A Four-Dimensional Spatiotemporal Cardiac Atlas Analysis Method and Its Application

Abstract

This invention discloses a four-dimensional spatiotemporal cardiac atlas analysis method and its applications. Specifically, this invention provides a method for constructing a four-dimensional spatiotemporal (4D) cardiac atlas, and a cardiovascular disease diagnosis and risk prediction model constructed based on the 4D cardiac atlas obtained by the method. The model of this invention can provide more accurate and efficient diagnostic and prediction results for cardiovascular diseases.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically, it relates to a four-dimensional spatiotemporal cardiac atlas analysis method and its application. Background Technology

[0002] Cardiovascular disease is the leading cause of death worldwide, making its early and accurate diagnosis and risk assessment of great clinical significance. Cardiac magnetic resonance imaging (CMR), with its excellent soft tissue resolution and multi-parameter imaging capabilities, has become the gold standard for assessing cardiac structure and function. The heart is a complex, four-dimensional, dynamic organ, and its spatial deformation and temporal evolution during the cardiac cycle contain rich pathophysiological information.

[0003] However, how to systematically quantify and extract highly discriminative spatiotemporal phenotypes from massive CMR image data is a key technical bottleneck currently facing the field of medical image processing and assisted diagnosis.

[0004] While existing technologies have made some progress in areas such as statistical analysis of cardiac structure, calculation of local functional indicators, and imaging-based disease prediction, they still mainly remain at the level of three-dimensional static modeling or local, low-dimensional dynamic analysis. There is still a lack of a unified four-dimensional cardiac modeling and phenotypic extraction technology framework that can simultaneously ensure anatomical consistency, temporal continuity, and the usability of deep learning at the large-scale population level.

[0005] Furthermore, in the process of constructing population-level four-dimensional cardiac atlases based on cardiac magnetic resonance imaging and using them for disease prediction, the field faces the following key technical problems and implementation difficulties:

[0006] (1) How to achieve spatiotemporal consistency alignment across individuals and cardiac phases in a large-scale sample when the heart has significant non-rigid periodic motion. The heart undergoes complex three-dimensional deformation during contraction and relaxation. There are significant differences between different individuals in terms of anatomical structure, heart rate and phase division. Without a stable spatiotemporal alignment mechanism, it is difficult to ensure comparability of the same anatomical location between different individuals and different phases, which leads to distortion of population statistical results.

[0007] (2) How to simultaneously characterize the static anatomical differences and dynamic motion features of the heart on the basis of unified alignment, and avoid mutual interference between the two types of information in high-dimensional representation. Traditional methods often focus only on a single phase or use overall indicators to describe dynamic changes, which makes it difficult to decouple and model cardiac structural remodeling and functional abnormalities, thus limiting the discriminative power and clinical interpretability of the extracted phenotypes.

[0008] (3) How to effectively model and learn high-dimensional trajectory data formed by large-scale cardiac anatomical landmarks within a complete cardiac cycle. This type of data is characterized by high dimensionality, significant noise accumulation, and complex temporal correlations. Directly using conventional deep learning models can easily introduce feature redundancy or lead to training instability, making it difficult to maintain sufficient sensitivity to minor but clinically significant motion abnormalities.

[0009] Therefore, there is an urgent need in this field to develop a population-level four-dimensional cardiac statistical atlas suitable for ultra-large populations, which can achieve stable alignment and comparable analysis of cardiac structures in different individuals and at different cardiac phases through a unified spatiotemporal reference framework. Summary of the Invention

[0010] This invention provides a population-level four-dimensional cardiac statistical atlas applicable to ultra-large populations (more than 50,000 cases). Through a unified spatiotemporal reference framework, it achieves stable alignment and comparable analysis of cardiac structures in different individuals and at different cardiac phases.

[0011] The purpose of this invention is also to achieve unified quantitative modeling of cardiac anatomical morphology features and motion trajectory features within the cardiac cycle based on the aforementioned four-dimensional cardiac atlas, and to characterize static structural differences and dynamic functional changes respectively through feature decoupling.

[0012] The purpose of this invention is also to develop a spatiotemporal deep phenotype system with good versatility and scalability, so that the extracted cardiac phenotypes can be reused in different sample sizes, different disease types and different downstream tasks.

[0013] The present invention also aims to utilize the spatiotemporal depth phenotype as an intermediate characterization to significantly improve the accuracy of automated diagnosis of cardiovascular diseases and the stability and reliability of long-term risk prediction.

[0014] To address the problems of existing technologies, this invention proposes a systematic solution: by introducing a non-rigid registration framework based on differential homeomorphism, a population average reference template is constructed under fixed control point constraints, thereby achieving dual consistency correspondence of cardiac anatomical landmarks in both spatial and temporal dimensions; based on this, multidimensional deep phenotypic extraction is performed, including decoupling two key cardiac phenotypes from the 4D cardiac atlas, namely morphological phenotype and motor phenotype, thereby overcoming the key technical bottlenecks in the characterization and prediction of large-scale four-dimensional cardiac data.

[0015] In a first aspect of the invention, a method for constructing a four-dimensional spatiotemporal (4D) cardiac atlas is provided, comprising the steps of:

[0016] (s1) Provide cardiac magnetic resonance imaging (CMR) data, which includes short-axis magnetic resonance images of the heart;

[0017] (s2) Automated image segmentation and anatomical structure extraction, including: sorting the short-axis magnetic resonance images of the heart in time so that they conform to a complete cardiac cycle; converting the sorted image files into a standard format file that can be processed by a computer; and then inputting the converted standard format file into a pre-trained deep convolutional neural network based on the nnU-Net architecture to achieve fully automated segmentation of the left ventricle, right ventricle and myocardial structure, and converting the two-dimensional image pixel information into a spatial mask with anatomical significance;

[0018] (s3) Generation and quality control of the three-dimensional heart mesh model, including: converting the segmented mask output into a three-dimensional heart mesh representation, smoothing the mesh, and removing low-quality samples with incomplete segmentation or topological errors through an automated quality control process.

[0019] (s4) Spatial rigid alignment and standardization of group morphology, including: using the iterative nearest point algorithm to perform spatial standardization processing on the heart mesh of all individuals generated in (s3) so that all heart models are in a unified standard coordinate system;

[0020] (s5) Construction of 3D cardiac atlas, including: under the framework of large-scale deformable differential homeomorphic metric mapping, using the Defermetrica method to perform accurate non-rigid registration of the cardiac mesh at end-diastole, and constructing a population average cardiac template at end-diastole (ED) by jointly optimizing the population average shape and the differential homeomorphic deformation mapping from the average shape to the shape of each individual heart; and simultaneously obtaining the deformation field parameters describing individual morphological differences, thereby forming a 3D cardiac statistical atlas at end-diastole;

[0021] (s6) Construction of 4D cardiac dynamic atlas, including: using the end-diastolic 3D cardiac atlas constructed in (s5) as a unified reference template, and under the condition of keeping the spatial distribution of control points fixed, non-rigid registration of cardiac morphology at different time phases within the same cardiac cycle is performed, thereby establishing a multi-temporal three-dimensional cardiac model sequence with consistent topological correspondence throughout the entire cardiac cycle, and then constructing a 4D cardiac dynamic atlas.

[0022] In another preferred embodiment, the cardiac magnetic resonance imaging data are derived from healthy individuals and patients with clinically diagnosed cardiovascular disease.

[0023] In another preferred embodiment, the cardiovascular disease is selected from the group consisting of: chronic ischemic heart disease (CHID), atrial fibrillation and atrial flutter (AF), angina pectoris (AP), acute myocardial infarction (AMI), cardiac complications and undefined heart disease (I51-HD), other arrhythmias (OCA), or combinations thereof.

[0024] In another preferred embodiment, step (s2) includes the following sub-steps:

[0025] (s2a) Time-sort the original DICOM format movie short-axis image files to ensure that they conform to a predetermined number of phase characteristics within a complete cardiac cycle;

[0026] (s2b) Use the dicom2niix tool to batch convert sorted DICOM files to NIfTI format files;

[0027] (s2c) The NIfTI format file is input into a pre-trained nnU-Net architecture deep convolutional neural network to achieve fully automatic segmentation of the left ventricle, right ventricle and myocardial structure, and to transform the two-dimensional image pixel information into a spatial mask with anatomical significance.

[0028] In another preferred embodiment, the predetermined number can be any positive integer, such as 20, 25, 30, 40, or 50.

[0029] In another preferred embodiment, the pre-trained nnU-Net architecture deep convolutional neural network is fully trained, effectively ensuring the accuracy and robustness of segmentation.

[0030] In another preferred embodiment, in step (s4), the standardization process includes:

[0031] By calculating the optimal translation and rotation transformations, the positional deviation of the subjects during scanning and the rigidity differences in the anatomical position of the heart are eliminated, so that all heart models are in a unified standard coordinate system.

[0032] In another preferred embodiment, step (s5) includes the following sub-step: optimizing a set of control points. q and its corresponding momentum μ , vertex coordinates x Convert to vector field X ( x ):

[0033] (1)

[0034] in, (2)

[0035] In the formula, X(x) This represents the vector field generated through transformation, used to convert the original vertex coordinates... x Mapped to the target shape; x These are the vertex coordinates of the heart mesh model; p The total number of control points (in this invention, denoted as ). p=2660 (For example) KThis represents a kernel function used to compute vertices. x With control points q The weight of mutual influence between them;

[0036] Equation (2) is a Gaussian kernel function used to define the smoothness of the deformation; y For the position variable in the kernel function; in equation (2), y Corresponding to control points q The position of σ; σ is the scale parameter of the kernel function, which determines the range of influence of the deformation, that is, the smoothness of the deformation in space.

[0037] In another preferred embodiment, step (s6) includes the following sub-steps:

[0038] (s6a) By performing phase non-rigid registration at time point t of each cardiac cycle, a set of differential homeomorphic transformations is estimated. This makes each phase surface S t Alignment with reference phase S0 (i.e., alignment with each time phase within the cardiac cycle) t Perform phase non-rigid registration once (for all cases).

[0039] (3);

[0040] (s6b) The alignment shape of all subjects at the same time phase is averaged, and the population mean shape at time t is defined as:

[0041] (4)

[0042] In the formula, N refers to the number of subjects.

[0043] In another preferred embodiment, step (s7) includes the following sub-steps:

[0044] (s7a) The 3D cardiac atlas was processed using principal component analysis to extract the main morphological variation patterns in the population, thereby obtaining the morphological phenotype;

[0045] (s7b) Using the ED phase as a reference state, the relative displacement trajectories of anatomical landmarks in subsequent phases are quantified; this is achieved by recording the coordinate displacement of each landmark during the cardiac cycle. This generates a dynamic trajectory feature set reflecting the heart's systolic and diastolic functions, thereby obtaining the motor phenotype.

[0046] In a second aspect of the invention, the use of a four-dimensional spatiotemporal (4D) cardiac atlas constructed using the method described in the first aspect of the invention is provided for:

[0047] (a) Constructing a diagnostic model for cardiovascular diseases;

[0048] (b) Construct a diagnostic system for cardiovascular diseases;

[0049] (c) Construct a cardiovascular disease risk prediction model;

[0050] (d) Construct a cardiovascular disease risk prediction system.

[0051] In a third aspect of the invention, a method for constructing a cardiovascular disease diagnostic model based on a four-dimensional spatiotemporal (4D) cardiac atlas constructed according to the method described in the first aspect of the invention is provided, comprising the steps of:

[0052] (Z1) Multidimensional deep phenotype extraction, including decoupling two key cardiac phenotypes from the 4D cardiac atlas, namely morphological phenotype and motor phenotype; the morphological phenotype is the pattern of cardiac morphological variation in the population, and the motor phenotype is a dynamic trajectory feature set reflecting cardiac systolic and diastolic functions;

[0053] (Z2) Spatiotemporal feature extraction, including: based on step (Z1), constructing a deep learning framework based on PointNet, directly encoding the disordered 3D point cloud of each phase of the heart, and extracting high-level spatial features; and introducing a temporal aggregation module to integrate cross-phase motion patterns, thereby obtaining a unified 4D spatiotemporal feature representation that simultaneously contains heart morphology information and motion information;

[0054] (Z3) Model construction, including inputting the 4D spatiotemporal features into the LightGBM classifier and the logistic regression model for training and testing, and selecting or outputting the model with the best performance as the cardiovascular disease diagnosis model;

[0055] The cardiovascular disease diagnostic model is used to diagnose whether the subject currently suffers from a specific cardiovascular disease.

[0056] In another preferred embodiment, the cardiac morphological variation pattern includes normal morphological variation patterns and abnormal morphological variation patterns selected from the group consisting of: chronic ischemic heart disease, atrial fibrillation, angina pectoris, acute myocardial infarction, cardiac complications and undefined heart diseases, other arrhythmias, or combinations thereof.

[0057] In another preferred embodiment, in step (Z3), different post-training parameters are obtained for different cardiovascular diseases.

[0058] In another preferred embodiment, when the cardiovascular disease diagnostic model is applied, its output is a judgment result on whether an individual has a specific cardiovascular disease at the current point in time.

[0059] In another preferred embodiment, the cardiovascular disease risk prediction model determines risk as follows: when the binary classification result output by the cardiovascular disease diagnosis model is 1, it indicates that the subject has a high risk of having the cardiovascular disease; otherwise, it indicates that the subject has a low risk of having the cardiovascular disease.

[0060] In another preferred embodiment, the cardiovascular disease is selected from the group consisting of: chronic ischemic heart disease (CHID), atrial fibrillation and atrial flutter (AF), angina pectoris (AP), acute myocardial infarction (AMI), cardiac complications and undefined heart disease (I51-HD), other arrhythmias (OCA), or combinations thereof.

[0061] In another preferred embodiment, the variation pattern refers to the extracted geometrical variation features in cardiac anatomy.

[0062] In another preferred embodiment, the mutation patterns include normal mutation patterns and abnormal mutation patterns.

[0063] In another preferred embodiment, the normal variation pattern refers to the differences in cardiac anatomy among healthy individuals caused by physiological factors such as gender, age, height, and weight.

[0064] In another preferred embodiment, the abnormal variation pattern refers to pathological changes in anatomical structures caused by cardiovascular disease.

[0065] In another preferred embodiment, a unified classification algorithm framework is used to train models for different types of cardiovascular diseases to obtain specific model parameters for each cardiovascular disease.

[0066] In another preferred embodiment, step (Z1) includes the following sub-steps:

[0067] (s7a) The 3D cardiac atlas was processed using principal component analysis to extract the main morphological variation patterns in the population, thereby obtaining the morphological phenotype;

[0068] (s7b) Using the ED phase as a reference state, the relative displacement trajectories of anatomical landmarks in subsequent phases are quantified; this is achieved by recording the coordinate displacement of each landmark during the cardiac cycle. This generates a dynamic trajectory feature set reflecting the heart's systolic and diastolic functions, thereby obtaining the motor phenotype.

[0069] In another preferred embodiment, the PointNet is an improved PointNet, and the improved PointNet is not only used to process a single frame of static point cloud, but also to perform frame-by-frame feature mapping on the five phases within a complete cardiac cycle, and to uniformly map the disordered three-dimensional point clouds of the left and right ventricles into a stable and alignable high-dimensional global spatial feature representation.

[0070] In a fourth aspect of the invention, a method for constructing a cardiovascular disease risk prediction model based on a four-dimensional spatiotemporal (4D) cardiac atlas constructed according to the method described in the first aspect of the invention is provided, comprising the steps of:

[0071] (Z1) Multidimensional deep phenotype extraction, including decoupling two key cardiac phenotypes from the 4D cardiac atlas, namely morphological phenotype and motor phenotype; the morphological phenotype is the normal variation pattern of cardiac morphology in the population, and the motor phenotype is the dynamic trajectory feature set reflecting cardiac systolic and diastolic functions;

[0072] (Z2) Spatiotemporal feature extraction, including: based on step (Z1), constructing a deep learning framework based on PointNet, directly encoding the disordered 3D point cloud of each phase of the heart, and extracting high-level spatial features; and introducing a temporal aggregation module to integrate cross-phase motion patterns, thereby obtaining a unified 4D spatiotemporal feature representation that simultaneously contains heart morphology information and motion information;

[0073] (Z3) Model construction includes: inputting the 4D spatiotemporal features and the results of whether the individuals corresponding to the 4D spatiotemporal features will develop corresponding cardiovascular diseases within a future preset time window into a machine learning model for training and testing, thereby constructing a cardiovascular disease risk prediction model;

[0074] The cardiovascular disease risk prediction model is used to predict whether a specific cardiovascular disease will occur in the subject within a preset time window.

[0075] In another preferred embodiment, the machine learning model includes a random forest model, a support vector machine model, and a gradient boosting tree model. The random forest model is preferably used to construct the cardiovascular disease risk prediction model in this invention, but the invention is not limited thereto. Those skilled in the art can choose other machine learning models that can achieve the same function according to specific application requirements.

[0076] In another preferred embodiment, the cardiovascular disease risk prediction model targets individuals who have not been diagnosed with the disease and uses a random forest model to prospectively determine whether the corresponding cardiovascular disease will occur within a preset future time window. Its output corresponds to the prediction of the disease occurrence event (i.e., the risk of the corresponding cardiovascular disease occurring within a preset future time window).

[0077] In another preferred embodiment, the cardiovascular risk prediction model focuses on prospective disease assessment, employing a random forest model to predict the risk of cardiovascular disease occurrence within a certain time window, and outputting a prediction of whether an individual will develop the corresponding cardiovascular disease in the future. In practical applications, corresponding diagnostic and prediction models are constructed and trained for different types of cardiovascular diseases, and their performance is independently evaluated, thereby achieving accurate diagnosis and risk stratification prediction for multiple disease scenarios.

[0078] In another preferred embodiment, steps (Z1) and (Z2) have a clear progressive and complementary relationship at the phenotypic modeling level. (Z1) focuses on the construction of explicit phenotypes by modeling the geometric changes of the cardiac mesh in the 4D cardiac atlas throughout the cardiac cycle, decomposing the complex morphological and kinematic information into parameterized representations with clear physical and physiological meanings. Essentially, this corresponds to the original three-dimensional point cloud representation of the cardiac geometry at each temporal phase. (Z2) further extracts implicit phenotypic features based on the explicit phenotypes obtained in (Z1). Specifically, utilizing the consistent topological correspondence of the 4D cardiac atlas throughout the cardiac cycle, the cardiac geometric vertices at each temporal phase are decoupled and represented as disordered three-dimensional point clouds, which are then used as input to the PointNet deep learning framework for training. Through end-to-end learning, PointNet can automatically mine the high-dimensional geometric and temporal structural features contained in the point cloud, thereby forming implicit phenotypic representations that are difficult to explicitly characterize through manual modeling. Therefore, (Z1) and (Z2) correspond to the construction process of explicit phenotype and implicit phenotype, respectively. (Z1) provides a standardized and structurally consistent point cloud input basis for (Z2), and (Z2) realizes deep feature learning of the complex geometry and motion pattern of the heart on this basis.

[0079] In another preferred embodiment, the dynamic trajectory refers to the relative displacement trajectory of cardiac anatomical landmarks in each phase.

[0080] In another preferred embodiment, step (Z1) includes the following sub-steps:

[0081] (s7a) The 3D cardiac atlas was processed using principal component analysis to extract the main morphological variation patterns in the population, thereby obtaining the morphological phenotype;

[0082] (s7b) Using the ED phase as a reference state, the relative displacement trajectories of anatomical landmarks in subsequent phases are quantified; this is achieved by recording the coordinate displacement of each landmark during the cardiac cycle. This generates a dynamic trajectory feature set reflecting the heart's systolic and diastolic functions, thereby obtaining the motor phenotype.

[0083] In a fifth aspect of the invention, a device for diagnosing or predicting the risk of cardiovascular disease is provided, comprising:

[0084] (a) An input module configured to input 4D spatiotemporal features of the object to be tested;

[0085] (b) An evaluation module configured to receive the 4D spatiotemporal features and input the 4D spatiotemporal features into a cardiovascular disease diagnostic model constructed using the method described in the third aspect of the present invention and / or a cardiovascular disease risk prediction model constructed using the method described in the fourth aspect of the present invention for diagnosis and / or risk prediction, and to obtain an evaluation result;

[0086] (c) Output the evaluation results.

[0087] In another preferred embodiment, the apparatus further includes (a0) a preprocessing module configured to perform the following operations and obtain 4D spatial features of the object under test:

[0088] (i) Acquire cardiac magnetic resonance imaging (CMR) data of the subject, including short-axis magnetic resonance images of the heart;

[0089] (ii) The short-axis magnetic resonance images of the heart are sorted by time to conform to a complete cardiac cycle; the sorted image files are converted into a standard format file that can be processed by a computer; the converted standard format file is then input into a pre-trained deep convolutional neural network based on the nnU-Net architecture to achieve fully automatic segmentation of the left ventricle, right ventricle and myocardial structure, and the two-dimensional image pixel information is converted into a spatial mask with anatomical significance.

[0090] (iii) Convert the segmented mask output into a three-dimensional cardiac mesh representation;

[0091] (iv) The three-dimensional cardiac mesh representation is spatiotemporally aligned with the 4D cardiac atlas constructed using the method described in the first aspect of the present invention, thereby projecting the cardiac motion trajectory of the subject under test onto a standard reference space.

[0092] (v) Decouple the phenotype from the projected standard reference space, the phenotype including cardiac morphology phenotype reflecting cardiac structure and motor phenotype reflecting the dynamic trajectory of myocardial function;

[0093] (vi) Input the morphological and kinematic phenotypes into the PointNet deep representation network to obtain the 4D spatiotemporal features of the object under test.

[0094] In another preferred embodiment, the apparatus further includes (g) a control module configured to control the operation of the modules.

[0095] In another preferred embodiment, the control includes inputting the result obtained from the previous module into the next module for further processing.

[0096] In a sixth aspect of the invention, an electronic device is provided, the electronic device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the method described in the first, third or fourth aspect of the invention.

[0097] In a seventh aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described in the first, third, or fourth aspects of the invention.

[0098] In another preferred embodiment, the computer-readable storage medium may be any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may include, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0099] In an eighth aspect of the invention, a computer program product is provided, comprising computer-executable instructions or a computer program that, when executed by a processor, implements the methods described in the first, third, or fourth aspects of the invention.

[0100] In another preferred embodiment, the computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0101] The main advantages of this invention include:

[0102] (a) The method of this invention achieves a refined characterization of the dynamic features of the entire cardiac cycle: compared to traditional 3D statistical maps that only focus on a single phase such as end-diastole, this invention captures the continuous motion trajectory of the heart during systole and diastole through differential homeomorphic mapping and phase non-rigid registration techniques. This four-dimensional modeling approach effectively compensates for the loss of functional information in static morphological analysis, and can reflect the spatial non-uniform distribution of myocardial motion, thereby identifying focal or early functional impairment.

[0103] (b) The method of this invention significantly improves the accuracy of disease diagnosis and risk prediction: by introducing a time-dimensional motor phenotype, the model demonstrates a qualitative leap in predictive performance for various cardiovascular diseases. Taking chronic ischemic heart disease as an example, the enhanced model based on the four-dimensional spatiotemporal phenotype of this invention achieves a prediction accuracy of 92.30% and an AUC value as high as 97.96%, while the baseline model based solely on three-dimensional static morphology has an accuracy of only 62.87%. In the prediction of acute myocardial infarction, the four-dimensional model effectively solves the problem of low specificity (only 22.04%) of traditional static models, significantly improving the specificity to 95.03%.

[0104] (c) The method of this invention possesses extremely high capacity for processing large-scale populations and is highly practical in engineering: This invention constructs a unified and stable spatiotemporal alignment framework, supporting the automated processing of over 50,000 samples. Through fully automatic segmentation using nnU-Net, non-rigid registration using Defermetrica, and automated quality control, it overcomes the bottlenecks of high computational complexity and unstable registration in large-scale population studies, ensuring the consistency of population statistical results.

[0105] (d) The method of this invention achieves decoupling and efficient representation of morphological and kinematic features: This invention innovatively decouples and models the structural differences (morphological principal components) of the heart from dynamic functional changes (displacement trajectory). Combined with a deep learning architecture oriented towards point cloud time series (PointNet + time aggregation module), it effectively filters redundancy and noise in high-dimensional trajectory data, improving the model's ability to discriminate subtle pathological features and its clinical interpretability.

[0106] It should be understood that, within the scope of this invention, the above-described technical features of this invention and the technical features specifically described below (such as in the embodiments) can be combined with each other to form new or preferred technical solutions. Due to space limitations, they will not be described in detail here. Attached Figure Description

[0107] Figure 1 The `ad` section in the diagram shows the overall flowchart of the four-dimensional cardiac atlas construction and downstream analysis in an embodiment of the present invention. Figure 1The data in Figure d shows the numerical distribution of various circulatory system diseases in a multicenter cardiac cine magnetic resonance imaging (MRI) dataset, including ischemic heart disease (CHID), hemorrhoids (HEM), varicose veins of the lower extremities (VVLE), atrial fibrillation and flutter (AF), angina pectoris (AP), cardiac complications and undefined heart disease (I51-HD), acute myocardial infarction (AMI), other arrhythmias (OCA), phlebitis and thrombophlebitis (P / TP), hypotension (HoTN), other peripheral vascular diseases (PVD), heart failure (HF), atrioventricular block and left bundle branch block (AVB), stroke (STROKE), paroxysmal tachycardia (PSVT), cerebral infarction (CI), non-rheumatic mitral valve disease (MVD), and other cerebrovascular diseases (CVD). Based on this, this invention selected six types of cardiovascular diseases with more than 1000 cases as research subjects.

[0108] Figure 2 The diagram in 'ac' shows the architecture of the four-dimensional cardiac feature extraction network in an embodiment of the present invention.

[0109] Figure 3 The results of visualization and principal component analysis of the cardiac model in this embodiment of the invention are shown.

[0110] Figure 4 The figure shown in 'ab' is a comparison chart of the superiority of 3D modeling and 4D modeling in the embodiments of the present invention.

[0111] Figure 5 This diagram illustrates the performance evaluation of different temporal keyframe sampling strategies in six cardiovascular disease diagnostic tasks according to embodiments of the present invention. The figure compares four temporal input configurations: 2-phase (2-ED-ES, blue), 3-phase (3-ED-ES-ED, green), 4-phase (4-ED-MD-MD-ED, orange), and 5-phase (5-ED-MD-ES-MD-ED, pink). Error bars represent standard deviations.

[0112] Figure 6 This diagram illustrates the performance evaluation of different temporal keyframe sampling strategies in six cardiovascular disease risk prediction tasks according to embodiments of the present invention. The figure compares four temporal input configurations: 2-phase (2-ED-ES, blue), 3-phase (3-ED-ES-ED, green), 4-phase (4-ED-MD-MD-ED, orange), and 5-phase (5-ED-MD-ES-MD-ED, pink). Error bars represent standard deviations.

[0113] Figure 7The results of SHAP analysis of morphological and kinematic features for disease diagnostic tasks are displayed and visualized using bar charts. Features are sorted by mean absolute SHAP value. "ED", "Mid1", "ES", "Mid2", and "nextED" correspond to end-diastole, first intermediate phase, end-systole, second intermediate phase, and next end-diastole, respectively; "X / Y / Z" represent displacement along the corresponding spatial directions. Detailed Implementation

[0114] Through extensive and in-depth research and rigorous screening, the inventors have developed, for the first time, a method for constructing a 4D cardiac dynamic atlas. Based on this 4D cardiac dynamic atlas, a model has been developed that can effectively diagnose and predict the risk of cardiovascular diseases in subjects. Experiments show that the cardiovascular disease diagnosis and risk prediction model constructed using the method of this invention significantly improves all performance indicators compared to models constructed based on 3D data. Based on this, this invention was completed.

[0115] Specifically, cardiovascular diseases in this invention include: chronic ischemic heart disease (CHID), atrial fibrillation and atrial flutter (AF), angina pectoris (AP), acute myocardial infarction (AMI), cardiac complications and undefined heart disease (I51-HD), other arrhythmias (OCA), or combinations thereof.

[0116] This invention targets various types of cardiovascular diseases, employing a unified classification algorithm framework for model training to obtain disease-specific model parameters. During the inference phase, the 4D spatiotemporal features of the test subject are input into the same model structure with different parameter configurations for separate discrimination, thereby achieving independent diagnosis and risk assessment for different cardiovascular diseases.

[0117] In this invention, the output of disease diagnosis is a judgment result indicating whether the subject has a specific cardiovascular disease at the current time point. Specifically, it is represented by a binary label corresponding to the disease type, used for disease type identification and assisted diagnosis. Disease risk prediction primarily targets individuals not currently diagnosed with the disease. By analyzing their 4D spatiotemporal cardiac characteristics, it prospectively predicts whether the corresponding cardiovascular disease will occur within a preset future time window. The output of the prediction model is also a binary label, its semantics indicating the probability of the disease occurring in the future time dimension, rather than the existence of the disease in the current state, used to support early warning and risk stratification management of diseases.

[0118] the term

[0119] To facilitate a clearer understanding of this disclosure, certain terms are first defined. As used herein, unless otherwise expressly specified herein, each of the following terms shall have the meaning given below. Other definitions are set forth throughout the application.

[0120] The term “about” can refer to a value or composition within an acceptable margin of error for a particular value or composition as determined by a person skilled in the art, depending in part on how the value or composition is measured or determined. For example, as used herein, the expression “about 100” includes all values ​​between 99 and 101.

[0121] As used herein, the terms “containing” or “including (comprise)” can be open-ended, semi-closed, or closed. In other words, the terms also include “consistently made of” or “composed of”.

[0122] As used herein, unless otherwise stated, any concentration range, percentage range, proportion range, or integer range shall be understood to include any integer value within the range and, where appropriate, its fractional value (e.g., one-tenth and one-hundredth of an integer).

[0123] As used herein, the term “and / or” refers to and covers any and all possible combinations of one or more of the related listed items.

[0124] As used in this article, the "differential homeomorphic transformation" is a very important type of transformation in mathematics and geometry. Simply put, it is a reversible, smooth (infinitely differentiable) transformation, and its inverse transformation is also smooth. It is used to align the shapes of different individuals' hearts (in a reference phase or other phases) to the same spatial template (reference shape). This ensures the establishment of spatial correspondence, and the deformation is smooth, reversible, and topologically preserving.

[0125] As used in this paper, a "phase surface" refers to the geometric surface of the heart at a specific moment (phase) during the cardiac cycle. It is the basic unit for constructing a spatiotemporal map. For each individual, there is a series of phase surfaces (e.g., 50) that describe the individual's cardiac motion. Therefore, it is necessary to align the phase surfaces of different individuals in both time and space for population statistical analysis.

[0126] As used in this article, a "reference phase" is a standard time point selected during time alignment, typically used as the reference zero point for the time axis. In cardiac motion analysis, a phase that is easily identifiable and of significant physiological importance is usually chosen as the reference phase. The most common reference phase is end-diastole (ED), the moment when the ventricles have finished filling and are about to begin contraction. After selecting a reference phase, other phases can be represented as time offsets relative to the reference phase (or standard time points can be obtained through interpolation, such as 50 phases). Simultaneously, in spatial registration, the heart shape of the reference phase is often used as a template (reference shape) for spatial alignment.

[0127] First, all individuals are aligned to a standard template (spatial alignment) on the surface of the reference phase using differential homeomorphic transformation. Then, for other phases, physiological semantics (such as end-diastolic feature points) are anchored to the reference phase, and combined with template-based deformation registration techniques, logical correspondences are established under a pre-defined standardized temporal sequence (such as 25 or 50 phases), achieving natural alignment of the time axis. In this way, the shape and motion characteristics of different individuals at the same spatiotemporal location (i.e., the same phase and the same anatomical location) can be compared.

[0128] Through this process, a four-dimensional (three-dimensional space + one-dimensional time) statistical shape model of the heart can be constructed, namely a spatiotemporal cardiac atlas. This atlas can describe the average pattern and range of variation of heart shape and movement in healthy individuals, and can then be used for disease diagnosis and risk prediction.

[0129] As used in this article, "anatomical landmarks" are key, well-defined reference points used for precise location and description of anatomical structures. They are typically specific locations on anatomical structures that are easily identifiable, repeatable, and have clear functional or morphological significance.

[0130] As used in this paper, the "fixed-point technique" is a constraint method used in non-rigid registration that guides and constrains the deformation process between images or shapes by defining a set of anatomically corresponding, relatively fixed key points.

[0131] The 4D cardiac feature extraction method and model construction of the present invention

[0132] This invention also provides a 4D cardiac feature extraction method and a model construction method. The specific process is as follows: Figure 2 As shown in ac. It includes the following steps:

[0133] (S1) Perform multidimensional deep phenotypic extraction. Specifically, decouple two types of explicit phenotypes—morphological phenotype and motor phenotype—from the 4D cardiac atlas. The morphological phenotype characterizes the variation patterns of cardiac geometry within a population. These variations include normal morphological variations in healthy individuals caused by physiological factors such as sex, age, and body size, as well as pathological changes in anatomical structures caused by cardiovascular diseases. These morphological variations do not directly correspond to specific disease names but are parametrically expressed as geometric structural changes. The motor phenotype analyzes the continuous deformation process of the heart throughout the cardiac cycle, extracting a dynamic trajectory feature set reflecting the heart's systolic and diastolic functions.

[0134] Building upon this, step (S2) is performed to conduct end-to-end spatiotemporal representation learning. Specifically, leveraging the consistent vertex correspondence of 4D cardiac atlases throughout the entire cardiac cycle, the geometric vertices of the heart surface at each temporal phase are decoupled into disordered 3D point cloud representations. Each temporal phase corresponds to a point cloud frame, and each point cloud frame consists of vertices at the same anatomical location. Subsequently, the point clouds of each temporal phase are input frame by frame into a PointNet-based spatial feature encoding module. Due to PointNet's insensitivity to the order of point cloud arrangement, it can effectively extract high-dimensional spatial geometric features from the disordered 3D point cloud, obtaining a global spatial feature representation corresponding to each temporal phase.

[0135] Subsequently, a temporal aggregation module is introduced to integrate the spatial features across phases. This module performs weighted fusion and nonlinear mapping on the spatial features of each phase, thereby modeling the continuous motion pattern of the heart throughout the cardiac cycle. This cross-phase motion pattern is semantically consistent with the dynamic trajectory phenotype extracted in step (S1), both essentially used to characterize the temporal functional features of the heart. Through temporal aggregation, a unified 4D spatiotemporal feature representation that simultaneously contains cardiac morphological and motion information is finally obtained.

[0136] In step (S3), a cardiovascular disease diagnostic model and a risk prediction model are constructed based on the 4D spatiotemporal features. The diagnostic model takes the current individual's 4D spatiotemporal features as input, trains using a LightGBM classifier and a logistic regression model, and evaluates the model's performance through cross-validation. The final output is a binary diagnostic result used to determine whether the individual currently suffers from a specific cardiovascular disease. The risk prediction model, on the other hand, targets individuals not diagnosed with the disease and uses a random forest model to prospectively determine whether the corresponding cardiovascular disease will occur within a preset future time window. Its output is also a binary classification result, but its semantics correspond to the prediction of disease occurrence events, rather than a judgment of the disease's existence in the current state.

[0137] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Experimental methods in the following embodiments, unless otherwise specified, are generally performed under conventional conditions, such as those described in Sambrook et al., Molecular Cloning: A Laboratory Manual (New York: Cold Spring Harbor Laboratory Press, 1989), or as recommended by the manufacturer. Unless otherwise stated, percentages and parts are weight percentages and parts by weight.

[0138] Example 1: Construction of a 4D Spatiotemporal Cardiac Atlas and Atlas-Based Analysis Method

[0139] This invention proposes a four-dimensional spatiotemporal cardiac atlas analysis method and its application. This technical solution targets large-scale cardiac magnetic resonance imaging data, and through a unified spatiotemporal alignment framework, jointly models the cardiac anatomy and dynamic motion information within the cardiac cycle to form an interpretable and scalable spatiotemporal depth phenotype, which can be used for downstream disease diagnosis and risk prediction. Specifically, this invention provides the following technical solution:

[0140] Step 1: Acquisition of Cardiac Imaging Data and Cohort Screening. First, high-quality short-axis CMR imaging data were screened from a large cohort, covering healthy individuals and patients with various cardiovascular diseases. This invention used short-axis cardiac magnetic resonance (CMR) images acquired by UK Biobank (UKB) between 2013 and 2022. The study cohort included two groups: 24,892 healthy individuals and 25,198 clinically diagnosed cardiovascular disease patients, with a male-to-female ratio of approximately 1.08:1. Participants ranged in age from 45 to 85 years and had a body mass index (BMI) ranging from 14.1 to 69.6.

[0141] CMR images used to construct cardiac atlases and train related models were obtained from 50,090 participants. Of these, 24,892 relatively healthy individuals were used to construct complete 4D cardiac atlases. Simultaneously, six common cardiovascular diseases were selected from the confirmed patient cohort, each with more than 1,000 cases. Sorted by sample size from largest to smallest, the selected diseases included: chronic ischemic heart disease (CHID), atrial fibrillation and atrial flutter (AF), angina pectoris (AP), and the remaining categories were cardiac complications and undefined heart disease (I51-HD), acute myocardial infarction (AMI), and other arrhythmias (OCA). For downstream analyses of each cardiovascular disease, data were uniformly divided into training, validation, and test sets in a 6:1:3 ratio. Detailed data distribution is shown below. Figure 1 As shown in d.

[0142] Step 2: Automated Image Segmentation and Anatomical Structure Extraction. The original DICOM format short-axis movie images were temporally sorted to ensure they conformed to 50 phase features within a complete cardiac cycle. Next, the DICOM files were batch-converted to NIfTI format using the dicom2niix tool for subsequent processing and input to deep learning models. Subsequently, a deep convolutional neural network based on the nnU-Net architecture was used for fully automated segmentation of the left ventricle, right ventricle, and myocardial structures, transforming the two-dimensional image pixel information into anatomically meaningful spatial masks (see...). Figure 1 (a) The model has been fully trained on thousands of samples, effectively ensuring the accuracy and robustness of segmentation.

[0143] Step 3: Generation and Quality Control of the 3D Heart Mesh Model. The segmented mask output is converted into a 3D heart mesh representation, which serves as the geometric input for subsequent atlas construction. At this stage, the mesh is smoothed (see...). Figure 1 (a) is used to eliminate interlayer alignment anomalies caused by CMR layer thickness and to remove low-quality samples with incomplete segmentation or topological errors through an automated quality control process.

[0144] Step 4: Spatial rigid alignment and standardization of group morphology. The iterative nearest-point algorithm is used to spatially standardize the heart mesh of all individuals (see...). Figure 1 (a) By calculating the optimal translation and rotation transformations, the positional deviation of the subjects during scanning and the rigidity differences in the anatomical position of the heart are eliminated, so that all heart models are in a unified standard coordinate system, providing an initial reference for subsequent fine non-rigid deformation.

[0145] Step 5: Construction of 3D Cardiac Statistical Atlas. Within the framework of a large-scale deformable differential homeomorphism metric mapping, the Defermetrica method is used to perform accurate non-rigid registration of the cardiac mesh at end-diastole. This is achieved by optimizing a set of control points. q and its corresponding momentum μ , vertex coordinates x Convert to vector field X ( x ):

[0146] (1)

[0147] (2)

[0148] In the formula, X(x) This represents the vector field generated through transformation, used to convert the original vertex coordinates... x Mapped to the target shape; x These are the vertex coordinates of the heart mesh model; p The total number of control points (in this invention, denoted as ).p=2660 (For example) K This represents a kernel function used to compute vertices. x With control points q The weight of mutual influence between them;

[0149] Equation (2) is a Gaussian kernel function used to define the smoothness of the deformation; y For the position variable in the kernel function; in equation (2), y Corresponding to control points q The position of σ; σ is the scale parameter of the kernel function, which determines the range of influence of the deformation, that is, the smoothness of the deformation in space.

[0150] This process (see) Figure 1 a) generates an average template of end-diastolic (ED) and maps the template to the deformation field of each individual.

[0151] Step 6: Spatiotemporal Consistency 4D Cardiac Dynamic Atlas Expansion. Using the constructed ED-phase 3D atlas as a baseline template, the atlas is expanded to the entire cardiac cycle using fixed control point techniques. By performing phase non-rigid registration at time point t of each cardiac cycle, a set of differential homeomorphic transformations is estimated. This makes each phase surface S t Align with reference phase S0:

[0152] (3)

[0153] Finally, the alignment shapes of all subjects at the same time phase are averaged, and the population mean shape at time t is defined as:

[0154] (4)

[0155] In the formula, N refers to the number of subjects.

[0156] The process (see) Figure 1 b) ensures the temporal continuity of the 4D atlas and the one-to-one correspondence with anatomical landmarks.

[0157] Step 7: Obtaining multidimensional deep phenotypes (morphological and kinematic features). Decoupling two key phenotypes from the 4D map:

[0158] Morphological phenotype: Principal component analysis was used to process the 3D map and extract the main morphological variation patterns in the population.

[0159] Motor phenotype: Using the ED phase as a reference state, the relative displacement trajectories of anatomical landmarks in subsequent phases are quantified. This is achieved by recording the coordinate displacement of each landmark during the cardiac cycle. This generates a dynamic trajectory feature set that reflects the heart's systolic and diastolic functions.

[0160] The aforementioned variation patterns refer to the extracted geometrical changes in cardiac anatomy, including normal and abnormal variation patterns. Normal variation patterns refer to differences in cardiac anatomy within healthy individuals caused by physiological factors such as sex, age, height, and weight. Abnormal variation patterns refer to pathological changes in anatomical structure caused by cardiovascular diseases.

[0161] The morphological and kinematic phenotypes essentially correspond to the original 3D point cloud representations of the cardiac geometry at various time phases.

[0162] Step 8: End-to-end spatiotemporal representation learning and prediction. Construct a deep learning framework based on PointNet (see...). Figure 2 This directly encodes the disordered 3D point cloud of each phase of the heart and extracts high-level spatial features. A temporal aggregation module is introduced to integrate cross-phase motion patterns, and finally, a classifier (see...) is used to... Figure 1 c) To achieve accurate diagnosis and risk prediction of cardiovascular diseases.

[0163] Specifically, high-level spatial feature extraction involves leveraging the consistent topological correspondence of 4D heart atlases throughout the entire cardiac cycle to decouple the geometric vertices of the heart at each temporal phase and represent them as disordered 3D point clouds. These point clouds are then used as input to the PointNet deep learning framework for training. Through end-to-end learning, PointNet can automatically mine the high-dimensional geometric and temporal structural features inherent in the point cloud, thereby forming implicit phenotypic representations that are difficult to explicitly characterize through manual modeling. Due to PointNet's permutation invariance, it can effectively extract the high-dimensional spatial geometric features of the heart for each frame from these disordered points.

[0164] At the spatial representation level, this invention modifies the PointNet encoder to adapt it for temporal conditions. Unlike traditional PointNet, which is only used to process static point clouds in a single frame, this invention uses an improved PointNet encoder to perform frame-by-frame feature mapping on the five phases within a complete cardiac cycle. This maps the disordered three-dimensional point clouds of the left and right ventricles into a stable and alignable high-dimensional global spatial feature representation, thereby providing a consistent feature foundation for cross-phase modeling.

[0165] In terms of temporal modeling, this invention introduces a temporal aggregator module for the first time. By combining linear mapping with an attention weight generation mechanism based on Softmax / Tanh, it automatically learns the relative contributions of different cardiac phases in disease discrimination tasks and adaptively weights and fuses the features of each phase. This temporal aggregation strategy breaks through the limitations of traditional methods that rely solely on end-diastolic and end-systolic keyframes for analysis. Instead, it effectively captures potential subtle dynamic anomalies in intermediate phases (such as Mid1 and Mid2) by jointly modeling the displacement and morphological evolution trajectories of all phases, thereby extending from static three-dimensional structural modeling to four-dimensional spatiotemporal joint representation.

[0166] For detailed procedures of steps 7 and 8, please refer to [link / reference]. Figure 2 The ac in [the context]. Specifically:

[0167] First, step 7 is performed to obtain the multidimensional deep phenotype. Specifically, two explicit phenotypes, morphological and motor phenotypes, are decoupled from the 4D cardiac atlas (this multidimensional deep phenotype serves as...). Figure 2 The input to the process shown comes from... Figure 1 (Results in b) The morphological phenotype is used to characterize the variation patterns of cardiac geometry in the population. These variations include normal morphological variations in healthy individuals caused by physiological factors such as sex, age, and body type, as well as pathological changes in anatomical structures caused by cardiovascular diseases. These morphological variations do not directly correspond to specific disease names but are parametrically expressed as geometric structural changes. The motor phenotype, on the other hand, analyzes the continuous deformation process of the heart throughout the cardiac cycle, extracting a dynamic trajectory feature set reflecting the heart's systolic and diastolic functions.

[0168] Building upon this, step 8 is performed to conduct end-to-end spatiotemporal representation learning. Specifically, leveraging the consistent vertex correspondence of 4D cardiac atlases throughout the entire cardiac cycle, the geometric vertices of the heart surface at each temporal phase are decoupled into disordered 3D point cloud representations. Each temporal phase corresponds to a point cloud frame, and each point cloud frame consists of vertices at the same anatomical location. Subsequently, the point clouds of each temporal phase are input frame by frame into a PointNet-based spatial feature encoding module. Because PointNet is insensitive to the order of point cloud arrangement, it can effectively extract high-dimensional spatial geometric features from the disordered 3D point cloud, obtaining a global spatial feature representation corresponding to each temporal phase.

[0169] Subsequently, a temporal aggregation module is introduced to integrate the cross-phase spatial features. This module performs weighted fusion and nonlinear mapping on the spatial features of each temporal phase, thereby modeling the continuous motion pattern of the heart throughout the cardiac cycle. This cross-phase motion pattern is semantically consistent with the dynamic trajectory phenotype extracted in step 7, both essentially used to characterize the temporal functional features of the heart. Through temporal aggregation, a unified 4D spatiotemporal feature representation that simultaneously contains cardiac morphological and motion information is finally obtained.

[0170] Step 9: Model Construction. Based on the aforementioned 4D spatiotemporal features, a cardiovascular disease diagnostic model and a risk prediction model are constructed respectively. The diagnostic model takes the current individual's 4D spatiotemporal features as input, uses a LightGBM classifier and a logistic regression model for training, and evaluates the model performance through cross-validation. Finally, it outputs a binary diagnostic result to determine whether the individual currently suffers from a specific cardiovascular disease.

[0171] The risk prediction model targets individuals who have not been diagnosed with the disease and uses a random forest model to prospectively determine whether the corresponding cardiovascular disease will occur within a preset time window. Its output is also a binary classification result, but its semantics correspond to the prediction of the disease occurrence event, rather than the judgment of the existence of the disease in the current state.

[0172] Example 2

[0173] In this embodiment, the population-scale cardiac statistical atlas constructed by this invention is used to quantitatively analyze the differences in cardiac morphology between healthy individuals and cardiovascular disease patients, in order to verify the effectiveness of the atlas in characterizing anatomical variations and pathological remodeling at the population level. The analysis is based on a three-dimensional cardiac atlas at end-diastole, uniformly characterizing the cardiac model after non-rigid registration, and using principal component analysis to extract the main morphological variation patterns in the population.

[0174] Experimental results are as follows Figure 3 As shown, in healthy individuals, the first three principal components explained 20.20%, 12.92%, and 10.80% of the overall morphological variation, respectively. Their corresponding anatomical meanings are, in order, the relative volume change of the right ventricle, the relative volume change of the left ventricle, and the geometric concentricity characteristics of the right ventricle. This indicates that the statistical atlas constructed in this invention can stably and interpretably reflect the population differences in the main cardiac structures.

[0175] Further application of this analytical method to the disease group sample revealed that the disease group, represented by patients with acute myocardial infarction, exhibited more significant local morphological abnormalities in the corresponding principal component directions, with their cardiac surface showing obvious irregular deformation and non-uniform remodeling characteristics in specific regions. Compared with the healthy group, this difference showed a clear separation trend in the statistical space.

[0176] The above analysis results show that the morphological principal component analysis method based on population-scale cardiac atlas of the present invention can effectively quantify and reveal the reconstruction law of cardiac structure under pathological conditions, and provide a reliable technical means for objective phenotypic characterization and population classification of cardiovascular diseases.

[0177] Example 3

[0178] 3.1 Performance Comparison of Different Models

[0179] In this embodiment, the population-scale four-dimensional cardiac statistical atlas constructed by this invention and its extracted spatiotemporal depth phenotype are applied to various cardiovascular disease diagnosis and risk prediction tasks to verify the effectiveness of this technical solution in actual clinical prediction scenarios. As a control, a baseline model based solely on end-diastolic 3D cardiac static morphological features is constructed and compared with the enhanced model using the 4D spatiotemporal phenotypic features of this invention under the same data partitioning and training strategy.

[0180] The four-dimensional spatiotemporal phenotype simultaneously includes static morphological information and landmark displacement trajectory information of each time phase relative to the initial phase within the cardiac cycle, thus comprehensively characterizing the dynamic deformation features of the heart during systole and diastole. Based on the above phenotypic features, predictive modeling is performed for various disease types, including chronic ischemic heart disease (CHID), atrial fibrillation (AF), angina pectoris (AP), cardiac complications (I51-HD), acute myocardial infarction (AMI), and other arrhythmias (OCA), and the model performance is evaluated on an independent test set.

[0181] The diagnostic performance of the model based solely on the 3D static morphological features of the heart at end-diastole and the enhanced model using the 4D spatiotemporal phenotypic features of this invention are summarized as follows: Figure 4 As shown in a. Table 1 shows the diagnostic results of the enhanced model of the invented 4D spatiotemporal phenotypic features in different disease cohorts. The classification performance of the proposed model was evaluated using accuracy, precision, recall, specificity, F1 score, and area under the receiver operating characteristic curve (AUC).

[0182] Table 1

[0183]

[0184] Note: All values ​​in the table are percentages (%).

[0185] As shown in Table 1, the model proposed in this invention is stable across all disease cohorts, with an accuracy ranging from 64.75% to 70.11% and an AUC value between 68.20% and 72.50%. Acute myocardial infarction (AMI) showed the best diagnostic performance (accuracy 70.11±1.29%, AUC 72.50±1.35%), while atrial fibrillation and atrial flutter (AF) and angina pectoris (AP) had relatively low recall rates, reflecting the differences in dynamic spatiotemporal phenotypes among different diseases.

[0186] And, as Figure 4 As shown in b and Table 2, the risk prediction results of the enhanced 4D spatiotemporal phenotypic feature model of this invention show a significant performance improvement compared to the model based solely on the 3D static morphological features of the heart at end-diastole. The enhanced 4D spatiotemporal phenotypic feature model of this invention achieves an accuracy exceeding 87% across all cohorts, with an AUC consistently above 96%, highlighting the strong discriminative power of the proposed framework in population-level risk assessment.

[0187] Table 2 shows the disease risk prediction results for different disease cohorts. The predictive performance of the proposed model was evaluated using accuracy, precision, recall, specificity, F1 score, and area under the receiver operating characteristic curve (AUC).

[0188] Table 2

[0189]

[0190] Note: All values ​​in the table are percentages (%).

[0191] 3.2 Contribution of Dynamic Motion Phenotype to the Model

[0192] To assess the contribution of dynamic motion phenotypes, this invention compares a baseline model using only 3D morphological features with an enhanced model that incorporates complete 4D motion information.

[0193] The results are as follows Figure 4 As shown, the overall performance of each disease cohort improved after incorporating motion features. The results indicate that clinically relevant disease features exist not only in static cardiac anatomy but also in dynamic spatiotemporal deformation patterns, and that introducing temporal motion information can significantly improve the accuracy and robustness of diagnosis and prognostic prediction.

[0194] Taking chronic ischemic heart disease as an example, the enhanced model using the four-dimensional spatiotemporal phenotype of this invention achieved a prediction accuracy of 92.30% and an area under the receiver operating characteristic (AUC) of 97.96% on the test set, which is significantly better than the baseline model that only uses three-dimensional static morphological features.

[0195] In other disease types, the four-dimensional augmented model also showed consistent and stable performance improvements in accuracy, sensitivity, specificity, and AUC evaluation metrics. In particular, it effectively improved the problem of low specificity in traditional static models in tasks such as acute myocardial infarction and other arrhythmias.

[0196] 3.2 Impact of different time-sampling strategies on disease diagnostic performance

[0197] Further experiments evaluated the impact of different time-sampling strategies on disease diagnostic performance.

[0198] The results are as follows Figures 5-6 As shown, increasing the number of cardiac phases sampled generally leads to stable performance improvements across most diseases and assessment metrics. Beyond end-diastolic and end-systolic frames, temporal sampling strategies incorporating intermediate phases demonstrate a better balance between sensitivity and specificity, reflected in higher F1 scores and AUCs across several disease categories. These results suggest that intermediate phase information helps capture subtle pathological motion patterns that are difficult to fully reflect with sparse phase selection. This performance difference further illustrates that dense temporal features covering the entire cardiac cycle can more accurately capture pathological abnormalities in cardiac motion, thereby improving the model's diagnostic effectiveness.

[0199] The above results demonstrate that the dynamic motion characteristics of the heart during a complete cardiac cycle contain important disease-discriminating information. By introducing the time dimension (3D + time) into the cardiac phenotype modeling process, this invention can more fully characterize pathological cardiac remodeling and functional abnormalities, thereby significantly improving the accuracy of cardiovascular disease diagnosis and the reliability of risk prediction, and possesses good clinical application value and promising prospects for promotion.

[0200] 3.3 SHAP Analysis

[0201] To better understand the contribution of different features to the model prediction, in this embodiment, the present invention further uses SHAP (SHapley Additive exPlanations) for ex post-hoc interpretability analysis.

[0202] The results are as follows Figure 7 As shown. SHAP analysis results indicate that the model significantly depends on the multi-phase displacement characteristics throughout the cardiac cycle and the three-dimensional spatial information of each phase in various diagnostic tasks.

[0203] It is worth noting that the features of the intermediate phases (Mid1 and Mid2), in addition to the traditional end-diastolic (ED) and end-systolic (ES) features, have consistently been among the most important contributing features in different diseases, while the importance of each phase feature varies in different diagnostic tasks.

[0204] These results demonstrate that the dynamic processes of cardiac contraction and relaxation contain rich, individual-specific information that static morphological features cannot fully capture. Compared to representations based solely on static structure (3D), four-dimensional cardiac atlases that integrate morphological and kinematic information can capture more comprehensive spatiotemporal patterns, thereby improving predictive performance and enhancing the biological interpretability of extracted features.

[0205] Example 4

[0206] Clinical personalized diagnosis and risk assessment application process. In a real-world clinical application scenario, we take the early screening of a suspected acute myocardial infarction (AMI) patient as an example.

[0207] First, the system needs to acquire the subject's full cardiac cycle short-axis cine CMR image sequence. The system automatically sorts and converts the original images in time, and calls the pre-trained nnU-Net model to perform fully automatic segmentation of the heart structure, generating a dynamic three-dimensional mesh model of the heart for each phase.

[0208] Subsequently, the individual's spatiotemporal alignment was performed using the population atlas constructed in this invention, projecting their cardiac motion trajectory onto a standard reference space. The system then extracted the patient's unique decoupled phenotype, namely, the morphological principal component features reflecting cardiac structure and the four-dimensional displacement trajectory of anatomical landmarks reflecting myocardial function.

[0209] Finally, these multidimensional deep phenotypes are input into the PointNet deep representation network, and the features across cardiac cycles are integrated using a temporal aggregation module to obtain a unified 4D spatiotemporal feature representation. Subsequently, the obtained 4D spatiotemporal features are input into either a cardiovascular disease diagnostic model or a cardiovascular disease risk prediction model. The model can then output the disease classification results and long-term risk prediction scores for the case in real time, providing clinicians with highly discriminative quantitative diagnostic evidence.

[0210] All documents mentioned in this invention are incorporated herein by reference as if each document were individually incorporated by reference. Furthermore, it should be understood that after reading the foregoing teachings of this invention, those skilled in the art can make various alterations or modifications to this invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A method for constructing a four-dimensional spatiotemporal (4D) cardiac atlas, characterized in that, Including the following steps: (s1) Provide cardiac magnetic resonance imaging (CMR) data, which includes short-axis magnetic resonance images of the heart; (s2) Automated image segmentation and anatomical structure extraction, including: sorting the short-axis magnetic resonance images of the heart in time so that they conform to a complete cardiac cycle; converting the sorted image files into a standard format file that can be processed by a computer; and then inputting the converted standard format file into a pre-trained deep convolutional neural network based on the nnU-Net architecture to achieve fully automated segmentation of the left ventricle, right ventricle and myocardial structure, and converting the two-dimensional image pixel information into a spatial mask with anatomical significance; (s3) Generation and quality control of the three-dimensional heart mesh model, including: converting the segmented mask output into a three-dimensional heart mesh representation, smoothing the mesh, and removing low-quality samples with incomplete segmentation or topological errors through an automated quality control process. (s4) Spatial rigid alignment and standardization of group morphology, including: using the iterative nearest point algorithm to perform spatial standardization processing on the heart mesh of all individuals generated in (s3) so that all heart models are in a unified standard coordinate system; (s5) 3D cardiac atlas construction, including: under the framework of large-scale deformable differential homeomorphic metric mapping, using the Defermetrica method to perform accurate non-rigid registration of the cardiac mesh at end-diastole, and constructing the population average cardiac template at end-diastole (ED) by jointly optimizing the population average shape and the differential homeomorphic deformation mapping from the average shape to the individual heart shape; and simultaneously obtaining the deformation field parameters describing the individual morphological differences, thereby constructing a 3D cardiac statistical atlas at end-diastole; (s6) Construction of 4D cardiac dynamic atlas, including: using the end-diastolic 3D cardiac atlas constructed in (s5) as a unified reference template, and under the condition of keeping the spatial distribution of control points fixed, performing non-rigid registration of cardiac morphology at different time phases within the same cardiac cycle, thereby establishing a multi-temporal three-dimensional cardiac model sequence with consistent topological correspondence within the entire cardiac cycle, thereby constructing the four-dimensional spatiotemporal cardiac atlas.

2. The method as described in claim 1, characterized in that, Step (s5) includes the following sub-step: optimizing a set of control points q and its corresponding momentum μ , vertex coordinates x Convert to vector field X ( x ): (1) in, (2) In the formula, X(x) This represents the vector field generated through transformation, used to convert the original vertex coordinates... x Mapped to the target shape; x These are the vertex coordinates of the heart mesh model; p This represents the total number of control points. K This represents a kernel function used to compute vertices. x With control points q The weight of mutual influence between them; Equation (2) is a Gaussian kernel function used to define the smoothness of the deformation; y For the position variable in the kernel function; in equation (2), y Corresponding to control points q The position of σ; σ is the scale parameter of the kernel function, which determines the range of influence of the deformation, that is, the smoothness of the deformation in space.

3. The method as described in claim 1, characterized in that, In (s6), the following sub-steps are included: (s6a) By performing phase non-rigid registration at time point t of each cardiac cycle, a set of differential homeomorphic transformations is estimated. This makes each phase surface S t Align with reference phase S0: (3)； (s6b) The alignment shape of all subjects at the same time phase is averaged, and the population mean shape at time t is defined as: (4) In the formula, N refers to the number of subjects.

4. A method for constructing a cardiovascular disease diagnostic model based on a four-dimensional spatiotemporal cardiac atlas constructed according to the method of claim 1, characterized in that, Including the following steps: (Z1) Multidimensional deep phenotype extraction, including decoupling two key cardiac phenotypes from the four-dimensional spatiotemporal cardiac atlas, namely morphological phenotype and motor phenotype; the morphological phenotype is the pattern of cardiac morphological variation in the population, and the motor phenotype is a dynamic trajectory feature set reflecting cardiac systolic and diastolic functions; (Z2) Spatiotemporal feature extraction, including: based on step (Z1), constructing a deep learning framework based on PointNet, directly encoding the disordered 3D point cloud of each phase of the heart, and extracting high-level spatial features; and introducing a temporal aggregation module to integrate cross-phase motion patterns, thereby obtaining a unified 4D spatiotemporal feature representation that simultaneously contains heart morphology information and motion information; (Z3) Model construction, including inputting the 4D spatiotemporal features into the LightGBM classifier and the logistic regression model for training and testing, and selecting or outputting the model with the best performance as the cardiovascular disease diagnosis model; The cardiovascular disease diagnostic model is used to diagnose whether the subject currently suffers from a specific cardiovascular disease.

5. The method as described in claim 4, characterized in that, The cardiovascular diseases are selected from the following groups: chronic ischemic heart disease (CHID), atrial fibrillation (AF), angina pectoris (AP), acute myocardial infarction (AMI), cardiac complications and undefined heart disease (I51-HD), other arrhythmias (OCA), or combinations thereof.

6. A method for constructing a cardiovascular disease risk prediction model based on a four-dimensional spatiotemporal (4D) cardiac atlas constructed according to the method of claim 1, characterized in that, Including the following steps: (Z1) Multidimensional deep phenotype extraction, including decoupling two key cardiac phenotypes from the 4D cardiac atlas, namely morphological phenotype and motor phenotype; the morphological phenotype is the normal variation pattern of cardiac morphology in the population, and the motor phenotype is the dynamic trajectory feature set reflecting cardiac systolic and diastolic functions; (Z2) Spatiotemporal feature extraction, including: based on step (Z1), constructing a deep learning framework based on PointNet, directly encoding the disordered 3D point cloud of each phase of the heart, and extracting high-level spatial features; and introducing a temporal aggregation module to integrate cross-phase motion patterns, thereby obtaining a unified 4D spatiotemporal feature representation that simultaneously contains heart morphology information and motion information; (Z3) Model construction includes: inputting the 4D spatiotemporal features and the results of whether the individuals corresponding to the 4D spatiotemporal features will develop corresponding cardiovascular diseases within a future preset time window into a machine learning model for training and testing, thereby constructing a cardiovascular disease risk prediction model; The cardiovascular disease risk prediction model is used to predict whether a subject will develop a specific cardiovascular disease within a preset time window.

7. A device for diagnosing or predicting the risk of cardiovascular diseases, characterized in that, include: (a) An input module configured to input 4D spatiotemporal features of the object to be tested; (b) An assessment module, which is configured to receive the 4D spatiotemporal features and input the 4D spatiotemporal features into a cardiovascular disease diagnostic model constructed using the method of claim 4 or a cardiovascular disease risk prediction model constructed using the method of claim 6 for diagnosis or risk prediction, and obtain an assessment result of whether the subject has a specific cardiovascular disease or whether there is a risk of a specific cardiovascular disease occurring within a future preset time window; (c) Output the evaluation results.

8. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the method of claim 1, 4, or 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of claim 1, 4, or 6.

10. A computer program product comprising computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, they implement the method of claim 1, 4 or 6.

Images