A deep learning-based aortic valve system modeling method, device, and medium

Abstract

The application relates to an aortic valve system modeling method, device and medium based on deep learning, wherein the method comprises the following steps: constructing and training a joint deep learning network model; acquiring CTA image data of a user's heart; performing stable structure modeling on the user's heart through a deformation field prediction branch in the trained joint deep learning network model; performing variable structure modeling on the user's heart through an implicit reconstruction branch in the network model; and outputting a three-dimensional geometric model of an aortic valve system based on the stable structure modeling and the variable structure modeling under the constraint of a joint optimization framework of the network model. Through the application, the deformation field prediction branch and the implicit reconstruction branch of the network model are utilized to complete unified modeling of an aorta, an aortic valve, a left ventricle and a calcified region, and under the constraint of the joint optimization framework, the model topological continuity between the stable structure and the variable structure of the heart is effectively improved, and the three-dimensional model space registration error is reduced.

Description

Technical Field

[0001] This application relates to the fields of intelligent analysis of medical images and biomedical engineering technology, and in particular to a deep learning-based method, device and medium for modeling aortic valve systems. Background Technology

[0002] The aortic valve is a crucial component of the heart, and its morphology and functional status have a critical impact on hemodynamic characteristics and the normal functioning of the cardiovascular system. With the development of imaging technology, cardiac CTA (Computed Tomography Angiography) is commonly used in clinical practice to assess the aorta and its valves for preoperative planning and efficacy prediction of aortic stenosis, calcification, and valve replacement surgery. Accurate three-dimensional geometric modeling of the aortic valve system is a fundamental step in achieving cardiac structural analysis, hemodynamic simulation, and individualized interventional surgical design, and is of great significance for clinical decision-making and interventional device design.

[0003] However, existing aortic valve system modeling methods mainly rely on traditional image segmentation and geometric reconstruction workflows, typically employing threshold segmentation, region growing, level sets, or morphological algorithms for structure extraction. These methods are highly sensitive to image noise, calcification artifacts, and resolution variations, resulting in significant uncertainty in segmentation results. During geometric reconstruction, model surfaces often exhibit breaks, wrinkles, or topological errors, making it difficult to accurately reflect the true spatial relationship between the valve and the aortic wall. Furthermore, traditional manual or semi-automatic reconstruction processes rely on human intervention, are time-consuming, and have poor repeatability, failing to meet clinical demands for high-precision, rapid, and automated 3D modeling.

[0004] In recent years, deep learning technology has made significant progress in the field of medical image analysis. Through end-to-end network structures, it can simultaneously learn image features and spatial morphological information, achieving high-precision modeling of complex anatomical structures. However, research on the aortic valve system still has shortcomings: on the one hand, the aortic valve and calcified regions have complex structures with large differences in leaflet morphology and blurred boundaries, making it difficult for models to take into account both overall morphological consistency and local geometric details; on the other hand, existing methods mostly focus on single-structure segmentation, lacking a holistic collaborative modeling mechanism for the aorta, valves, left ventricle, and calcified lesions.

[0005] Currently, no effective solution has been proposed for improving the 3D modeling effect of the aortic valve system in related technologies. Summary of the Invention

[0006] This application provides a deep learning-based aortic valve system modeling method, device, and medium to at least address the problem of how to improve the three-dimensional modeling effect of the aortic valve system in related technologies.

[0007] In a first aspect, embodiments of this application provide a three-dimensional method for an aortic valve system based on deep learning, the method comprising:

[0008] Construct and train a joint deep learning network model for aortic valve system modeling;

[0009] Acquire CTA image data of the user's heart, wherein the CTA image data covers the aorta, aortic valve, left ventricle and calcified areas;

[0010] Based on the CTA image data, the stable structure model of the aorta, aortic valve and left ventricle of the user's heart is modeled by the deformation field prediction branch in the trained joint deep learning network model.

[0011] The implicit reconstruction branch in the trained joint deep learning network model is used to model the variable structure of the calcified region of the user's heart.

[0012] Under the joint optimization framework constraints of the joint deep learning network model, based on the stable structure modeling of the deformation field prediction branch and the variable structure modeling of the implicit reconstruction branch, a three-dimensional geometric model of the aortic valve system is output.

[0013] In some embodiments, building and training a joint deep learning network model for aortic valve system modeling includes:

[0014] Based on single-case standard cardiac CTA imaging data, a template geometric model of the aortic valve system is constructed to provide morphological priors and spatial references in the joint deep learning network model.

[0015] A joint deep learning network model for aortic valve system modeling is constructed, wherein the joint deep learning network model includes a deformation field prediction branch and an implicit reconstruction branch;

[0016] The deformation field prediction branch learns the three-dimensional deformation field from the template space to the user's individual anatomical space at the voxel level based on the morphological prior of the template geometry model, and extends the three-dimensional deformation field to the continuous full space through interpolation; the implicit reconstruction branch learns the implicit functional representation of calcified variable structures.

[0017] The CTA image data of the training samples are segmented and labeled to establish a standard three-dimensional geometric model of the aorta, aortic valve, left ventricle and calcified structure, which is used to provide supervision signals and evaluation benchmarks during the model training phase.

[0018] The joint deep learning network model is trained using the CTA image data of the training samples to obtain a trained joint deep learning network model.

[0019] In some embodiments, the joint deep learning network model is trained using the training sample CTA image data to obtain a trained joint deep learning network model, including:

[0020] Based on the training sample CTA image data, the joint deep learning network model is trained by a joint optimization loss function to ensure the consistency of the reconstruction results of the deformation field prediction branch and the implicit reconstruction branch in the network model in terms of space, topology and anatomy, thereby obtaining a trained joint deep learning network model. The joint optimization loss function includes registration consistency constraint, symbolic distance field constraint, grid consistency constraint and contact consistency constraint.

[0021] In some embodiments, the method includes:

[0022] The training of the joint deep learning network model is performed in the PyTorch deep learning framework, and the Adam optimization algorithm is used for parameter updates. The combination of multiple constraints in the joint optimization loss function is a weighted combination.

[0023] The registration consistency constraint is used to ensure the morphological matching between the template geometric model and the user CTA image;

[0024] The symbolic distance field constraint is used to improve the geometric accuracy of the implicit reconstruction results;

[0025] The grid consistency constraint is used to maintain topological continuity between the aorta, valves, and left ventricle;

[0026] The contact consistency constraint is used to prevent non-physical penetration between the valve leaflet and the aortic wall.

[0027] In some embodiments, constructing a joint deep learning network model for aortic valve system modeling includes:

[0028] Using the 3D U-Net network as the backbone network architecture of the joint deep learning network model, a joint deep learning network model for aortic valve system modeling was constructed.

[0029] The three-dimensional U-Net network serves as a shared encoder for the network model, used to extract multi-scale features from image data step by step, and to establish spatial and anatomical feature representations of the aortic valve system. Each coding unit includes three-dimensional convolution, normalization, and nonlinear activation operations to extract structural texture information at different scales.

[0030] In some embodiments, the method includes:

[0031] The joint deep learning network model includes a deformation field prediction branch;

[0032] The deformation field prediction branch takes the multi-scale features extracted by the three-dimensional U-Net network as input, uses the deformation decoder to upsample and fuse high-resolution features step by step, and maintains the integrity of spatial information through multi-scale skip connections to predict the three-dimensional deformation field from the template geometric model to the user individual.

[0033] The three-dimensional deformation field is used to drive the template geometry model to transform in space for individualized morphological reconstruction of the aorta, aortic valve, and left ventricle.

[0034] In some embodiments, the method includes:

[0035] The joint deep learning network model includes an implicit reconstruction branch;

[0036] The implicit reconstruction branch takes the multi-scale features extracted by the three-dimensional U-Net network as input and uses an implicit decoder based on a multilayer perceptron to output continuous implicit function values ​​at the corresponding positions of the calcified regions.

[0037] The continuous implicit function values ​​are used to characterize the fine morphology and boundary distribution of the calcified region, so as to reconstruct the calcified region using the zero-level set method.

[0038] In some embodiments, acquiring CTA imaging data of the user's heart includes:

[0039] Acquire CTA image data of the user's heart and perform standardized preprocessing on the CTA image data to obtain preprocessed CTA image data. The standardized preprocessing includes resampling, normalization, and region of interest extraction.

[0040] In a second aspect, embodiments of this application provide a computer device, including 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 described in the first aspect above.

[0041] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0042] Compared to related technologies, this application provides a deep learning-based aortic valve system modeling method, device, and medium. The method involves constructing and training a joint deep learning network model for aortic valve system modeling; acquiring CTA image data of the user's heart; performing stable structure modeling of the aorta, aortic valve, and left ventricle of the user's heart through the deformation field prediction branch in the trained joint deep learning network model; performing variable structure modeling of the calcified region of the user's heart through the implicit reconstruction branch in the network model; and outputting a three-dimensional geometric model of the aortic valve system based on the stable structure modeling of the deformation field prediction branch and the variable structure modeling of the implicit reconstruction branch under the constraints of the joint optimization framework of the network model. This achieves unified modeling of the aorta, aortic valve, left ventricle, and calcified region using the deformation field prediction branch and the implicit reconstruction branch of the network model. Under the constraints of the joint optimization framework, it effectively improves the model topological continuity between the stable and variable structures of the heart, reduces the spatial registration error of the three-dimensional model, and solves the problem of how to improve the three-dimensional modeling effect of the aortic valve system. Attached Figure Description

[0043] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0044] Figure 1 This is a flowchart illustrating the steps of a deep learning-based three-dimensional method for aortic valve systems according to an embodiment of this application.

[0045] Figure 2 This is a schematic diagram of the structure of a joint deep learning network model according to an embodiment of this application;

[0046] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of this application. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0048] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0049] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0050] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0051] This application provides a three-dimensional method for aortic valve systems based on deep learning. Figure 1 This is a flowchart illustrating the steps of a deep learning-based three-dimensional method for aortic valve systems according to embodiments of this application, as follows: Figure 1 As shown, the method includes the following steps:

[0052] Step S102: Construct and train a joint deep learning network model for aortic valve system modeling;

[0053] Step S102 specifically includes the following steps:

[0054] Step S1021: Based on single-case standard cardiac CTA image data, construct a template geometric model of the aortic valve system to provide morphological priors and spatial references in the joint deep learning network model.

[0055] Specifically, step S1021 involves constructing a template image and a template geometric model of the aortic valve system based on a single standard cardiac CTA image, serving as a morphological prior and spatial reference. The template geometric model is generated through segmentation, surface reconstruction, and mesh optimization, and includes key anatomical structures such as the aortic root, valve leaflets, and left ventricular outflow tract.

[0056] It should be noted that step S1021 is the template geometric model construction stage. Based on CTA images of healthy subjects, this stage generates a standard template geometric model containing the aortic root, valve leaflets, and left ventricular outflow tract through segmentation, surface reconstruction, and mesh smoothing. This template geometric model serves as a morphological prior, providing not only spatial coordinate references but also guiding the spatial mapping learning of deformation field prediction branches during subsequent network training, thus providing a reference for the overall structural consistency of the aortic valve system.

[0057] Step S1022: Construct a joint deep learning network model for aortic valve system modeling, wherein the joint deep learning network model includes a deformation field prediction branch and an implicit reconstruction branch.

[0058] The deformation field prediction branch is based on the morphological prior of the template geometry model. It learns the three-dimensional deformation field from the template space to the user's individual anatomical space at the voxel level and extends the three-dimensional deformation field to the continuous full space through interpolation. The implicit reconstruction branch learns the implicit functional representation of calcified variable structures.

[0059] Specifically, in step S1022, Figure 2 This is a schematic diagram of the structure of a joint deep learning network model according to an embodiment of this application, such as... Figure 2 As shown, a joint deep learning network model for aortic valve system modeling is constructed using a 3D U-Net network as the backbone network architecture. The 3D U-Net network serves as the shared encoder of the network model, used to extract multi-scale features from image data step by step, and to establish spatial and anatomical feature representations of the aortic valve system. Each encoding unit includes 3D convolution, normalization, and nonlinear activation operations to extract structural texture information at different scales.

[0060] It should be noted that step S1022 constructs a joint deep learning network model with a dual-branch structure. This network model consists of a shared encoder, a deformation field prediction branch, and an implicit reconstruction branch. The input is CTA image data, and the output is a three-dimensional geometric model of the aorta, aortic valve, left ventricle, and calcified region. The network uses a three-dimensional U-Net as its backbone architecture, combined with multi-scale feature extraction, skip connections, and attention mechanisms to achieve accurate modeling of global morphological constraints and local geometric details.

[0061] Specifically, step S1022, as follows: Figure 2 As shown, the joint deep learning network model includes a deformation field prediction branch. The deformation field prediction branch takes the multi-scale features extracted by the 3D U-Net network as input, uses the deformation decoder to upsample and fuse high-resolution features step by step, and maintains the integrity of spatial information through multi-scale skip connections to predict the 3D deformation field from the template geometric model to the user's individual. The 3D deformation field is used to drive the template geometric model to transform in space to perform individualized morphological reconstruction of the aorta, aortic valve and left ventricle.

[0062] It should be noted that during the encoding stage, the network model extracts multi-scale features from the image step-by-step through a multi-layer 3D convolutional structure of the 3D U-Net, establishing a spatial and anatomical representation of the aortic valve system. Each encoding unit includes 3D convolution, normalization, and nonlinear activation operations to extract structural texture information at different scales. The deformation field prediction branch takes the multi-scale features of the shared encoder as input, utilizes the deformation decoder to upsample and fuse high-resolution features step-by-step, and maintains the integrity of spatial information through multi-scale skip connections, thereby predicting the 3D deformation field from the template geometric model to the individual patient. The predicted deformation field is used to drive the template geometric model to transform in space, achieving individualized morphological reconstruction of the aorta, aortic valve, and left ventricle.

[0063] Specifically, step S1022, as follows: Figure 2 As shown, the joint deep learning network model includes an implicit reconstruction branch. The implicit reconstruction branch takes the multi-scale features extracted by the 3D U-Net network as input and uses an implicit decoder based on a multilayer perceptron to output continuous implicit function values ​​at the corresponding positions of the calcified regions. The continuous implicit function values ​​are used to characterize the subtle morphology and boundary distribution of the calcified regions, so as to reconstruct the calcified regions through the zero-level set method.

[0064] It should be noted that the implicit reconstruction branch, based on the features of the shared encoder, reconstructs the calcified structure by combining 3D convolutional feature fusion with a multilayer perceptron implicit decoder. This branch takes spatial coordinates and local image features as input and outputs continuous implicit function values ​​at the corresponding locations, used to characterize the subtle morphology and boundary distribution of the calcified region. The output of the implicit reconstruction extracts the surface of the calcified region using the zero-level set method, forming a smooth and structurally continuous 3D model.

[0065] It should be further explained that these two branches collaboratively model under the constraint of the joint optimization loss function in subsequent step S1025 (i.e., the joint optimization framework in step S110), achieving feature sharing through a shared encoder portion, ensuring the consistency of the deformation field and implicit representation in spatial semantics. The network introduces a channel attention mechanism in high-level features to enhance the response to key structural regions. To prevent deformation folding or calcification penetration during reconstruction, constraints are imposed on structural contact boundaries during network training, thus maintaining overall anatomical consistency while considering local geometric accuracy. The network model constructed in this way can simultaneously complete the collaborative reconstruction of stable and variable structures of the aortic valve system with a single input CTA image, providing a high-precision geometric basis for subsequent model training and clinical applications.

[0066] Step S1023: Segment and label the CTA image data of the training samples to establish a three-dimensional geometric model standard for the aorta, aortic valve, left ventricle and calcified structures, which is used to provide supervision signals and evaluation benchmarks during the model training phase.

[0067] Specifically, in step S1023, for the training sample CTA image data, a gold standard geometric model of the aorta, aortic valve, left ventricle and calcification structure is generated through expert annotation or semi-automatic segmentation, which serves as a supervision signal for network model training and a benchmark for model evaluation.

[0068] It should be noted that in the geometric model gold standard construction stage of step S1023, different structures employ differentiated annotation methods to ensure annotation accuracy and consistency: the aorta, left ventricle, and calcified regions are entirely manually segmented to ensure spatial coherence between cavity boundaries and calcification distribution, and a smooth 3D surface mesh is extracted from the segmented voxels using the marching cubes algorithm; due to the complex boundaries and diverse morphologies of the aortic valve leaflets, a semi-automated boundary annotation and key point annotation method is used. The operator first automatically identifies the positions of the valve annulus and valve tip based on the CTA grayscale gradient, and then manually corrects the key points and leaflet boundary curves to accurately define the valve morphology. Finally, after topological verification and spatial alignment, the 3D geometric models of all structures form a high-precision geometric model gold standard, which is used as a supervision signal for deep learning network training.

[0069] It should be further noted that the training sample CTA image data was divided into training and validation sets in a 4:1 ratio. Each training sample included standardized CTA image data and the corresponding gold standard geometric model. The training data was augmented through random rotation, translation, scaling, and intensity perturbation to improve the network's generalization ability and adaptability to anatomical differences among patients. All data was input into the network in the form of 3D blocks to ensure high-resolution feature learning under limited GPU memory conditions.

[0070] Step S1025: Train the joint deep learning network model using training sample CTA image data to obtain the trained joint deep learning network model.

[0071] Specifically, step S1025 involves training a joint deep learning network model based on the training sample CTA image data using a joint optimization loss function. This ensures the consistency of the reconstruction results of the deformation field prediction branch and the implicit reconstruction branch in the network model in terms of space, topology, and anatomy, resulting in a well-trained joint deep learning network model. The joint optimization loss function includes registration consistency constraints, symbolic distance field constraints, grid consistency constraints, and contact consistency constraints.

[0072] The training of the joint deep learning network model is performed within the PyTorch deep learning framework. The Adam optimization algorithm is used for parameter updates, and the combination of multiple constraints in the joint optimization loss function is a weighted combination. Among these, the registration consistency constraint, used to ensure morphological matching between the template geometric model and the user's CTA image, is expressed as:

[0073]

[0074] Among them, S p S is the predicted surface point set after deformation. g Let x and y be the point set of the gold standard surface, where x and y are the x and y coordinates of points within the point set.

[0075] The symbolic distance field constraint, used to improve the geometric accuracy of implicit reconstruction results, is expressed as:

[0076]

[0077] Where S(x) is the voxel-level SDF of the calcification branch output, S * (x) represents the true SDF calculated from the gold standard surface. The weight function is used to reinforce the boundary. These are preset coefficients.

[0078] Mesh consistency constraints, used to maintain topological continuity between the aorta, valves, and left ventricle, are represented as:

[0079]

[0080] Where M is the number of nodes. To predict surface vertices and normals, The gold standard vertex and normal are paired with it, and α is a preset coefficient.

[0081] Contact consistency constraints, used to prevent non-physical penetration between the valve leaflets and the aortic wall, are represented as follows:

[0082]

[0083] Where K is the number of contact candidate points, d k For contact candidate point p k The minimum distance to another surface, δ max β and γ are preset coefficients representing the maximum allowable gap.

[0084] It should be noted that the network training and joint optimization employ a multi-stage joint optimization strategy. The first stage fixes the implicit reconstruction branch, optimizing only the deformation field prediction branch to quickly align the template geometric model with the individual patient's structure. The second stage unfreezes the implicit reconstruction branch, enabling implicit learning of calcification structures. The third stage trains both branches simultaneously, achieving joint optimization of morphological consistency and geometric details through gradient sharing. The training batch size is set to 4, and the initial learning rate is 1×10⁻⁶. -3 The model was dynamically decayed using a cosine annealing strategy, and converged after 200 training epochs. Throughout the training process, the loss function value of the validation set was monitored in real time. Registration consistency constraints ensured morphological matching between the template and the individual space, mesh consistency constraints maintained structural continuity, signed distance field constraints improved reconstruction accuracy, and contact consistency constraints prevented interleaf crossing or penetration. The best-performing model weights were automatically saved, and stable convergence was achieved through a multi-stage optimization strategy, ensuring that the output 3D model met medical accuracy requirements in terms of spatial consistency and anatomical rationality.

[0085] Step S104: Obtain CTA image data of the user's heart, wherein the CTA image data covers the aorta, aortic valve, left ventricle and calcified areas.

[0086] Specifically, step S104 involves acquiring CTA image data of the user's heart and performing standardized preprocessing on the CTA image data to obtain preprocessed CTA image data. The standardized preprocessing includes resampling, normalization, and region of interest extraction.

[0087] In step S104, optionally, cardiac CTA image data from real clinical cases can be selected as the research sample. The images are acquired by a 128-slice spiral CT scanner (such as Siemens SOMATOM Definition Flash), with a slice thickness of 0.5 mm, covering the aortic root to the left ventricular outflow tract. The acquired images are exported in DICOM format, and after resampling (standardized to 0.5 mm voxel spacing), intensity normalization (using vessel peak grayscale as the standardization reference), and ROI cropping, standardized input data is obtained.

[0088] It should be noted that when acquiring cardiac CTA images covering the aorta, aortic valve, left ventricle, and calcified areas, the slice thickness should not exceed 0.625 mm to ensure spatial resolution for subsequent modeling. The acquired raw data is resampled to unify the pixel spacing to isotropic voxels (0.5 mm × 0.5 mm × 0.5 mm), and grayscale normalization and intensity normalization are performed to eliminate differences caused by different scanning equipment and parameters. Subsequently, the region of interest (ROI) from the aortic root to the left ventricular outflow tract is extracted to provide concentrated and effective data for the model input.

[0089] Step S106: Based on CTA image data, the stable structure model of the aorta, aortic valve and left ventricle of the user's heart is performed by using the deformation field prediction branch in the trained joint deep learning network model.

[0090] Step S108: Through the implicit reconstruction branch in the trained joint deep learning network model, the variable structure model of the calcified region of the user's heart is performed.

[0091] Step S110: Under the joint optimization framework constraint of the joint deep learning network model, based on the stable structure modeling of the deformation field prediction branch and the variable structure modeling of the implicit reconstruction branch, the three-dimensional geometric model of the aortic valve system is output.

[0092] It should be noted that for steps S104 to S110 above, after the model training is completed, inputting new patient cardiac CTA images into the trained network model can achieve automated three-dimensional reconstruction of the aortic valve system. This step requires no manual intervention and can complete the entire process from image input to output of a complete geometric model within minutes.

[0093] During the inference phase, the input CTA images undergo the same standardized preprocessing as in the training phase, including resampling, intensity normalization, and ROI extraction, ensuring consistency in image size and grayscale range. The processed images are then input to a shared encoder, where the network automatically extracts multi-scale features and passes them to two branches for collaborative inference. The deformation field prediction branch uses learned morphological mapping relationships to generate a patient-specific three-dimensional deformation field, and accordingly accurately transforms the template geometric model into patient space, obtaining individualized geometric structures of the aorta, aortic valve, and left ventricle. Simultaneously, the implicit reconstruction branch performs continuous implicit function estimation on calcified regions, achieving refined reconstruction of the calcified structure through surface extraction of the zero-level set of the implicit function.

[0094] After deformation field and implicit function inference are completed, the network model automatically performs structural merging and geometric post-processing operations. First, spatial alignment and topological correction are performed on the surfaces of each structure in the aortic valve system to ensure continuous and non-intersecting contact boundaries between different structures. Then, a lightweight mesh smoothing algorithm is used to eliminate surface noise, and the leaflet thickness and calcification morphology are corrected according to preset anatomical boundary rules, making the model more closely resemble the actual anatomical structure in local geometry. The output 3D geometric model can be exported as a standard format file (such as STL, VTK, or OBJ), supporting direct use in mainstream 3D medical visualization platforms or finite element analysis software. Users can visualize, rotate, slice, and measure the model within the system as needed, or import it into the hemodynamic simulation module for applications such as quantitative assessment of valvular stenosis, interventional pathway planning, and stent expansion simulation analysis.

[0095] Based on the method steps provided in the embodiments of this application described above, this application has at least the following beneficial effects:

[0096] (1) The multi-structure collaborative modeling framework proposed in this application can simultaneously realize the unified geometric modeling of the aorta, aortic valve, left ventricle and calcified region, solving the problems of independent structural modeling, spatial misalignment and topological discontinuity in traditional methods, and significantly improving the overall anatomical consistency and geometric integrity of the model.

[0097] (2) This application can automatically complete three-dimensional geometric reconstruction directly from cardiac CTA images without manual segmentation and post-processing, which significantly improves modeling efficiency and accuracy. Through the joint optimization mechanism of deep learning network, the model has high reconstruction accuracy in both overall morphological fitting and local geometric details, and can accurately reflect the morphology and calcification distribution characteristics of valve leaflets.

[0098] (3) This application has good versatility and scalability. The device has a simple structure, is compatible with multiple image formats and operating platforms, and can be deployed in medical imaging workstations, scientific research servers or cloud computing environments. The generated three-dimensional geometric model can be directly applied to quantitative assessment of valvular lesions, preoperative planning, stent expansion simulation and hemodynamic analysis, and has significant clinical application value and promotion potential.

[0099] It should be noted that the steps shown in the above process or in the flowchart of the accompanying figures can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0100] This embodiment provides a computer device for implementing the aforementioned deep learning-based aortic valve system modeling method. The device can be a medical image processing workstation, a research server, or a high-performance computing platform with GPU acceleration capabilities. The device includes a processor, a memory, and a communication interface. The memory stores an executable computer program, which, when executed by the processor, implements all the steps of the aforementioned aortic valve system modeling method.

[0101] The processor can be a general-purpose central processing unit (CPU), or it can include a graphics processing unit (GPU), a tensor processing unit (TPU), or a neural network acceleration chip. The processor performs tasks such as data reading, deep learning inference, model training, and 3D reconstruction calculations. In training mode, the processor calls a deep learning framework stored in memory, loads a template geometric model and patient CTA image data, and achieves joint optimization of deformation field prediction and implicit reconstruction by performing forward and backward propagation of the network. In inference mode, the processor automatically generates 3D geometric models of the aorta, aortic valve, left ventricle, and calcified structures based on the input CTA images.

[0102] The memory includes high-speed RAM, SSD hard disk, or other non-volatile storage media for storing the operating system, deep learning model parameters, template data, geometric model gold standard, and inference result files. The memory can also store a medical image data management module, a model visualization module, and a post-processing module. The data management module is responsible for image import, resampling, and normalization preprocessing; the visualization module provides model display and measurement functions; and the post-processing module performs smoothing, topology correction, and format conversion of the reconstructed model.

[0103] in addition, Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of this application, such as... Figure 3 As shown, a computer device is provided, which may be a server, and its internal structure diagram may be as follows. Figure 3As shown, the computer device includes a processor, a network interface, internal memory, and non-volatile memory connected via an internal bus. The non-volatile memory stores the operating system, computer programs, and a database. The processor provides computing and control capabilities, the network interface communicates with external terminals via a network connection, the internal memory provides an environment for the operation of the operating system and computer programs, the computer programs are executed by the processor to implement a deep learning-based aortic valve system modeling method, and the database stores data.

[0104] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0105] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is used to implement the steps of the above-described deep learning-based aortic valve system modeling method.

[0106] When the program is executed by the processor in the computer device, it performs the following functions in sequence: reads and standardizes the patient's CTA images and template geometric model through the data acquisition module; calls the deep learning model module to construct a two-branch network structure and generate three-dimensional geometric reconstruction results of the aorta, aortic valve, left ventricle and calcified area; executes the joint optimization module to achieve collaborative training of the deformation field prediction branch and the implicit reconstruction branch; and automatically outputs an individualized three-dimensional model of the aortic valve system during the inference stage and saves it in a standard file format (such as STL, VTK, etc.) required for visualization and subsequent simulation analysis.

[0107] The program in this computer-readable storage medium can be installed on medical imaging workstations, scientific computing servers, or cloud-based image analysis platforms, supporting GPU parallel computing and automatic batch processing. Through this embodiment of the computer-readable storage medium, computing devices can quickly and stably achieve high-precision modeling of the aortic valve system without manual operation, providing intelligent data support for valvular disease analysis, preoperative planning, and biomechanical simulation.

[0108] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0109] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0110] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A three-dimensional method for aortic valve systems based on deep learning, characterized in that, The method includes: Construct and train a joint deep learning network model for aortic valve system modeling; Acquire CTA image data of the user's heart, wherein the CTA image data covers the aorta, aortic valve, left ventricle and calcified areas; Based on the CTA image data, the stable structure model of the aorta, aortic valve and left ventricle of the user's heart is modeled by the deformation field prediction branch in the trained joint deep learning network model. The implicit reconstruction branch in the trained joint deep learning network model is used to model the variable structure of the calcified region of the user's heart. Under the joint optimization framework constraints of the joint deep learning network model, based on the stable structure modeling of the deformation field prediction branch and the variable structure modeling of the implicit reconstruction branch, a three-dimensional geometric model of the aortic valve system is output.

2. The method according to claim 1, characterized in that, The construction and training of a joint deep learning network model for aortic valve system modeling includes: Based on single-case standard cardiac CTA imaging data, a template geometric model of the aortic valve system is constructed to provide morphological priors and spatial references in the joint deep learning network model. A joint deep learning network model for aortic valve system modeling is constructed, wherein the joint deep learning network model includes a deformation field prediction branch and an implicit reconstruction branch; The deformation field prediction branch learns the three-dimensional deformation field from the template space to the user's individual anatomical space at the voxel level based on the morphological prior of the template geometry model, and extends the three-dimensional deformation field to the continuous full space through interpolation; the implicit reconstruction branch learns the implicit functional representation of calcified variable structures. The CTA image data of the training samples are segmented and labeled to establish a standard three-dimensional geometric model of the aorta, aortic valve, left ventricle and calcified structure, which is used to provide supervision signals and evaluation benchmarks during the model training phase. The joint deep learning network model is trained using the CTA image data of the training samples to obtain a trained joint deep learning network model.

3. The method according to claim 2, characterized in that, The joint deep learning network model is trained using the CTA image data from the training samples, resulting in a trained joint deep learning network model including: Based on the training sample CTA image data, the joint deep learning network model is trained by a joint optimization loss function to ensure the consistency of the reconstruction results of the deformation field prediction branch and the implicit reconstruction branch in the network model in terms of space, topology and anatomy, thereby obtaining a trained joint deep learning network model. The joint optimization loss function includes registration consistency constraint, symbolic distance field constraint, grid consistency constraint and contact consistency constraint.

4. The method according to claim 3, characterized in that, The method includes: The training of the joint deep learning network model is performed in the PyTorch deep learning framework, and the Adam optimization algorithm is used for parameter updates. The combination of multiple constraints in the joint optimization loss function is a weighted combination. The registration consistency constraint is used to ensure the morphological matching between the template geometric model and the user CTA image; The symbolic distance field constraint is used to improve the geometric accuracy of the implicit reconstruction results; The grid consistency constraint is used to maintain topological continuity between the aorta, valves, and left ventricle; The contact consistency constraint is used to prevent non-physical penetration between the valve leaflet and the aortic wall.

5. The method according to claim 2, characterized in that, Constructing a joint deep learning network model for aortic valve system modeling includes: Using a 3D U-Net network as the backbone network architecture of a joint deep learning network model, a joint deep learning network model for aortic valve system modeling is constructed. The three-dimensional U-Net network serves as a shared encoder for the network model, used to extract multi-scale features from image data step by step, and to establish spatial and anatomical feature representations of the aortic valve system. Each coding unit includes three-dimensional convolution, normalization, and nonlinear activation operations to extract structural texture information at different scales.

6. The method according to claim 5, characterized in that, The method includes: The joint deep learning network model includes a deformation field prediction branch; The deformation field prediction branch takes the multi-scale features extracted by the three-dimensional U-Net network as input, uses the deformation decoder to upsample and fuse high-resolution features step by step, and maintains the integrity of spatial information through multi-scale skip connections to predict the three-dimensional deformation field from the template geometric model to the user individual. The three-dimensional deformation field is used to drive the template geometry model to transform in space for individualized morphological reconstruction of the aorta, aortic valve, and left ventricle.

7. The method according to claim 5, characterized in that, The method includes: The joint deep learning network model includes an implicit reconstruction branch; The implicit reconstruction branch takes the multi-scale features extracted by the three-dimensional U-Net network as input and uses an implicit decoder based on a multilayer perceptron to output continuous implicit function values ​​at the corresponding positions of the calcified regions. The continuous implicit function values ​​are used to characterize the fine morphology and boundary distribution of the calcified region, so as to reconstruct the calcified region using the zero-level set method.

8. The method according to claim 1, characterized in that, Obtaining CTA imaging data of the user's heart includes: Acquire CTA image data of the user's heart and perform standardized preprocessing on the CTA image data to obtain preprocessed CTA image data. The standardized preprocessing includes resampling, normalization, and region of interest extraction.

9. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images