Dynamic risk early warning method based on multi-phase patch registration and morphological functional parameter extraction of time series difference evolution network

Abstract

The application relates to a dynamic risk early warning method based on a time sequence difference evolution network of multi-phase plaque registration and morphological functional parameter extraction. The method comprises the following steps: acquiring basic body information data, coronary CTA image data and a synchronous electrocardiogram signal; performing phase standardization on each coronary CTA image data based on the synchronous electrocardiogram signal, and extracting a blood vessel morphology from each homophase image data to obtain an anatomical topology graph data; aligning the homophase image data by combining a longitudinal deformation field estimation network and the anatomical topology graph data to obtain aligned image data of each phase; extracting morphological functional parameters of a plaque region of a patient based on the aligned image data of each phase to obtain a morphological functional parameter matrix; extracting a time sequence difference feature vector, and inputting the time sequence difference feature vector and the basic body information data into an early warning time sequence difference evolution network to obtain a high-risk probability. The method can improve the accuracy and timeliness of risk assessment.

Description

Technical Field

[0001] This invention belongs to the field of coronary artery imaging, and in particular relates to a dynamic risk warning method based on a temporal differential evolution network of multi-temporal plaque registration and morphological and functional parameter extraction. Background Technology

[0002] With the development of coronary imaging technology, multiphase computed tomography (CTA) can non-invasively acquire images of coronary plaques at different time points in patients. Traditionally, clinicians or researchers mainly rely on manual or semi-automatic methods to perform qualitative observation and simple quantification of images to assess the risk of plaque progression.

[0003] Traditional methods typically involve first roughly overlapping CTA images from different time points using rigid registration (such as manual alignment based on bony landmarks or vascular branches), then manually delineating plaque boundaries on selected slices, calculating the rate of change in plaque area or volume, and making an empirical judgment based on the patient's clinical risk factors. This process relies on the operator's experience and often ignores anatomical differences caused by cardiac pulsation and the continuity of vascular topology.

[0004] However, traditional methods lack dual constraints on vascular anatomy and topology and cardiac cycle phase in the registration step, making it difficult to eliminate spatial misalignment across temporal phases and resulting in large errors in the extracted morphological and functional parameters. At the same time, relying solely on the rate of change of static parameters cannot capture the dynamic differential patterns of plaque evolution, making the accuracy and timeliness of risk assessment insufficient and failing to meet the clinical need for early warning of high-risk plaques. Summary of the Invention

[0005] Therefore, it is necessary to provide a dynamic risk early warning method based on a temporal differential evolution network with multi-temporal patch registration and morphological function parameter extraction, which can improve the accuracy and timeliness of risk assessment, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a dynamic risk early warning method based on a temporal differential evolution network using multi-temporal patch registration and morphological-functional parameter extraction, including:

[0007] Acquire basic physical information data of the patient, coronary CTA image data at least two time phases, and synchronized electrocardiogram signals;

[0008] Phase standardization of coronary CTA image data based on synchronous electrocardiogram signals is performed to obtain in-phase image data; and vascular morphology is extracted from each in-phase image data to obtain anatomical topology data for each time phase; the anatomical topology data includes at least a set of vascular centerline nodes;

[0009] Using the earliest anatomical topology data as a reference, and combining the longitudinal deformation field estimation network and the anatomical topology data, the in-phase image data of each time phase are aligned to obtain the aligned image data of each time phase.

[0010] Based on aligned image data from different time phases, morphological and functional parameters of the patient's plaque region are extracted to obtain a morphological and functional parameter matrix; wherein, the morphological and functional parameters include at least the plaque volume percentage value.

[0011] The differential features of the morphological function parameter matrix are extracted to obtain the temporal differential feature vector. The temporal differential feature vector and basic body information data are then input into the early warning temporal differential evolution network to obtain the high-risk probability of the patient.

[0012] Secondly, this application also provides a dynamic risk early warning device based on a temporal differential evolution network of multi-temporal patch registration and morphological-functional parameter extraction, comprising:

[0013] The data acquisition module is used to acquire the patient's basic physical information data, coronary CTA image data at least two time phases, and synchronous electrocardiogram signals;

[0014] The anatomical extraction module is used to perform phase normalization on each coronary CTA image data based on synchronous electrocardiogram signals to obtain in-phase image data; and to extract vascular morphology from each in-phase image data to obtain anatomical topology map data for each time phase; the anatomical topology map data includes at least a set of vascular centerline nodes;

[0015] The time-phase alignment module is used to align the in-phase image data of each time phase with the earliest time-phase anatomical topology data as a reference, combined with the longitudinal deformation field estimation network and the anatomical topology data of each time phase, to obtain the aligned image data of each time phase.

[0016] The parameter extraction module is used to extract morphological and functional parameters of the patient's plaque region based on aligned image data at various time phases, and obtain a morphological and functional parameter matrix; wherein, the morphological and functional parameters include at least the plaque volume percentage value;

[0017] The risk analysis module is used to extract the differential features of the morphological and functional parameter matrix to obtain the temporal differential feature vector. The temporal differential feature vector and basic body information data are then input into the early warning temporal differential evolution network to obtain the patient's high-risk probability.

[0018] Thirdly, this application also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements any of the dynamic risk warning methods of the first aspect of this application based on a temporal differential evolution network of multi-temporal patch registration and morphological function parameter extraction.

[0019] Fourthly, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the dynamic risk warning methods of the first aspect of this application based on a temporal differential evolution network of multi-temporal patch registration and morphological-functional parameter extraction.

[0020] The aforementioned dynamic risk warning method based on a temporal differential evolution network derived from multi-temporal plaque registration and morphological-functional parameter extraction acquires basic patient information, coronary CTA image data from at least two temporal phases, and synchronized electrocardiogram (ECG) signals. Based on the synchronized ECG signals, the coronary CTA image data are phase-normalized to obtain in-phase image data. Vascular morphology is extracted from each in-phase image data to obtain anatomical topology data for each temporal phase. The anatomical topology data includes at least a set of vascular centerline nodes. The earliest temporal anatomical topology data is used as a reference. By combining longitudinal deformation field estimation network and anatomical topology data, in-phase image data from different time phases are aligned to obtain aligned image data for each time phase. Based on the aligned image data, morphological and functional parameters of the plaque region in patients are extracted to obtain a morphological and functional parameter matrix. Among them, the morphological and functional parameters include at least the plaque volume percentage value. The differential features of the morphological and functional parameter matrix are extracted to obtain a temporal differential feature vector. The temporal differential feature vector and basic body information data are input into the early warning temporal differential evolution network to obtain the high-risk probability of the patient. Cardiac phase standardization and vascular topology constraints reduce multi-temporal registration errors and ensure spatial consistency of parameter extraction. At the same time, differential features are used to characterize the dynamic evolution pattern, combined with a deep temporal network for risk prediction, effectively improving the accuracy and clinical interpretability of risk assessment. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a dynamic risk warning method based on a temporal differential evolution network derived from multi-temporal patch registration and morphological-functional parameter extraction, provided as an embodiment of this application;

[0023] Figure 2 A schematic diagram of the structure of a dynamic risk warning device based on a temporal differential evolution network of multi-temporal patch registration and morphological function parameter extraction, provided in one embodiment of this application;

[0024] Figure 3 This is a schematic diagram of the structure of a computer device for a dynamic risk warning method based on a temporal differential evolution network for multi-temporal patch registration and morphological-functional parameter extraction, as provided in one embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] In one embodiment, such as Figure 1 As shown, a dynamic risk early warning method based on a temporal differential evolution network derived from multi-temporal patch registration and morphological-functional parameter extraction is provided. This embodiment illustrates the application of this method to an early warning terminal. It is understood that this method can also be applied to a server, or to a system including both an early warning terminal and a server, and is implemented through the interaction between the early warning terminal and the server. In this embodiment, the method includes the following steps S101-S105, wherein:

[0027] S101 acquires basic physical information data of the patient, coronary CTA image data at least two time phases, and synchronized electrocardiogram signals.

[0028] Specifically, the early warning terminal acquires the patient's basic physical information, coronary CTA image data from at least two time phases, and synchronized electrocardiogram signals. The basic physical information characterizes the patient's basic physiological condition, including at least age and gender, and may also include a history of hypertension, diabetes, smoking, etc., and can be represented in numerical vector form. This basic physical information can be exported from an electronic medical record system. The time phase represents the sampling time of the coronary CTA image data. The coronary CTA image data is a three-dimensional voxel image, and each time phase can include multiple coronary CTA image data points. Coronary CTA image data can be denoted as ,in It is the number of voxels representing the image height; It is the number of voxels in the image width; It refers to the number of slices (number of voxel layers). The relative sampling time label of coronary CTA image data is used to characterize the offset from the start time of the synchronous electrocardiogram (ECG) signal acquired at the same time phase, in milliseconds. The CT value of each voxel in the coronary CTA image data is expressed in Hounsfield Units (HU), used to characterize the degree of attenuation of X-rays by the corresponding body tissue. The synchronous ECG signal is a one-dimensional time-series signal recorded synchronously during each coronary CTA image data acquisition, containing multiple R-wave peak times, used to determine the reconstructed phase of each phase image within the cardiac cycle.

[0029] S102, based on the synchronous electrocardiogram signal, the phase of each coronary CTA image data is standardized to obtain in-phase image data; and the vascular morphology is extracted from each in-phase image data to obtain anatomical topology data of each time phase; the anatomical topology data includes at least the set of vascular centerline nodes.

[0030] Specifically, the early warning terminal uses synchronized electrocardiogram signals to process coronary CTA image data from multiple different phases at each time stage, selecting or interpolating to generate coronary CTA image data under a unified phase. This results in coronary CTA image data where all time stages are at the same cardiac phase, referred to as in-phase image data, which can be denoted as... ,in, It is any phase of time. It is the number of voxels representing the image height; It is the number of voxels in the image width; This refers to the number of slices (voxel layers), where each voxel value is a CT value expressed in Henle units. Based on this, the warning terminal extracts vascular morphology from the in-phase image data of each time phase, obtaining anatomical topology data for each phase. The anatomical topology data can be a graphical representation of the coronary artery vascular tree structure, which can be denoted as... ,in It is a set of nodes along the vascular centerline, where each node represents the three-dimensional spatial coordinates (in mm) of a sampling point along the vascular centerline. The anatomical topology data may also include a set of connected edges and a set of local radii. The set of connected edges represents the topological connections between the nodes along the vascular centerline, while the set of local radii represents the lumen radius (in mm) perpendicular to the vascular direction at the corresponding location.

[0031] S103, using the earliest anatomical topology data as a reference, and combining the longitudinal deformation field estimation network and the anatomical topology data, aligns the in-phase image data of each time phase to obtain aligned image data of each time phase.

[0032] Specifically, the early warning terminal uses the earliest time-phase anatomical topology map data as a fixed reference and combines it with a longitudinal deformation field estimation network to spatially align the in-phase image data of each subsequent time phase, obtaining aligned image data for each time phase. The longitudinal deformation field estimation network is introduced during the alignment process to eliminate non-rigid misalignment caused by complex factors. The longitudinal deformation field estimation network is a pre-trained deep learning model. Its input is a pair of in-phase images from two adjacent time phases after rigid coarse registration, and its output is a three-dimensional displacement field describing local non-rigid deformation, used to compensate for residual distortions that cannot be corrected by rigid transformation. For example, the early warning terminal can use a graph matching algorithm between the anatomical topology map data to solve the rigid transformation matrix, perform a rigid transformation on the in-phase images of subsequent time phases to obtain coarsely registered images; then, input the coarsely registered image pairs from adjacent time phases into the longitudinal deformation field estimation network to obtain the local non-rigid displacement field; finally, combine the rigid transformation and the displacement field to transform the coarsely registered images, obtaining aligned image data for each time phase. Aligning image data unifies the spatial coordinates of all phases to the reference space of the earliest phase, enabling precise correspondence between the positions of the same anatomical voxel at different time points.

[0033] S104. Based on the aligned image data of each time phase, extract the morphological and functional parameters of the patient's plaque region to obtain a morphological and functional parameter matrix; wherein, the morphological and functional parameters include at least the plaque volume percentage value.

[0034] Specifically, the early warning terminal segments the aligned image data for each time phase, separating the lumen region, wall region, and plaque region of the coronary artery. Based on the segmentation results, it extracts morphological and functional parameters of the plaque region, obtaining a morphological and functional parameter matrix. These parameters include at least the plaque volume percentage. The plaque region is defined as a sub-region within the vessel wall that has undergone pathological thickening or compositional changes. Based on the segmentation results, the plaque volume percentage is defined as the percentage of plaque volume to the total vessel volume, i.e., the sum of the lumen region volume and the wall region volume. The morphological and functional parameter matrix can be represented as follows: ,in The total number of phases, The Line corresponding to time phase Morphological and functional parameters , The morphological-functional parameters include the number of parameters, and The morphological-functional parameter matrix characterizes the geometric and biological state of plaques at different time phases. For example, morphological-functional parameters may also include parameters such as the percentage of non-calcified plaque volume, lipid core loading index, and minimum fibrous cap thickness.

[0035] S105, extract the differential features of the morphological function parameter matrix to obtain the temporal differential feature vector, and input the temporal differential feature vector and basic body information data into the early warning temporal differential evolution network to obtain the patient's high-risk probability.

[0036] Specifically, the early warning terminal extracts temporal difference features from the morphological and functional parameter matrix to obtain a temporal difference feature vector. This temporal difference feature vector prominently expresses the dynamic evolution trend of plaques, eliminating static differences between patients. The early warning terminal concatenates the temporal difference feature vector with basic physical information data and inputs it into a pre-trained early warning temporal difference evolution network to obtain the patient's high-risk probability for early warning. The high-risk probability can be denoted as... This is used to characterize the probability of a patient experiencing high-risk events within a pre-defined time period, such as rapid plaque progression, high-risk phenotypic transformation, or major adverse cardiovascular events. Differential features can include absolute and relative differential features, representing the magnitude and rate of change of parameters, respectively. The early warning temporal differential evolution network is used to capture the long-range dependencies of temporal differential feature vectors. Specifically, it can be constructed using a temporal model such as a Transformer encoder and trained using historical data.

[0037] This embodiment provides a dynamic risk warning method based on a temporal differential evolution network for multi-temporal plaque registration and morphological-functional parameter extraction. By acquiring multi-temporal coronary CTA image data, synchronized electrocardiogram (ECG) signals, and basic body information data, the method first unifies the images from different temporal phases to the same cardiac phase using ECG information to obtain in-phase image data. Then, vascular anatomical topology is extracted from the in-phase image data to obtain anatomical topology data as geometric constraints. Combined with a pre-trained longitudinal deformation field estimation network, the in-phase image data is precisely aligned to the baseline space to obtain aligned image data. Next, the morphological-functional parameter matrix of the plaque is extracted from the aligned image data, and the temporal differential features of the parameter matrix are calculated to obtain a temporal differential feature vector. Finally, the temporal differential feature vector and basic body information data are input into the warning temporal differential evolution network to obtain a high-risk probability. Cardiac phase standardization and vascular topology constraints reduce multi-temporal registration errors and ensure spatial consistency of parameter extraction. Simultaneously, the use of differential features to characterize dynamic evolution patterns, combined with a deep temporal network for risk prediction, effectively improves the accuracy and clinical interpretability of risk assessment.

[0038] In one embodiment, the coronary CTA image data carries a relative sampling time label. Phase normalization is performed on each coronary CTA image data based on synchronized electrocardiogram signals to obtain in-phase image data, including:

[0039] S201: For each time phase, extract the peak time from the synchronized electrocardiogram signal, and calculate the reconstructed phase label of each coronary CTA image data based on the peak time and the relative sampling time label.

[0040] Specifically, for each phase of the synchronized electrocardiogram (ECG) signal, the warning terminal uses a peak detection algorithm to extract the times when all R-wave peaks occur. The time between two R-wave peak occurrences represents one cardiac cycle. The warning terminal calculates the actual cardiac cycle length based on the extracted peak times and divides the synchronized ECG signal into several cardiac cycles. For coronary CTA image data with different relative sampling time labels in the same phase, the warning terminal determines the cardiac cycle of the coronary CTA image data using its relative sampling time label, the start time of the synchronized ECG signal, and the cardiac cycle length. It also calculates the offset time of the relative sampling time label relative to the start peak time of the cardiac cycle, divides this offset time by the cardiac cycle length, and obtains a more accurate reconstructed phase label, which can be denoted as... ,in For phase index, It is the relative sampling time label of coronary CTA image data.

[0041] S202, for each time phase, calculate the absolute value of the difference between the reconstructed phase label and the preset target diastolic phase to obtain the phase difference value of each reconstructed phase label; and determine the reconstructed phase label corresponding to the smallest phase difference value as the original reference phase label of the time phase.

[0042] Specifically, the preset target diastolic phase is a standardized value for phase uniformity, used to indicate which phase of the cardiac cycle each in-phase image data represents. It can be set according to cardiac motion and image clarity, and can be denoted as... The default setting is 75%. For coronary CTA image data with different reconstructed phase labels for each time phase, the warning terminal calculates the absolute value of the difference between its reconstructed phase label and the target diastolic phase, obtaining the phase difference value for each reconstructed phase label. The reconstructed phase label corresponding to the smallest phase difference value is determined as the original reference phase label for that time phase. Specifically, for any given time phase... Its original reference phase label can be denoted as .

[0043] S203, for each time phase, compare the phase difference value of the original reference phase label with the preset phase difference threshold to obtain the phase change comparison result; when the phase change comparison result is that the phase difference value is less than the phase difference threshold, the coronary CTA image data of the selected time phase is determined as the same phase image data of the time phase.

[0044] Specifically, the preset phase difference threshold is a standard used to determine whether the original reference phase label and the target diastolic phase are close. The default setting is 2%, but it can be adjusted according to actual work requirements. For each time phase, the warning terminal compares the phase difference value of its original reference phase label with the preset phase difference threshold to obtain the phase change comparison result for that time phase. When the phase difference comparison result is less than the phase difference threshold, it indicates that the difference between the existing closest image for that time phase and the target diastolic phase is sufficiently small, requiring no additional processing. The warning terminal directly selects the coronary CTA image data corresponding to the original reference phase label as the in-phase image data for that time phase.

[0045] S204, or when the phase change comparison result is that the phase difference is greater than or equal to the phase difference threshold, linear interpolation is performed on the coronary CTA image data of two adjacent reconstructed phase labels of the target diastolic phase to obtain the same-phase image data of the time phase.

[0046] Specifically, for each time phase, when the phase difference between the original reference phase label and the phase difference threshold is greater than or equal to the phase difference threshold, it indicates that the image corresponding to the original reference phase label has a large phase difference with the target's diastolic phase. The warning terminal selects two reconstructed phase labels in that time phase that are located on either side of the original reference phase label and are closest to it. These can be denoted as follows: and (satisfy (and acquire the corresponding coronary CTA images) and Then, an image of the target phase is generated according to the linear interpolation formula: interpolation coefficients To obtain that time phase In-phase image data .

[0047] This embodiment provides a dynamic risk warning method based on a temporal differential evolution network using multi-temporal patch registration and morphological-functional parameter extraction. It extracts peak time from synchronized electrocardiogram signals and combines this with relative sampling time labels to obtain reconstructed phase labels for each image. The difference between each reconstructed phase label in each temporal phase and a preset target diastolic phase is calculated. The label corresponding to the smallest difference is selected as the original reference phase label and compared with a preset threshold. When the difference is less than the threshold, the original image is used directly; when the difference is greater than or equal to the threshold, the target phase image is generated by linear interpolation using images of two adjacent reconstructed phase labels. This accurately unifies all temporal images to the same cardiac diastolic phase. This effectively eliminates geometric differences caused by inconsistent cardiac pulsation phases across different temporal phases, providing a phase consistency basis for subsequent cross-temporal spatial alignment and parameter extraction, reducing registration errors, and improving the accuracy of dynamic risk warning.

[0048] In one embodiment, vascular morphology is extracted from each in-phase image data to obtain anatomical topology data for each time phase, including:

[0049] S301, input the in-phase image data into the preset blood vessel segmentation network to obtain the blood vessel centerline probability map and blood vessel radius regression map for each time phase.

[0050] Specifically, the early warning terminal will display the in-phase image data of each time phase. The data is input into a pre-defined vessel segmentation network to obtain a vessel centerline probability map and a vessel radius regression map for that time phase. The pre-defined vessel segmentation network is a segmentation model based on a 3D convolutional neural network, which can employ a 3D U-Net or V-Net architecture. It is trained using a large amount of historical coronary CTA image data labeled with vessel centerlines and radii, and can employ a joint loss function, including cross-entropy loss for centerline segmentation and mean squared error loss for radius regression, optimized through backpropagation. Any time phase The probability diagram of the vascular centerline can be denoted as: The value of each voxel is a probability value for the vessel midline, representing the probability that the voxel belongs to the vessel midline, with a value range of [value range missing]. Where 0 indicates not belonging, and 1 indicates absolutely belonging; the blood vessel radius regression diagram can be denoted as... The value of each voxel is its local radius, which represents the local radius perpendicular to the blood vessel at the voxel's location, in mm.

[0051] S302, non-maximum suppression is applied to the vascular centerline probability map to obtain the candidate centerline probability map.

[0052] Specifically, for any time phase Probability map of the center line of blood vessels The early warning terminal uses a non-maximum suppression algorithm to eliminate fuzzy responses in the vascular centerline probability map, refining the broad high-probability region into candidate centerlines with a single voxel width, thus obtaining the candidate centerline probability map. Specifically, for the probability map of the vascular centerline. For each voxel in the algorithm, its vessel centerline probability value is compared with the vessel centerline probability values ​​of all voxels in its 3D spatial neighborhood (e.g., 26 neighborhoods). Only if the voxel's vessel centerline probability value is the maximum value in its 3D spatial neighborhood is its vessel centerline probability value retained; otherwise, its vessel centerline probability value is set to zero. After non-maximum suppression, a candidate centerline probability map is obtained. The candidate centerline probability plot retains only local extreme points and removes redundant responses.

[0053] S303, skeletonize the candidate centerline probability graph to obtain the set of blood vessel centerline nodes and the set of node connected edges.

[0054] Specifically, for any time phase Probability plot of candidate center lines The early warning terminal first performs binarization processing. Specifically, voxels with a vessel centerline probability value greater than 0.5 are assigned a vessel centerline probability value of 1, while the rest are assigned a value of 0, resulting in a binary candidate centerline probability map. The candidate centerlines in this binary candidate centerline probability map may still have a certain width, meaning that the vessel centerline probability values ​​of multiple adjacent voxels are simultaneously marked as 1. Therefore, the early warning terminal performs a morphological skeletonization algorithm on the binary candidate centerline probability map to gradually peel away the outer voxels until a centerline skeleton of the width of a single voxel is obtained. Specifically, a distance-transform-based thinning algorithm can be used. This algorithm first calculates the Euclidean distance from each foreground voxel (i.e., voxels with a vessel centerline probability of 1) in the binary candidate centerline probability map to the nearest background voxel (i.e., voxels with a vessel centerline probability of 0), resulting in a distance-transform map. Voxels with larger distance values ​​are closer to the ridge line in the central region. Then, iteratively, peripheral voxels, i.e., edge voxels with smaller Euclidean distance values, are stripped while maintaining topological connectivity and not deleting key voxels that would break connections. Finally, only the ridge line voxels with the largest local distance values ​​are retained, forming the skeleton of each voxel's width. After skeletonization, each retained voxel is a node on the vessel centerline, and its spatial coordinates can be obtained by multiplying the voxel's index in the 3D mesh by the voxel's physical spacing, in mm. These spatial coordinates are then integrated to form a set of vessel centerline nodes. ,in Indicates the first The three-dimensional coordinates of each blood vessel centerline node are determined. Simultaneously, the spatial adjacency relationships between the blood vessel centerline nodes are traversed. For example, if two blood vessel centerline nodes are adjacent within a 26-neighborhood, it indicates a connection between them, thus obtaining the set of node-connected edges. Each edge Indicates the nodes of the blood vessel centerline and the central line node of blood vessels They are directly connected in terms of vascular topology.

[0055] S304: Extract the vascular radius regression map based on the vascular centerline node set to obtain the node local radius set; Based on the vascular centerline node set, the node connected edge set, and the node local radius set, anatomical topology map data for each time phase is formed.

[0056] Specifically, for any time phase The early warning terminal utilizes the set of nodes along the centerline of blood vessels. Regression plot of blood vessel radius Extract the local radius value of the voxel position corresponding to each vessel centerline node, and the local radius values ​​of all vessel centerline nodes constitute the node local radius set. ,in It is the first The local radius values ​​of each blood vessel centerline node are determined. Finally, the early warning terminal sets up the blood vessel centerline nodes. Node connected edge set and the set of local radii of nodes Combined to form this phase The anatomical topology data can be denoted as... This graph data structure fully describes the geometric skeleton (node ​​coordinates), topological connections (edges), and local diameter information (radius) of the coronary artery tree, providing anatomical constraints for subsequent cross-temporal spatial alignment.

[0057] This embodiment provides a dynamic risk warning method based on a temporal differential evolution network for multi-temporal patch registration and morphological-functional parameter extraction. The method involves inputting in-phase image data into a preset vascular segmentation network to obtain a vascular centerline probability map and a vascular radius regression map. Non-maximum suppression and skeletonization are applied to the vascular centerline probability map to obtain a set of vascular centerline nodes and a set of node-connected edges. Then, the local radius values ​​of each vascular centerline node are extracted from the vascular radius regression map to obtain a set of node local radii. Finally, these are combined to form an anatomical topology map containing the vascular centerline node set, the node-connected edge set, and the node local radius set. This achieves automated and high-precision conversion from 3D CTA images to a vascular treemap structure. The extracted anatomical topology map preserves the geometric morphology and topological connectivity of the vessels, providing reliable input data for subsequent coarse registration based on graph matching, effectively improving the accuracy and robustness of multi-temporal image registration.

[0058] In one embodiment, using the earliest temporal anatomical topology data as a reference, and combining the longitudinal deformation field estimation network and each anatomical topology data, the in-phase image data of each temporal phase are aligned to obtain aligned image data of each temporal phase, including:

[0059] S401, using the earliest anatomical topology data as a reference, rigidly registers the anatomical topology data of subsequent time phases to obtain the rigid transformation matrix of each time phase.

[0060] Specifically, the early warning terminal will be the earliest time phase, i.e. Anatomical topology data As a fixed reference diagram. For each subsequent time phase Dissect its topology data and Graph matching is performed. Specifically, a graph matching algorithm can be used for registration. This algorithm treats each vessel centerline node as a point cloud in three-dimensional space and uses the coordinates and local radius values ​​of these nodes as features. The Hungarian algorithm is then used to find the most likely node correspondence between the two graphs. Based on the successfully matched node pairs, the rigid transformation matrix for that phase is obtained, which can be denoted as... The rigid transformation matrix contains rotational components. Translation components This minimizes the mean square error between the coordinates of the vessel centerline nodes in subsequent time phases after transformation and the coordinates of the corresponding vessel centerline nodes in the reference image.

[0061] S402, starting from the second time phase, transform the in-phase image data based on the rigid transformation matrix to obtain coarse registration image data for each time phase.

[0062] Specifically, starting from the second phase, for each subsequent phase... The early warning terminal transmits its in-phase image data As the original image, its rigid transformation matrix is... inverse matrix This process operates on the voxel coordinates of the original image, mapping each voxel in the original image to a target space referenced by the earliest temporal phase space. Specifically, a trilinear interpolation algorithm can be used to estimate the CT value at each integer coordinate location in the target space. For any integer coordinate point in the target space... The early warning terminal first calculates the floating-point coordinates of the point in the original image through the inverse transformation of the rigid transformation matrix. Then, find the area surrounding the original image. Eight adjacent integer voxels (forming a cubic grid) are respectively in Linear interpolation is performed in three directions, and the weighted average is used to calculate the value at that coordinate point. This process is repeated for all target spatial coordinate points to obtain the transformed 3D image, called the coarse registration image data, denoted as . The size and resolution of the target space are set to be consistent with the in-phase image data of the earliest time phase.

[0063] S403, starting from the second time phase, input the coarse registration image data of the time phase and the coarse registration image data of the previous time phase into the longitudinal deformation field estimation network to obtain the local non-rigid displacement field data of the time phase.

[0064] Specifically, from the second phase Initially, the early warning terminal processes adjacent time phases sequentially. For each time phase... Compared to its previous period The early warning terminal coarsely registered its image data and The data are input into the longitudinal deformation field estimation network to obtain the local non-rigid displacement field data for that time phase. Local non-rigid displacement field data This includes every voxel in the coarse-registered image data. 3D offset vector , indicating the phase The voxel moves to the phase in the coarsely registered image data. The required positional offset for each voxel. The longitudinal deformation field estimation network is a deformable registration model based on a 3D convolutional neural network, which can be trained using historical longitudinal coronary CTA image data from patients.

[0065] S404, for the earliest time phase, the in-phase image data of the time phase is determined as the aligned image data of the time phase; starting from the second time phase, the coarse registration image is transformed based on the local non-rigid displacement field data and the rigid transformation matrix to obtain the aligned image data of each time phase.

[0066] Specifically, for the earliest phase The early warning terminal directly sets the aligned image data to aligned image data. For each subsequent time phase... The early warning terminal first utilizes a rigid transformation matrix and local non-rigid displacement field Through displacement field accumulation and composite transformation, the time phase is generated. to reference phase The complete deformation mapping is obtained. Specifically, the rigid transformation matrix is ​​concatenated with the local non-rigid displacement fields between adjacent time phases to obtain the final transformation. ,for Then, directly combine the rigid transformation and the displacement field. Then, coarsely register their images. Applying this transformation, the final aligned image data for that phase can be obtained through trilinear interpolation resampling. .

[0067] The longitudinal deformation field estimation network includes:

[0068] The input layer receives coarse registration image data of the current time phase and coarse registration image data of the previous time phase, concatenates the coarse registration image data of the current time phase and coarse registration image data of the previous time phase to obtain a three-dimensional input vector, and transmits the three-dimensional input vector to the feature coding layer.

[0069] Specifically, the input layer receives coarsely registered image data of the previous time phase and coarsely registered image data of the previous time phase, which can be respectively recorded as fixed images. and moving images The input layer concatenates the two along the channel dimension to form a three-dimensional input vector, which is then fed into the feature encoding layer.

[0070] The feature encoding layer is used to downsample the three-dimensional input vector through multiple three-dimensional convolutional kernels, extract multi-scale spatial features, obtain a compressed feature map, and transmit the compressed feature map to the displacement field decoding layer.

[0071] Specifically, the feature encoding layer consists of multiple downsampling modules stacked sequentially. Each downsampling module contains a 3D convolutional layer, a normalization layer, a ReLU activation layer, and a max-pooling layer connected in sequence. The max-pooling layer halves the spatial resolution of the input feature map, i.e., the height, width, and number of slices are all reduced by half, while doubling the number of feature channels, for example, increasing from 32 channels to 64, 128, and 256 channels progressively. The first downsampling module receives the 3D input vector, processes it to obtain the first compressed feature map, and inputs the first compressed feature map into the second downsampling module, until the last downsampling module outputs the final compressed feature map. The compressed feature maps output by all downsampling modules form a compressed feature map set, which is transmitted to the displacement field decoding layer. Let the feature encoding layer have a total of The downsampling module then passes through the first... After one module, the spatial resolution of the output feature map is equal to that of the input coarsely registered image data. The number of channels is 1 / 3 of the initial number of channels. The compressed feature maps output by each downsampling module constitute a multi-scale feature map set, denoted as . ,in Corresponding to the Compressed feature maps output by each downsampling module.

[0072] The displacement field decoding layer is used to upsample the compressed feature map to obtain local non-rigid displacement field data of the time phase.

[0073] Specifically, the displacement field decoding layer consists of multiple upsampling modules stacked sequentially, with the number of upsampling modules being the same as the number of downsampling modules in the feature encoding layer. Each upsampling module contains an upsampling layer, a 3D convolutional layer, a normalization layer, and a ReLU activation layer connected in sequence. The upsampling layer doubles the spatial resolution of the input feature map, i.e., doubling the height, width, and number of slices, while halving the number of feature channels. Within each upsampling module, the decoding layer also receives compressed feature maps with the same spatial resolution from the feature encoding layer via skip connections. (in This is the sequence number of the current upsampling module. (Corresponding to the first level of upsampling). Specifically, for the first level of upsampling... Each upsampling module first upsamples the feature map output by the previous module to obtain a feature map with doubled spatial resolution; then, it compares this feature map with the corresponding compressed feature map passed through the skip connections. The data is stitched together along the channel dimension; then, feature fusion is performed sequentially through a 3D convolutional layer, a normalization layer, and a ReLU activation layer. Through progressive upsampling and feature fusion, the spatial resolution of the feature map output by the decoding layer is restored to the same level as the coarsely registered image data. The last upsampling module of the displacement field decoding layer outputs local non-rigid displacement field data, which can be denoted as... The aforementioned longitudinal deformation field estimation network is trained as follows: A pre-defined registration training dataset is acquired, containing historical longitudinal CTA image pairs from multiple patients. Each pair of upsampling modules originates from different sampling time points of the same patient and has undergone phase normalization and rigid coarse registration, serving as fixed and moving images, respectively. An initial network with the input layer, feature encoding layer, and displacement field decoding layer is constructed, and the weight parameters of each convolutional layer are randomly initialized. Image pairs from the training dataset are input into the network, and the predicted displacement field is obtained through forward propagation. The moving image is then distorted according to the predicted displacement field using a spatial transformation layer to obtain the deformed moving image. Based on the similarity between the fixed image and the deformed moving image, and the smoothness regularization term of the predicted displacement field, an unsupervised loss function is calculated. With the goal of minimizing the loss, the convolutional kernel weight parameters in the network are iteratively updated using a backpropagation algorithm and an optimizer until the loss converges or the preset number of iterations is reached, resulting in the trained longitudinal deformation field estimation network.

[0074] This embodiment provides a dynamic risk warning method based on a temporal differential evolution network for multi-temporal patch registration and morphological functional parameter extraction. It obtains a rigid transformation matrix through graph matching rigid registration using the earliest temporal anatomical topology map as a reference. The rigid matrix is ​​then used to transform in-phase images to obtain coarse-registered images. Adjacent coarse-registered images are input into a longitudinal deformation field estimation network to obtain local non-rigid displacement fields. Finally, the earliest temporal image is directly assigned values, and subsequent temporal images are obtained by combining rigid transformation and displacement fields. The longitudinal deformation field estimation network employs an end-to-end deformable registration architecture with input layer stitching, feature encoding layer downsampling to extract multi-scale features, and displacement field decoding layer upsampling to output the displacement field. This method significantly improves the cross-temporal registration accuracy of multi-temporal coronary CTA images through the combined effect of anatomical topology constraints and a deep deformation field network. It effectively compensates for local non-rigid deformations that rigid transformation cannot correct, ensuring spatial consistency and quantitative accuracy of subsequent morphological functional parameter extraction, thereby reducing the risk of false positives due to registration errors.

[0075] In one embodiment, based on aligned image data from each time phase, morphological functional parameters of the patient's plaque region are extracted to obtain a morphological functional parameter matrix, including:

[0076] S501, for each time phase, based on the preset lumen-wall separation model, the aligned image data is segmented to obtain the lumen mask data and wall mask data for that time phase; in the lumen mask data, voxels with a value of 1 represent the lumen, and voxels with a value of 1 represent the wall.

[0077] Specifically, for each time phase The early warning terminal aligns it with the image data. The data is input into a preset lumen-wall separation model to obtain the lumen mask data for that time phase. and pipe wall mask data The pre-defined lumen-wall separation model can employ a 3D U-Net architecture and be trained on a large number of historical coronary CTA images with labeled lumen and wall regions. Lumen mask data. A voxel with a median value of 1 indicates that it belongs to the coronary artery lumen (blood flow region), while a value of 0 indicates that it does not; vessel wall mask data. A voxel with a median value of 1 indicates that it belongs to the coronary artery wall (including normal wall tissue and plaque), while a value of 0 indicates that it does not belong to it. The segmentation results of the lumen and wall provide a spatial localization basis for subsequent plaque region identification and parameter calculation.

[0078] S502, for each time phase, based on a preset patch CT value threshold range, the tube wall mask data is extracted to obtain the patch region mask data for that time phase. Based on the patch region mask data, tube wall mask data, and lumen mask data, the patch volume percentage value for each time phase is calculated. In the patch region mask data, voxels with a value of 1 represent patch regions. The expression for the patch volume percentage value is:

[0079] ;

[0080] In the formula, It is any phase of time. That is the phase The percentage value of plaque volume, That is the phase The number of voxels with a value of 1 in the patch region mask data. That is the phase The number of voxels with a value of 1 in the lumen mask data. That is the phase The number of voxels with a value of 1 in the lumen mask data.

[0081] Specifically, the preset plaque CT value threshold range is used to characterize the CT values ​​of voxels belonging to the plaque region. This range can be set according to actual work requirements, typically from -30 HU to 130 HU. This plaque CT value threshold range covers the typical CT value range for non-calcified plaques and lipid cores. For each time phase... The early warning terminal filters pipe wall mask data using a preset threshold range for patch CT values. Specifically, for pipe wall mask data... For each voxel with a median value of 1, the early warning terminal checks its alignment with the image data. If the corresponding CT value falls within the threshold range, then the voxel is marked as a patch region, forming patch region mask data. Voxels with a value of 1 represent plaque regions. Then, the plaque volume percentage is calculated. It is defined as the percentage of the total plaque volume to the total blood vessel volume (the sum of the lumen volume and the vessel wall volume). In the formula, It represents the number of voxels with a value of 1 in the patch area mask data, indicating the number of voxels in the patch; It represents the number of voxels with a value of 1 in the pipe wall mask data; This is the number of voxels with a value of 1 in the lumen mask data. This is due to the physical volume of all voxels. The same applies to the numerator and denominator in the actual calculation definition. Since the volume ratio can be directly expressed using voxel counts, and then multiplied by 100% to obtain the percentage, we get the plaque volume percentage value. The plaque volume percentage value can quantitatively reflect the degree of plaque occupancy in blood vessels.

[0082] S503, for each time phase, based on a preset threshold range for non-calcified plaque CT values, the tube wall mask data is extracted to obtain the non-calcified plaque region mask data for that time phase. Based on the non-calcified plaque region mask data, tube wall mask data, and lumen mask data, the percentage of non-calcified plaque volume for each time phase is calculated. In the non-calcified plaque region mask data, voxels with a value of 1 represent non-calcified plaques. The expression for the percentage of non-calcified plaque volume is:

[0083] ;

[0084] In the formula, It is any phase of time. That is the phase The percentage of non-calcified plaque volume. That is the phase The number of voxels with a value of 1 in the mask data for non-calcified plaque regions. That is the phase The number of voxels with a value of 1 in the lumen mask data. That is the phase The number of voxels with a value of 1 in the lumen mask data.

[0085] Specifically, the preset CT value threshold range for non-calcified plaques is used to characterize the CT values ​​of voxels belonging to non-calcified plaque regions. This range can be set according to actual work requirements, typically between 30 HU and 130 HU, but can also be adjusted based on specific needs. For each time phase... The early warning terminal filters pipe wall mask data using a preset threshold range for non-calcified plaque CT values. Specifically, for pipe wall mask data... For each voxel with a median value of 1, the early warning terminal checks its alignment with the image data. If the corresponding CT value falls within the threshold range, then the voxel is marked as a non-calcified plaque region, forming the non-calcified plaque region mask data. Voxels with a value of 1 represent non-calcified plaques. The volume percentage of non-calcified plaques is then calculated. It is defined as the percentage of non-calcified plaque volume in the total vascular volume. In the formula, This is the number of voxels with a value of 1 in the mask data of non-calcified plaque areas, with the same denominator as in the PAV formula (voxel count of total vascular volume). This parameter is used to assess the proportion of non-calcified components in plaques, which are generally considered to have a higher risk.

[0086] S504: For each time phase, based on a preset lipid value threshold range, lipid region segmentation is performed on the aligned image data to obtain lipid region mask data.

[0087] Specifically, the preset lipid value threshold range is used to characterize the CT values ​​of voxels belonging to the lipid region. This range can be set according to actual work requirements, typically from -30 HU to 30 HU, but can also be adjusted based on specific needs. For each time phase... The early warning terminal filters pipe wall mask data using a preset lipid value threshold range. Specifically, for pipe wall mask data... For each voxel with a median value of 1, the early warning terminal checks its alignment with the image data. If the corresponding CT value falls within the threshold range, then the voxel is marked as the lipid core region, forming the lipid region mask data. A voxel with a value of 1 indicates that it belongs to the lipid core.

[0088] S505: For each time phase, based on the set of vascular centerline nodes and the preset segmentation length, the aligned image data is segmented to obtain segmentation window data. For each segmentation window data, based on lipid region mask data and plaque region mask data, the lipid volume ratio data of each segmentation window data of the time phase is calculated.

[0089] Specifically, the preset segmentation length can be set according to the resolution of coronary CTA image data in actual work, with a default setting of 4mm. For each time phase... The early warning terminal first analyzes the topology map data. Obtain the set of nodes of the blood vessel centerline. The warning terminal uses a preset segmentation length as the window width and slides along the vessel centerline nodes with a fixed step size, such as 1 mm, to crop the data of each window from the aligned image data, resulting in a series of segmented window data. For each segmented window data, the warning terminal statistically analyzes the mask data belonging to the lipid region within that segmented window data. The number of voxels with a median value of 1 and mask data belonging to patch regions. The number of voxels with a median value of 1 is counted, and the ratio between the two is calculated to obtain the lipid volume percentage data for that segmentation window. Each voxel in the lipid volume percentage data is a real number between 0 and 1, and the lipid volume percentage data reflects the abundance of the lipid core within a local vascular segment.

[0090] S506: For each time phase, the lipid core load index of the phase is calculated based on the maximum lipid volume percentage data of the phase. Based on the plaque volume percentage value, the non-calcified plaque volume percentage value and the lipid core load index, the morphological and functional parameters of the phase are composed, and a morphological and functional parameter matrix is ​​composed based on the morphological and functional parameters of each phase.

[0091] Specifically, for each time phase The early warning terminal selects the maximum value from the lipid volume percentage data of all segmented windows, multiplies this maximum value by 1000, and obtains the lipid core load index for that phase, which can be denoted as... Multiplying by 1000 aims to align the numerical range with the standard range of lipid core loading index measured by near-infrared spectroscopy, typically between 0 and 1000 for maxLCBI4mm. Then, the plaque volume percentage, non-calcified plaque volume percentage, and lipid core loading index for each phase are arranged in a fixed order to form a morphological-functional parameter for that phase. The morphological-functional parameters of all phases are stacked in chronological order to obtain the morphological-functional parameter matrix. ,in The total number of time phases, the matrix's th phase Line corresponding to time phase Morphological and functional parameters.

[0092] This embodiment provides a dynamic risk warning method based on a temporal differential evolution network using multi-temporal plaque registration and morphological-functional parameter extraction. The method segments the lumen and wall masks from aligned images using a pre-defined lumen-wall separation model. Plaque regions are extracted using plaque CT value threshold ranges, and plaque volume percentages are calculated. Similarly, non-calcified plaque regions are extracted using non-calcified plaque CT value threshold ranges, and non-calcified plaque volume percentages are calculated. A lipid core mask is extracted using lipid value threshold ranges. A sliding window along the vessel centerline is then used to calculate the lipid volume percentage for each window, and the maximum value is taken as the lipid core load index. Finally, these three parameters are combined into a morphological-functional parameter matrix. Precise voxel-level segmentation and quantitative calculations are performed in the aligned space, ensuring spatial consistency and repeatability of parameter extraction. The extracted morphological and functional parameters reflect plaque load, composition, and high-risk characteristics from different dimensions, providing high-quality input data for subsequent differential feature construction and dynamic risk warning.

[0093] In one embodiment, the difference features of the morphological function parameter matrix are extracted to obtain a temporal difference feature vector, including:

[0094] S601, based on the morphological function parameter matrix, starting from the second phase, calculates the difference between the morphological function parameters of the previous phase and the morphological function parameters of the previous phase to obtain the absolute difference value of the phase.

[0095] Specifically, starting from the second time phase, the early warning terminal calculates the morphological and functional parameters of that time phase in the morphological and functional parameter matrix. Morphological and functional parameters of the previous time phase The difference is used to obtain the absolute difference value of the time phase. Specifically, let... Indicates the first The morphological functional parameters of the first phase Item parameter value ( ; ), At least the plaque volume percentage value should be included. The warning terminal starts from the second time phase ( Starting with the morphological-functional parameters, the absolute difference between each parameter in each time phase and the same parameter in the previous time phase is calculated sequentially, thus obtaining the absolute difference of that parameter in the morphological-functional parameters for that time phase. The absolute difference value of each parameter in the same time phase is formed by splicing together the absolute differences of that time phase. .

[0096] S602, based on the morphological function parameter matrix, starting from the second time phase, adds the morphological function parameters of the previous time phase to the preset non-zero coefficients to obtain the relative difference reference value of the previous time phase, and divides the absolute difference value of the time phase to the relative difference reference value of the previous time phase to obtain the relative difference value of the time phase.

[0097] Specifically, the preset non-zero coefficient can be a very small positive number. Its function is to prevent division by zero errors, and it can be denoted as... Starting from the second phase, for any given phase... The early warning terminal uses the morphological and functional parameters of the previous time phase. The relative difference reference value of the previous time phase is obtained by adding the preset non-zero coefficients, and then the relative difference value of the time phase is obtained by dividing the absolute difference value of the time phase by the relative difference reference value of the previous time phase. Specifically, for any time phase... ,set up Indicates the first The morphological functional parameters of the first phase Item parameter value ( ; ), At least a percentage of patch volume should be included. The early warning terminal calculates the... The morphological functional parameters of the first phase The sum of the parameter values ​​and the preset non-zero coefficients yields the first... The morphological functional parameters of the first phase Baseline value of the parameter The relative difference reference value for each parameter in the same time phase is formed by piecing together the reference values ​​of various parameters in that time phase. The absolute difference value of that parameter is then extracted from the absolute difference value. Through the formula: Calculated time phase The first morphological function parameter Relative difference of parameter values The relative differences of various parameters at the same time phase are spliced ​​together to form the relative difference value of that time phase. .

[0098] S603, based on the absolute difference value and the relative difference value of each time phase, forms a time-series difference feature vector.

[0099] Specifically, the early warning terminal concatenates the absolute differences and relative differences of all time phases in a preset order to obtain a time-series difference feature vector, which can be denoted as... For example, the preset order can be the order of time phases from first to last.

[0100] This embodiment provides a dynamic risk warning method based on a temporal differential evolution network (TDER) using multi-temporal plaque registration and morphological-functional parameter extraction. The method calculates the absolute and relative differences between adjacent temporal phases from the morphological-functional parameter matrix and combines them into a temporal differential feature vector. This transforms static morphological parameters into dynamic evolutionary features, highlighting the plaque's progression rate and acceleration over time. It effectively eliminates the interference of baseline differences between patients, enabling the subsequent TDER to more sensitively capture the high-risk phenotype of rapid progression, thereby improving the discriminative power and generalization performance of the warning model.

[0101] In one embodiment, the early warning time-series differential evolution network is obtained through the following steps:

[0102] S701. Obtain the preset evolutionary training data, which includes historical time-series difference feature vectors, historical basic body information data, and historical risk labels; and construct the model training set and model test set by using the historical time-series difference feature vectors and historical basic body information data as input feature vectors and the historical risk labels as the feature values ​​corresponding to the input feature vectors.

[0103] Specifically, the early warning terminal acquires pre-set evolutionary training data. This data can be obtained by processing a large number of historical patient cases that have undergone multi-temporal CTA examinations, registration, parameter extraction, and differential feature calculations. Each evolutionary training data point represents the situation of a historical patient, including historical temporal differential feature vectors, historical basic physical information data, and historical risk labels. For the first... The evolutionary training data, including its historical time-series difference eigenvectors, can be denoted as: Its form is similar to the current temporal difference feature vector, and it is obtained by extracting difference features from the morphological and functional parameter matrix of historical patients. The included historical basic physical information data can be denoted as... This is the basic physical information data of the historical patient, and its format is similar to that of the current patient's basic physical information data. The historical risk labels included can be denoted as... The historical risk label is a binary label that can be used to annotate whether the patient has experienced a preset high-risk event during the follow-up period, such as annual plaque volume percentage progression ≥1% or the occurrence of a major adverse cardiovascular event, where 1 indicates the occurrence of a high-risk event and 0 indicates no occurrence. For each evolutionary training dataset, the early warning terminal concatenates its historical time-series differential feature vector and historical basic physical information data as the input feature vector, and uses the historical risk label as the feature value corresponding to the input feature vector. The evolutionary training data is then divided into a model training set and a model test set according to a preset ratio. For example, the preset ratio can be set according to actual work, and the default setting is model training set: model test set = 8:2.

[0104] S702 constructs an initial temporal difference evolution network based on an input embedding layer, a Transformer encoder, and a classification output layer. Each input feature vector in the model training set is then input into the initial temporal difference evolution network to obtain the predicted risk probability corresponding to the input feature vector.

[0105] Specifically, the early warning terminal constructs an initial temporal difference evolution network. This network consists of a sequentially connected input embedding layer, a Transformer encoder, and a classification output layer. The input embedding layer maps the input feature vector to a high-dimensional feature space (e.g., 256-dimensional) through a linear transformation and superimposes learnable positional codes to preserve the order information of each feature component. The Transformer encoder, composed of multiple stacked multi-head self-attention layers and feedforward neural network layers (e.g., four layers, each with eight attention heads), captures long-range dependencies and interaction patterns between different elements in the input feature vector. The classification output layer maps the feature vector output by the Transformer encoder to a scalar value between 0 and 1 by passing it through a fully connected layer followed by a sigmoid activation function. The early warning terminal sequentially inputs each input feature vector from the model training set into this initial temporal difference evolution network to obtain the predicted risk probability corresponding to the input feature vector.

[0106] S703, based on the predicted risk probability and historical risk labels, the model loss value is calculated. The expression for the model loss value is:

[0107] ;

[0108] In the formula, It is the model loss value. It is the number of input feature vectors in the model training set. It is the index of any input feature vector. It is the first Historical risk labels of each input feature vector It is the first The predicted risk probability of each input feature vector.

[0109] Specifically, the early warning terminal uses a binary cross-entropy loss function to measure the difference between the predicted risk probability and the actual historical risk label. In the formula, It is the model loss value, which is a dimensionless scalar. It represents the number of input feature vectors in the model training set, which can be obtained by counting the number of input feature vectors in the model training set. It is the index of any input feature vector in the model training set. It is the first Historical risk labels of each input feature vector It is the first The predicted risk probability of each input feature vector can be obtained through S702. This loss function penalizes the consistency between the predicted probability and the true label: when the true label is 1, the loss term is... The closer the predicted risk probability is to 1, the smaller the loss; when the true label is 0, the loss term is... The closer the predicted risk probability is to 0, the smaller the loss will be.

[0110] S704, combined with the model test set, with the goal of minimizing the model loss, iteratively updates the initial temporal difference evolution network until the preset training stopping condition is met, thus obtaining the pre-trained early warning temporal difference evolution network.

[0111] Specifically, the early warning terminal optimizes by minimizing the model loss value. It uses the gradient of each trainable parameter in the initial temporal difference evolution network relative to the loss value and utilizes an optimizer, such as the Adam optimizer, with a learning rate typically set to... or These parameters are iteratively updated. In each subsequent iteration, the model loss value can be calculated and the parameters updated using a small batch of input feature vectors from the model training set, or the model loss value can be calculated using the input feature vectors from the entire model training set. The warning terminal repeats the above process until a preset training stopping condition is met, resulting in a trained warning temporal difference evolution network. The preset training stopping condition can be: the model loss value on the model test set no longer decreases for several consecutive epochs (e.g., 10 epochs) (early stopping); it can also be reaching a preset maximum number of iterations (e.g., 200 epochs); or it can be that the loss value on the model test set drops below a preset threshold.

[0112] This embodiment provides a dynamic risk warning method based on a temporal differential evolution network using multi-temporal patch registration and morphological-functional parameter extraction. It acquires evolutionary training data containing historical temporal differential feature vectors, historical basic physical information, and historical risk labels. An initial network consisting of an input embedding layer, a Transformer encoder, and a classification output layer is constructed. The predicted risk probability is obtained through forward propagation. The difference between the predicted value and the true label is calculated using a binary cross-entropy loss function. The network parameters are iteratively optimized through backpropagation with the goal of minimizing the loss until convergence. This provides a complete and reproducible supervised training process for the temporal differential evolution network for risk warning, enabling the network to learn the complex mapping relationship between differential features and high-risk events from historical data. This allows it to output accurate risk probabilities when facing new patients, ensuring the generalization ability and clinical reliability of the warning model.

[0113] The aforementioned dynamic risk warning method based on a temporal differential evolution network using multi-temporal plaque registration and morphological-functional parameter extraction involves acquiring basic patient information, coronary CTA image data from at least two temporal phases, and synchronized electrocardiogram (ECG) signals. Based on the synchronized ECG signals, the coronary CTA image data are phase-normalized to obtain in-phase image data. Vascular morphology is extracted from each in-phase image data to obtain anatomical topology data for each temporal phase. The anatomical topology data includes at least a set of vascular centerline nodes. The earliest temporal anatomical topology data is used as the basis for the analysis. Based on the reference, and combining the longitudinal deformation field estimation network and anatomical topology data, in-phase image data of each time phase are aligned to obtain aligned image data for each time phase. Based on the aligned image data of each time phase, morphological and functional parameters of the patient's plaque region are extracted to obtain a morphological and functional parameter matrix. These morphological and functional parameters include at least the plaque volume percentage value. Differential features of the morphological and functional parameter matrix are extracted to obtain a temporal differential feature vector. This temporal differential feature vector and basic body information data are input into an early warning temporal differential evolution network to obtain the patient's high-risk probability. Cardiac phase standardization and vascular topology constraints reduce multi-temporal registration errors and ensure spatial consistency of parameter extraction. Simultaneously, differential features are used to characterize dynamic evolution patterns, combined with a deep temporal network for risk prediction, effectively improving the accuracy and clinical interpretability of risk assessment.

[0114] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0115] Based on the same inventive concept, this application also provides a dynamic risk warning device for implementing the aforementioned dynamic risk warning based on a temporal differential evolution network using multi-temporal patch registration and morphological-functional parameter extraction. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the dynamic risk warning device based on a temporal differential evolution network using multi-temporal patch registration and morphological-functional parameter extraction provided below can be found in the limitations of the dynamic risk warning method based on a temporal differential evolution network using multi-temporal patch registration and morphological-functional parameter extraction described above, and will not be repeated here.

[0116] In one exemplary embodiment, such as Figure 2 As shown, a dynamic risk early warning device 200 based on a temporal differential evolution network of multi-temporal patch registration and morphological-functional parameter extraction is provided, comprising:

[0117] The data acquisition module 201 is used to acquire the patient's basic physical information data, coronary CTA image data of at least two time phases, and synchronous electrocardiogram signals;

[0118] The anatomical extraction module 202 is used to perform phase normalization on each coronary CTA image data based on the synchronous electrocardiogram signal to obtain in-phase image data; and to extract the vascular morphology from each in-phase image data to obtain anatomical topology map data for each time phase; the anatomical topology map data includes at least a set of vascular centerline nodes;

[0119] The temporal alignment module 203 is used to align the in-phase image data of each time phase with the earliest anatomical topology map data as a reference, combined with the longitudinal deformation field estimation network and each anatomical topology map data, to obtain the aligned image data of each time phase.

[0120] The parameter extraction module 204 is used to extract morphological and functional parameters of the patient's plaque region based on the aligned image data of each time phase, and obtain a morphological and functional parameter matrix; wherein, the morphological and functional parameters include at least the plaque volume percentage value;

[0121] The risk analysis module 205 is used to extract the differential features of the morphological function parameter matrix to obtain the temporal differential feature vector, and input the temporal differential feature vector and basic body information data into the early warning temporal differential evolution network to obtain the patient's high-risk probability.

[0122] Furthermore, since the coronary CTA image data carries a relative sampling time label, the anatomical extraction module can also be used for:

[0123] For each time phase, the peak time is extracted from the synchronized electrocardiogram signal, and the reconstructed phase label of each coronary CTA image data is calculated based on the peak time and the relative sampling time label.

[0124] For each time phase, the absolute value of the difference between the reconstructed phase label and the preset target diastolic phase is calculated to obtain the phase difference value of each reconstructed phase label; and the reconstructed phase label corresponding to the smallest phase difference value is determined as the original reference phase label of the time phase.

[0125] For each time phase, the phase difference value of the original reference phase label is compared with the preset phase difference threshold to obtain the phase change comparison result; when the phase change comparison result is that the phase difference value is less than the phase difference threshold, the coronary CTA image data of the selected time phase is determined as the same phase image data of the time phase.

[0126] Alternatively, when the phase difference comparison result is greater than or equal to the phase difference threshold, linear interpolation is performed on the coronary CTA image data of two adjacent reconstructed phase labels of the target diastolic phase to obtain the same-phase image data of the time phase.

[0127] Furthermore, the anatomical extraction module can also be used for:

[0128] Input the in-phase image data into the preset blood vessel segmentation network to obtain the blood vessel centerline probability map and blood vessel radius regression map for each time phase;

[0129] Non-maximum suppression is applied to the probability map of the vessel centerline to obtain the candidate centerline probability map;

[0130] The candidate centerline probability graph is skeletonized to obtain the set of nodes and the set of connected edges of the blood vessel centerlines.

[0131] The vascular radius regression map is extracted based on the vascular centerline node set to obtain the node local radius set; based on the vascular centerline node set, the node connected edge set, and the node local radius set, the anatomical topology map data for each time phase is formed.

[0132] Furthermore, the time-relative alignment module can also be used for:

[0133] Using the earliest anatomical topology data as a reference, rigid registration is performed on the anatomical topology data of subsequent time phases to obtain the rigid transformation matrix of each time phase.

[0134] Starting from the second time phase, the in-phase image data is transformed based on the rigid transformation matrix to obtain coarse registration image data for each time phase;

[0135] Starting from the second time phase, the coarse registration image data of the time phase and the coarse registration image data of the previous time phase are input into the longitudinal deformation field estimation network to obtain the local non-rigid displacement field data of the time phase.

[0136] For the earliest time phase, the in-phase image data of the time phase is determined as the aligned image data of the time phase; starting from the second time phase, the coarse registration image is transformed based on the local non-rigid displacement field data and the rigid transformation matrix to obtain the aligned image data of each time phase.

[0137] The longitudinal deformation field estimation network includes:

[0138] The input layer receives the coarse registration image data of the current time phase and the coarse registration image data of the previous time phase, concatenates the coarse registration image data of the current time phase and the coarse registration image data of the previous time phase to obtain a three-dimensional input vector, and transmits the three-dimensional input vector to the feature coding layer.

[0139] The feature encoding layer is used to downsample the three-dimensional input vector through multiple three-dimensional convolutional kernels, extract multi-scale spatial features, obtain a compressed feature map, and transmit the compressed feature map to the displacement field decoding layer.

[0140] The displacement field decoding layer is used to upsample the compressed feature map to obtain local non-rigid displacement field data of the time phase.

[0141] Furthermore, the parameter extraction module can also be used for:

[0142] For each time phase, based on a preset lumen-wall separation model, the aligned image data is segmented to obtain lumen mask data and wall mask data for that time phase; in the lumen mask data, voxels with a value of 1 represent the lumen, and voxels with a value of 1 represent the wall.

[0143] For each time phase, based on a preset CT value threshold range for plaques, the tube wall mask data is extracted to obtain the plaque region mask data for that time phase. Based on the plaque region mask data, tube wall mask data, and lumen mask data, the plaque volume percentage value for each time phase is calculated. In the plaque region mask data, voxels with a value of 1 represent plaque regions. The expression for the plaque volume percentage value is as follows:

[0144] ;

[0145] In the formula, It is any phase of time. That is the phase The percentage value of plaque volume, That is the phase The number of voxels with a value of 1 in the patch region mask data. That is the phase The number of voxels with a value of 1 in the lumen mask data. That is the phase The number of voxels with a value of 1 in the lumen mask data;

[0146] For each time phase, based on a preset threshold range for non-calcified plaque CT values, the tube wall mask data is extracted to obtain the non-calcified plaque region mask data for that time phase. Based on the non-calcified plaque region mask data, tube wall mask data, and lumen mask data, the percentage of non-calcified plaque volume for each time phase is calculated. In the non-calcified plaque region mask data, voxels with a value of 1 represent non-calcified plaques. The expression for the percentage of non-calcified plaque volume is:

[0147] ;

[0148] In the formula, It is any phase of time. That is the phase The percentage of non-calcified plaque volume. That is the phase The number of voxels with a value of 1 in the mask data for non-calcified plaque regions. That is the phase The number of voxels with a value of 1 in the lumen mask data. That is the phase The number of voxels with a value of 1 in the lumen mask data;

[0149] For each time phase, lipid regions are segmented from the aligned image data based on a preset lipid value threshold range to obtain lipid region mask data;

[0150] For each time phase, the aligned image data is segmented based on the set of blood vessel centerline nodes and the preset segmentation length to obtain segmentation window data. For each segmentation window data, the lipid volume ratio data of each segmentation window data of the time phase is calculated based on lipid region mask data and plaque region mask data.

[0151] For each time phase, the lipid core load index of the phase is calculated based on the maximum lipid volume percentage data of the phase. Based on the plaque volume percentage value, the non-calcified plaque volume percentage value and the lipid core load index, the morphological and functional parameters of the phase are composed, and a morphological and functional parameter matrix is ​​composed based on the morphological and functional parameters of each phase.

[0152] Furthermore, the risk analysis module can also be used for:

[0153] Based on the morphological function parameter matrix, starting from the second phase, the difference between the morphological function parameters of the phase and the morphological function parameters of the previous phase is calculated to obtain the absolute difference value of the phase.

[0154] Based on the morphological function parameter matrix, starting from the second time phase, the morphological function parameters of the previous time phase and the preset non-zero coefficients are added to obtain the relative difference reference value of the previous time phase. Then, the absolute difference value of the time phase and the relative difference reference value of the previous time phase are divided to obtain the relative difference value of the time phase.

[0155] The time-series difference feature vector is formed based on the absolute difference value and the relative difference value of each time phase.

[0156] Furthermore, the device may also include a model building module, which may also be used for:

[0157] Obtain the preset evolutionary training data, which includes historical time-series difference feature vectors, historical basic body information data, and historical risk labels; and construct the model training set and model test set by using the historical time-series difference feature vectors and historical basic body information data as input feature vectors and the historical risk labels as the corresponding feature values ​​of the input feature vectors.

[0158] An initial temporal difference evolution network is constructed based on an input embedding layer, a Transformer encoder, and a classification output layer. Each input feature vector in the model training set is then input into the initial temporal difference evolution network to obtain the predicted risk probability corresponding to the input feature vector.

[0159] Based on the predicted risk probability and historical risk labels, the model loss value is calculated, and the expression for the model loss value is:

[0160] ;

[0161] In the formula, It is the model loss value. It is the number of input feature vectors in the model training set. It is the index of any input feature vector. It is the first Historical risk labels of each input feature vector It is the first The predicted risk probability of each input feature vector;

[0162] Using the model test set as the training objective, the initial temporal difference evolution network is iteratively updated until the preset training stopping condition is met, resulting in a trained early warning temporal difference evolution network.

[0163] In one embodiment, such as Figure 3 A computer device is provided, comprising:

[0164] At least one processor 301, and a memory 302 communicatively connected to at least one of the processors 301: the memory stores application code executable by at least one of the processors, the application code being executed by at least one of the processors to enable at least one of the processors to execute the dynamic risk warning method based on the temporal differential evolution network of multi-temporal patch registration and morphological functional parameter extraction as described above.

[0165] Computer equipment may also include: sensor 303.

[0166] The processor 301, memory 302 and sensor 303 can be connected via a bus or other means, with the bus being an example in the figure.

[0167] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0168] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0169] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A dynamic risk early warning method based on a temporal differential evolution network using multi-temporal patch registration and morphological-functional parameter extraction, characterized in that, The method includes: Acquire basic physical information data of the patient, coronary CTA image data at least two time phases, and synchronized electrocardiogram signals; Based on the synchronized electrocardiogram signal, the coronary CTA image data of each coronary artery are phase-normalized to obtain in-phase image data; and the vascular morphology is extracted from each in-phase image data to obtain the anatomical topology map data of each time phase; the anatomical topology map data includes at least a set of vascular centerline nodes; Using the earliest anatomical topology data of the aforementioned time phase as a reference, and combining the longitudinal deformation field estimation network and each of the aforementioned anatomical topology data, the in-phase image data of each of the aforementioned time phases are aligned to obtain the aligned image data of each of the aforementioned time phases; Based on the aligned image data of each of the aforementioned time phases, morphological and functional parameters of the plaque region of the patient are extracted to obtain a morphological and functional parameter matrix; wherein, the morphological and functional parameters include at least the plaque volume percentage value; The differential features of the morphological function parameter matrix are extracted to obtain the temporal differential feature vector. The temporal differential feature vector and the basic body information data are then input into the early warning temporal differential evolution network to obtain the high-risk probability of the patient.

2. The method according to claim 1, characterized in that, The coronary CTA image data carries a relative sampling time label. The phase normalization of each coronary CTA image data based on the synchronized electrocardiogram signal to obtain in-phase image data includes: For each of the aforementioned time phases, the peak time is extracted from the synchronized electrocardiogram signal, and based on the peak time and the relative sampling time label, the reconstructed phase label of each of the coronary CTA image data is calculated; For each of the time phases, the absolute value of the difference between the reconstructed phase label and the preset target diastolic phase is calculated to obtain the phase difference value of each reconstructed phase label; and the reconstructed phase label corresponding to the smallest phase difference value is determined as the original reference phase label of the time phase. For each time phase, the phase difference value of the original reference phase label is compared with a preset phase difference threshold to obtain a phase change comparison result; when the phase change comparison result is that the phase difference value is less than the phase difference threshold, the coronary CTA image data of the selected time phase is determined as the same phase image data of the selected time phase. Alternatively, when the phase change comparison result is that the phase difference is greater than or equal to the phase difference threshold, linear interpolation is performed on the coronary CTA image data of the two adjacent reconstructed phase labels of the target diastolic phase to obtain the in-phase image data of the time phase.

3. The method according to claim 1, characterized in that, The step of extracting blood vessel morphology from each of the in-phase image data to obtain anatomical topology data for each of the time phases includes: The in-phase image data is input into a preset blood vessel segmentation network to obtain the blood vessel centerline probability map and blood vessel radius regression map for each time phase. Non-maximum suppression is applied to the blood vessel centerline probability map to obtain a candidate centerline probability map; The candidate centerline probability graph is skeletonized to obtain the set of nodes and the set of connected edges of the blood vessel centerline; The blood vessel radius regression map is extracted based on the set of blood vessel centerline nodes to obtain the set of local radii of nodes; the anatomical topology map data for each time phase is formed based on the set of blood vessel centerline nodes, the set of connected edges of nodes, and the set of local radii of nodes.

4. The method according to claim 3, characterized in that, The method involves using the earliest anatomical topology data of the aforementioned time phase as a reference, combining the longitudinal deformation field estimation network and each of the anatomical topology data, to align the in-phase image data of each of the aforementioned time phases, resulting in aligned image data of each of the aforementioned time phases, including: Using the earliest anatomical topology data of the aforementioned time phase as a reference, rigid registration is performed on the anatomical topology data of each subsequent time phase to obtain the rigid transformation matrix of each time phase; Starting from the second phase, the in-phase image data is transformed based on the rigid transformation matrix to obtain coarse registration image data for each phase; Starting from the second time phase, the coarse registration image data of the time phase and the coarse registration image data of the previous time phase are input into the longitudinal deformation field estimation network to obtain the local non-rigid displacement field data of the time phase. For the earliest time phase, the in-phase image data of that time phase is determined as the aligned image data of that time phase; starting from the second time phase, the coarse registration image is transformed based on the local non-rigid displacement field data and the rigid transformation matrix to obtain the aligned image data of each time phase; The longitudinal deformation field estimation network includes: The input layer is used to receive the coarse registration image data of the current time phase and the coarse registration image data of the previous time phase, concatenate the coarse registration image data of the current time phase and the coarse registration image data of the previous time phase to obtain a three-dimensional input vector, and transmit the three-dimensional input vector to the feature coding layer. The feature encoding layer is used to downsample the three-dimensional input vector through multiple three-dimensional convolutional kernels, extract multi-scale spatial features, obtain a compressed feature map, and transmit the compressed feature map to the displacement field decoding layer. The displacement field decoding layer is used to upsample the compressed feature map to obtain the local non-rigid displacement field data of the time phase.

5. The method according to claim 1, characterized in that, Based on the aligned image data of each of the aforementioned time phases, morphological and functional parameters of the plaque region of the patient are extracted to obtain a morphological and functional parameter matrix, including: For each of the aforementioned time phases, based on a preset lumen-wall separation model, the aligned image data is segmented to obtain lumen mask data and wall mask data for that time phase; in the lumen mask data, voxels with a value of 1 represent lumens, and voxels with a value of 1 in the lumen mask data represent walls. For each time phase, based on a preset plaque CT value threshold range, the tube wall mask data is extracted to obtain plaque region mask data for that time phase. Based on the plaque region mask data, the tube wall mask data, and the lumen mask data, the plaque volume percentage value for each time phase is calculated. In the plaque region mask data, voxels with a value of 1 represent plaque regions. The expression for the plaque volume percentage value is: ； In the formula, It is any phase of time. That is the phase The percentage value of plaque volume, That is the phase The number of voxels with a value of 1 in the patch region mask data. That is the phase The number of voxels with a value of 1 in the lumen mask data. That is the phase The number of voxels with a value of 1 in the lumen mask data; For each time phase, based on a preset threshold range for non-calcified plaque CT values, the vessel wall mask data is extracted to obtain non-calcified plaque region mask data for that time phase. Based on the non-calcified plaque region mask data, the vessel wall mask data, and the lumen mask data, the non-calcified plaque volume percentage value for each time phase is calculated. In the non-calcified plaque region mask data, voxels with a value of 1 represent non-calcified plaques. The expression for the non-calcified plaque volume percentage value is: ； In the formula, It is any phase of time. That is the phase The percentage of non-calcified plaque volume. That is the phase The number of voxels with a value of 1 in the mask data for non-calcified plaque regions. That is the phase The number of voxels with a value of 1 in the lumen mask data. That is the phase The number of voxels with a value of 1 in the lumen mask data; For each of the aforementioned time phases, lipid region segmentation is performed on the aligned image data based on a preset lipid value threshold range to obtain lipid region mask data; For each of the aforementioned time phases, the aligned image data is segmented based on the set of blood vessel centerline nodes and a preset segmentation length to obtain segmentation window data. For each segmentation window data, the lipid volume ratio data of each segmentation window data of the aforementioned time phase is calculated based on the lipid region mask data and the plaque region mask data. For each of the time phases, the lipid core load index of the time phase is calculated based on the largest lipid volume percentage data of the time phase. Based on the plaque volume percentage value, the non-calcified plaque volume percentage value and the lipid core load index, the morphological function parameters of the time phase are formed, and the morphological function parameter matrix is ​​formed based on the morphological function parameters of each of the time phases.

6. The method according to claim 1, characterized in that, The step of extracting the difference features of the morphological function parameter matrix to obtain the temporal difference feature vector includes: Based on the morphological function parameter matrix, starting from the second phase, the difference between the morphological function parameter of the phase and the morphological function parameter of the previous phase is calculated to obtain the absolute difference value of the phase. Based on the morphological function parameter matrix, starting from the second time phase, the morphological function parameters of the previous time phase and the preset non-zero coefficients are added to obtain the relative difference reference value of the previous time phase. Then, the absolute difference value of the time phase and the relative difference reference value of the previous time phase are divided to obtain the relative difference value of the time phase. The time-series difference feature vector is formed based on the absolute difference value and the relative difference value of each time phase.

7. The method according to claim 1, characterized in that, The early warning time-series differential evolution network is obtained through the following steps: Obtain preset evolutionary training data, which includes historical time-series difference feature vectors, historical basic body information data, and historical risk labels; and construct a model training set and a model test set by using the historical time-series difference feature vectors and historical basic body information data as input feature vectors and the historical risk labels as the feature values ​​corresponding to the input feature vectors. An initial temporal difference evolution network is constructed based on an input embedding layer, a Transformer encoder, and a classification output layer. Each input feature vector in the model training set is then input into the initial temporal difference evolution network to obtain the predicted risk probability corresponding to the input feature vector. Based on the predicted risk probability and the historical risk label, the model loss value is calculated, and the expression for the model loss value is: ； In the formula, It is the model loss value. It is the number of input feature vectors in the model training set. It is the index of any input feature vector. It is the first Historical risk labels of each input feature vector It is the first The predicted risk probability of each input feature vector; Using the model test set as a training objective, the initial temporal differential evolution network is iteratively updated until a preset training stopping condition is met, resulting in the trained early warning temporal differential evolution network.

8. A dynamic risk early warning device based on a temporal differential evolution network of multi-temporal patch registration and morphological-functional parameter extraction, characterized in that, The device includes: The data acquisition module is used to acquire the patient's basic physical information data, coronary CTA image data at least two time phases, and synchronous electrocardiogram signals; The anatomical extraction module is used to perform phase normalization on each of the coronary CTA image data based on the synchronous electrocardiogram signal to obtain in-phase image data; and to extract vascular morphology from each of the in-phase image data to obtain anatomical topology map data for each of the time phases; the anatomical topology map data includes at least a set of vascular centerline nodes; The time-phase alignment module is used to align the in-phase image data of each time phase with reference to the earliest anatomical topology map data, combined with the longitudinal deformation field estimation network and each anatomical topology map data, to obtain aligned image data of each time phase. The parameter extraction module is used to extract morphological and functional parameters of the plaque region of the patient based on the aligned image data of each of the aforementioned time phases, to obtain a morphological and functional parameter matrix; wherein, the morphological and functional parameters include at least the plaque volume percentage value; The risk analysis module is used to extract the differential features of the morphological function parameter matrix to obtain the temporal differential feature vector, and input the temporal differential feature vector and the basic body information data into the early warning temporal differential evolution network to obtain the high-risk probability of the patient.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

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

Images