Method for three-dimensional visualization of heart scar, control device and storage medium

Abstract

The invention provides a method for three-dimensional visualization of heart scars, a control device and a storage medium, and belongs to the technical field of three-dimensional heart modeling and visual display. The method comprises the following steps: matching a heart structure obtained by segmenting a pre-trained nnU-Net and a binary tag graph corresponding to a scar tissue with a myocardial shape shown by the binary tag graph of the heart structure by using a standard myocardial grid model as a template; and fitting deformation parameters of a standard myocardial grid model by using a deep learning regression model, and applying the deformation parameters to a standard grid to complete grid deformation so as to fit individual myocardial morphology. The heart shape prior is effectively utilized based on the template deformation method, so that the reconstructed three-dimensional model has higher anatomical consistency and fineness. And superposing the fitted three-dimensional myocardial grid model and the reconstructed scar tissue grid under the same coordinate for registration and superposition to realize image three-dimensional transparent myocardial scar visualization so as to accurately restore a continuous three-dimensional scar surface.

Description

Technical Field

[0001] This application belongs to the field of three-dimensional cardiac modeling and visualization technology, and specifically relates to a method, control device and storage medium for three-dimensional visualization of cardiac scars. Background Technology

[0002] Late-onset gadolinium-enhanced cardiac magnetic resonance imaging (LGE-CMR) is an important imaging technique for qualitative and quantitative assessment of myocardial scarring or fibrosis, reliably detecting myocardial infarction or fibrosis. However, in clinical practice, multiple two-dimensional LGE slices are often used to obtain whole-heart images. These slices are typically 6–8 mm thick and spaced 3–5 mm apart, and patient respiration or heartbeat causes motion artifacts between different slices. Therefore, two-dimensional LGE images have anisotropic resolution, and scar structures within the myocardium are often broken or distorted in different slices. Direct three-dimensional reconstruction based on sparse slices will produce severe step effects and morphological distortions.

[0003] In related technologies, 2D LGE images are segmented layer by layer or interpolated using voxels before 3D reconstruction. However, it is difficult to accurately restore continuous 3D scar surfaces, especially when the slice spacing is large and image noise or artifacts are significant. Meanwhile, deep learning segmentation techniques (e.g., U-Net and its variants) have made significant progress in myocardial and scar segmentation, but most studies focus on 2D or sparse 3D segmentation. Subsequent high-precision 3D visualization still relies on traditional interpolation and isosurface algorithms, making it difficult to directly obtain smooth and anatomically consistent 3D models. Summary of the Invention

[0004] The purpose of this application is to provide a method, control device, and storage medium for three-dimensional visualization of cardiac scars, with the aim of accurately reproducing a continuous three-dimensional scar surface.

[0005] According to a first aspect of this application, a method for three-dimensional visualization of cardiac scars is provided, the method comprising: acquiring an LGE image for describing a cardiac magnetic resonance imaging sequence; inputting the LGE image into a trained nnU-Net model, outputting segmented cardiac structures and binary label images corresponding to scar tissue; fitting the binary label images of the cardiac structures to a preset standard myocardial mesh model, and reconstructing the binary label images corresponding to the scar tissue; and overlaying the fitted three-dimensional myocardial mesh model and the reconstructed scar tissue mesh in a preset coordinate system, and performing visualization rendering to obtain a visualized cardiac model.

[0006] In an optional implementation, after acquiring LGE images for describing cardiac magnetic resonance imaging sequences, the method further includes: normalizing the LGE images to a reference coordinate system based on shape, suppressing noise using Gaussian filtering or nonlocal mean filtering; and mapping LGE images under different scanning conditions to a standard intensity range.

[0007] In an optional implementation, before fitting the binary labeled image of the heart structure to a preset standard myocardial mesh model, the method further includes: calculating the centroid of the heart structure and constructing a covariance matrix based on the binary labeled image of the heart structure; determining the major axis and minor axis directions of the left ventricle through principal component analysis to establish a reference three-dimensional orthogonal coordinate system; and registering the reference three-dimensional orthogonal coordinate system with a standard four-chamber heart plane to initially align the binary labeled image of the heart structure with the standard myocardial mesh model.

[0008] In an optional implementation, fitting the binary labeled image of the heart structure to a preset standard myocardial mesh model includes: using a deep learning regression model to fit the deformation parameters of the standard myocardial mesh model based on the binary labeled image of the heart structure. The standard myocardial mesh model and the binary labeled image of the heart structure serve as inputs to the regression model, and the output is the deformation parameters of the standard myocardial mesh model. During the training of the regression model, the difference between the binary labeled image of the heart structure and the standard myocardial mesh model is used as a supervision signal. By calculating the nearest distance between the grid vertices of the standard myocardial mesh model and the target segmentation boundary shown in the binary labeled image of the heart structure, as well as the normal consistency of the standard myocardial mesh model, a loss function reflecting anatomical differences and mesh properties is constructed.

[0009] In an optional implementation, the reconstruction of the binary label image corresponding to the scar tissue includes: resampling the binary label image corresponding to the scar tissue, and using an isosurface extraction algorithm to extract isosurfaces from the resampled binary label image to obtain an initial three-dimensional mesh; optimizing the initial three-dimensional mesh, wherein hole filling technology is used to fill holes and isolated regions in the initial three-dimensional mesh caused by threshold extraction, a mesh simplification algorithm is used to reduce the surface area of ​​the initial three-dimensional mesh, and mesh smoothing filtering is applied to remove surface roughness noise of the initial three-dimensional mesh; and extracting the isosurfaces of the optimized three-dimensional mesh and correcting the vertex normals to obtain the reconstructed three-dimensional surface model of the scar tissue.

[0010] In an optional implementation, before overlaying the fitted three-dimensional myocardial mesh model with the reconstructed scar tissue mesh, the method further includes: acquiring myocardial endocardium, myocardial epicondyle label data and partition labels from the binary label map of the heart structure; converting the fitted three-dimensional myocardial mesh model and the reconstructed scar tissue mesh into a voxel mesh of uniform resolution; determining the affiliation of each voxel based on the positional relationship within the voxel mesh and assigning corresponding partition labels to identify the overlapping area of ​​myocardium and scar; and calculating the scar volume percentage and total scar percentage of each partition to form scar distribution data.

[0011] In an optional implementation, the visualization rendering to obtain a visualized heart model includes: calculating the thickness value corresponding to each scar mesh vertex, and normalizing the thickness value before mapping it onto a gradient color spectrum.

[0012] In an optional implementation, the visualization rendering to obtain a visualized heart model further includes: dividing the left ventricle into multiple ring layers along a first direction and dividing them into corresponding sector angle regions; and statistically analyzing scar indicators in each sector angle region and displaying the corresponding values ​​as color blocks or numerical values ​​on a polar coordinate plane to form a two-dimensional image similar to a bullseye.

[0013] According to a second aspect of this application, a control device is provided, the control device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the computer program to implement the method described above.

[0014] According to a third aspect of this application, a machine-readable storage medium is provided, on which instructions are stored, causing a machine to perform the methods described above.

[0015] Through the above technical solution, this application embodiment uses a pre-constructed standard myocardial mesh model as a template to match the myocardial shape shown in the binary label map of the heart structure obtained by pre-trained nnU-Net segmentation with the corresponding binary label map of the heart structure. A deep learning regression model is used to fit the deformation parameters of the standard myocardial mesh model and apply them to the standard mesh to complete the mesh deformation to fit the individual myocardial morphology. The template deformation method effectively utilizes prior knowledge of the heart shape, resulting in a reconstructed three-dimensional geometric model with higher anatomical consistency and precision. The fitted three-dimensional myocardial mesh model and the reconstructed scar tissue mesh are superimposed and registered in the same spatial coordinates, achieving three-dimensional transparent myocardial scar visualization to accurately restore the continuous three-dimensional scar surface.

[0016] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures and processes shown in the description and the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of a method for three-dimensional visualization of cardiac scars provided in an exemplary embodiment of this application.

[0019] Figure 2 This is an example flowchart illustrating a method for three-dimensional visualization of cardiac scars provided in an exemplary embodiment of this application.

[0020] Figure 3 This is an example schematic diagram of three-dimensional visualization of cardiac scars provided by an exemplary embodiment of this application.

[0021] Figure 4 This is a schematic diagram of a bullseye image of a cardiac scar provided in an exemplary embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Figure 1 This is a schematic flowchart of a method for three-dimensional visualization of cardiac scars provided by an exemplary embodiment of this application. The method may include the following steps: Step S110: Acquire LGE images to describe cardiac magnetic resonance imaging sequences.

[0024] Among them, sparse 2D delayed gadolinium enhancement (LGE) is an imaging sequence for cardiac magnetic resonance imaging. It is used to visualize myocardial necrosis or scar tissue after intravenous injection of gadolinium contrast agent. Under LGE sequence, myocardial scar areas appear as high signals and can be used to assess the location, distribution pattern, and degree of fibrosis. However, LGE images consist of multiple two-dimensional slices with large intervals.

[0025] In a preferred embodiment of this application, after step S110, the method may further include: normalizing the LGE image to a reference coordinate system based on shape, and suppressing noise using Gaussian filtering or nonlocal mean filtering; and mapping LGE images under different scanning conditions to a standard intensity range.

[0026] Please refer to Figure 2 For example, standardizing the acquired LGE images (multiple two-dimensional slices with large intervals) can include: normalizing the two-dimensional slices to the same reference coordinate system based on their shape, and using methods such as Gaussian filtering or nonlocal mean filtering to suppress noise and improve segmentation quality; normalizing the pixel intensity of the LGE images (volume data) to unify LGE images under different scanning conditions to a standard intensity range to ensure the consistency and robustness of subsequent deep learning model inputs.

[0027] Step S120: Input the LGE image into the trained nnU-Net model, and output the segmented heart structure and the corresponding binary label map of the scar tissue.

[0028] Please refer to Figure 2 For example, inputting a pre-processed LGE image into a trained nnU-Net model can segment regions such as the left ventricular myocardium, left ventricular cavity, right ventricle, and scar tissue, generating binary label maps. nnU-Net is a self-configuring deep learning medical image segmentation framework (based on the U-Net structure) that can automatically adjust the network structure and parameters according to data characteristics and performs excellently in various cardiac MRI segmentation tasks. The self-configurable nature of nnU-Net allows it to adaptively select the optimal network structure and loss function for the characteristics of LGE images, ensuring the accuracy and robustness of the segmentation results. The resulting segmented binary label maps are used for subsequent coordinate system construction and mesh deformation.

[0029] Step S130: Fit the binary label map of the heart structure to the preset standard myocardial mesh model, and reconstruct the binary label map corresponding to the scar tissue.

[0030] In a preferred embodiment of this application, before step S130, the method may further include: calculating the centroid of the heart structure and constructing a covariance matrix based on the binary label map of the heart structure; determining the major axis direction and minor axis direction of the left ventricle through principal component analysis to establish a reference three-dimensional orthogonal coordinate system; and registering the reference three-dimensional orthogonal coordinate system with the standard four-chamber heart plane so that the binary label map of the heart structure is initially aligned with the standard myocardial mesh model.

[0031] Principal component analysis (PCA) is an unsupervised dimensionality reduction method (based on linear transformation). It maps high-dimensional data to a low-dimensional space by maximizing the variance of the data projection. It can extract the most representative feature directions in the data and is widely used in tasks such as data preprocessing, visualization and compression.

[0032] Please refer to Figure 2 For example, based on the binary labeled map of the heart structure—that is, the voxel data of the segmented left ventricle, right ventricle, and heart chambers—the centroids are calculated and a covariance matrix is ​​constructed. Principal component analysis (PCA) is used to extract the major direction, which is determined as the major axis direction of the left ventricle (i.e., the direction from the apex of the left ventricle to the base of the left ventricle). Combining the relative positions of the right and left ventricles, the direction from the center of the left ventricle to the center of the right ventricle is determined as the minor axis direction (i.e., the x-axis). The cross product of these two directions is the y-axis, which can be used to establish a reference three-dimensional orthogonal coordinate system (a standardized spatial reference system, for example, a three-dimensional coordinate system based on the major and minor axes, used to align an individual heart with a standard heart model). This reference three-dimensional orthogonal coordinate system is then registered with the standard four-chamber plane (a classic view of the four chambers of the heart, including the left atrium, left ventricle, right atrium, and right ventricle, used for registration and coordinate system establishment). Using the reference three-dimensional orthogonal coordinate system as the target, the initial heart mesh is aligned to the standard four-chamber heart plane through rigid registration, ensuring that subsequent template deformation operations are performed under a unified geometric reference system.

[0033] In a preferred embodiment of this application, fitting the binary labeled image of the heart structure to a preset standard myocardial mesh model may include: using a deep learning regression model to fit the deformation parameters of the standard myocardial mesh model based on the binary labeled image of the heart structure. The standard myocardial mesh model and the binary labeled image of the heart structure serve as inputs to the regression model, and the output is the deformation parameters of the standard myocardial mesh model. During the training of the regression model, the difference between the binary labeled image of the heart structure and the standard myocardial mesh model is used as a supervision signal. By calculating the nearest distance between the grid vertices of the standard myocardial mesh model and the target segmentation boundary shown in the binary labeled image of the heart structure, as well as the normal consistency of the standard myocardial mesh model, a loss function reflecting anatomical differences and mesh properties is constructed.

[0034] In this context, the standard myocardial mesh model or standard myocardial mesh (canonical mesh) represents a three-dimensional mesh template of the average myocardial geometry, serving as the initial template for subsequent deformation. Mesh morphing is the process of adjusting the vertices of the template (i.e., the standard myocardial mesh model) to match the individual anatomical structure. In this embodiment, a deep learning method can be used to fit the affine parameters (deformation parameters) from the standard myocardial mesh model to the binary labeled map of the heart structure. Compared to pure interpolation reconstruction, this embodiment can significantly reduce geometric errors caused by missing slices.

[0035] Among them, normal consistency (loss) is a loss function used to optimize the surface smoothness and geometric consistency (based on the surface normal direction). By minimizing the angle difference between the normal vectors of adjacent triangular facets or point pairs, it improves the geometric continuity of the reconstructed surface or generated model.

[0036] Please refer to Figure 2 The standard myocardial mesh model or standard myocardial mesh is used as a deformation template. In this embodiment, the standard myocardial mesh can be derived from the statistical average myocardial shape of multiple CT segmentation labels in the MMWHS (Multi-Modality Whole Heart Segmentation) dataset. Based on the binary label map of the segmented heart structure, a deep learning method is used to fit the deformation parameters of the standard myocardial mesh model. For example, a loss function reflecting anatomical differences and the properties of the mesh itself can be constructed by calculating the nearest distance between the mesh vertices of the standard myocardial mesh model and the target segmentation boundary shown in the binary label map of the heart structure, as well as the normal consistency of the mesh itself. The difference index is used as the supervision signal of the regression network. The input of the regression network is the standard myocardial mesh model and the binary label map of the heart structure, and the output is the deformation parameters of the standard myocardial mesh model (which may include affine parameters and non-rigid deformation coefficients, such as weights based on harmonic deformation basis functions). During training, mesh deformation can be optimized by minimizing multiple loss functions, including Chamfer distance (a geometric distance metric that measures the similarity between two sets of point clouds, effectively measuring the degree of shape matching by calculating and summing the distances from each point to the nearest point in the other set) and normal consistency loss, to make it fit the myocardial boundary of the individual segment (i.e., the binary label map of the heart structure) while maintaining a smooth and self-intersecting topology. After this deformation process, the standard myocardial mesh model is transformed into a 3D mesh that matches the geometry of the individual myocardium, achieving individualized mesh generation, such as... Figure 3 As shown, by using shape priors, mesh deformation can generate 3D structures that are as detailed as LGE images but smoother, effectively overcoming the accuracy loss caused by simple interpolation.

[0037] In a preferred embodiment of this application, reconstructing the binary label image corresponding to scar tissue may include: resampling the binary label image corresponding to scar tissue, and using an isosurface extraction algorithm to extract isosurfaces from the resampled binary label image to obtain an initial three-dimensional mesh; optimizing the initial three-dimensional mesh, wherein hole filling technology is used to fill holes and isolated regions in the initial three-dimensional mesh caused by threshold extraction, a mesh simplification algorithm is used to reduce the surface area of ​​the initial three-dimensional mesh, and mesh smoothing filtering is applied to remove surface roughness noise of the initial three-dimensional mesh; and extracting the isosurfaces of the optimized three-dimensional mesh and correcting the vertex normals to obtain the reconstructed three-dimensional surface model of the scar tissue.

[0038] Please refer to Figure 2 For example, the binary label map corresponding to scar tissue (i.e., the voxel data of the segmented scar region) is resampled (oversampled), for instance, by increasing the voxel mesh resolution through trilinear interpolation to improve the smoothness of subsequent isosurface extraction. The Flying Edges isosurface extraction algorithm in VTK is used to extract isosurfaces from the binary label map corresponding to scar tissue to obtain the initial 3D mesh. The Flying Edges algorithm is a highly efficient and parallel isosurface extraction algorithm for volume data (isosurfaces are contour lines with the same scalar field value), used to extract smooth surfaces from 3D segmentation labels. The Flying Edges algorithm is implemented in VTK as vtkFlyingEdges3D, designed for large-scale volume data, and is 1-2 orders of magnitude faster than traditional Marching Cubes.

[0039] After isosurface extraction, an unprocessed, coarse polygonal mesh is obtained. To ensure mesh coherence and smoothness, the initial 3D mesh is optimized. This optimization may include: using hole-filling techniques (e.g., VTK's vtkFillHolesFilter or similar operations) to fill holes and isolated regions in the initial 3D mesh caused by threshold extraction; using mesh simplification algorithms to reduce the facets of the initial 3D mesh, such as mesh simplification based on quadratic error metrics, reducing the number of triangles while preserving the shape as much as possible; applying mesh smoothing filters (e.g., Laplacian smoothing or Taubin smoothing) to remove surface roughness noise, making the mesh smoother. Vertex normals are recalculated and corrected to enhance visualization. After the above processing, a reconstructed 3D surface model of the scar tissue is obtained, as shown below. Figure 3 As shown, the three-dimensional surface model has a coherent mesh topology and a delicate surface morphology, which can reflect the deformation characteristics of scar tissue.

[0040] Step S140: Under the preset coordinate system, the fitted three-dimensional myocardial mesh model is superimposed with the reconstructed scar tissue mesh, and then visualized and rendered to obtain a visualized heart model.

[0041] Please refer to Figure 2 For example, the reconstructed 3D myocardial mesh model and scar tissue mesh are set in the same coordinate system so that they can be correctly superimposed and displayed. For instance, the DICOM spatial transformation parameters given during image acquisition are applied to the mesh data, mapping the 3D myocardial mesh model and scar tissue mesh from the image coordinate system to the physical space coordinate system. The fitted 3D myocardial mesh model and the reconstructed scar tissue mesh can be correctly aligned in the same 3D space, intuitively showing the spatial distribution of scars within the myocardium. Figure 3 As shown.

[0042] In a preferred embodiment of this application, before step S140, the method may further include: acquiring the myocardial endocardium, myocardial epicondyle label data and partition labels from the binary label map of the heart structure; converting the fitted three-dimensional myocardial mesh model and the reconstructed scar tissue mesh into a voxel mesh of uniform resolution, determining the affiliation of each voxel according to the positional relationship within the voxel mesh, and assigning corresponding partition labels to identify the overlapping area of ​​myocardium and scar; and statistically analyzing the scar volume ratio and total scar ratio of each partition to form scar distribution data.

[0043] Endocardium / Epicardium Separation can be understood as the process of precisely distinguishing myocardial mesh or voxel data into an inner layer (endocardium) and an outer layer (epidermal). It is crucial for assessing myocardial wall thickness, scar permeability, and performing biomechanical simulations.

[0044] In this embodiment, a three-dimensional myocardial mesh model and a scar tissue mesh are partitioned, labeled, and quantitatively analyzed to achieve accurate assessment of myocardial infarction lesions. For partitioning, endocardial and epicardial label data from a binary label map of the heart structure, as well as partition labels conforming to, for example, the American Heart Association (AHA) 17-partition standard, can be obtained. For instance, the left ventricle can be divided into four levels along the Z-axis (i.e., from the apex of the left ventricle to the base of the left ventricle) using polar coordinate transformation: base, middle, apex, and apical cap. Each level is further subdivided according to circumferential angles to generate 17 representative partition labels, achieving systematic spatial partitioning of the myocardium. For the separation of the endocardium and epicardium, preliminary endocardial and epicardial labels are obtained using the distance from a vertex to the center of the three-dimensional myocardial mesh model. By calculating the nearest neighbor distance between the vertices of the three-dimensional myocardial mesh model and combining it with a neighborhood voting algorithm for outlier processing, the membrane properties of each vertex are accurately identified. Building upon this foundation, to support quantitative analysis, the three-dimensional myocardial mesh model and scar tissue mesh were converted into a voxel mesh with uniform resolution. The affiliation of each voxel was determined based on its positional relationship within the voxel mesh, and corresponding partition labels were assigned. Furthermore, the overlapping areas of myocardium and scar tissue were identified. Additionally, the proportion of scar volume in each partition and the total proportion of scar tissue were statistically analyzed to generate detailed partitioned scar distribution data, assisting clinicians in intuitively assessing and making decisions regarding the extent and severity of lesions.

[0045] In a preferred embodiment of this application, visualization rendering to obtain a visualized heart model may include: calculating the thickness value corresponding to each scar mesh vertex, and normalizing the thickness value and mapping it onto a gradient color spectrum.

[0046] Scar local thickness can be understood as the depth of scar coverage in the myocardial thickness direction measured in the 3D reconstruction results, reflecting the degree of scar penetration and the severity of the lesion. Please refer to [reference needed]. Figure 3 For example, scar thickness can be rendered using a red-blue gradient color scheme, where red represents areas of greater local thickness and blue represents areas of thinner thickness. For instance, the thickness value corresponding to each scar grid vertex can be calculated (e.g., measuring the distance from a point on the scar surface to the myocardium or the scar layer thickness), normalized, and then mapped onto a red-blue gradient color map. This makes areas of high thickness appear red and areas of low thickness appear blue, visually reflecting the scar depth distribution.

[0047] In a preferred embodiment of this application, visualization rendering to obtain a visualized heart model may further include: dividing the left ventricle into multiple ring layers along a first direction and dividing them into corresponding sector angle regions; and statistically analyzing scar indicators in each sector angle region and displaying the corresponding values ​​as color blocks or numerical values ​​on a polar coordinate plane to form a two-dimensional image similar to a bullseye.

[0048] In this embodiment, a bull's eye plot can be further generated to display zonal statistics. The bull's eye plot is a two-dimensional polar coordinate plot used to visualize the distribution of cardiac function or lesions on, for example, an AHA 17 zonal model. It projects the three-dimensional cardiac structure onto a two-dimensional plane and uses color coding to visually display the quantitative indicators of each zonal region.

[0049] Please refer to Figure 4 For example, a bullseye map can be implemented by mapping the left ventricle into 17 zones. This can include dividing the left ventricle along its long axis (i.e., the first direction) into multiple ring layers (e.g., basal, middle, and apical layers) and further subdividing them into corresponding sector-shaped angular regions, each representing a specific area of ​​the ventricle. Within each zone, relevant scar indicators (e.g., scar volume or average thickness) are summarized and their values ​​are displayed as color blocks or numerical values ​​on a polar coordinate plane, forming a two-dimensional image resembling a bullseye to assess scar distribution in each zone.

[0050] In a preferred embodiment of this application, visualization rendering can be performed using VTK's PyVista library. PyVista provides a high-level rendering interface for 3D datasets, used to load polygon meshes and set lighting, materials, and color mapping. Interactive rendering can be achieved through PyVista, for example, by setting effects such as smooth shading, normal interpolation, and transparency, thereby enriching the visualization. Furthermore, plotting libraries such as Matplotlib can be used to generate statistical charts (e.g., bullseye charts, bar charts, etc.), achieving multi-view, multi-modal visualization integration. The above visualization strategy enables the collaborative display of the myocardial 3D model, scar thickness distribution, and statistical analysis charts in a unified coordinate system, providing intuitive and rich information on cardiac structure and lesion distribution.

[0051] Accordingly, this application embodiment uses a pre-constructed standard myocardial mesh model as a template to match the myocardial shape shown in the binary label map of the heart structure obtained by pre-trained nnU-Net segmentation with the corresponding binary label map of the heart structure. A deep learning regression model is used to fit the deformation parameters of the standard myocardial mesh model and apply them to the standard mesh to complete the mesh deformation to fit the individual myocardial morphology. The template-based deformation method effectively utilizes prior knowledge of the heart shape, resulting in a reconstructed 3D geometric model with higher anatomical consistency and precision. For the binary label map of the segmented scar tissue, the Flying Edges isosurface extraction algorithm can be used to generate an initial 3D scar surface. After obtaining the surface, the triangular mesh is geometrically smoothed to remove jagged edges and step artifacts, ensuring a smooth and continuous scar surface while preserving key pathological features. This technical approach, combining deep learning with classical algorithms, fully leverages the advantages of deep learning in modeling nonlinear deformations and capturing structural priors, while also utilizing the maturity of classical geometric algorithms in terms of stability and computational efficiency. The embodiments of this application not only improve the accuracy and anatomical consistency of surface reconstruction, but also significantly enhance the adaptability and robustness of the overall system in handling different types of cardiac tissue (such as myocardium and scar tissue), thereby providing a more reliable three-dimensional structural basis for subsequent cardiac function analysis and individualized treatment. Furthermore, by superimposing the fitted three-dimensional myocardial mesh model with the reconstructed scar tissue mesh in the same spatial coordinates, three-dimensional transparent myocardial scar visualization is achieved, accurately restoring the continuous three-dimensional scar surface.

[0052] Furthermore, the proportion of scar tissue in different cardiomyocyte regions is calculated, and based on this, the spatial location and local thickness of the scar are color-coded (e.g., thin scars are represented in red, and thick scars in high-risk regions are represented in blue-green), thus visually reflecting the extent and severity of the lesion. This visualization can be used for clinical diagnosis and can also assist in intraoperative navigation and treatment planning.

[0053] The embodiments of this application also achieve the following technical effects: 1) Improved reconstruction accuracy: Combining deep learning segmentation and template mesh deformation, the three-dimensional shape of individual myocardium can be accurately restored. Template deformation utilizes global shape priors, resulting in higher anatomical consistency and detail fidelity of the reconstructed model. Experiments show that the scar surface generated by this method is smoother and more continuous than traditional interpolation methods, and better matches manual annotation. 2) Full-process automation: From segmentation, registration, deformation to reconstruction, no manual intervention or parameter fine-tuning is required. The nnU-Net segmentation module provides good adaptability and repeatability for processing various LGE data. 3) Applicability to multiple diseases: The framework provided in the embodiments of this application can handle myocardial scars caused by different etiologies, such as scars formed after acute or chronic myocardial infarction, fibrosis caused by myocarditis, and left atrial scars in patients with atrial fibrillation. The multi-structure segmentation and three-dimensional reconstruction technology used in the embodiments of this application can be widely applied to various clinical scenarios. 4) Visualization and evaluation capabilities: The three-dimensional visualization results intuitively show the spatial distribution and thickness of the scar, assisting doctors in quantitatively assessing the degree of lesion. Color coding makes lesions and healthy tissue easily distinguishable, which can be used for preoperative planning and intraoperative navigation, and has high clinical application value. 5) Improve post-processing reconstruction quality. Adaptive smoothing or implicit surface methods (e.g., multi-level unified implicit models) can be introduced to simultaneously ensure smoothness and feature fidelity, reducing the loss of important details. 6) Multimodal fusion. Combining with other MRI sequences (e.g., coronary artery enhancement or cardiac cinematic functional imaging) can provide more structural information for segmentation, improving the robustness of segmentation and reconstruction. 7) Deep learning direct mesh prediction. Methods such as graph convolutional networks can be used to directly predict the shape of the cardiac mesh from sparse images to further improve resolution and connectivity.

[0054] This application also provides a control device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the above-described method.

[0055] This application also provides a machine-readable storage medium storing instructions that cause a machine to perform the above-described method.

[0056] It should be noted that the control device and machine-readable storage medium described above can implement the method for three-dimensional visualization of cardiac scars provided in the above embodiments. For specific implementation methods, please refer to the description of the method for three-dimensional visualization of cardiac scars in the above embodiments, which will not be repeated here.

[0057] It is understood that the circuit structures, names, and parameters described in the above embodiments are merely examples. Those skilled in the art can also make readily conceived combinations and adjustments to the structural features of the above embodiments according to their needs, and the concept of this application should not be limited to the specific details of the above examples.

[0058] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for three-dimensional visualization of cardiac scars, characterized in that, The method includes: Acquire LGE images to describe cardiac magnetic resonance imaging sequences; The LGE image is input into the trained nnU-Net model, which outputs the segmented heart structure and the corresponding binary label image of the scar tissue. The binary labeled image of the heart structure is fitted to a preset standard myocardial mesh model, and the binary labeled image corresponding to the scar tissue is reconstructed; and Under a preset coordinate system, the fitted three-dimensional myocardial mesh model is superimposed with the reconstructed scar tissue mesh and then visualized and rendered to obtain a visualized heart model.

2. The method for three-dimensional visualization of cardiac scars according to claim 1, characterized in that, After acquiring LGE images for describing cardiac magnetic resonance imaging sequences, the method further includes: The LGE image is normalized to a reference coordinate system based on its shape, and noise suppression is performed using Gaussian filtering or nonlocal mean filtering; and Map LGE images under different scanning conditions to a standard intensity range.

3. The method for three-dimensional visualization of cardiac scars according to claim 1, characterized in that, Before fitting the binary labeled map of the heart structure to a preset standard myocardial mesh model, the method further includes: Based on the binary labeled map of the heart structure, its centroid is calculated and a covariance matrix is ​​constructed. Principal component analysis is used to determine the major and minor axes of the left ventricle, thus establishing a baseline three-dimensional orthogonal coordinate system. The reference three-dimensional orthogonal coordinate system is registered with the standard four-chamber heart plane so that the binary label map of the heart structure is initially aligned with the standard myocardial mesh model.

4. The method for three-dimensional visualization of cardiac scars according to claim 1, characterized in that, The step of fitting the binary labeled image of the heart structure to a preset standard myocardial mesh model includes: Based on the binary label map of the heart structure, the deformation parameters of the standard myocardial mesh model are fitted using a deep learning regression model. The standard myocardial mesh model and the binary label map of the heart structure are used as inputs to the regression model, and the output is the deformation parameters of the standard myocardial mesh model. During the training of the regression model, the difference between the binary label map of the cardiac structure and the standard myocardial mesh model is used as a supervision signal. By calculating the nearest distance between the grid vertices of the standard myocardial mesh model and the target segmentation boundary shown in the binary label map of the heart structure, as well as the normal consistency of the standard myocardial mesh model, a loss function reflecting anatomical differences and mesh properties is constructed.

5. The method for three-dimensional visualization of cardiac scars according to claim 1, characterized in that, The reconstruction of the binary label map corresponding to the scar tissue includes: The binary label image corresponding to the scar tissue is resampled, and the isosurface is extracted from the resampled binary label image using an isosurface extraction algorithm to obtain an initial three-dimensional mesh; The initial 3D mesh is optimized by employing a hole-filling technique to fill in holes and isolated regions caused by threshold extraction, using a mesh simplification algorithm to reduce the surface area of ​​the initial 3D mesh, and applying a mesh smoothing filter to remove surface roughness noise from the initial 3D mesh; and The isosurfaces of the optimized 3D mesh are extracted, and the vertex normals are corrected to obtain the 3D surface model of the reconstructed scar tissue.

6. The method for three-dimensional visualization of cardiac scars according to claim 1, characterized in that, Before overlaying the fitted three-dimensional myocardial mesh model with the reconstructed scar tissue mesh, the method further includes: Obtain the label data of the endocardium and epicardium of the heart structure from the binary label map of the heart structure; The fitted 3D myocardial mesh model and the reconstructed scar tissue mesh are converted into a voxel mesh of uniform resolution. The affiliation of each voxel is determined based on its positional relationship within the voxel mesh, and a corresponding partition label is assigned to identify the overlapping areas of the myocardium and scar tissue. The percentage of scar volume in each zone and the percentage of total scar volume were statistically analyzed to form scar distribution data.

7. The method for three-dimensional visualization of cardiac scars according to claim 1, characterized in that, The visualization rendering process to obtain a visualized heart model includes: Calculate the thickness value corresponding to each scar mesh vertex, and normalize the thickness value before mapping it onto the gradient color map.

8. The method for three-dimensional visualization of cardiac scars according to claim 7, characterized in that, The process of performing visualization rendering to obtain a visualized heart model also includes: The left ventricle is divided into multiple ring layers along the first direction and further subdivided into corresponding sector-shaped angular regions; and Scar indicators are statistically analyzed in each sector angle region, and the corresponding values ​​are displayed on a polar coordinate plane in the form of color blocks or numerical values ​​to form a two-dimensional image.

9. A control device, characterized in that, The control device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for three-dimensional visualization of cardiac scars according to any one of claims 1-8.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions that cause the machine to perform the method for three-dimensional visualization of cardiac scars according to any one of claims 1-8.

Images