonlinear transformation functions; commonly used nonlinear functions include multilayer perceptrons, convolutional neural networks, and attention pooling. Residual connection techniques preserve original feature information through skip connections, preventing the loss of important information during feature fusion. Batch normalization stabilizes numerical computation during the fusion process, accelerates convergence, and improves the model's generalization ability. Dropout technology prevents overfitting by randomly discarding some features, improving the robustness of the fusion model. Visual analysis of feature fusion uses heatmaps or feature maps to show the contribution distribution of features at different scales, helping to understand the working principle of the fusion mechanism. Quality evaluation metrics include feature discrimination, information preservation, and computational complexity. Finally, a fusion feature response map integrating multi-scale information is obtained, which contains a comprehensive feature description of the embolic lesion, providing a rich information foundation for subsequent segmentation processing.

[0086] The spatial consistency constraint processing of Conditional Random Fields (CRF) is based on the fused feature response map, using a probabilistic graphical model to model spatial relationships and optimize consistency. CRF is an undirected probabilistic graphical model capable of establishing spatial dependencies and label consistency constraints between pixels, effectively solving the problems of spatial discontinuity and label inconsistency in segmentation results. The spatial consistency constraint is established based on the similarity assumption between neighboring pixels; adjacent pixels with similar density values ​​tend to have the same label, while different label assignments are allowed when density differences are large. A univariate potential function calculates the local label probability of each pixel based on the fused feature response map. The univariate potential function encourages pixel labels to be consistent with the feature response; pixels with high response values ​​tend to be labeled as embolic. A binary potential function establishes label consistency constraints between adjacent pixels, considering distance, density differences, and gradient information between pixels. The binary potential function in the Potts model penalizes adjacent pixels with inconsistent labels, with the penalty strength inversely proportional to pixel similarity. Higher-order potential functions consider the joint label configuration of multiple pixels, capturing more complex spatial patterns and structural constraints. The graph cutting algorithm solves for the optimal label placement in a conditional random field using the minimum-cut maximum-flow theorem, transforming the label assignment problem into a minimum-cut problem in the graph. The belief propagation algorithm iteratively updates the marginal probabilities of pixels through a message-passing mechanism until it converges to a stable probability distribution. The mean-field approximation simplifies the inference process through the independence assumption and is suitable for real-time processing of large-scale images. Energy function optimization employs numerical optimization algorithms such as gradient descent or quasi-Newton methods to find the label placement that minimizes the total energy. Parameter learning learns the parameters of the potential function from the training data using methods such as structured perceptrons or conditional likelihood maximization. Regularization techniques prevent overfitting and improve the model's generalization ability; commonly used regularization methods include weight decay and early stopping strategies. Finally, a spatially continuous segmentation result that satisfies spatial consistency constraints is obtained, eliminating isolated misclassified points and discontinuous segmentation regions.

[0087] The distance constraint verification process for the vessel centerline performs anatomical and geometric constraints verification on the spatial continuity of the segmentation results. Distance constraint verification is based on the tubular geometry of the vessel; the embolized region must be located inside the vessel and its distance from the vessel centerline cannot exceed the vessel radius. The constraint verification process first calculates the Euclidean distance from each segmented pixel to the nearest vessel centerline, using a nearest neighbor search algorithm or distance transformation algorithm for efficient computation. The vessel radius is estimated based on the local diameter information calculated during vessel centerline extraction, and the vessel radius value at any location is obtained through interpolation. Distance constraint judgment compares the distance from a pixel to the centerline with the local vessel radius; pixels with a distance less than the radius are located inside the vessel, while pixels with a distance greater than the radius are located outside the vessel and need to be discarded. Constraint violations are handled by using morphological operations or region growing algorithms to correct the segmentation boundaries, ensuring that all embolized regions are located inside the vessel boundaries. Boundary smoothing is performed using Gaussian filtering or bilateral filtering to eliminate jagged boundaries caused by distance constraints, maintaining the naturalness of the segmentation contour. Multi-level validation considers the geometric characteristics of different vascular levels. The main pulmonary artery has a larger radius, allowing for greater distance tolerance, while subsegmental arteries have smaller radii, requiring stricter distance constraints. Tolerance parameters are set considering factors such as image resolution, reconstruction error, and pathological deformation. An appropriate tolerance range balances the strictness of constraints with the integrity of segmentation. Visualization of constraint validation uses color coding to display regions that meet and violate constraints, helping to verify the rationality of the constraint settings. Statistical analysis calculates the proportion and distribution of constraint violations, assessing the anatomical rationality of the segmentation results. Iterative optimization gradually improves segmentation quality through multiple constraint validations and boundary adjustments until the preset quality standards are met. Finally, a lesion segmentation boundary that satisfies both probability distribution characteristics and vascular geometric constraints is obtained, accurately describing the spatial extent and morphological characteristics of the embolic lesion.

[0088] In one specific embodiment, the process of executing step S105 may specifically include the following steps:

[0089] Based on the lesion segmentation boundary, a vascular branch connectivity graph is constructed and topological relationship verification is performed to obtain the embolism propagation path constraint conditions.

[0090] The constraints of the embolism propagation path are input into a graph theory algorithm for abnormal segmentation region identification, resulting in isolated embolism markers that violate anatomical rules;

[0091] Based on isolated embolism markers, anomaly region removal is performed using a connectivity analysis algorithm to obtain connectivity-corrected segmentation results;

[0092] Based on the connectivity correction segmentation results, hemodynamic constraint verification was performed to obtain the consistency check results of upstream and downstream vascular embolism.

[0093] The consistency of embolism between upstream and downstream vessels was analyzed by iterative boundary optimization to obtain the final segmentation result of the embolism lesion.

[0094] Specifically, the process of verifying the topological relationships of the vascular branch connectivity graph is based on constructing a graph structure representation of the vascular system according to the lesion segmentation boundary and performing topological rationality checks. The connectivity graph construction process first represents vascular branches as nodes in the graph, with each node containing attributes such as the spatial location, hierarchical identifier, geometric parameters, and embolism status of the vascular branch. The connections between vascular branches are represented as edges in the graph, with attributes including connection type, blood flow direction, bifurcation angle, and distance information. The graph construction is based on the hierarchical structure of the vascular tree, starting from the root node of the main pulmonary artery and recursively establishing parent-child and sibling relationships according to the anatomical branching pattern. The analysis of the embolism propagation path is based on the pathophysiological mechanism of thromboembolism; thrombi typically propagate gradually from upstream vessels to downstream vessels, following the laws of blood flow direction and gravity. The establishment of propagation path constraints considers the principle of embolism continuity; embolic areas must be distributed along connected vascular branches, and isolated emboli cannot appear in leaps. The topology verification algorithm traverses the vascular connectivity graph through depth-first search or breadth-first search to check the connectivity of embolism distribution and the rationality of the propagation path. Path length constraints are based on the physical dimensions of the embolism; the embolism cannot exceed the length limit of the vessel branch, and excessively long emboli may indicate segmentation errors. Bifurcation point constraints verify the distribution pattern of emboli at vessel bifurcation points; emboli typically accumulate near bifurcation points but should not simultaneously block all downstream branches. Blood flow direction constraints ensure that embolism propagation conforms to the physical laws of blood flow; embolism distributions propagating against flow are marked as abnormal. Hierarchical constraints verify the transmission relationship of emboli between different vessel levels; the embolic status of a higher-level vessel affects the embolic probability of a lower-level vessel. Spatial distance constraints check the spatial continuity between adjacent embolic regions; excessively large gaps indicate possible segmentation omissions or misidentifications. Finally, embolism propagation path constraints are obtained, describing the reasonable propagation path and distribution pattern of emboli in the vascular tree.

[0095] The graph theory algorithm for anomaly segmentation and identification processes inputs the constraints of embolism propagation paths into graph analysis algorithms for automatic identification and labeling of anomaly regions. The anomaly identification process employs multiple graph theory algorithms to analyze the topological characteristics and connectivity properties of the embolism distribution. Connected component analysis decomposes the vascular connectivity graph into several connected subgraphs, each representing an independent embolism propagation region. Isolated node detection algorithms identify independent embolism nodes not connected to any embolism region; these isolated emboli are often errors in the segmentation algorithm. Shortest path algorithms calculate the shortest propagation path from the embolism region to the vessel root; embolism regions with abnormally long paths are marked as suspicious. Cut set analysis identifies key vascular branches that interrupt the embolism propagation path; the existence of cut sets indicates discontinuities in embolism propagation. Centrality analysis calculates the importance of vascular nodes in the embolism propagation network; nodes with abnormal centrality may have been incorrectly labeled. Subgraph isomorphism algorithms compare the actual embolism distribution with typical embolism patterns; regions significantly deviating from typical patterns are marked as anomalies. Graph coloring algorithms verify the consistency of embolism states in adjacent vascular branches; regions violating coloring constraints indicate incorrect labeling. The path coverage algorithm checks whether the embolic region can be reached via a continuous vascular path; isolated regions that cannot be reached are marked as anomalous. The graph matching algorithm verifies the rationality of the pairing between embolic distribution and vascular branches; mismatched regions indicate possible segmentation errors. The topological sorting algorithm sorts the embolic regions according to the blood flow direction; regions with abnormal sorting violate the temporal logic of embolism propagation. The anomaly scoring mechanism integrates multiple graph theory indicators to calculate the degree of anomalousness for each region; regions exceeding a threshold are marked as isolated emboli that violate anatomical rules. Finally, isolated embolism labels identifying the location and type of all anomalous segmentation regions are obtained.

[0096] The connectivity analysis algorithm's abnormal region removal process corrects and optimizes the segmentation results based on isolated embolism markers through connectivity checks and region merging algorithms. The abnormal region removal process first analyzes the type and severity of isolated embolisms, employing corresponding processing strategies based on different anomaly types. Small-area isolated embolisms are removed using a direct deletion strategy; isolated regions smaller than a preset threshold are considered noise or artifacts and directly removed from the segmentation results. Large-area isolated embolisms require further analysis of their causes, which may include missed connection paths in the segmentation algorithm or the presence of genuine multiple embolisms. The path search algorithm attempts to find possible connection paths between isolated embolisms and major embolic regions, establishing connectivity by relaxing constraints or introducing intermediate nodes. The region merging algorithm merges spatially adjacent and feature-similar embolic regions, eliminating artificial separation caused by over-refinement of the segmentation. Connectivity repair uses morphological operations such as dilation, erosion, and connection to repair minor breaks between embolic regions. A distance threshold is set to determine the maximum allowable distance for merging embolic regions; regions that are too close together are considered to belong to the same embolism. Similarity metrics assess the rationality of region merging based on the similarity of density values, texture features, and shape features. Boundary smoothing eliminates irregular boundaries caused by region deletion and merging, preserving the naturalness of the segmentation contour. Connectivity verification confirms that outlier regions have been correctly handled by recalculating the connected components of the corrected segmentation result. Quality assessment calculates connectivity, integrity, and consistency indices for the segmentation results before and after correction to verify the correction effect. An iterative processing mechanism allows for multiple rounds of anomaly detection and correction until the segmentation result meets connectivity requirements. Finally, a connectivity-corrected segmentation result is obtained that eliminates isolated outlier regions and satisfies topological connectivity.

[0097] The hemodynamic constraint verification process verifies and checks the consistency of blood flow physical laws based on the connectivity correction segmentation results. Hemodynamic constraint verification is based on the fundamental principles of blood circulation; the embolism state of upstream vessels directly affects the blood flow and contrast agent distribution in downstream vessels. The constraint verification process first establishes a blood flow model of the vascular tree, calculating blood flow resistance and pressure distribution based on vessel diameter, length, and bifurcation structure. Complete embolism constraint verifies the state of downstream vessels when the upstream vessel is completely blocked; no contrast agent should pass through completely embolized branches, and all downstream branches should lack contrast agent filling. Partial embolism constraint considers the impact of partial occlusion of the upstream vessel on downstream blood flow; partial embolism leads to reduced but not complete interruption of downstream blood flow. Pressure gradient constraint is based on fluid dynamics principles; blood flows from high-pressure areas to low-pressure areas, and embolism alters the pressure distribution within the vessel. Flow conservation constraint verifies the flow balance relationship at vessel bifurcation points; the blood flow entering the bifurcation point equals the total outflow. Time delay constraint considers the time effect of blood flow propagation; the impact of upstream embolism requires a certain amount of time to propagate downstream. Collateral circulation constraints consider the existence of alternative blood flow paths; collateral circulation may be compensatorily enhanced when the main vessel is embolized. Contrast agent concentration constraints verify the rationality of contrast agent concentrations in different vascular branches, showing a gradient from high concentration upstream of the embolism to low concentration inside the embolism and a decreasing concentration downstream. Vessel wall stress constraints consider the mechanical effects of embolism on the vessel wall; embolism leads to local vascular dilation and increased wall stress. A consistency check algorithm compares the logical consistency of embolism states in related vascular branches, identifying anomalous combinations that violate hemodynamic principles. Quantitative assessment uses blood flow simulation software to calculate changes in blood flow parameters, verifying the degree of conformity between embolism distribution and hemodynamic theory. Finally, the upstream and downstream vessel embolism consistency check results, reflecting the intrinsic consistency of the embolism state in the vascular system, are obtained.

[0098] The iterative boundary optimization process repeatedly optimizes the consistency check results of upstream and downstream embolism to obtain the optimal segmentation boundary. The boundary optimization process uses an iterative algorithm to gradually adjust the boundary position of the embolic region, ensuring that the segmentation result simultaneously satisfies image features, anatomical constraints, and hemodynamic constraints. The optimization objective function comprehensively considers multiple evaluation indicators, including image fit, boundary smoothness, connectivity consistency, and hemodynamic rationality. The gradient descent algorithm determines the direction and step size of boundary adjustment by calculating the gradient of the objective function, iteratively moving boundary points until convergence to a local optimum. The active contour model drives the boundary evolution towards the true embolic boundary through the principle of energy minimization; internal energy maintains the smoothness of the boundary, while external energy attracts the boundary to move towards the image edge. The level set method handles topological changes through implicit curve evolution, automatically handling boundary splitting and merging. The graph cut algorithm transforms boundary optimization into a minimum cut problem in the graph, solving for the optimal boundary using the maximum flow minimum cut theorem. Deformable template matching matches typical embolic shapes using deformable models, constraining shape changes during the boundary optimization process. Multi-resolution optimization employs a coarse-to-fine strategy, first performing global optimization at low resolution, then refining local details at high resolution. Constraint optimization addresses various geometric and physical constraints, integrating them into the optimization objective using the Lagrange multiplier method or penalty function method. Convergence is determined based on metrics such as boundary changes, objective function changes, or gradient norm; the algorithm is considered convergent when these metrics remain below a threshold for several consecutive iterations. Multi-startpoint optimization involves multiple optimizations using different initial boundaries, selecting the result with the optimal objective function value as the final boundary. Post-processing steps include boundary smoothing, isolated point removal, and small region merging to further improve segmentation quality. Quality assessment evaluates segmentation accuracy using metrics such as Dice coefficient, Hausdorff distance, sensitivity, and specificity. Finally, a segmentation result of the embolic lesion, optimized through multiple iterations and satisfying all constraints, is obtained, describing the spatial distribution, morphological characteristics, and severity of pulmonary embolism.

[0099] The above describes the image recognition-based pulmonary embolism lesion segmentation method in the embodiments of this application. The following describes the image recognition-based pulmonary embolism lesion segmentation system in the embodiments of this application. Please refer to [link to relevant documentation]. Figure 2 One embodiment of the pulmonary embolism lesion segmentation system based on image recognition in this application includes:

[0100] The extraction module is used to perform hierarchical extraction of vascular tree data from CTPA three-dimensional image data through vascular diameter gradient analysis to obtain multi-level vascular topology data.

[0101] The modeling module is used to model the intravascular density distribution based on the multi-level vascular topology data using a Weibull mixture model to obtain embolism density characteristic parameters.

[0102] The estimation module is used to perform embolism probability estimation on blood vessel pixels based on the embolism density feature parameters through variational Bayesian inference to obtain a lesion probability distribution map.

[0103] The fusion module is used to perform multi-scale feature fusion processing on the lesion probability distribution map to obtain the lesion segmentation boundary;

[0104] The optimization module is used to optimize the vascular connectivity constraints based on the lesion segmentation boundary to obtain the final embolized lesion segmentation result.

[0105] above Figure 2 The image recognition-based pulmonary embolism lesion segmentation system in this embodiment of the invention is described in detail from the perspective of modular functional entities. The image recognition-based pulmonary embolism lesion segmentation device in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0106] Reference Figure 3 This invention also provides an image recognition-based pulmonary embolism lesion segmentation device, which can be a server, and its internal structure can be as follows: Figure 3 As shown, the image recognition-based pulmonary embolism lesion segmentation device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computational and control capabilities. The memory of the image recognition-based pulmonary embolism lesion segmentation device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the image recognition-based pulmonary embolism lesion segmentation device stores the data corresponding to this embodiment. The network interface of the image recognition-based pulmonary embolism lesion segmentation device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0107] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the image recognition-based pulmonary embolism lesion segmentation device to which the present invention is applied.

[0108] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the image recognition-based pulmonary embolism lesion segmentation method.

[0109] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0110] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an image recognition-based pulmonary embolism lesion segmentation device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0111] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for segmenting pulmonary embolism lesions based on image recognition, characterized in that, The method includes: By performing vascular diameter gradient analysis, vascular tree hierarchical extraction was performed on CTPA three-dimensional image data to obtain multi-level vascular topology data. Based on the multi-level vascular topology data, the density distribution within the blood vessels is modeled using a Weibull mixture model to obtain embolism density characteristic parameters. This includes: extracting density values ​​from the cross-sectional areas of vascular branches based on the multi-level vascular topology data to obtain layered vascular density sampling data; inputting the layered vascular density sampling data into the Weibull mixture model for component fitting to obtain shape parameters, scale parameters, and initial values ​​for the mixture weights; iteratively optimizing the shape parameters, scale parameters, and initial values ​​for the mixture weights using an expectation-maximization algorithm to obtain a convergent Weibull parameter combination; calculating the probability density functions of the contrast agent, thrombus, and vessel wall components based on the convergent Weibull parameter combination to obtain a density distribution model; and establishing a vascular hierarchy-Weibull parameter mapping relationship based on the density distribution model to obtain embolism density characteristic parameters. Based on the embolism density feature parameters, variational Bayes inference is used to estimate the embolism probability of blood vessel pixels to obtain a lesion probability distribution map. The lesion probability distribution map is subjected to multi-scale feature fusion processing to obtain the lesion segmentation boundary; The vascular connectivity constraints are optimized based on the lesion segmentation boundary to obtain the final embolic lesion segmentation result.

2. The method for segmenting pulmonary embolism lesions based on image recognition according to claim 1, characterized in that, The process involves extracting vascular tree hierarchically from CTPA three-dimensional image data using vascular diameter gradient analysis to obtain multi-level vascular topology data, including: Anisotropic diffusion filtering was performed on the CTPA three-dimensional image data to obtain the noise-reducing contrast agent distribution data; Based on the noise-reducing contrast agent distribution data, Hessian matrix eigenvalue decomposition was performed to obtain the coordinates of the vascular centerline skeleton. Based on the coordinates of the vascular centerline skeleton, the local diameter change rate and curvature feature are calculated to obtain a set of vascular geometric parameters; The set of vascular geometric parameters is input into a hierarchical discriminant function for branch classification processing to obtain three-level classification labels for the main pulmonary artery, segmental branch arteries, and subsegmental branch arteries. The three-level classification labels are processed to construct topological relationships, resulting in multi-level vascular topology data.

3. The method for segmenting pulmonary embolism lesions based on image recognition according to claim 1, characterized in that, The step of inputting the layered vascular density sampling data into a Weibull mixture model for component fitting processing to obtain shape parameters, scale parameters, and initial values ​​of mixture weights includes: The density value statistical analysis of the layered blood vessel density sampling data is performed to obtain the blood vessel density distribution characteristics; Based on the aforementioned blood vessel density distribution characteristics, a Weibull probability density function fitting process is performed to obtain the initial Weibull model parameters; The initial Weibull model parameters are input into the maximum likelihood estimation algorithm for parameter optimization to obtain optimized Weibull parameter values; Based on the optimized Weibull parameter values, parameter analysis is performed to obtain the shape and scale parameters of the Weibull distribution. The shape parameters and scale parameters are processed by component weight calculation to obtain the initial values ​​of shape parameters, scale parameters and mixed weights.

4. The method for segmenting pulmonary embolism lesions based on image recognition according to claim 1, characterized in that, The process of estimating the embolism probability of blood vessel pixels based on the embolism density feature parameters using variational Bayesian inference to obtain a lesion probability distribution map includes: The embolism density feature parameters are input into the variational posterior distribution and processed by a Bayesian classification model to obtain the prior distribution of embolism probability. Based on the prior distribution of embolism probability, the likelihood function is calculated on the density value of blood vessel pixels to obtain the pixel-level embolism likelihood value. Based on the pixel-level embolism likelihood values, the parameters are updated using a variational lower bound optimization algorithm to obtain a set of convergent variational parameters. The convergent variational parameter set is input into the posterior probability calculation module for embolism probability inference processing to obtain the pixel embolism probability value. Spatial mapping and uncertainty quantification are performed on the embolism probability values ​​of the pixels to obtain the lesion probability distribution map.

5. The method for segmenting pulmonary embolism lesions based on image recognition according to claim 1, characterized in that, The step of performing multi-scale feature fusion processing on the lesion probability distribution map to obtain the lesion segmentation boundary includes: Based on the lesion probability distribution map, multi-scale convolution kernel feature extraction processing is performed to obtain large-scale morphological features, medium-scale texture features and small-scale boundary features. The large-scale morphological features, medium-scale texture features, and small-scale boundary features are input into an attention mechanism for weight allocation to obtain a scale-adaptive feature weight vector. The multi-scale features are weighted and fused according to the scale-adaptive feature weight vector to obtain a fused feature response map. Based on the fused feature response map, spatial consistency constraint processing is performed using a conditional random field to obtain a spatial continuity segmentation result; The spatial continuity segmentation results are verified by the distance constraint of the blood vessel centerline to obtain the lesion segmentation boundary.

6. The method for segmenting pulmonary embolism lesions based on image recognition according to claim 1, characterized in that, The optimization of vascular connectivity constraints based on the lesion segmentation boundary to obtain the final embolic lesion segmentation result includes: Based on the lesion segmentation boundary, a vascular branch connectivity graph is constructed and topological relationship verification is performed to obtain the embolism propagation path constraint conditions. The constraints of the embolism propagation path are input into a graph theory algorithm for abnormal segmentation region identification, resulting in isolated embolism markers that violate anatomical rules; Based on the isolated embolism markers, anomaly region removal processing is performed using a connectivity analysis algorithm to obtain connectivity-corrected segmentation results; Based on the connectivity correction segmentation results, hemodynamic constraint verification processing is performed to obtain the consistency check results of upstream and downstream vascular embolism. The consistency check results of upstream and downstream vascular embolism were subjected to iterative boundary optimization to obtain the final embolism lesion segmentation results.

7. A pulmonary embolism lesion segmentation system based on image recognition, characterized in that, For implementing the image recognition-based pulmonary embolism lesion segmentation method as described in any one of claims 1-6, the image recognition-based pulmonary embolism lesion segmentation system comprises: The extraction module is used to perform hierarchical extraction of vascular tree data from CTPA three-dimensional image data through vascular diameter gradient analysis to obtain multi-level vascular topology data. The modeling module is used to model the intravascular density distribution based on the multi-level vascular topology data using a Weibull mixture model to obtain embolism density feature parameters. This includes: extracting density values ​​from the cross-sectional areas of vascular branches based on the multi-level vascular topology data to obtain layered vascular density sampling data; inputting the layered vascular density sampling data into the Weibull mixture model for component fitting to obtain shape parameters, scale parameters, and initial values ​​for the mixture weights; iteratively optimizing the shape parameters, scale parameters, and initial values ​​for the mixture weights using an expectation-maximization algorithm to obtain a converged Weibull parameter combination; calculating the probability density functions of the contrast agent, thrombus, and vessel wall components based on the converged Weibull parameter combination to obtain a density distribution model; and establishing a vascular hierarchy-Weibull parameter mapping relationship based on the density distribution model to obtain embolism density feature parameters. The estimation module is used to perform embolism probability estimation on blood vessel pixels based on the embolism density feature parameters through variational Bayesian inference to obtain a lesion probability distribution map. The fusion module is used to perform multi-scale feature fusion processing on the lesion probability distribution map to obtain the lesion segmentation boundary; The optimization module is used to optimize the vascular connectivity constraints based on the lesion segmentation boundary to obtain the final embolized lesion segmentation result.

8. A pulmonary embolism lesion segmentation device based on image recognition, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the image recognition-based pulmonary embolism lesion segmentation method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to perform the image recognition-based pulmonary embolism lesion segmentation method as described in any one of claims 1 to 6.

Images