Cardiovascular system simulation modeling method for left ventricular remodeling

Abstract

The invention provides a cardiovascular system simulation modeling method for left ventricle reconstruction. The cardiovascular system simulation modeling method aims at achieving the integrated process of structure extraction, three-dimensional reconstruction and simulation modeling under driving of a multi-modal medical image. According to the method, a dual-path structure difference fusion model is constructed, semantic information and detail features are extracted respectively, and a hierarchical difference feedback mechanism is introduced to improve segmentation precision; in the three-dimensional modeling stage, a structure grid consistent with topology is generated through a topology interaction rule and occupancy prediction, and a topology maintenance loss function is introduced to optimize a reconstruction result; further combining individualized fiber direction and regional parameters, constructing a finite element model with spatial attributes, and realizing joint modeling of a left ventricle structure and a response behavior; the method improves the accuracy of complex structure recognition and the geometric consistency of modeling, and is suitable for the fields of image processing, three-dimensional modeling, engineering simulation and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a cardiovascular system simulation modeling method for left ventricular remodeling. Background Art

[0002] As a complex three-dimensional organ, the heart's motion state and morphological changes have obvious time dependence and tissue heterogeneity. With the development of medical imaging technology, multimodal imaging methods such as magnetic resonance imaging (MRI), computed tomography (CT), and echocardiography (Echo) have been widely used in the visualization and anatomical reconstruction of cardiac structure. In order to digitally model and dynamically analyze cardiac structure, image segmentation, three-dimensional reconstruction, fiber orientation modeling, and finite element simulation have become key technical steps in this field. In existing technologies, cardiac image segmentation methods mainly rely on traditional threshold methods, region growing methods, or single-path convolutional neural network models, which cannot fully express the boundary differences and structural hierarchical relationships between multimodal images, resulting in low accuracy and stability in anatomical structure extraction. In addition, three-dimensional modeling of myocardial tissue mostly uses surface reconstruction or mesh splicing, which lacks topological consistency constraints and is prone to introducing artifacts and geometric defects, affecting the continuity and anatomical rationality of subsequent models.

[0003] In terms of simulation modeling, finite element models are usually constructed based on ideal geometry or regular grids, and the fiber direction is mostly set using empirical formulas. There is a lack of linkage with individual data, making it difficult to truly reflect the mechanical heterogeneity and spatial variability of myocardial tissue in different regions. In addition, most existing modeling processes do not fully integrate the error feedback mechanism of image segmentation and structural modeling, and cannot achieve closed-loop optimization of the segmentation-reconstruction-simulation process.

[0004] In summary, existing cardiac image modeling technologies generally have problems such as insufficient structural segmentation accuracy, poor reconstruction topology consistency, single fiber direction modeling, and static simulation process. There is an urgent need for a complete process that combines multimodal image analysis and simulation modeling to improve the accuracy of cardiac structure expression and the credibility of simulation analysis. Summary of the Invention

[0005] The present invention proposes a cardiovascular system simulation modeling method for left ventricular reconstruction, aiming to construct a structurally accurate, topologically consistent, and fiber-oriented three-dimensional model of the left ventricle from multimodal medical images, and to carry out high-reliability simulation analysis based on this. The core innovation lies in proposing a dual-path structural difference fusion model and a topology-preserving three-dimensional reconstruction mechanism to achieve high-precision extraction and topological consistency modeling of key cardiac structures in multimodal medical images. The proposed dual-path model is collaboratively modeled through semantic enhancement and high-frequency detail recovery, and a hierarchical difference feedback mechanism is introduced to improve the expression accuracy of anatomical boundaries and the robustness of structural recognition. In the three-dimensional modeling stage, a topology-preserving loss function is introduced based on topological interaction rules and occupancy prediction to effectively ensure the structural continuity and geometric rationality of the myocardial model. Further, combined with individualized fiber direction modeling and regional heterogeneity parameter setting, the fine construction and mechanical response simulation of the left ventricular finite element model are achieved, forming a complete image-driven cardiovascular modeling and simulation analysis process.

[0006] The present invention provides a cardiovascular system simulation modeling method for left ventricular remodeling, the method comprising the following steps:

[0007] Step S1: Acquire MRI, CT, or echocardiographic image sequences in the original DICOM format, select the resting and systolic phase sequences of each image, align the spatial structure, and unify the anatomical reference system; in terms of time, synchronize the different modal images to the key frames of the same cardiac cycle using ECG gating synchronization technology to obtain a multimodal cardiac image dataset;

[0008] Step S2: Construct a dual-path structural difference fusion model, process the multimodal cardiac image dataset through the dual-path structural difference fusion model to perform automatic semantic segmentation on the image, extract the structural data of the left ventricle, myocardium, aortic root and coronary artery entrance, and obtain cardiac structure mask data; the dual-path structural difference fusion model includes a data preprocessing unit, a feature modeling unit, a difference enhancement unit, a fusion prediction unit, a mask generation unit, an encoder and a dual-branch decoder; the dual-branch decoder includes a DC path and a DelPU path;

[0009] Step S3: constructing a myocardial three-dimensional reconstruction algorithm, combining the cardiac structure mask data, processing the cardiac structure mask data through the myocardial three-dimensional reconstruction algorithm, and generating a 3D model of the myocardial region;

[0010] Step S4: Based on the 3D model of the myocardial region, the true fiber direction is reconstructed using DTI-MRI data, the myocardial spiral angle distribution is defined using the Streeter model, the myocardial layered structure is established, and the fiber tensor field is obtained;

[0011] Step S5: Based on the 3D model of the myocardial region and the fiber tensor field, a finite element mesh model is constructed. Regional heterogeneity parameters are set to distinguish the infarct area, the edge area, and the normal area. The fiber direction is mapped to generate a ventricular finite element model including fiber orientation and mechanical properties.

[0012] Step S6: Based on the ventricular finite element model, set the ventricular cavity pressure boundary conditions, simulate the complete cardiac cycle, obtain the stress, strain and wall motion distribution of the left ventricle at different time points, and obtain the simulation results;

[0013] Step S7: Based on the simulation results, simulate the long-term left ventricular deformation evolution process and evaluate the reconstruction trend and functional change indicators.

[0014] Furthermore, the process of processing the multimodal cardiac image dataset through the dual-path structural difference fusion model to obtain cardiac structure mask data specifically includes the following:

[0015] The data preprocessing unit slides and slices the 3D image sequences in the multimodal cardiac image dataset into image blocks of 16×64×64 size. Structural labels, including the left ventricular cavity, myocardium, aortic root, and coronary artery inlet, are added to the annotated images. Unlabeled images are marked as "unsupervised samples" and their spatial distribution is normalized to obtain an image block queue.

[0016] The feature modeling unit inputs the image block queue into the encoder to extract spatial-semantic embedding features. The DC path (semantic enhancement path) uses deep transposed convolution to restore the global structural information of the spatial-semantic embedding features layer by layer, enhancing semantic consistency. The DelPU path (detail recovery path) constructs a Laplacian pyramid to perform hierarchical modeling of the spatial-semantic embedding features, capturing high-frequency details and enhancing the ability to express cardiac structural boundaries and microstructures.

[0017] The difference enhancement unit introduces a hierarchical difference learning mechanism to construct the difference tensors of the feature maps corresponding to the DC path and the DelPU path, calculate the feature residuals between the decoder outputs layer by layer, and feed the feature residuals back to the encoder, thereby realizing dynamic modeling and adaptive enhancement control of multi-scale structural differences. This mechanism effectively captures the boundary differences and local microstructural features of the left ventricle, myocardium, and aortic root tissue in multimodal data, improving the structural expression ability of the segmentation model and the accuracy of anatomical region recognition. It also obtains the dual-pathway structural prediction distribution, including the DC path structure prediction probability map and the DelPU path structure prediction probability map.

[0018] The fusion prediction unit combines the dual-path structure prediction distribution and introduces the η-mixable loss function to perform differential analysis on the dual-path structure prediction distribution. It generates nonlinear weighting coefficients through soft distance mapping and adopts a fixed-share update strategy to fuse the current and historical path outputs to form a fused structure prediction graph. To ensure the robustness of this distributed weighted fusion strategy under different modal and structural states, a dynamic regret bound based on the Mixability theory is introduced to constrain the upper limit of the cumulative fusion error of the fused structure prediction graph and reduce its feature dimension dependence to obtain an optimized structure prediction graph.

[0019] The mask generation unit combines the optimized structure prediction map and performs argmax operation to generate a hard mask, retaining the main structures including the left ventricle, myocardium, aorta, and coronary artery entrance, extracting credible areas, removing small artifacts, and generating cardiac structure mask data.

[0020] Furthermore, the process of processing the cardiac structure mask data using the myocardial 3D reconstruction algorithm to generate a 3D model of the myocardial region specifically includes the following steps:

[0021] Step S31: constructing a voxel-level label map based on the cardiac structure mask data, extracting the initial structural surface mesh of the voxel-level label map using the Marching Cubes algorithm, and then performing Laplacian smoothing and topological repair to generate a preliminary repaired structural mesh;

[0022] Step S32: Preset a topological interaction rule set, jointly input the preliminary repair structure mesh and the voxel-level label map, perform pair-by-pair detection on the inclusion and exclusion relationships between structures based on the preset topological interaction rule set, perform spatial neighborhood convolution using a 6-connected voxel kernel, identify topological violation areas, and obtain a topological violation point cloud;

[0023] Step S33: Using the topological violation point cloud as a constraint source, the spatial occupancy state in the preliminary repair structure mesh is determined, and a positive and negative sample set between structures is constructed based on the topological interaction rule set. A topology preservation loss function is introduced to jointly supervise the topological consistency and anatomical rationality of the preliminary repair structure mesh, and the optimized structure mesh is output.

[0024] Step S34: performing global mesh reconstruction and region filling operations in combination with the optimized structural mesh to construct a 3D model of the myocardial region.

[0025] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0026] The present invention achieves high-precision extraction and semantic modeling of the left ventricle and its related structures in multimodal images, improving the structural expression ability and image segmentation stability of complex anatomical regions; through the dual-path structural difference fusion mechanism, it effectively integrates semantic context information and high-frequency boundary details, and solves the problems of boundary blur and structural confusion in the multimodal fusion process of conventional image processing methods, providing high-quality segmentation output for subsequent three-dimensional modeling.

[0027] The present invention adopts a topology-preserving reconstruction mechanism to perform three-dimensional modeling and mesh repair on the structural mask, introduces topological consistency constraints and occupancy prediction models, effectively solves problems such as geometric discontinuity, topological errors and local distortion in the traditional reconstruction process, and significantly enhances the model's expressive ability in terms of spatial integrity and structural coherence; this mechanism ensures the model's reconstruction reliability when facing complex geometric shapes and anatomical structures, and is suitable for fine structure modeling tasks in multiple scenarios.

[0028] Furthermore, the present invention constructs a finite element simulation model with configurable regional attributes, integrates fiber direction reconstruction and parametric heterogeneity modeling technology, and achieves highly consistent expression of structure-attribute joint modeling; by unifying image processing, structural modeling and response simulation processes, it improves the efficiency and accuracy of complex structural system modeling and dynamic analysis, and enhances its application value in engineering fields such as virtual modeling, digital entity construction, and structural simulation evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is the stress gradient distribution diagram proposed in Example 6;

[0030] Figure 2 This is the three-cycle motion trajectory diagram of the apex proposed in Example 6.

[0031] Figure 1 In the figure, the principal stress gradient distribution of the stress gradient distribution diagram (t=800 ms) is shown, where red indicates high stress and yellow indicates low stress;

[0032] Figure 2 In the figure, the starting point (green) and the ending point (red) mark the beginning of the first cycle and the end of the last cycle; the fluctuation of the curve reflects the dynamic changes of the apex during the periodic contraction and relaxation. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0034] In a first embodiment, the present invention provides a cardiovascular system simulation modeling method for left ventricular remodeling, the method comprising the following steps:

[0035] Step S1: Acquire MRI, CT, or echocardiographic image sequences in the original DICOM format, select the resting and systolic phase sequences of each image, align the spatial structure, and unify the anatomical reference system; in terms of time, synchronize the different modal images to the key frames of the same cardiac cycle using ECG gating synchronization technology to obtain a multimodal cardiac image dataset;

[0036] Step S2: Construct a dual-path structural difference fusion model, process the multimodal cardiac image dataset through the dual-path structural difference fusion model to perform automatic semantic segmentation on the image, extract the structural data of the left ventricle, myocardium, aortic root and coronary artery entrance, and obtain cardiac structure mask data; the dual-path structural difference fusion model includes a data preprocessing unit, a feature modeling unit, a difference enhancement unit, a fusion prediction unit, a mask generation unit, an encoder and a dual-branch decoder; the dual-branch decoder includes a DC path and a DelPU path;

[0037] Step S3: constructing a myocardial three-dimensional reconstruction algorithm, combining the cardiac structure mask data, processing the cardiac structure mask data through the myocardial three-dimensional reconstruction algorithm, and generating a 3D model of the myocardial region;

[0038] Step S4: Based on the 3D model of the myocardial region, the true fiber direction is reconstructed using DTI-MRI data, the myocardial spiral angle distribution is defined using the Streeter model, the myocardial layered structure is established, and the fiber tensor field is obtained;

[0039] Step S5: Based on the 3D model of the myocardial region and the fiber tensor field, a finite element mesh model is constructed. Regional heterogeneity parameters are set to distinguish the infarct area, the edge area, and the normal area. The fiber direction is mapped to generate a ventricular finite element model including fiber orientation and mechanical properties.

[0040] Step S6: Based on the ventricular finite element model, set the ventricular cavity pressure boundary conditions, simulate the complete cardiac cycle, obtain the stress, strain and wall motion distribution of the left ventricle at different time points, and obtain the simulated mechanical response;

[0041] Step S7: Based on the simulated mechanical response, the long-term left ventricular deformation evolution process is simulated to evaluate the remodeling trend and functional change indicators.

[0042] Example 2: This example is based on Example 1. In this example, a process of processing a multimodal cardiac image dataset using a dual-path structural difference fusion model to obtain cardiac structure mask data specifically includes the following:

[0043] The data preprocessing unit slides and slices the 3D image sequences in the multimodal cardiac image dataset into image blocks of 16×64×64 size. Structural labels, including the left ventricular cavity, myocardium, aortic root, and coronary artery inlet, are added to the annotated images. Unlabeled images are marked as "unsupervised samples" and their spatial distribution is normalized to obtain an image block queue.

[0044] The feature modeling unit inputs the image block queue into the encoder to extract spatial-semantic embedding features. The DC path (semantic enhancement path) uses deep transposed convolution to restore the global structural information of the spatial-semantic embedding features layer by layer, enhancing semantic consistency. The DelPU path (detail recovery path) constructs a Laplacian pyramid to perform hierarchical modeling of the spatial-semantic embedding features, capturing high-frequency details and enhancing the ability to express cardiac structural boundaries and microstructures.

[0045] The difference enhancement unit introduces a hierarchical difference learning mechanism to construct the difference tensors of the feature maps corresponding to the DC path and the DelPU path respectively, calculates the feature residuals between the decoder outputs layer by layer, and feeds the feature residuals back to the encoder, thereby realizing dynamic modeling and adaptive enhancement control of multi-scale structural differences; this mechanism effectively captures the boundary differences and local microstructural features of the left ventricle, myocardium, and aortic root tissue in multimodal data, improving the structural expression ability of the segmentation model and the accuracy of anatomical region recognition; the dual-pathway structure prediction distribution is obtained, including the DC path structure prediction probability map and the DelPU path structure prediction probability map, and the formula used is as follows:

[0046] Difference tensor calculation formula:

[0047] ;

[0048] in, Represents the index of the current decoding layer, Represents the index of the current image block, Indicates the Layer, The difference tensor at iterations; Indicates the DC path Layer decoding operations, Indicates the first Layer decoding operations; Indicates the first Layer, The feature map output by the iteration is Indicates the first Layer, The feature map output by the iteration; Indicates that the The DC feature map of the layer is decoded and restored to the layer scale; Indicates that the The DelPU feature map of the layer is decoded and restored to the layer scale;

[0049] The fusion prediction unit combines the dual-path structure prediction distribution and introduces the η-mixable loss function to perform differential analysis on the dual-path structure prediction distribution. Nonlinear weighting coefficients are generated through soft distance mapping, and a fixed-share fixed sharing update strategy is adopted to fuse the current and historical path outputs to form a fused structure prediction graph. To ensure the robustness of this distributed weighted fusion strategy under different modal and structural states, a dynamic regret bound based on the Mixability theory is introduced to constrain the upper limit of the cumulative fusion error of the fused structure prediction graph and reduce its feature dimension dependence to obtain the optimized structure prediction graph. The formula used is as follows:

[0050] Dynamic regret bound formula:

[0051] ;

[0052] in, represents the total time step, Expressing dynamic regret, represents the dimension of the feature space, represents the logarithmic term, represents the asymptotic upper bound symbol; represents the influence of time growth, Indicates the impact item of path change;

[0053] This product term is a non-stationary control factor in Mixability theory, used to characterize the coupling relationship between the stability and drift amplitude of the predicted path as it evolves over time. It also serves as an adaptive error tolerance threshold in the fusion strategy. Based on this control mechanism, the present invention introduces a soft distance mapping mechanism and a fixed-share fixed sharing update strategy, which automatically biases the path fusion weight toward the current optimal performance path under conditions of a high number of path changes. At the same time, the overall structural consistency is maintained through exponential weighting of historical path distribution, preventing the model from falling into performance degradation caused by local drift.

[0054] In the η-mixable loss function, η is the mixability coefficient, which controls the curvature or combination strength of the loss; mixable means that the loss function satisfies the mixability condition;

[0055] The formula is as follows:

[0056] ;

[0057] in, represents the miscibility coefficient, Represents the path index, represents the total number of candidate paths, represents the structure prediction result after fusion, Indicates the The output result of the structure prediction path; represents the loss value of η-mixable, Indicates the The weight of a path, Indicates the The path has its individual loss value The exponentially weighted penalty term under ;

[0058] The fixed-share update strategy is a weighted update mechanism based on online learning and expert advice. It aims to balance new and historical information in non-stationary environments or path drift, ensuring the model's short-term adaptability and long-term stability.

[0059] The mask generation unit combines the optimized structure prediction map and performs argmax operation to generate a hard mask, retaining the main structures including the left ventricle, myocardium, aorta, and coronary artery entrance, extracting credible areas, removing small artifacts, and generating cardiac structure mask data.

[0060] Example 3: This example is based on Example 1. In this example, the process of processing a multimodal cardiac image dataset to obtain cardiac structure mask data specifically includes the following:

[0061] The data preprocessing unit slides and slices the 3D image sequences in the multimodal cardiac image dataset into image blocks of 16×64×64 size. Structural labels, including the left ventricular cavity, myocardium, aortic root, and coronary artery inlet, are added to the annotated images. Unlabeled images are marked as "unsupervised samples" and their spatial distribution is normalized to obtain an image block queue.

[0062] The feature modeling unit inputs the image block queue into the encoder to extract the spatial-semantic embedding features; the decoder extracts the information of the spatial-semantic embedding features and generates the path structure prediction distribution;

[0063] The mask generation unit combines the path structure prediction distribution and performs argmax operation to generate a hard mask, retaining the main structures including the left ventricle, myocardium, aorta, and coronary artery entrance, extracting credible areas, removing small artifacts, and generating cardiac structure mask data.

[0064] Example 4: This example is based on Example 2. In this example, the process of processing cardiac structure mask data using a myocardial 3D reconstruction algorithm to generate a 3D model of the myocardial region specifically includes the following steps:

[0065] Step S31: constructing a voxel-level label map based on the cardiac structure mask data, extracting the initial structural surface mesh of the voxel-level label map using the Marching Cubes algorithm, and then performing Laplacian smoothing and topological repair to generate a preliminary repaired structural mesh;

[0066] Step S32: Preset a topological interaction rule set, jointly input the preliminary repair structure mesh and the voxel-level label map, perform pair-by-pair detection on the inclusion and exclusion relationships between structures based on the preset topological interaction rule set, perform spatial neighborhood convolution using a 6-connected voxel kernel, identify topological violation areas, and obtain a topological violation point cloud;

[0067] Step S33: Using the topological violation point cloud as a constraint source, the spatial occupancy state in the preliminary repair structure grid is determined, and a positive and negative sample set between structures is constructed based on the topological interaction rule set. A topology preservation loss function is introduced to jointly supervise the topological consistency and anatomical rationality of the preliminary repair structure grid, and the optimized structure grid is output. The formula used is as follows:

[0068] Topology preservation loss function formula:

[0069] ;

[0070] in, represents the topology preservation loss, Indicates the number of query points, including positive and negative sample sets, represents the query point, extracted from the topological violation point cloud; Indicates the point The predicted occupancy probability, Indicates a point The true occupancy label, represents the logarithmic loss term predicted as a positive sample, Represents the logarithmic loss term predicted as a negative sample;

[0071] Step S34: performing global mesh reconstruction and region filling operations in combination with the optimized structural mesh to construct a 3D model of the myocardial region.

[0072] Example 5: This example is based on Example 2. In this example, the process of processing cardiac structure mask data and generating a 3D model of the myocardial region specifically includes the following steps:

[0073] Step R1: Initial mesh extraction: Based on the cardiac structure mask data, a 3D voxel label map is constructed; the structural surface mesh corresponding to the voxel map is extracted using the Marching Cubes algorithm to obtain the initial myocardial structure mesh model;

[0074] Step R2: Mesh smoothing and repair: Laplacian smoothing is performed on the initial myocardial structure mesh model to eliminate jagged edges and local noise. A topological repair algorithm based on geometric rules is then used to repair and streamline small-area holes, non-manifold boundaries, and redundant triangles to obtain a smoothly connected myocardial structure mesh.

[0075] Step R3: Mesh quality assessment and optimization: Perform geometric consistency assessment on the smoothly connected myocardial structure mesh, including normal consistency, curvature continuity, and triangle density assessment; generate a quality-optimized mesh structure;

[0076] Step R4: Three-dimensional structural model reconstruction: Combined with the quality-optimized grid structure, a regular interpolation and region filling algorithm is used to reconstruct a complete three-dimensional model of the myocardial region to construct a 3D model of the myocardial region.

[0077] Example 6: This example is based on Example 5.

[0078] Step S6: Based on the ventricular finite element model, set the ventricular cavity pressure boundary conditions, simulate the complete cardiac cycle, obtain the stress, strain and wall motion distribution of the left ventricle at different time points, and obtain the simulation results;

[0079] Simulation results:

[0080] Stress distribution: At the end of cardiac cycle (t=800 ms), the average principal stress of the inner layer of the left ventricular wall is 16.2 kPa, and that of the outer layer is 8.7 kPa. The stress gradient distribution diagram is shown in the attached figure. Figure 1 As shown;

[0081] Strain and motion analysis:

[0082] Maximum apical displacement amplitude: 12.1 mm;

[0083] Peak radial strain in the middle layer of the ventricular wall: +18.5%; circumferential strain: −22.7%;

[0084] The simulated EF (ejection fraction) value was 56.3%, which was in good agreement with the MRI measurement value (54.8%).

[0085] The trajectory of the apex's three cycles is shown in the attached figure. Figure 2 As shown;

[0086] Step S7: Based on the simulation results, simulate the long-term left ventricular deformation evolution process and evaluate the remodeling trend and functional change indicators;

[0087] Simulating the long-term evolution of left ventricular deformation:

[0088] Simulation period: 1 heartbeat per cycle (T = 800ms), with a total simulation of 180,000 time steps (representing 6 months of cumulative stress-induced changes);

[0089] Simulation results:

[0090] Left ventricular end-diastolic volume (EDV) increased by 17.2%;

[0091] The overall ventricular wall thickness decreased by 12.5% ​​in the non-infarcted area and increased by 3.8% in the border area;

[0092] EF decreased to 47.9%, indicating weakened systolic function;

[0093] The apical area of ​​the heart develops a fan-shaped bulge and deformation.

[0094] The present invention and its embodiments are described above. Such description is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in this field are inspired by it and do not depart from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. A cardiovascular system simulation modeling method for left ventricular remodeling, characterized by: The method comprises the following steps: Step B1: Acquire a multimodal cardiac image dataset; Step B2: Constructing a dual-path structural difference fusion model, processing the multimodal cardiac image dataset through the dual-path structural difference fusion model to obtain cardiac structure mask data; the dual-path structural difference fusion model includes a data preprocessing unit, a feature modeling unit, a difference enhancement unit, a fusion prediction unit, a mask generation unit, an encoder, and a dual-branch decoder; Step B3: Construct a myocardial 3D reconstruction algorithm, process the cardiac structure mask data through the myocardial 3D reconstruction algorithm, and generate a 3D model of the myocardial region.

2. A cardiovascular system simulation modeling method for left ventricular remodeling according to claim 1, characterized in that: The dual-branch decoder includes a DC path and a DelPU path.

3. A cardiovascular system simulation modeling method for left ventricular remodeling according to claim 2, characterized in that: The data preprocessing unit slides and slices the three-dimensional image sequence in the multimodal cardiac image data set to obtain an image block queue.

4. A cardiovascular system simulation modeling method for left ventricular remodeling according to claim 3, characterized in that: The feature modeling unit inputs the image block queue into the encoder and extracts spatial-semantic embedding features; The DC path recovers the global structural information of spatial-semantic embedding features layer by layer through deep transposed convolution; the DelPU path constructs a Laplacian pyramid to hierarchically model the spatial-semantic embedding features and capture high-frequency details.

5. The cardiovascular system simulation modeling method for left ventricular remodeling according to claim 4, characterized in that: The difference enhancement unit introduces a hierarchical difference learning mechanism to construct the difference tensors of the DC path and the DelPU path respectively, obtain the feature residuals, and feed the feature residuals back to the encoder to obtain the dual-path structure prediction distribution.

6. The cardiovascular system simulation modeling method for left ventricular remodeling according to claim 5, characterized in that: The fusion prediction unit introduces the η-mixable loss function to perform differential analysis on the dual-path structure prediction distribution, generates nonlinear weighting coefficients through soft distance mapping, and adopts a fixed-share fixed sharing update strategy for fusion to form a fused structure prediction graph. A dynamic regret bound based on the Mixability theory is introduced to constrain the upper limit of the cumulative fusion error of the fused structure prediction graph, reduce the dependency on feature dimensions, and obtain an optimized structure prediction graph.

7. A cardiovascular system simulation modeling method for left ventricular remodeling according to claim 6, characterized in that: The mask generation unit combines the optimized structure prediction map to generate cardiac structure mask data.

8. The cardiovascular system simulation modeling method for left ventricular remodeling according to claim 1, characterized in that: The process of generating a 3D model of the myocardial region includes the following steps: Step B31: constructing a voxel-level label map based on the cardiac structure mask data and generating a preliminary repair structure mesh; Step B32: The preliminary repaired structure grid and the voxel-level label map are jointly input for pair-wise detection. A 6-connected voxel kernel is used for spatial neighborhood convolution to identify topological violation areas and obtain a topological violation point cloud. Step B33: Use the topological violation point cloud as a constraint source to determine the spatial occupancy state of the preliminary repair structure mesh; introduce a topology preservation loss function to jointly supervise the topological consistency and anatomical rationality of the preliminary repair structure mesh, and output the optimized structure mesh; Step B34: Combine the optimized structural mesh to construct a 3D model of the myocardial region.

Images