Ar cardiac digital twin method and system for individualized medical treatment

Abstract

This invention proposes an AR-based digital twin method and system for personalized medicine, comprising: acquiring initial structure-aware features of local nodes in each medical institution, extracting corresponding structure-aware feature vectors using a structure-guided encoder, performing local training on the local nodes based on the structure-guided encoder and performing federated aggregation, generating an updated structure-guided encoder to decode the local modal features, generating the final structure-aware features and a set of structure-aware features; mapping the set of structure-aware features to a unified individual latent representation; acquiring a time variable, combining it with the individual latent representation as input, outputting a predicted complete heart structure model at time t, and constructing a periodic guided-structural deformation neural network; acquiring the patient's real-time physiological signals, mapping them to a time variable, combining them with the complete heart structure model, inferring the current twin heart state, and displaying it in a high frame rate rendering on an augmented reality device.

Description

Technical Field

[0001] This invention belongs to the field of digital twins, and particularly relates to an AR cardiac digital twin method and system for personalized medicine. Background Technology

[0002] With the development of personalized medicine, research on constructing patient-specific cardiac models based on digital twin technology has gradually become a key direction in precision medicine. Digital twins, by reconstructing the structure and functional state of individual organs in a virtual space, can achieve disease risk prediction, surgical plan simulation, and postoperative functional assessment, and are particularly suitable for modeling and analyzing organs like the heart, which have complex structures and dramatic dynamic functional changes. However, several technical bottlenecks still exist in practical applications. First, most existing cardiac digital twin systems rely on static, single-modal imaging data (such as CT or MRI) to construct three-dimensional cardiac models, making it difficult to express the dynamic deformation of the patient's heart caused by heartbeat and respiration throughout the complete cardiac cycle, resulting in insufficient temporal consistency and physiological realism in the models. Second, since patient medical images are often scattered across different hospitals or devices, the current centralized processing methods of image data face significant privacy risks and are also limited by "data silos," making it difficult to construct personalized models with broad generalization capabilities. Furthermore, although some studies have attempted to display cardiac digital twin models in augmented reality (AR) environments for preoperative guidance or clinical teaching, most systems remain at the level of static projection of the model, lacking the ability to link with the patient's real-time physiological state (such as electrocardiogram, blood flow velocity, etc.), and thus failing to meet the clinical application needs for real-time interaction and dynamic updates.

[0003] Therefore, current cardiac digital twins still have significant technological gaps in the four core aspects of "personalization, dynamism, privacy protection, and AR visualization interaction," and a systemic breakthrough is urgently needed. Summary of the Invention

[0004] The purpose of this invention is to propose an AR digital twin method and system for personalized medicine, which constructs a complete closed-loop system covering data acquisition, model building, privacy protection and augmented reality interaction, and focuses on solving the limitations of existing technologies in terms of insufficient dynamism, data security, modeling accuracy and real-time interactivity.

[0005] To achieve the above objectives, a first aspect of the present invention provides an AR (Augmented Reality) cardiac digital twin method for personalized medicine, the method comprising the following steps:

[0006] The initial structure-aware features of the local nodes of each medical institution are obtained, and the corresponding structure-aware feature vectors are extracted using a structure-guided encoder. The local nodes are trained locally based on the structure-guided encoder and federated aggregation is performed to generate an updated structure-guided encoder to decode the local modal features, thereby generating the final structure-aware features and structure-aware feature set.

[0007] The structure-aware feature set is mapped to a unified latent representation of individuals;

[0008] The time variable is obtained and combined with the individual latent representation as input to output the complete cardiac structure model at the predicted time t, thus constructing a period-guided-structural deformation neural network.

[0009] The patient's real-time physiological signals are acquired, mapped to time variables, and combined with the complete heart structure model to infer the current state of the twin heart, which is then displayed in a high frame rate rendering on an augmented reality device.

[0010] Furthermore, the structure-guided encoder enhances the feature extraction of the functional areas of the heart chambers and valves by using preset response weights for key cardiac regions, so that the output structure-aware features focus on the key parts of the heart's anatomical structure.

[0011] Further, the step of performing local training and federated aggregation on the local nodes based on the structure-guided encoder includes:

[0012] Obtain the model parameters of the structure-guided encoder, and use them as the encoder parameters for local training at any edge center;

[0013] Global aggregation is performed based on the model parameters to generate aggregated global model parameters.

[0014] Furthermore, mapping the structure-aware feature set to a unified individual latent representation includes:

[0015] Obtain the final structural-aware features of different modalities;

[0016] By using a lightweight linear alignment transformation, all structural-aware features of different modalities are mapped to a unified alignment domain, generating intermediate features in the unified representation space after the corresponding modal mapping.

[0017] Based on the intermediate features in the unified representation space after the corresponding modality mapping, and combined with the structural balance regularization term, the latent encoding corresponding to the intermediate features in the unified representation space after the corresponding modality mapping is obtained; the structural balance regularization term is used to constrain the contribution of all modalities to the latent encoding to not deviate from the overall semantic center.

[0018] Furthermore, the periodic guided-structural deformation neural network is constructed as follows:

[0019] Obtain the resting state of the heart structure as the ground state;

[0020] By combining time variables, a dynamic model is constructed to predict the deformation field of cardiac tissue at different times, generating a continuous dynamic three-dimensional model of the heart.

[0021] Furthermore, the periodic guided-structural deformation neural network is generated by training with supervised data from 4D images;

[0022] The loss function of the periodic guided-structural deformation neural network includes reconstruction error, periodic consistency term, and structural elasticity constraint term. The periodic consistency term is obtained by calculating the periodic consistency loss of the deformation field at different times. The structural elasticity constraint term ensures that the heart tissue maintains elasticity and reasonable topological structure during deformation by penalizing the difference in distance between adjacent points at time t and the resting state, thus avoiding unreasonable deformation.

[0023] Furthermore, the patient's real-time physiological signal is an electrocardiogram or a heart rate variability curve; the augmented reality device is an AR device frame rate controller.

[0024] Furthermore, the process of acquiring the patient's real-time physiological signals, mapping them to time variables, combining them with the complete heart structure model, inferring the current state of the twin heart, and displaying them in a high frame rate rendering on an augmented reality device includes:

[0025] Map the patient's real-time physiological signals at any given time to a normalized time variable;

[0026] The calling process of the dynamic model is controlled by the normalized time variable, and the current three-dimensional structure of the heart is output; wherein, the current three-dimensional structure of the heart is streamed into the AR device frame rate controller;

[0027] The AR device frame rate controller receives the current three-dimensional structure of the heart and supports real-time interactive operation.

[0028] Furthermore, the real-time interactive operations include: spatial positioning and rotation, layered display of anatomical structures, disease prompts, playback function, and multi-user sharing mode.

[0029] A second aspect of the invention provides an AR (Augmented Reality) cardiac digital twin system for personalized medicine, the system comprising:

[0030] The federated feature encoding module is used to obtain the initial structure-aware features of the local nodes of each medical institution, and use the structure-guided encoder to extract the corresponding structure-aware feature vectors. The local nodes are trained locally based on the structure-guided encoder and federated aggregation is performed to generate an updated structure-guided encoder to decode the local modal features and generate the final structure-aware features and structure-aware feature set.

[0031] A semantic alignment module is used to map the structure-aware feature set into a unified latent representation of individuals;

[0032] The dynamic modeling module is used to acquire time variables, combine them with the individual latent representation as input, and output the predicted complete heart structure model at time t, thus constructing a periodic guided-structural deformation neural network.

[0033] The AR interaction module is used to acquire the patient's real-time physiological signals, map them into time variables, combine them with the complete heart structure model, infer the current state of the twin heart, and display them in high frame rate rendering on the augmented reality device.

[0034] The beneficial technical effects of the present invention are at least as follows:

[0035] This invention introduces a dynamic reconstruction mechanism that can express continuous temporal changes in the heart, enabling the digital twin model to realistically reproduce the multi-scale morphological changes of the patient's heart during the cardiac cycle. Addressing the problem of ineffective fusion of multimodal medical images, this invention proposes a cross-modal feature unified mapping mechanism that integrates structural and functional information from different imaging modalities such as CT, MRI, and ultrasound, significantly improving model accuracy. To address patient data privacy leaks and data silos, the system designs a distributed data collaborative modeling architecture, allowing multiple medical institutions to jointly construct cardiac twins without sharing original data, effectively balancing data security and modeling effectiveness. Finally, addressing the lack of dynamic interaction in existing AR systems, this invention establishes a dynamic visualization mechanism driven by physiological signal feedback, enabling the digital twin to respond in real-time to changes in the patient's physiological state within AR glasses, achieving a truly "real-time, dynamic, and personalized" virtual-real interactive presentation of the heart. Through the synergistic effect of these key innovations, this invention constructs an efficient, intelligent, and secure digital twin solution for clinical diagnosis and treatment, preoperative planning, teaching, and remote monitoring. Attached Figure Description

[0036] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0037] Figure 1This is a flowchart of the AR cardiac digital twin method for personalized medicine according to the present invention. Detailed Implementation

[0038] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0039] like Figure 1 As shown in the embodiment of the present invention, an AR-based digital twin method for personalized medicine is provided, the method comprising:

[0040] S1. Obtain the initial structure-aware features of the local nodes of each medical institution, and use the structure-guided encoder to extract the corresponding structure-aware feature vector. Perform local training based on the structure-guided encoder on the local nodes and perform federated aggregation to generate an updated structure-guided encoder to decode the local modal features, and generate the final structure-aware features and structure-aware feature set.

[0041] Specifically, this step aims to address two key challenges currently facing medical imaging data in personalized medicine: first, the heterogeneity of multimodal data (CT, MRI, and ultrasound differ fundamentally in image resolution, tissue imaging, and spatiotemporal distribution); and second, the privacy issue arising from the inability to share patient data across multiple centers. Therefore, this invention proposes a multimodal feature extraction method with a "structure-guided attention mechanism" and "individual difference-aware federated optimization." Its output is the modal feature representation f. m This will serve as the unified individualized latent representation z in step two. p The input basis.

[0042] At each local node of a medical institution, this invention designs a structure-guided feature encoder E for each modality m (e.g., CT, MRI, ultrasound). m (I m ;θ m ), guided by structural diagram A m Enhanced response to key cardiac anatomical regions. The final modal characteristics are represented as follows:

[0043]

[0044] in,

[0045] I m : Input medical images (e.g., CT images or ultrasound images) for modality m;

[0046] E mThe structure-guided encoder on mode m has the parameter θ. m ;

[0047] : The output of the i-th channel convolutional layer under modality m;

[0048] A m (i): The weight of the i-th channel in the structural response graph, used to emphasize areas such as heart chambers and valves;

[0049] C: Total number of channels;

[0050] f m ′: The d-dimensional initial structure-aware feature vector extracted from modality m, used for fusion in step two.

[0051] To address the issue that traditional federated averaging methods do not consider structural differences, this invention proposes an "Individual Difference-Aware Federated Aggregation Mechanism" (IF-FedOpt), whose aggregation objective function is:

[0052]

[0053] in,

[0054] θ m : Encoder parameters completed by the m-th center (hospital) during local training;

[0055] θ: Aggregated global model parameters;

[0056] α m : The structural complexity weight of the m-th center, for example, obtained based on the variance statistics of heart morphology;

[0057] λ: Regularization term weight coefficient;

[0058] Ω div ({θ m}): A regularization term used to maintain the differences between the central models and prevent the feature representation from being over-averaged.

[0059] Furthermore, after all centers have completed local training and federated aggregation, the updated model E is used. m For local mode I m Encode the feature f and output it. m This will serve as the input for modality fusion and unified representation in the next step. This feature possesses:

[0060] Structural sensitivity (focusing on cardiac functional regions); ability to preserve individual differences (guiding individual cardiac expression); cross-modal comparability (facilitating subsequent unification of the latent space z). p Construction); Privacy protection compliance (no risk of raw data leakage).

[0061] S2. Map the structure-aware feature set to a unified individual latent representation.

[0062] Specifically, the task of this step is to process the structure-aware feature set {f} from the output of step one. m} is mapped to a unified individual latent representation z p This provides an expressive foundation for subsequent dynamic cardiac modeling. This invention proposes an individual semantic construction method based on a "modality-aligned distributional embedding" (MAD-Embedding), which has the following two stages:

[0063] I. Multimodal Feature Alignment Transformation (Intramodal Difference Normalization)

[0064] Different modes f m The feature distribution in which it exists exhibits a systematic shift. This invention addresses this by applying a lightweight linear alignment transformation T. m (·) Map the features of all modalities to a unified alignment domain, and let the final feature be... Then we have:

[0065]

[0066] in:

[0067] f m The structure-aware feature vector of modality m output in step one;

[0068] Learnable linear mapping parameters for intramodal distribution alignment;

[0069] : Intermediate features in the unified representation space after modal m mapping.

[0070] This alignment process does not require the original image, but only relies on the f passed in step one. m Its purpose is to solve the problem of misalignment in modal representation space, so that different modalities have a unified structural basis.

[0071] II. Construction of Individual Latent Representations (Aggregation to Generate Latent Representations)

[0072] After all modes have been transformed to the alignment domain, this invention proposes a structure weight-guided aggregation mechanism to generate the latent code z. p Assume each mode has a learnable weight β. m The final individual code is defined as follows:

[0073]

[0074] in:

[0075] β m : The structural contribution weight of modality m is automatically learned from patient samples during the training phase;

[0076] Ψ reg (·): Structural balance regularization term, used to prevent a certain modality from dominating the expression and improve the ability to co-construct multiple modalities;

[0077] μ: Regularization coefficient (e.g., set to 0.1);

[0078] z p The final constructed individualized latent representation is the basis for subsequent modeling of the network G. dyn Input.

[0079] Furthermore, Ψ reg (·) can be designed as in To represent the modal average, constrain all mode pairs z. p The contribution cannot deviate from the overall semantic center;

[0080] Compared to ordinary average fusion or attention mechanisms, this scheme introduces modal feature learnable mapping + regularized guided fusion, which can more accurately construct individual heart morphology expressions.

[0081] Output z p This will be used as the dynamic modeling network G in step three. dyn (z p The unique structure input of z(t) is used to predict the three-dimensional heart state at any time t. p It is generated uniformly after multimodal fusion and retains individual structural differences, thus possessing stronger non-rigid time series prediction capabilities, which aligns with the overall goal of "individual dynamic twin modeling" in this invention.

[0082] S3. Obtain the time variable, combine it with the individual latent representation as input, and output the complete heart structure model at the predicted time t to construct a periodic guided-structural deformation neural network.

[0083] Specifically, the task of this step is to: based on the unified individual latent code z output in step two... p Construct a time-driven cardiac dynamics modeling function G dyn (z p (,t), used to predict the three-dimensional structural state H of a patient's heart at any time t. t .

[0084] The inputs for this step include:

[0085] z p The individualized unified latent representation output in step two incorporates multimodal structural semantic information;

[0086] t: Normalized time variable t∈[0,1], representing a time point within the cardiac cycle (e.g., systole / diastole);

[0087] The output is H t That is, the three-dimensional morphological representation of the heart predicted at time t (which can be a point cloud or a dense grid).

[0088] This invention will be based on an end-to-end function G dyn (z p The prediction is achieved using the function t. This function is not a general neural network, but rather a specialized modeling network proposed in this invention for the heart, a physiological organ with "strong physical structure and periodic motion": Cyclic-Aware Structural Deformation Network (CSD-Net).

[0089] Specifically, the model architecture and main function design:

[0090] This invention will z p The network is jointly input with the time variable t, and a series of time-aware modules are used to predict the deformation field D of cardiac tissue at time t. t This leads to the formation of the heart structure H. t The basic model structure is as follows:

[0091] H t =G dyn (z p ,t)=H0+D t =H0+f θ (z p ,t) (5)

[0092] in,

[0093] H0: A three-dimensional structural reference model of an individual heart in the resting state (accessible via z-axis) p (Decoded);

[0094] f θ (z p ,t): Structural deformation prediction function, outputting the point displacement field D at time t. t ;

[0095] H t The predicted complete heart structure model at time t is used for subsequent AR demonstrations.

[0096] Compared with the traditional direct generation of H t Unlike other methods, this scheme introduces the ground state H0 to ensure a relatively stable, continuous, and controllable deformation process; D tIt can be constructed using multilayer perceptrons and convolutional deformable blocks; during the training phase... Supervised data from 4D imagery; G dyn It is the generation center of the entire patent system. It connects the upstream and downstream of steps two and four, and serves as the intermediate bridge between system data, modeling, and display.

[0097] Furthermore, during the training process, this invention proposes a joint loss function oriented towards the spatiotemporal characteristics of the heart, including a periodic consistency term. With structural elastic constraints Both constitute a complete loss target:

[0098]

[0099] First item: Based on the reconstruction error, align with the real 3D structure;

[0100] The second item (periodic consistency item):

[0101]

[0102] To ensure the continuity of the start and end structure of the cardiac cycle and effectively suppress the problem of "cycle jump";

[0103] The third item (structural elasticity):

[0104]

[0105] in:

[0106] This represents a locally connected region (such as a node on the ventricular wall) defined in the cardiac point cloud.

[0107] : The coordinates of the i-th point at time t;

[0108] : The coordinates of the i-th point in the static ground state;

[0109] This ensures that the local topology is not severely damaged during deformation, especially in the valve boundary region.

[0110] This is a special regularization term introduced in this patent specifically for the "cycle-closed-loop characteristics" of the heart, which does not appear in general 3D modeling; Simulating the physiological elasticity of the myocardium and the rationality of strongly constraining non-rigid deformation, this approach is particularly suitable for dynamic but tightly packed organs like the heart. Compared to traditional image-time-based RNN modeling, this scheme constructs a composite modeling logic of "anatomical structure + physiological attributes," which is a highly patented design.

[0111] In summary, this step starts from the input (structural latent encoding z) p From modeling (time-driven deformation function G) dyn ), to output (structural prediction H at any time) t This forms a closed loop.

[0112] S4. Acquire the patient's real-time physiological signals, map them to time variables, combine them with the complete heart structure model, infer the current twin heart state, and display them in a high frame rate rendering on an augmented reality device.

[0113] Specifically, the core task of this step is to use the dynamic model G trained in step three. dyn (z p The system receives real-time physiological signals from the patient (t), maps them to a time variable t′, and uses this data to infer the current twin heart state H. t′ It is then rendered and displayed at a high frame rate on augmented reality (AR) devices to achieve closed-loop interactive feedback between humans, machines, and pathology.

[0114] Input is: Individual code z p Dynamic model G dyn (z p ,t); real-time acquired physiological signals S(t), such as electrocardiogram (ECG) or heart rate variability curve (HRV) AR device frame rate controller (e.g., supporting 50fps rendering).

[0115] Furthermore, real-time time variable calculation (signal mapping):

[0116] A periodic normalized mapping mechanism, PHI-TMap (Physiological-to-IndexTimeMapping), is proposed to map physiological signals at any time to normalized time variables t′∈[0,1].

[0117]

[0118] in:

[0119] τ: The time interval between the current moment and the previous R peak (in milliseconds);

[0120] T RR Total duration of the current heartbeat cycle;

[0121] t′: Normalized time variable, representing the position of the current physiological rhythm in the modeling time axis.

[0122] Furthermore, dynamic twin generation and AR rendering:

[0123] G is controlled by t′ dyn The call process outputs the current three-dimensional structure of the heart:

[0124] H t′ =G dyn (z p ,t′) (10)

[0125] Variable description:

[0126] z p Patient-specific structural semantic encoding;

[0127] t′: The modeling time index corresponding to the current physiological rhythm;

[0128] H t′ The three-dimensional structural state of the heart is used for projection onto the AR terminal.

[0129] This structure streams data into the AR rendering engine, refreshing 50 times per second to achieve high-fidelity dynamic display.

[0130] Furthermore, AR real-time interactive control:

[0131] AR system with H t′ For input, the following operations are supported:

[0132] Spatial positioning and rotation (view at any angle);

[0133] Layered anatomical structure display (atrial / ventricular / valvular switching);

[0134] Disease indications (such as color-changing markings in valvular regurgitation areas);

[0135] Replay function (replays the heart status of the past 5 cycles);

[0136] Multi-user sharing mode (doctors and patients can view the same model simultaneously).

[0137] This invention also provides an AR (Augmented Reality) cardiac digital twin system for personalized medicine, the system comprising:

[0138] The federated feature encoding module is used to obtain the initial structure-aware features of the local nodes of each medical institution, and use the structure-guided encoder to extract the corresponding structure-aware feature vectors. The local nodes are trained locally based on the structure-guided encoder and federated aggregation is performed to generate an updated structure-guided encoder to decode the local modal features and generate the final structure-aware features and structure-aware feature set.

[0139] A semantic alignment module is used to map the structure-aware feature set into a unified latent representation of individuals;

[0140] The dynamic modeling module is used to acquire time variables, combine them with the individual latent representation as input, and output the predicted complete heart structure model at time t, thus constructing a periodic guided-structural deformation neural network.

[0141] The AR interaction module is used to acquire the patient's real-time physiological signals, map them into time variables, combine them with the complete heart structure model, infer the current state of the twin heart, and display them in high frame rate rendering on the augmented reality device.

[0142] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0143] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0144] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0145] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An AR-based digital twin method for the heart in personalized medicine, characterized in that, The method includes the following steps: The initial structure-aware features of the local nodes of each medical institution are obtained, and the corresponding structure-aware feature vectors are extracted using a structure-guided encoder. The local nodes are trained locally based on the structure-guided encoder and federated aggregation is performed to generate an updated structure-guided encoder to decode the local modal features, thereby generating the final structure-aware features and structure-aware feature set. The structure-aware feature set is mapped to a unified latent representation of individuals; The time variable is obtained, combined with the individual's latent representation, and used as input to output the predicted time. A complete cardiac structure model was constructed, and a periodic guided-structural deformation neural network was built. The patient's real-time physiological signals are acquired, mapped to time variables, and combined with the complete heart structure model to infer the current state of the twin heart, which is then displayed in a high frame rate rendering on an augmented reality device. Wherein, mapping the structure-aware feature set into a unified individual latent representation includes: Obtain the final structural-aware features of different modalities; By using a lightweight linear alignment transformation, all structural-aware features of different modalities are mapped to a unified alignment domain, generating intermediate features in the unified representation space after the corresponding modal mapping. Based on the intermediate features in the unified representation space after the corresponding modality mapping, and combined with the structural balance regularization term, the latent encoding corresponding to the intermediate features in the unified representation space after the corresponding modality mapping is obtained; the structural balance regularization term is used to constrain the contribution of all modalities to the latent encoding to not deviate from the overall semantic center. The periodic guided-structural deformation neural network is constructed as follows: Obtain the resting state of the heart structure as the ground state; By combining time variables, a dynamic model is constructed to predict the deformation field of cardiac tissue at different times, generating a continuous dynamic three-dimensional model of the heart. The periodic guided-structural deformation neural network is generated by training with supervised data from 4D images; The loss function of the periodic guided-structural deformation neural network includes reconstruction error, periodic consistency term, and structural elasticity constraint term. The periodic consistency term is obtained by calculating the periodic consistency loss of the deformation field at different times. The structural elasticity constraint term ensures that the heart tissue maintains elasticity and reasonable topological structure during deformation by penalizing the difference in distance changes between adjacent points at time t and the resting state, thus avoiding unreasonable deformation.

2. The AR cardiac digital twin method for personalized medicine according to claim 1, characterized in that, The structure-guided encoder enhances the feature extraction of the functional areas of the heart chambers and valves by using preset response weights for key areas of the heart, so that the output structure-aware features focus on the key parts of the heart's anatomical structure.

3. The AR cardiac digital twin method for personalized medicine according to claim 1, characterized in that, The step of performing local training and federated aggregation on the local nodes based on the structure-guided encoder includes: Obtain the model parameters of the structure-guided encoder, and use them as the encoder parameters for local training at any edge center; Global aggregation is performed based on the model parameters to generate aggregated global model parameters.

4. The AR cardiac digital twin method for personalized medicine according to claim 1, characterized in that, The patient's real-time physiological signals are electrocardiograms or heart rate variability curves; the augmented reality device is an AR device frame rate controller.

5. The AR cardiac digital twin method for personalized medicine according to claim 4, characterized in that, The process of acquiring the patient's real-time physiological signals, mapping them to a time variable, combining them with the complete heart structure model, inferring the current state of the twin heart, and displaying them in a high frame rate rendering on an augmented reality device includes: Map the patient's real-time physiological signals at any given time to a normalized time variable; The calling process of the dynamic model is controlled by the normalized time variable, and the current three-dimensional structure of the heart is output; wherein, the current three-dimensional structure of the heart is streamed into the AR device frame rate controller; The AR device frame rate controller receives the current three-dimensional structure of the heart and supports real-time interactive operation.

6. The AR cardiac digital twin method for personalized medicine according to claim 5, characterized in that, The real-time interactive operations include: spatial positioning and rotation, layered display of anatomical structures, disease prompts, playback function, and multi-user sharing mode.

7. A system for implementing the AR cardiac digital twin method for personalized medicine as described in claim 1, characterized in that, The system includes: The federated feature encoding module is used to obtain the initial structure-aware features of the local nodes of each medical institution, and use the structure-guided encoder to extract the corresponding structure-aware feature vectors. The local nodes are trained locally based on the structure-guided encoder and federated aggregation is performed to generate an updated structure-guided encoder to decode the local modal features and generate the final structure-aware features and structure-aware feature set. A semantic alignment module is used to map the structure-aware feature set into a unified latent representation of individuals; The dynamic modeling module is used to acquire time variables, combine them with the individual latent representations as input, and output the predicted time. A complete cardiac structure model was constructed, and a periodic guided-structural deformation neural network was built. The AR interaction module is used to acquire the patient's real-time physiological signals, map them into time variables, combine them with the complete heart structure model, infer the current state of the twin heart, and display them in high frame rate rendering on the augmented reality device.

Images