A method for detecting cardiovascular disease based on image analysis

Abstract

The application relates to the technical field of medical image analysis, and discloses a cardiovascular disease detection method based on image analysis. The method acquires a medical image sequence of a target patient, extracts a blood vessel region contour to generate an initial blood vessel topology graph. Then, the initial topology graph is registered with a standard cardiovascular model, the curvature deviation degree and the pipe diameter variation coefficient of each blood vessel branch are calculated, and a blood vessel elasticity feature matrix is generated in combination with abnormal displacement nodes. The curvature deviation degree, the pipe diameter variation coefficient and the elasticity feature matrix are subjected to multi-scale fusion, a blood vessel structure abnormality index is output, and a blood flow dynamics parameter set is generated. According to the gradient distribution of the parameter set on a three-dimensional model, a risk area is divided, a texture symbiotic matrix analysis is carried out on pixel clusters in the high-risk area, and texture fingerprints of calcified plaques and lipid deposits are extracted. According to the peak value distribution, the type and severity grade of the cardiovascular disease are determined. The application realizes precise and objective detection and risk assessment of the cardiovascular disease.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, specifically to a method for detecting cardiovascular diseases based on image analysis. Background Technology

[0002] The high incidence of cardiovascular disease necessitates long-term medical treatment and care for a large number of patients. This not only severely declines the patients' quality of life but also places enormous economic and human burdens on their families. Patients may require frequent medical visits, long-term medication, and even surgery, all of which impose a heavy financial burden on their families. Furthermore, family members must dedicate significant time and energy to caring for the patient, disrupting their normal lives and work.

[0003] From a societal perspective, the widespread prevalence of cardiovascular disease places an enormous strain on medical resources. To meet the treatment needs of cardiovascular disease patients, hospitals require substantial investments in medical equipment, medications, and professional medical personnel. In some large hospitals, cardiology and cardiovascular surgery departments are often overcrowded, with bed shortages and heavy workloads for medical staff. This over-concentration of medical resources in cardiovascular disease treatment also affects the supply of medical services for other diseases, leading to an imbalance in the allocation of medical resources. Simultaneously, the large number of patients who lose their ability to work due to cardiovascular disease negatively impacts socio-economic development, reducing the labor supply, lowering productivity, and hindering sustained economic growth.

[0004] In the face of the severe challenges posed by cardiovascular diseases, timely and accurate detection is crucial for the effective treatment and prevention of these diseases. However, existing traditional detection methods have many limitations and are insufficient to meet the needs of clinical practice.

[0005] Electrocardiography (ECG), a commonly used method for detecting cardiovascular diseases, works by placing electrodes on the body surface to record changes in the heart's electrical activity. When heart cells are excited, they produce changes in electrical potential, which can be recorded by electrodes. The waveform of an ECG reflects the changes in the electrical potential of myocardial cells in different parts of the heart. ECG has definite value in diagnosing arrhythmias and conduction blocks, and characteristic ECG changes and evolutions are also a reliable and practical method for diagnosing myocardial infarction. However, ECG also has significant limitations. Its ability to detect subtle lesions and early diseases is limited. For occult coronary artery disease, due to the mild degree of myocardial ischemia, ECG may not be able to capture obvious abnormalities, easily leading to missed diagnoses. Similarly, ECG cannot provide accurate diagnostic information for microvascular dysfunction, because microvascular lesions usually do not cause significant changes in the heart's macroscopic electrical activity.

[0006] Echocardiography, also known as echocardiography, utilizes the physical properties of ultrasound waves. A probe emits high-frequency sound waves, which propagate within the body's tissues and are reflected at different tissue interfaces. The reflected sound waves are received by the probe and converted into electrical signals, which are then processed by a computer to form an image of the heart. This image can display the heart's structure, function, and blood flow in real time, providing crucial information for clinical diagnosis and treatment. It can be used to detect cardiac structural abnormalities, assess cardiac function and hemodynamics, diagnose congenital heart disease, and monitor changes in the condition of patients with heart disease. However, it also has limitations. It is insufficient in assessing changes in cardiovascular function over time and cannot comprehensively and accurately monitor dynamic changes in the cardiovascular system. For some functional disorders, such as changes in myocardial function caused by early myocardial ischemia, echocardiography may not be able to detect them in time because the heart's structural morphology may not have changed significantly at this stage, and ultrasound images alone cannot accurately determine subtle abnormalities in myocardial function. In addition, the diagnostic results of echocardiography are greatly affected by the doctor's experience and skills, and are highly subjective. Different doctors may reach different diagnostic conclusions. Its resolution is also limited by the accuracy of the equipment and the doctor's operating level, which limits the detection of some subtle lesions. The detection angle is also limited, making it difficult to obtain all the information about the lesion from multiple angles. Summary of the Invention

[0007] The purpose of this invention is to provide a cardiovascular disease detection method based on image analysis to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides a cardiovascular disease detection method based on image analysis, the method comprising:

[0009] Acquire medical image sequences of the target patient, extract the vascular region contours in each frame of the image and generate an initial vascular topology map;

[0010] The blood vessel motion trajectory is constructed based on the displacement vector and grayscale difference of the blood vessel region in adjacent frames of images, and abnormal displacement nodes are marked.

[0011] Spatial registration was performed between the initial vascular topology map and the standard cardiovascular model, and the curvature deviation and diameter variation coefficient of each vascular branch were calculated.

[0012] The vascular elasticity feature matrix is ​​generated based on the distribution density of abnormal displacement nodes and the mechanical parameters of vascular branches.

[0013] Multi-scale fusion of curvature deviation, diameter variation coefficient and vascular elasticity feature matrix is ​​performed to output vascular structural abnormality index;

[0014] Temporal fluctuation features of blood flow signals are extracted from image sequences and combined with vascular structural abnormality index to generate a set of hemodynamic parameters;

[0015] Based on the gradient distribution of hemodynamic parameter sets on the three-dimensional vascular model, high-risk lesion areas and low-risk normal areas are divided.

[0016] Texture co-occurrence matrix analysis was performed on pixel clusters within high-risk lesion areas to extract texture fingerprints of calcified plaques and lipid deposits;

[0017] Based on the coupling relationship between texture fingerprint and vascular structural abnormality index, a lesion probability prediction matrix is ​​constructed;

[0018] The type and severity of cardiovascular disease are determined based on the peak distribution of the lesion probability prediction matrix.

[0019] Preferably, the step of extracting the vascular region contours from each frame of image and generating an initial vascular topology map includes:

[0020] An adaptive threshold segmentation algorithm is used to extract the binarized mask of the blood vessel region in each frame of the image;

[0021] Morphological erosion is performed on the binarized mask of the blood vessel region in adjacent frames to eliminate artifact interference and generate a connected component label map.

[0022] Extract the spatial coordinates of all vessel centerlines from the connected component labeled graph, and construct an initial vessel topology map using cubic spline interpolation.

[0023] Preferably, the step of constructing the vascular motion trajectory based on the displacement vector and grayscale difference of the vascular region in adjacent frames includes:

[0024] Calculate the Euclidean distance and orientation angle between the matching vessel centerlines in adjacent image frames;

[0025] When the Euclidean distance exceeds a preset multiple of the blood vessel diameter and the directional angle is greater than a critical threshold, the node is marked as an abnormal displacement node.

[0026] Spatiotemporal clustering is performed on all abnormal displacement nodes to generate a fracture region map of the blood vessel movement trajectory.

[0027] Preferably, the spatial registration of the initial vascular topology map with the standard cardiovascular model includes:

[0028] Feature point matching is performed between the main branches of the initial vascular topology map and the anatomical structures of the standard cardiovascular model.

[0029] The spatial deformation of vascular branches is corrected by a thin-plate spline transformation algorithm, and the registered vascular curvature field is output.

[0030] The curvature deviation of the midline of each branch from that of the standard model is calculated in the registered vascular curvature field.

[0031] Preferably, the generation of the vascular elastic feature matrix based on the distribution density of abnormal displacement nodes and the mechanical parameters of vascular branches includes:

[0032] Measure the periodic contraction amplitude and phase delay of the blood vessel wall around the abnormal displacement node;

[0033] The local Young's modulus was derived based on the linear relationship between the contraction amplitude and intravascular pressure.

[0034] The Young's modulus data of all abnormal displacement nodes are integrated to generate a vascular elasticity feature matrix.

[0035] Preferably, the multi-scale fusion of curvature deviation, diameter variation coefficient, and vascular elasticity feature matrix includes:

[0036] Wavelet decomposition of the vascular structural abnormality index yields high-frequency detail components and low-frequency contour components.

[0037] Convolution operation is performed between high-frequency detail components and the vascular elasticity feature matrix to enhance the local lesion feature response;

[0038] The weighting of low-frequency contour components and hemodynamic parameter sets is optimized using the backpropagation algorithm.

[0039] Preferably, the risk zones include:

[0040] Calculate the Laplacian operator at each vertex of the surface of a three-dimensional vascular model for the set of hemodynamic parameters.

[0041] When the Laplace operator exceeds the dynamic threshold and the vascular structural abnormality index continues to increase, it is identified as a high-risk lesion area.

[0042] The high-risk lesion area is expanded using a region growing algorithm until the fluid shear force mutation point at the blood vessel bifurcation is encountered.

[0043] Preferably, the extraction of texture fingerprints from calcified plaques and lipid deposits includes:

[0044] Multiple sets of pixel grayscale sequences were extracted along the normal direction of the blood vessel wall within high-risk lesion areas;

[0045] Each group of pixel grayscale sequences is processed by the Gabor filter bank to extract the texture response spectrum of different frequency bands;

[0046] After dimensionality reduction by principal component analysis, texture fingerprints of calcified plaques and lipid deposits were generated.

[0047] Preferably, the construction of the lesion probability prediction matrix includes:

[0048] Establish a joint probability density function for texture fingerprint encoding and vascular structural abnormality index;

[0049] Monte Carlo sampling method was used to simulate the spatial distribution patterns of different lesion types;

[0050] A lesion probability prediction matrix is ​​generated based on the clustering degree of various lesions in the sampling results.

[0051] Preferably, determining the type and severity of cardiovascular disease based on the peak distribution of the lesion probability prediction matrix includes:

[0052] Identify the spatial geometric features of continuous peak regions in the lesion probability prediction matrix;

[0053] When the peak region exhibits a ring-shaped distribution and the texture fingerprint encoding conforms to calcification characteristics, it is determined to be atherosclerosis;

[0054] When the peak region exhibits a linear distribution and the vascular elasticity characteristic matrix value is abnormal, it is determined to be a vascular fibrosis lesion.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] This image-based cardiovascular disease detection method overcomes the limitations of traditional detection techniques, providing a comprehensive and precise analysis of the cardiovascular system from multiple dimensions. It goes beyond the detection of a single indicator, comprehensively considering multiple key elements such as vascular topology, motion trajectory, and hemodynamics. By acquiring medical image sequences of target patients, the contours of vascular regions in each frame are extracted to generate an initial vascular topology map, clearly revealing the morphology and layout of the cardiovascular system and providing a solid foundation for subsequent analysis. Based on the displacement vectors and grayscale differences of vascular regions in adjacent frames, vascular motion trajectories are constructed, and abnormal displacement nodes are marked, enabling real-time tracking of dynamic changes in blood vessels and capturing subtle anomalies in vascular movement. The initial vascular topology map is spatially registered with a standard cardiovascular model, and the curvature deviation and diameter variation coefficient of each vascular branch are calculated, providing a quantitative analysis of blood vessels from a geometric morphology perspective and accurately revealing vascular deformation and dilation. In terms of hemodynamics, the temporal fluctuation characteristics of blood flow signals in the image sequences are extracted and combined with a vascular structural anomaly index to generate a hemodynamic parameter set, gaining a deeper understanding of the flow characteristics and changing patterns of blood flow in the cardiovascular system. This multi-dimensional analytical approach enables doctors to comprehensively understand the cardiovascular status, perceive subtle changes in the cardiovascular system, and discover lesions that are easily missed by traditional methods. In detecting early coronary artery disease, traditional methods may only focus on the degree of coronary artery stenosis, while this method, by comprehensively analyzing vascular topology, motion trajectory, and hemodynamics, can detect early atherosclerotic plaques in the coronary arteries, microscopic damage to the vessel wall, and abnormal changes in hemodynamics, thereby achieving early diagnosis and intervention. Through spatial registration, parameter calculation, and other techniques, this method can accurately quantify changes in blood vessels, providing accurate data for diagnosis. When assessing the degree of vascular lesions, traditional methods often rely on the doctor's subjective judgment, which involves significant errors and uncertainties. This method, by calculating parameters such as curvature deviation and coefficient of variation in vessel diameter, can accurately quantify the degree of vascular tortuosity and changes in vessel diameter, providing doctors with objective and accurate diagnostic evidence, making the diagnostic results more reliable and scientific, and providing strong support for subsequent treatment planning.

[0057] Early detection is crucial in the prevention and treatment of cardiovascular diseases, and this image analysis-based detection method demonstrates unique advantages in this regard. By marking abnormal displacement nodes, it can keenly capture abnormal changes in vascular movement. Abnormal vascular displacement is often an early signal of cardiovascular disease; for example, decreased elasticity of the vessel wall and the appearance of plaques within the vessel can lead to abnormal displacement of the vessel during movement. This method can promptly detect these abnormal displacement nodes and conduct in-depth analysis, thereby achieving early warning of cardiovascular diseases. The temporal fluctuation characteristics of blood flow signals in image sequences are extracted and analyzed, and combined with a vascular structural abnormality index to generate a hemodynamic parameter set, further uncovering early signs of disease from a hemodynamic perspective. Changes in hemodynamics often precede obvious changes in vascular morphology. Through detailed analysis of blood flow signals, abnormalities in parameters such as blood flow velocity, flow rate, and pressure can be detected in the early stages of disease, providing important clues for early diagnosis. Texture co-occurrence matrix analysis is performed on pixel clusters within high-risk lesion areas to extract texture fingerprints of calcified plaques and lipid deposits, enabling the identification of lesions at a very early stage. Calcified plaques and lipid deposition are important pathological features of cardiovascular diseases, and early detection of these features is crucial for disease diagnosis and treatment. By deeply studying the coupling relationship between texture fingerprints and vascular structural abnormality indices, a lesion probability prediction matrix can be constructed, enabling more accurate prediction of disease occurrence and development trends. In detecting early myocardial ischemia, this method can analyze vascular motion trajectories and hemodynamic parameters to identify vascular motion abnormalities and insufficient blood perfusion in the myocardial supply area. Combined with texture analysis, it can detect early changes in myocardial tissue, thus making an accurate diagnosis before obvious symptoms appear. Early detection is of immeasurable importance for disease treatment and patient recovery. Early detection of cardiovascular diseases allows patients to receive treatment in the early stages, significantly improving cure rates. Many cardiovascular diseases, in their early stages, have smaller lesions and milder symptoms; timely intervention and treatment can effectively control disease progression and even achieve complete cure. Early treatment can reduce treatment costs. Compared to the middle and late stages of disease, early treatment requires fewer medical resources, uses simpler treatment methods, and is less expensive, reducing the economic burden on patients and their families. Early diagnosis and treatment can also reduce the occurrence of complications. If cardiovascular diseases are not treated promptly, they often lead to a series of serious complications, such as heart failure, arrhythmia, and myocardial infarction. These complications not only increase the difficulty of treatment but also pose a greater threat to the patient's life and health. Early detection and treatment can effectively prevent the occurrence of complications, improve the patient's quality of life, and safeguard their life and health. Attached Figure Description

[0058] Figure 1 This is a schematic diagram illustrating the working principle of the image analysis-based cardiovascular disease detection method described in this invention.

[0059] Figure 2 A flowchart for extracting the contours of blood vessel regions in each frame of an image and generating an initial blood vessel topology map;

[0060] Figure 3 A flowchart for spatial registration of the initial vascular topology map with a standard cardiovascular model. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Please see Figure 1This invention provides an image analysis-based method for cardiovascular disease detection. The method involves processing medical image sequences of a target patient to achieve automated detection and assessment of cardiovascular diseases. The medical image sequences typically originate from modalities such as computed tomography (CT) angiography or magnetic resonance angiography (MRI), containing multiple consecutive image frames. The method first acquires these image sequences and extracts the contours of vascular regions frame by frame to generate an initial vascular topology map. This initial vascular topology map represents the spatial connectivity of vascular branches. Based on the motion information of vascular regions in adjacent image frames, displacement vectors and grayscale differences are calculated to construct vascular motion trajectories, and abnormal displacement nodes are marked on the trajectories. Abnormal displacement nodes reflect abnormal movement of the vascular wall. The initial vascular topology map is further spatially registered with a standard cardiovascular model, which is constructed based on anatomical data from healthy individuals. After registration, the curvature deviation and diameter variation coefficient of each vascular branch are calculated. The curvature deviation measures the difference between the degree of vascular tortuosity and the standard model, while the diameter variation coefficient describes the fluctuation of the vascular diameter. Combining the distribution density of abnormal displacement nodes and the mechanical parameters of vascular branches, a vascular elasticity feature matrix is ​​generated, which quantifies the elastic properties of the vascular wall. Multi-scale fusion of curvature deviation, coefficient of variation of vessel diameter, and vascular elasticity feature matrix is ​​performed to integrate feature information from different spatial resolutions, outputting a vascular structural abnormality index. This index provides a comprehensive assessment of vascular morphology. Temporal fluctuation features of blood flow signals are extracted from image sequences, including periodic changes in blood flow velocity and pressure. Blood flow signals are combined with the vascular structural abnormality index to generate a hemodynamic parameter set, which describes the interaction between blood flow and vascular structure. The gradient distribution of the hemodynamic parameter set is analyzed on a 3D vascular model, revealing the spatial rate of change of the parameters. Based on the gradient distribution, high-risk lesion regions and low-risk normal regions are defined, with high-risk lesion regions indicating potential pathological changes. Texture co-occurrence matrix analysis is performed on pixel clusters within high-risk lesion regions, extracting spatial relationship features of pixels to identify texture fingerprints of calcified plaques and lipid deposits. Texture fingerprints serve as digital signatures for specific lesions. Based on the coupling relationship between texture fingerprints and the vascular structural abnormality index, a lesion probability prediction matrix is ​​constructed, quantifying the likelihood of lesions occurring at different locations. Finally, the type and severity of cardiovascular disease are determined based on the peak distribution of the lesion probability prediction matrix, which includes the geometric pattern of the high-probability region in the matrix.

[0063] Example 1: See Figure 2Acquisition of medical image sequences is accomplished through medical imaging equipment, such as coronary CT scanners, which generate multiple consecutive image frames. Each frame requires preprocessing to reduce noise and artifacts. Preprocessing operations include Gaussian filtering and nonlocal mean denoising algorithms. Gaussian filtering smooths image details, while nonlocal mean denoising preserves edge information. Image sequences typically contain data from different phases of the cardiac cycle, ensuring sufficient temporal resolution to capture vascular motion. An adaptive thresholding algorithm is used to extract vascular region contours. This algorithm dynamically calculates the segmentation threshold based on the local pixel grayscale distribution. The threshold calculation is based on the mean and standard deviation of pixels within a window, with the window size adapting to changes in vascular size. The adaptive thresholding algorithm processes images with different contrasts and outputs a binary mask, in which vascular regions are marked as foreground pixels. Morphological erosion is then performed on the vascular region binary masks of adjacent frames. This morphological erosion operation uses circular structuring elements to eliminate small artifacts and noise points. The radius of the structuring element is adjusted according to the vascular diameter. Morphological erosion reduces false detection areas and improves the connectivity of vascular regions. The connected component labeling graph is generated by scanning a binarized mask and assigning a unique label to each connected blood vessel region. The connected component labeling algorithm is based on region growing or boundary tracking methods to ensure that each blood vessel branch is independently identified. The connected component labeling graph stores the topological relationships of the blood vessel regions.

[0064] The spatial coordinates of all vessel centerlines in the connected component labeled graph are extracted. The vessel centerlines are extracted using a skeletonization algorithm, which preserves the center pixels through iterative erosion. The vessel centerline represents the central axis path of the vessel. An initial vessel topology map is constructed using cubic spline interpolation. Cubic spline interpolation processes the centerline coordinates to generate a smooth and continuous curve. The initial vessel topology map contains nodes and edges; nodes represent the coordinates of the center points, and edges represent the connection relationships between vessel segments. The initial vessel topology map is stored as a graph data structure to support subsequent spatial analysis. Constructing the vessel motion trajectory requires analyzing the motion changes of the vessel region in adjacent image frames. The Euclidean distance and orientation angle of the matching vessel centerlines in adjacent image frames are calculated. The Euclidean distance measures the linear displacement between center points, and the orientation angle is calculated based on vector dot product and cross product. Matching the vessel centerlines is achieved through feature point correspondence, such as using SIFT or SURF feature descriptors. When the Euclidean distance exceeds a preset multiple of the vessel diameter and the orientation angle is greater than a critical threshold, this node is marked as an abnormal displacement node. The preset multiple threshold is set according to the dynamic range of the vessel diameter, and the critical threshold is a fixed angle value. Abnormal displacement nodes indicate abnormal movement or local deformation of the vessel wall. Spatiotemporal clustering is performed on all abnormal displacement nodes. Spatiotemporal clustering algorithms such as DBSCAN group nodes that are adjacent in time and space. Spatiotemporal clustering considers the timestamps and spatial coordinates of the nodes to generate a fracture region map of the blood vessel movement trajectory. The fracture region map visualizes the areas of discontinuous motion and provides input for elasticity analysis.

[0065] The adaptive thresholding segmentation algorithm relies on local image statistics to calculate the gray-level characteristics within the neighborhood of each pixel. The size of the neighborhood affects the segmentation accuracy; smaller neighborhoods capture details but are susceptible to noise, while larger neighborhoods provide strong smoothing effects but may blur boundaries. The adaptive thresholding algorithm addresses gray-level inhomogeneity in the image, improving the robustness of vessel extraction. Morphological erosion operations utilize a sliding window with structuring elements to remove boundary pixels, enhancing vessel shape. Circular structuring elements simulate isotropic erosion, adapting to the circular cross-section of vessels. During connected component labeling, a binary mask is scanned to identify connected components, and the labeling algorithm ensures a unique identifier for each region, facilitating subsequent centerline extraction. Vessel centerline extraction employs a thinning algorithm, iteratively removing boundary pixels until a single pixel width is reached, ensuring the centerline accurately represents the geometric axis of the vessel. Cubic spline interpolation fits the centerline point sequence, with spline curves ensuring first- and second-order continuity. The initial vessel topology is constructed as a graph model, with node attributes including coordinates and connectivity information. In motion trajectory analysis, inter-frame registration is achieved through optical flow or feature matching. Euclidean distance and orientation angle calculations are based on registered point pairs. Abnormal displacement node detection uses threshold comparison, with threshold settings based on statistical experience or training data. The spatiotemporal clustering algorithm DBSCAN is based on density clustering to identify anomalous clustering regions, and the fracture region map is output as image or vector data.

[0066] Medical image sequences may contain patient motion or device artifacts. Preprocessing steps reduce these effects. Adaptive thresholding adapts to local contrast variations, and morphological erosion optimizes the quality of the binary mask. Connected component labeling ensures the integrity of vascular regions, and centerline extraction provides a simplified representation. Motion trajectory construction adds a temporal dimension, abnormal displacement nodes focus on pathological regions, and spatiotemporal clustering enhances the reliability of anomaly detection. During implementation, algorithm parameters need to be tuned to adapt to different image modalities, such as the differences between CT and MRI. Computational efficiency is optimized through parallel processing, such as GPU-accelerated image manipulation. In the specific implementation of the adaptive thresholding algorithm, a local threshold is calculated for each pixel. The local threshold is based on the mean and standard deviation of the pixel values ​​in the neighborhood. The neighborhood window size is set according to the image resolution, such as a 5x5 or 7x7 pixel window. The adaptive thresholding algorithm handles regions with large grayscale variations, avoiding the limitations of global thresholding. The morphological erosion operation uses structuring element convolution, where the structuring element is a binary matrix. After erosion, isolated points in the binary mask are removed. Connected component labeling uses a two-pass scanning algorithm: the first pass assigns temporary labels, and the second pass parses equivalent labels to generate the final labeled map.

[0067] Vessel centerline extraction is achieved using the Zhang-Suen thinning algorithm, which iteratively erodes boundary pixels while preserving the central skeleton. Cubic spline interpolation calculates the curves between control points, ensuring a smooth centerline. The initial vessel topology map stores a node list and adjacency matrix. In motion analysis, inter-frame registration uses phase correlation or feature matching, Euclidean distance calculates the straight-line distance between point pairs, and the directional angle is obtained using the vector angle formula. After anomaly labeling, the spatiotemporal clustering algorithm DBSCAN groups nodes based on distance metrics, generating bright discontinuous regions in the fracture region map. Temporal consistency of the image sequence is crucial; preprocessing aligns inter-frame displacements, and adaptive threshold segmentation improves segmentation consistency. Morphological erosion enhances vessel shape, and connected component labeling handles branch connections. Accurate centerline extraction ensures the central axis is captured, and spline interpolation smooths the path. Motion trajectory construction depends on registration accuracy, and anomaly detection thresholds need validation. Clustering algorithm parameters such as epsilon and minPts affect the results, and fracture maps are used for visualization. The entire process is automated, reducing manual intervention. The adaptive threshold segmentation algorithm performs critically in low-contrast images, with local thresholds adapting to changes in dark or bright areas. Morphological corrosion parameters, such as the size of structuring elements, affect the degree of corrosion, requiring a balance between noise reduction and shape preservation. Connected component labeling handles complex branches, and algorithms address overlapping regions. Centerline extraction avoids branch breakage, and spline interpolation handles sparse points. In motion analysis, registration errors may propagate, and robust thresholding is needed for anomaly detection. Spatiotemporal clustering processes dynamic data, and fracture maps indicate motion anomalies. Implementation must consider computational load and optimize algorithm speed.

[0068] Extracting the vascular region contour is a fundamental step, and an adaptive thresholding segmentation algorithm provides a highly adaptive method. Morphological erosion cleanses the binary mask, and connected component labeling establishes regional relationships. Vessel centerline generation simplifies geometry, and cubic spline interpolation ensures continuity. The initial vascular topology map supports subsequent registration. Motion trajectory analysis adds temporal information, and abnormal displacement nodes identify pathological points. Spatiotemporal clustering discovers patterns, and the fracture region atlas outputs abnormal regions. The method integrates image processing techniques to achieve automated analysis. The adaptive thresholding segmentation algorithm avoids global thresholding issues through local computation, and morphological erosion uses circular structuring elements to maintain shape. The connected component labeling algorithm efficiently processes large images, and the centerline extraction and refinement algorithm preserves topology. Cubic spline interpolation smooths curves, and the topology map stores the graph structure. Motion trajectories are constructed through inter-frame comparisons, and Euclidean distance and direction angle quantify motion. Abnormal node labeling is based on thresholds, and spatiotemporal clustering (DBSCAN) discovers clusters. The fracture atlas displays discontinuous areas, providing input for elasticity analysis. The workflow processes multi-frame data to ensure temporal consistency.

[0069] Noise in medical image sequences can affect segmentation. Preprocessing uses Gaussian filtering to smooth noise and non-local means denoising to preserve structure. Adaptive threshold segmentation dynamically adjusts, and morphological erosion removes small objects. Connected component labeling identifies regions, and centerline extraction creates a skeleton. Spline interpolation creates smooth paths, and a topology graph represents the network. Motion analysis calculates displacement, and thresholding is used for anomaly detection. Clustering groups anomalies, and fracture maps are visualized. The method emphasizes automation to reduce subjectivity. In the adaptive threshold segmentation algorithm, a local threshold is calculated for each pixel, based on neighborhood statistics to avoid uneven illumination. Morphological erosion applies structuring elements to shrink the erosion region. Connected component labeling uses a disjoint-set data structure for efficient label processing. Centerline extraction uses iterative erosion, and spline interpolation uses control points. Motion trajectory registration uses Euclidean distance for simple and effective frame point matching. Direction angles are calculated using vector angles, and anomaly marking is compared with thresholds. Spatiotemporal clustering density is based on [data structure not specified], and fracture maps are generated into images. Implementation requires handling large datasets, so memory usage should be optimized. The accuracy of vessel extraction affects subsequent steps; adaptive threshold segmentation adapts to local variations. Morphological corrosion cleans the mask, connected component labeling establishes connections. Centerline extraction extracts the central axis, spline interpolation smooths the surface. Topological graph stores relationships, motion analysis adds a time dimension. Anomaly nodes identify problem points, clustering discovers patterns. Fracture maps guide in-depth analysis.

[0070] Example 2: See Figure 3 The spatial registration process aligns the initial vascular topology map with a standard cardiovascular model. The initial vascular topology map is the result of vascular centerline extraction and cubic spline interpolation. The standard cardiovascular model is an average anatomical atlas constructed based on medical imaging data from a large number of healthy individuals. The standard cardiovascular model contains three-dimensional geometric information of the major cardiac vascular branches, such as the standard dimensions and spatial orientation of the left main coronary artery, the left anterior descending artery, the circumflex artery, and their major branches. The feature point matching operation selects the main branches of the initial vascular topology map and their corresponding anatomical structures in the standard cardiovascular model for corresponding point identification. Feature points include bifurcation points, curvature extrema, and markers with significant geometric features. Bifurcation points are the locations where a vessel splits into two or more sub-branches, and curvature extrema are the points with the maximum or minimum curvature on the vascular centerline. These feature points possess anatomical stability and repeatability. Feature point matching is optimized using the Iterative Closest Point Algorithm or its improved version. The Iterative Closest Point Algorithm minimizes the average distance error between matching point pairs by iteratively calculating the rigid body transformation between two point sets. The matching process requires setting a distance tolerance and a maximum number of iterations to prevent infinite loops.

[0071] The spatial deformation of vascular branches is corrected using a thin-plate spline transformation algorithm. Thin-plate spline transformation is a non-rigid registration technique that fits a smooth, non-linear spatial transformation field based on the displacement of a set of control points, which are the previously matched feature point pairs. The algorithm obtains the transformation function by solving an optimization problem that minimizes bending energy. This transformation function smoothly deforms the standard cardiovascular model to fit the actual shape of the initial vascular topology while maintaining the integrity of the topology. The registered vascular curvature field is output as a scalar field defined at each point on the vascular centerline, representing the curvature value at each point. Curvature calculation is based on the Frenet framework in differential geometry, obtained through the first and second derivatives of the centerline parametric equations. The curvature deviation of the central axis of each vascular branch from the standard cardiovascular model is calculated in the registered vascular curvature field. This curvature deviation is a quantitative indicator used to measure the difference between the curvature of the patient's blood vessels and the healthy standard model. Calculating curvature deviation requires point-by-point comparison between the registered vessel centerline and the standard cardiovascular model centerline. The absolute or squared value of the difference in curvature values ​​at corresponding points is calculated, and then the result is integrated or averaged along the entire vessel branch path. The integration path is the arc length of the vessel centerline. The coefficient of variation (COP) for vessel diameter is calculated simultaneously as a supplementary morphological parameter. The COP describes the variation of vessel diameter along its length. The COP is obtained by measuring the local diameter of the vessel and calculating the ratio of its standard deviation to its mean. The local diameter can be derived from the perpendicular distance between the vessel contour and the centerline.

[0072] The construction of a standard cardiovascular model is a data-intensive process involving the acquisition, segmentation, 3D reconstruction, and statistical averaging of image data from healthy volunteers. The model is typically stored as a point cloud or mesh. The accuracy of feature point matching directly affects the accuracy of subsequent registration. Iterative nearest-point algorithms are sensitive to initial positions, sometimes requiring manual assistance or the use of more robust feature descriptors to provide a good initial correspondence. The core of the thin-plate spline transformation algorithm is solving a system of linear equations to determine the coefficients of the transformation function. The bending energy term ensures the smoothness of the transformation, preventing unnatural and drastic deformation. The registered vascular curvature field contains not only curvature values ​​but may also include vector information such as curvature direction, providing a geometric basis for subsequent mechanical analysis. Calculating curvature deviation requires solving the point correspondence problem, i.e., identifying which points on the two centerlines are anatomically corresponding. This is typically achieved through arc-length parameterization or nearest-point projection methods. The calculation of the coefficient of variation of vessel diameter relies on accurate vessel wall boundary detection. The vessel wall boundary can be re-extracted from the original medical images through edge detection or active contour models to ensure the accuracy of diameter measurement. The entire spatial registration process aims to eliminate anatomical differences between individuals and geometric distortions during the imaging process, so that the vascular morphology parameters of different patients are comparable.

[0073] The initial vascular topology may exhibit complex spatial deformations due to the patient's specific physiological conditions or pathological state. The thin-plate spline transformation algorithm can effectively handle such nonlinear deformations, and its smoothness constraints avoid excessive local distortion, preserving the overall topological structure of the vessels. Generating the vascular curvature field is a crucial preprocessing step in computational fluid dynamics simulations and hemodynamic analysis; high curvature regions are often associated with abnormal wall shear stress distributions. Curvature deviation, as a comprehensive indicator, can highlight morphological abnormalities such as tortuosity and angularity of vessels, which may be related to diseases such as atherosclerosis and vasospasm. The coefficient of variation in vessel diameter focuses more on the regularity of the vascular lumen; significant variation may indicate the presence of local stenosis or dilatational lesions. Computational efficiency is a crucial factor to consider when performing spatial registration, especially when dealing with high-resolution 3D vascular models. Both the iterative nearest-point algorithm and the thin-plate spline transformation algorithm can be accelerated through optimized data structures and parallel computing. The selection of feature points needs to cover a sufficient number of anatomical landmarks to ensure uniformity and accuracy of registration across the entire vascular tree. Standard cardiovascular models may require multiple versions for different populations, such as different ages, sexes, or ethnicities, to improve the universality and accuracy of registration. Validation of registration results typically requires visual inspection or calculation of landmark positioning errors.

[0074] In the thin-plate spline transform algorithm, the smoothing parameter controls the flexibility of the transformation. A larger smoothing parameter results in a more rigid transformation, suitable for global shape alignment, while a smaller smoothing parameter allows for more local deformation to accommodate differences in detail. The choice of parameter needs to be adjusted according to the specific application scenario and the anatomical characteristics of the blood vessels. The registered vascular curvature field can be exported as image data or a numerical list for subsequent quantitative analysis. The calculated results of curvature deviation and diameter variation coefficient can be mapped back onto the 3D vascular model for visualization, helping clinicians to intuitively identify areas of morphological abnormalities. The successful implementation of spatial registration provides a standardized and comparable geometric basis for subsequent vascular elasticity analysis, hemodynamic calculations, and lesion risk assessment. It places patient-specific vascular geometry within a unified anatomical reference frame, enabling automated, image-based disease detection and quantification.

[0075] Example 3: The generation of the vascular elasticity feature matrix relies on abnormal displacement nodes identified from the vascular motion trajectory and the inherent mechanical properties of the vascular branches. Abnormal displacement nodes are marked during the vascular motion trajectory construction process; these nodes represent the spatial locations of abnormal vascular wall motion during the cardiac cycle. The distribution density of abnormal displacement nodes is calculated by statistically analyzing the number of abnormal displacement nodes per unit length of a given vascular segment. The distribution density reflects the spatial aggregation degree of abnormal vascular wall motion. Obtaining mechanical parameters requires measuring the periodic contraction amplitude and phase delay of the vascular wall. The periodic contraction amplitude is quantified by analyzing the maximum displacement of a specific point of the vascular wall relative to its equilibrium position in time-series images. The phase delay is determined by comparing the time difference between the time point of vascular wall contraction motion and the peak value of the R wave in the synchronously recorded electrocardiogram. The local Young's modulus is derived based on a simplified biomechanical model that assumes the vascular wall is a linearly elastic material with a linear relationship between stress and strain. The periodic contraction amplitude corresponds to strain, while the intravascular pressure corresponds to stress. Intravascular pressure data can be obtained from non-invasive blood pressure measurements, image-based hemodynamic simulations, or typical physiological pressure values ​​of the vascular segment. The local Young's modulus E is estimated by the following formula:

[0076]

[0077] in: This represents the local Young's modulus. Indicates local intravascular pressure. This indicates the radius of the blood vessel at the end of diastole. Indicates the amplitude of periodic contraction. This indicates the thickness of the blood vessel wall. It can be measured from high-resolution medical images. The local Young's modulus values ​​calculated from the locations of all abnormal displacement nodes are integrated and organized into a two-dimensional structure according to their corresponding spatial coordinates, namely the vascular elasticity feature matrix. The rows of the vascular elasticity feature matrix usually correspond to different vascular segments, and the columns correspond to mechanical parameters such as Young's modulus. The value of the matrix element is the estimated value of the local Young's modulus at a specific location.

[0078] The goal of multi-scale fusion is to integrate features from different sources, with different physical meanings, and at different spatial scales into a comprehensive vascular structural anomaly index. These input features include curvature deviation and diameter variation coefficients characterizing the geometric morphology, and the vascular elasticity feature matrix characterizing the mechanical properties. The multi-scale fusion process begins with wavelet decomposition of the initially calculated vascular structural anomaly index, which is a scalar field distributed along the vessel centerline or mapped onto the vessel surface. Wavelet decomposition uses a selected mother wavelet function, such as the Daubechies wavelet family, to decompose the vascular structural anomaly index into coefficients on a series of sub-bands of different frequencies. High-frequency detail components capture subtle local changes, noise, and edge information of the vascular structure, while low-frequency contour components carry the global shape and slow changing trends of the vascular structure. The high-frequency detail components are then convolved with the vascular elasticity feature matrix. The essence of the convolution operation is to use a sliding window function to perform a weighted summation on the vascular elasticity feature matrix. This operation aims to enhance the response of regions that exhibit abnormalities in both geometric morphology and mechanical properties. For example, a region with both high curvature deviation and high Young's modulus may produce a significantly high output value after convolution. The low-frequency profile component is then processed in conjunction with a set of hemodynamic parameters extracted from the blood flow signal. This set of hemodynamic parameters includes parameters such as blood flow velocity, wall shear stress, and pressure gradient.

[0079] The integration of low-frequency contour components with hemodynamic parameters is achieved through a weighting mechanism, which adjusts the contribution of different features to the final fusion result. The weighting mechanism is optimized using a backpropagation algorithm, which iteratively minimizes a pre-defined objective function to adjust the weight values. The objective function can be defined as the difference between the fusion result and a certain ideal state, although in an unsupervised setting, the ideal state may be defined by statistical properties or physical constraints. After multi-scale fusion processing, an enhanced vascular structural anomaly index containing multimodal information is output, which more comprehensively reflects the combined state of the blood vessel in terms of morphology, mechanics, and hemodynamics. The calculation of the distribution density of abnormal displacement nodes needs to consider the length normalization of the vessel segment to avoid bias caused by different vessel segment lengths. Mechanistic parameter measurements require sufficiently high temporal resolution in the image data to accurately capture the periodic motion of the vessel wall. The derivation formula for the local Young's modulus is based on a thin-walled cylindrical mechanical model, a commonly used simplified model; the mechanical behavior of actual biological tissues is more complex. The dimension of the vascular elastic feature matrix depends on the number of abnormal displacement nodes and the range of the analyzed blood vessels.

[0080] The choice of wavelet decomposition layer number requires a trade-off. Too many layers may lead to computational complexity and excessively fragmented high-frequency information, while too few layers may fail to effectively separate features at different scales. The size of the sliding window in convolution operations needs to match the feature scale of the vascular structure; an excessively large window will smooth out local details, while an excessively small window may be overly sensitive to noise. When optimizing weights in the backpropagation algorithm, the learning rate and number of iterations affect the convergence speed and the final result. The vascular structural abnormality index output by the multi-scale fusion process is an indicator that enhances lesion-related features, facilitating subsequent risk area classification. The generation of the vascular elasticity feature matrix establishes a bridge connecting abnormal vascular motion with wall mechanical properties; the measurement of periodic contraction amplitude and phase delay depends on the accuracy of image registration and tracking algorithms. The parameters in the local Young's modulus formula need to be accurately obtained from the image data, as the estimation of intravascular pressure introduces uncertainty. Wavelet transform in multi-scale fusion provides a multi-resolution analysis tool, enabling simultaneous analysis of signal characteristics in the frequency and spatial domains. Convolution is a linear operation; introducing nonlinear activation functions can also be considered to enhance feature representation capabilities. Backpropagation is typically used in supervised learning, where the definition of the objective function needs to be based on domain knowledge or heuristic rules.

[0081] Example 4: The division between high-risk lesion areas and low-risk normal areas is based on the spatial gradient characteristics of the hemodynamic parameter set on the surface of a 3D vascular model. The 3D vascular model originates from the patient-specific vascular geometry reconstruction results after spatial registration, and its surface is composed of numerous triangular facet vertices. The hemodynamic parameter set includes physical quantities such as flow velocity, wall shear stress, and pressure calculated from medical image sequences. These parameter values ​​are mapped to each vertex of the 3D vascular model. The Laplacian operator of the hemodynamic parameter set at each vertex of the 3D vascular model surface is calculated. The Laplacian operator is a differential operator characterizing the local curvature of a scalar field, and its discrete form can be solved by the weighted sum of the differences between the parameter values ​​of a vertex and its adjacent vertices. The magnitude of the Laplacian operator reflects the degree of local variation of hemodynamic parameters at the vertices of the triangular facets. The dynamic threshold is set based on the statistical distribution characteristics of the hemodynamic parameter set across the entire vascular model surface. For example, the 95th percentile of all vertex Laplacian operator values ​​is taken as the initial threshold, and dynamically adjusted in conjunction with the local trend of the vascular structural abnormality index. The vascular structural abnormality index is obtained from the multi-scale fusion output. The judgment of its continuous increase is based on trend analysis of the index values ​​of multiple consecutive sampling points along the vascular centerline or surface path, for example, using a linear regression slope greater than zero as a criterion. When the Laplacian operator value at the vertex of a triangular facet exceeds the dynamic threshold and the vascular structural abnormality index of its local area shows a continuous increasing trend, the vertex and its adjacent area are identified as high-risk lesion areas.

[0082] The initially identified high-risk lesion regions are expanded using a region growing algorithm, which uses vertices exceeding a threshold as seed points. The similarity criterion for region growing is based on the proximity of hemodynamic parameters between vertices, such as the difference in flow velocity or wall shear stress being less than a predetermined tolerance. The growth process iteratively incorporates adjacent vertices that meet the similarity criteria into the high-risk region. The stopping boundary of region growing is determined by the fluid shear force abrupt change point at the vessel bifurcation. Fluid shear force abrupt changes are identified by calculating the gradient magnitude of the wall shear stress vector at the vertex. When the gradient magnitude exceeds a threshold set based on global shear force distribution statistics, the point is marked as an abrupt change point, and the growth process stops there. The final generated high-risk lesion region is a connected surface patch on the surface of the 3D vascular model, exhibiting abnormal hemodynamic and structural features.

[0083] Texture fingerprinting of calcified plaques and lipid deposits was extracted in high-risk lesion areas. Texture analysis focused on the pixel grayscale patterns of vascular wall tissue in medical imaging. Multiple sets of pixel grayscale sequences were extracted along the normal direction of the inner surface of the vascular wall. This extraction was achieved by sampling along a direction line perpendicular to the vascular wall surface in the 3D image data, collecting pixel grayscale values ​​within a certain depth along each direction line to form a one-dimensional grayscale signal. Each acquired pixel grayscale sequence was processed using a Gaussian-tuned sine wave filter bank, which consists of a series of Gaussian-tuned sine wave filters with different center frequencies and directions, simulating the human visual system's perception of texture. Each Gaussian-tuned filter was convolved with the pixel grayscale sequence to obtain the filter's response output at specific frequencies and directions, extracting texture response spectra in different frequency bands. The texture response spectrum is a distribution map of the filter response energy as a function of frequency and direction. Principal component analysis (PCA) was used to reduce the dimensionality of the high-dimensional texture response spectrum data. PCA identified the orthogonal directions with the largest variance in the data, retaining the first few principal components, which were sufficient to represent most of the texture feature information. After dimensionality reduction, a low-dimensional feature vector is generated, namely the texture fingerprint encoding of calcified plaques and lipid deposits. Texture fingerprint encoding is a digital signature that quantitatively describes the local texture characteristics of the blood vessel wall. Refer to Table 1, which lists the key parameters involved in the texture fingerprint extraction process and their typical values ​​or the basis for their settings.

[0084] Table 1: Texture Fingerprint Extraction Parameter Table

[0085] Parameter name Setting basis or typical values Pixel grayscale sequence sampling depth Covering the visible wall thickness from the intima to the adventitia of the blood vessel lumen, for example, sampling 15 pixels. Gabor filter center frequency range Covers the fundamental frequency of the texture to high-frequency components, such as 0.1 to 0.4 times the Nyquist frequency. Number of Gabor filter directions To cover textures at different angles, typically four or eight directions are set. Principal component analysis retained component number Based on a cumulative variance contribution rate greater than 90%, 3 to 5 principal components are typically retained.

[0086] The computation of the Laplacian operator relies on the mesh connectivity of the 3D vascular model, typically requiring the construction of an adjacency list for the vertices. A dynamic thresholding mechanism allows the partitioning criteria to adapt to physiological variations among different patients. Trend analysis of the vascular structural abnormality index requires selecting an appropriate window size to balance noise sensitivity and the reliability of trend identification. The similarity tolerance parameter in the region growing algorithm affects the final range and shape of high-risk regions and requires careful selection. Detection of fluid shear stress mutation points relies on accurate calculation of the wall shear stress field, which is usually based on computational fluid dynamics simulations. The sampling depth of the pixel grayscale sequence needs to be reasonably set according to the image resolution and vascular wall thickness. The parameters of the Gabor filter bank cover different spatial frequencies and orientations to capture texture features of varying properties. Principal component analysis effectively reduces the dimensionality of feature data while retaining the most discriminative information; the generated texture fingerprint encoding facilitates subsequent pattern recognition and classification operations.

[0087] Example 5: The construction of the lesion probability prediction matrix begins with the analysis of the coupling relationship between texture fingerprint encoding and the vascular structural abnormality index. Texture fingerprint encoding is extracted from high-risk lesion areas, characterizing the microscopic texture characteristics of the local vascular wall. The vascular structural abnormality index is derived from multi-scale fusion output, reflecting the macroscopic morphology and mechanical state of the blood vessel. Establishing the joint probability density function of texture fingerprint encoding and vascular structural abnormality index is the core step in quantifying their statistical correlation. The joint probability density function is constructed using a non-parametric kernel density estimation method. Kernel density estimation uses a Gaussian kernel function to smooth the sample data points, transforming discrete sample points into a continuous probability distribution. The bandwidth parameter of kernel density estimation is selected by minimizing the integral square error criterion to ensure that the probability density function is neither over-smoothed nor overfitted to noise. Texture fingerprint encoding samples and corresponding vascular structural abnormality index samples constitute two-dimensional observation data points, with each data point corresponding to a specific spatial location on the surface of a three-dimensional vascular model.

[0088] Monte Carlo sampling was employed to simulate the spatial distribution patterns of different lesion types. Monte Carlo sampling randomly extracted a large number of sample points from the aforementioned joint probability density function. Each sample point contains a virtual texture fingerprint encoding value and a virtual vascular structural abnormality index value. The simulation process set different sampling region preferences for different potential lesion types. For example, for calcified plaques, sampling tended to be performed in areas with high intensity of high-frequency components in the texture fingerprint encoding; for lipid deposition, sampling might focus more on areas with moderate vascular structural abnormality indices but exhibiting specific texture patterns. Monte Carlo sampling generated tens of thousands of sample points, which formed a simulated lesion distribution cloud map in the 3D vascular model space. A lesion probability prediction matrix was generated based on the clustering degree of various lesions in the sampling results. Clustering degree calculation discretized the space of the 3D vascular model into regular voxel grids or local regions based on surface grids. For each voxel or surface region, the number of Monte Carlo sample points falling within it and labeled as a specific lesion type was counted. The ratio of the number of Monte Carlo sample points to the total number of sample points is the probability estimate of the occurrence of this type of lesion in the spatial location of the 3D vascular model. The lesion probability prediction matrix is ​​a three-dimensional data array, where the value of each element represents the probability of a lesion occurring at the corresponding location in three-dimensional space. The dimension of the matrix is ​​consistent with the spatial discretization accuracy of the vascular model.

[0089] The type and severity of cardiovascular diseases are determined based on the peak distribution of the lesion probability prediction matrix. Peak distribution refers to the spatial arrangement of local maxima in the lesion probability prediction matrix whose probability values ​​are significantly higher than those of the surrounding areas. Identifying the spatial geometric features of continuous peak regions in the lesion probability prediction matrix requires the application of spatial clustering algorithms, such as density-based noise clustering. Spatial clustering algorithms group points that are spatially close and have high probability values ​​together, forming continuous peak regions. For each identified continuous peak region, its spatial geometric features are calculated. These features include the region's volume, shape factor, orientation relative to the vessel principal axis, and the uniformity of the peak probability distribution. When the peak region exhibits a ring-shaped distribution and the texture fingerprint encoding conforms to calcification characteristics, it is identified as atherosclerosis. A ring-shaped distribution means that the peak region presents an approximately circular geometric shape on the vessel cross-section, enclosing the vessel lumen. For the texture fingerprint encoding to conform to calcification characteristics, predefined numerical conditions must be met, such as a specific principal component score exceeding a threshold. This threshold is obtained by learning from training samples of known calcified plaques. When the peak regions exhibit a linear distribution and the vascular elastography matrix values ​​are abnormal, it is diagnosed as vascular fibrosis. A linear distribution means the peak regions extend along the longitudinal axis of the vessel, forming a long, strip-like geometric shape. Abnormal vascular elastography matrix values ​​are manifested by local Young's modulus values ​​consistently exceeding the reference range for normal vascular tissue. The reference range is derived from statistical data of healthy individuals. Severity is determined based on a comprehensive assessment of the total volume of the continuous peak regions, the average peak probability, and the distribution range of the peak regions along major vascular branches. For example, the lesion is classified into mild, moderate, and severe levels.

[0090] The construction of the lesion probability prediction matrix links microscopic texture information with macroscopic structural mechanical information through a probabilistic statistical model. The joint probability density function captures the potential statistical dependencies between these two heterogeneous data sets. Monte Carlo sampling provides a powerful computational tool for simulating the space of possible outcomes under complex statistical relationships. The lesion probability prediction matrix is ​​essentially a spatial probability map that maps abstract statistical learning results back to a concrete anatomical space, greatly enhancing the intuitiveness and clinical interpretability of the results. Disease type determination rules based on peak spatial geometric features integrate morphological knowledge and data-driven pattern recognition, improving the pathological relevance of classification. The comprehensive assessment of severity considers multiple dimensions of lesion burden, intensity, and extent, resulting in a more comprehensive evaluation.

[0091] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting cardiovascular diseases based on image analysis, characterized in that, include: Acquire medical image sequences of the target patient, extract the vascular region contours in each frame of the image and generate an initial vascular topology map; The blood vessel motion trajectory is constructed based on the displacement vector and grayscale difference of the blood vessel region in adjacent frames of images, and abnormal displacement nodes are marked. Spatial registration was performed between the initial vascular topology map and the standard cardiovascular model, and the curvature deviation and diameter variation coefficient of each vascular branch were calculated. The vascular elasticity feature matrix is ​​generated based on the distribution density of abnormal displacement nodes and the mechanical parameters of vascular branches. Multi-scale fusion of curvature deviation, diameter variation coefficient and vascular elasticity feature matrix is ​​performed to output vascular structural abnormality index; Temporal fluctuation features of blood flow signals are extracted from image sequences and combined with vascular structural abnormality index to generate a set of hemodynamic parameters; Based on the gradient distribution of hemodynamic parameter sets on the three-dimensional vascular model, high-risk lesion areas and low-risk normal areas are divided. Texture co-occurrence matrix analysis was performed on pixel clusters within high-risk lesion areas to extract texture fingerprints of calcified plaques and lipid deposits; Based on the coupling relationship between texture fingerprint and vascular structural abnormality index, a lesion probability prediction matrix is ​​constructed; The type and severity of cardiovascular disease are determined based on the peak distribution of the lesion probability prediction matrix; The extracted texture fingerprints of calcified plaques and lipid deposits include: Multiple sets of pixel grayscale sequences were extracted along the normal direction of the blood vessel wall within high-risk lesion areas; Each group of pixel grayscale sequences is processed by the Gabor filter bank to extract the texture response spectrum of different frequency bands; After dimensionality reduction by principal component analysis, texture fingerprints of calcified plaques and lipid deposits were generated.

2. The cardiovascular disease detection method based on image analysis according to claim 1, characterized in that, The step of extracting the contours of blood vessel regions from each frame of image and generating an initial blood vessel topology map includes: An adaptive threshold segmentation algorithm is used to extract the binarized mask of the blood vessel region in each frame of the image; Morphological erosion is performed on the binarized mask of the blood vessel region in adjacent frames to eliminate artifact interference and generate a connected component label map. Extract the spatial coordinates of all vessel centerlines from the connected component labeled graph, and construct an initial vessel topology map using cubic spline interpolation.

3. The cardiovascular disease detection method based on image analysis according to claim 2, characterized in that, The step of constructing the blood vessel motion trajectory based on the displacement vector and grayscale difference of the blood vessel region in adjacent frames of images includes: Calculate the Euclidean distance and orientation angle between the matching vessel centerlines in adjacent image frames; When the Euclidean distance exceeds a preset multiple of the blood vessel diameter and the directional angle is greater than a critical threshold, the node is marked as an abnormal displacement node. Spatiotemporal clustering is performed on all abnormal displacement nodes to generate a fracture region map of the blood vessel movement trajectory.

4. The cardiovascular disease detection method based on image analysis according to claim 3, characterized in that, The spatial registration of the initial vascular topology map with the standard cardiovascular model includes: Feature point matching is performed between the main branches of the initial vascular topology map and the anatomical structures of the standard cardiovascular model. The spatial deformation of vascular branches is corrected by a thin-plate spline transformation algorithm, and the registered vascular curvature field is output. The curvature deviation of the midline of each branch from that of the standard model is calculated in the registered vascular curvature field.

5. The cardiovascular disease detection method based on image analysis according to claim 4, characterized in that, The generation of the vascular elastic feature matrix based on the distribution density of abnormal displacement nodes and the mechanical parameters of vascular branches includes: Measure the periodic contraction amplitude and phase delay of the blood vessel wall around the abnormal displacement node; The local Young's modulus was derived based on the linear relationship between the contraction amplitude and intravascular pressure. The Young's modulus data of all abnormal displacement nodes are integrated to generate a vascular elasticity feature matrix.

6. The cardiovascular disease detection method based on image analysis according to claim 5, characterized in that, The multi-scale fusion of curvature deviation, diameter variation coefficient, and vascular elasticity feature matrix includes: Wavelet decomposition of the vascular structural abnormality index yields high-frequency detail components and low-frequency contour components. Convolution operation is performed between high-frequency detail components and the vascular elasticity feature matrix to enhance the local lesion feature response; The weighting of low-frequency contour components and hemodynamic parameter sets is optimized using the backpropagation algorithm.

7. The cardiovascular disease detection method based on image analysis according to claim 6, characterized in that, The risk zones include: Calculate the Laplacian operator at each vertex of the surface of a three-dimensional vascular model for the set of hemodynamic parameters. When the Laplace operator exceeds the dynamic threshold and the vascular structural abnormality index continues to increase, it is identified as a high-risk lesion area. The high-risk lesion area is expanded using a region growing algorithm until the fluid shear force mutation point at the blood vessel bifurcation is encountered.

8. The cardiovascular disease detection method based on image analysis according to claim 7, characterized in that, The constructed lesion probability prediction matrix includes: Establish a joint probability density function for texture fingerprint encoding and vascular structural abnormality index; Monte Carlo sampling method was used to simulate the spatial distribution patterns of different lesion types; A lesion probability prediction matrix is ​​generated based on the clustering degree of various lesions in the sampling results.

9. The cardiovascular disease detection method based on image analysis according to claim 8, characterized in that, The determination of the type and severity of cardiovascular disease based on the peak distribution of the lesion probability prediction matrix includes: Identify the spatial geometric features of continuous peak regions in the lesion probability prediction matrix; When the peak region exhibits a ring-shaped distribution and the texture fingerprint encoding conforms to calcification characteristics, it is determined to be atherosclerosis; When the peak region exhibits a linear distribution and the vascular elasticity characteristic matrix value is abnormal, it is determined to be a vascular fibrosis lesion.

Images