Method for constructing a 3D anatomical model of the heart and a 3D mapping system for the heart

Abstract

A method for constructing a three-dimensional anatomical model of the heart and a three-dimensional mapping system for the heart are provided, wherein an atrial template model and point cloud data of the atrial to be modeled are acquired, the point cloud data is the position coordinate information of some feature points in the atrial to be modeled, some feature points in the atrial to be modeled are mapped to the atrial template model using an affine transformation network, the point cloud data mapped to the atrial template model is input to a prediction network, the displacement of control feature points in the atrial template model is acquired by prediction, and the atrial template model is deformed based on the displacement of the control feature points to acquire a three-dimensional anatomical model of the atrial to be modeled. Thus, the present invention makes it possible to construct a three-dimensional anatomical model of the atrial to be modeled using a trained neural network based on the position coordinate information of some feature points in the atrial to be modeled, and to achieve fully automated construction of a three-dimensional anatomical model of the atrial without requiring manual intervention.

Description

[Technical Field]

[0001] The present invention relates to the field of electrocardiography, and more specifically, to a method for constructing a three-dimensional anatomical model of the heart and a three-dimensional mapping system for the heart. [Background technology]

[0002] Cardiac mapping systems involve implanting electrode catheters in the heart, locating the catheters using magnetic or electric fields, manipulating the catheters to ensure they adhere to the inner walls of the heart chambers and acquire spatial positional data, and then manipulating the catheters to fully sample the spatial coordinate information of each point on the inner wall of the atrial cavity. This process yields high-density point cloud data, which can then be used to construct a three-dimensional anatomical model of the atrium. However, access to certain areas of the atrium, such as the left and right pulmonary veins and atrial appendages, is difficult, significantly increasing the difficulty of catheter manipulation. Furthermore, the density of acquired spatial points may be insufficient, potentially leading to inaccurate three-dimensional anatomical models. Additionally, directly constructing a three-dimensional anatomical model of the atrium using point cloud data often requires manual correction by experts, such as highlighting vein locations and the elevation between the left superior vein and left atrial appendage, or creating holes for the mitral and tricuspid valves.

[0003] Therefore, currently, constructing a three-dimensional anatomical model of the heart requires reliance on a large number of sampled, high-density spatial points, which presents challenges such as inaccuracies in model construction and the need for manual correction by experts. [Overview of the project]

[0004] To address the above-mentioned problems, the present invention provides a method for constructing a three-dimensional anatomical model of the heart and a three-dimensional mapping system for the heart.

[0005] According to a first embodiment, one embodiment provides a method for constructing a three-dimensional anatomical model of the heart based on point cloud data. The method is A step of obtaining an atrial template model, wherein the atrial template model is a three-dimensional mesh model having a predetermined atrial boundary, the atrial template model is provided with a plurality of predetermined control feature points, and the atrial template model is deformed according to the displacement of the control feature points. A step of acquiring point cloud data of a to-model atrium within the heart, wherein the point cloud data of the to-model atrium is position coordinate information of some feature points in the to-model atrium; The steps include: mapping some of the feature points in the atrium to be modeled to the atrial template model using an affine transformation network, and obtaining point cloud data mapped to the atrial template model; The method includes the steps of inputting the point cloud data mapped to the atrial template model into a prediction network and obtaining the displacement of control feature points in the atrial template model by prediction, and deforming the atrial template model based on the displacement of the control feature points to obtain a three-dimensional anatomical model of the atrium to be modeled.

[0006] According to a second aspect, in one embodiment, a three-dimensional mapping system for the heart is provided. The system is A data acquisition device for obtaining point cloud data of the atrium to be modeled within the heart, A processor for performing the method for constructing a three-dimensional anatomical model of the heart described in any one of the above embodiments and outputting the three-dimensional anatomical model of the atrium to be modeled, The system includes a display device for displaying a three-dimensional anatomical model of the atrium to be modeled.

[0007] According to a third aspect, in one embodiment, a computer-readable storage medium is provided. The medium stores a program that can be performed by a processor according to the method of any one of the above embodiments.

[0008] According to the method for constructing a 3D anatomical model of the heart and the 3D mapping system for the heart in the above embodiment, an atrial template model and point cloud data of the atrium to be modeled are acquired. The point cloud data is the position coordinate information of some feature points in the atrium to be modeled. An affine transformation network is used to map some feature points in the atrium to be modeled to the atrial template model. The point cloud data mapped to the atrial template model is input to a prediction network, and the displacement of control feature points in the atrial template model is obtained by prediction. Based on the displacement of the control feature points, the atrial template model is deformed to obtain a 3D anatomical model of the atrium to be modeled. Thus, the present invention can construct a 3D anatomical model of the atrium to be modeled using a trained neural network based on the position coordinate information of some feature points in the atrium to be modeled, and can achieve fully automated construction of a 3D anatomical model of the atrium without requiring manual intervention. [Brief explanation of the drawing]

[0009] [Figure 1] This is a flowchart illustrating a method for constructing a three-dimensional anatomical model of the heart based on point cloud data according to an embodiment of the present invention. [Figure 2] This is a specific flowchart illustrating a method for constructing a three-dimensional anatomical model of the heart according to one embodiment. [Figure 3] This is a schematic diagram of a three-dimensional heart mapping system according to an embodiment of the present invention. [Modes for carrying out the invention]

[0010] Hereinafter, the present invention will be described in more detail with reference to the drawings through specific embodiments. Similar components in different embodiments have related similar component numbers. In the following detailed description, many details are described to make the present application easier to understand. However, those skilled in the art can easily understand that some features may be omitted under different circumstances or replaced by other components, materials, and methods. In some cases, some operations related to the present application are not illustrated or described in this specification in order to avoid burying the core part of the present application in excessive explanations. For those skilled in the art, it is not necessary to describe these related operations in detail, and based on the description in the specification and general knowledge in the technical field, the related operations can be fully grasped.

[0011] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable way to form various embodiments. Also, the steps or actions in the description of the method can be reordered or adjusted in a way that is obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for clearly explaining specific embodiments and do not mean a necessary order unless otherwise specified that a specific order must be followed.

[0012] The numbering of components in this specification, such as "first", "second", etc., is only used to distinguish the described objects and has no order or technical meaning. The "connection" and "coupling" described in this application include both direct and indirect connections (couplings) unless otherwise specified. The documents related to the embodiments of the present invention are as follows.

[0013] Reference 1: High-resolution 3D surface construction algorithm (Lorensen, WE, & Cline, HE (1987). Marching cubes: A high-resolution 3D surface construction algorithm. Proceedings of the 14th annual conference on Computer graphics and interactive techniques.) Reference 2: Linear subspace design method for real-time shape deformation (Wang, Y., Jacobson, A., Barbic, J., & Kavan, L. (2015). Linear subspace design for real-time shape deformation. ACM Transactions on Graphics (TOG), 34, 1 - 11.); Reference 3: Method for generating cardiac meshes based on deep learning simulations of patient images (Kong, F., & Shadden, SC (2022). Learning Whole Heart Mesh Generation From Patient Images for Computational Simulations. IEEE Transactions on Medical Imaging, 42, 533-545.) Reference 4: Deep Hierarchical Feature Learning on Point Sets in a Metric Space (Qi, C., Yi, L., Su, H., & Guibas, LJ (2017). PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. NIPS.) Reference 5: Deep Learning Method for Point Sets for 3D Classification and Segmentation (Qi, C., Su, H., Mo, K., & Guibas, LJ (2016). PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 77-85.)

[0014] In embodiments of the present invention, a complete and relatively accurate three-dimensional anatomical model of the atrium can be reconstructed based on a pre-constructed atrial template model and a small amount of positional coordinate information of feature points within the atrium that has been collected.

[0015] Referring to Figure 1, an embodiment of the present invention provides a method for constructing a three-dimensional anatomical model of the heart based on point cloud data. Hereinafter, this will be abbreviated as the method for constructing a three-dimensional anatomical model of the heart. The method for constructing a three-dimensional anatomical model of the heart includes steps 101 to 104. These will be described in detail below.

[0016] Step 101: Obtain an atrial template model. The atrial template model is a three-dimensional mesh model having a predetermined atrial boundary, and the atrial template model is provided with a plurality of predetermined control feature points, and the atrial template model deforms according to the displacement of the control feature points.

[0017] The atrial template model in this embodiment is a pre-constructed, known, and satisfying atrial three-dimensional mesh model template that meets the requirements for subsequent electrophysiological analysis. While it can be constructed using various existing methods, this embodiment provides the following construction method.

[0018] A CT scan image of the heart is selected, and the anatomical structures of the left and right atria of the heart are separated from this CT scan image, and the separated anatomical structures are converted into a 3D mesh model. In this embodiment, isosurfaces can be extracted using the Marching cube method (Reference 1) to obtain a 3D mesh model representing the atrial boundary, and this 3D mesh model is the atrial template model. Several keypoints need to be selected as control feature points on the mesh surface of the atrial template model. A specific linear mapping relationship exists between the positions of these control feature points and each point of the mesh of the atrial template model, so that the control feature points are used to control the deformation of the mesh of the atrial template model. That is, the displacement of the control feature points and the displacement of each mesh vertex of the atrial template model have a linear mapping relationship calculated by Biharmonic coordinates (Reference 2). In one embodiment, a linear mapping relationship (W∈R) relating each control feature point and the displacement of each mesh vertex is obtained by solving an optimization problem relating to the biharmonic equations. (n×c) ) can be calculated. Here, n represents the number of mesh vertices in the atrial template model, c represents the number of control feature points in the atrial template model, and R (n×c) This represents the mapping set between the mesh vertices of the atrial template model and the control feature points in the atrial template model. The mesh vertices of the atrial template model can be smoothly deformed according to the coordinate changes of the control feature points, and the coordinates of the mesh vertices of the atrial template model are P∈R (n×3) Toshi, R (n×3) Q∈R represents the set of position coordinate information for the mesh vertices of the atrial template model, and the coordinates of the control feature points are Q∈R. (c×3) Toshi, R (c×3) If we consider the set of position coordinate information for control feature points, then the coordinates of the mesh vertices and the coordinates of the control feature points satisfy P=WQ. Matrix S∈R (c×n) Assuming that represents the selection of Q from P, the linear constraint SP = Q must be satisfied.

[0019] Furthermore, control feature points can be selected in various ways depending on the situation. For example, multiple points on the mesh surface of the atrial template model can be uniformly selected as control feature points, or points at important locations in important atria (e.g., pulmonary vein inlet, atrial apex, etc.) can be selected as control feature points.

[0020] Step 102: Obtain point cloud data of the atrium to be modeled within the heart. Here, the point cloud data of the atrium to be modeled is position coordinate information of some feature points in the atrium to be modeled. The atrium to be modeled may be the left atrium or the right atrium of the heart. This embodiment uses an electrode catheter and / or electrode pads to collect position coordinate information of feature points within the atrium to be modeled. Position coordinate information of feature points within the atrium may be collected using the electrode catheter alone, or the position coordinate information of feature points within the atrium may be collected using a combination of the electrode catheter and electrode pads. The electrode catheter is a catheter having one or more electrodes, and position coordinate information of feature points within the atrium is collected by inserting the catheter into the cardiac chamber and moving the electrodes. The electrode pads are attached to the chest or back of the subject to position the electrodes mounted on the catheter, thereby enabling the electrode catheter to collect more accurate position coordinate information of feature points.

[0021] Furthermore, the characteristic points of the modeled atrium collected in this embodiment are few in number, and there is no need to manipulate the electrode catheter to collect them intensively.

[0022] Step 103: Using an affine transformation network, some feature points in the atrium to be modeled are mapped to an atrial template model, and point cloud data mapped to the atrial template model is obtained. The atrial template model and point cloud data obtained in Steps 101 and 102 are used as input to the affine transformation network, and the affine transformation network outputs point cloud data mapped to the atrial template model.

[0023] Before inputting the point cloud data into the affine transformation network, it is necessary to perform normalization processing on the point cloud data obtained in step 102. By subtracting the center-of-mass coordinates of the point cloud data from the coordinates of each feature point in the point cloud data and dividing by the maximum radius from the center of mass, the normalized point cloud data can be obtained. The normalized point cloud data is input into the affine transformation network. In this embodiment, a fully-connected neural network is used to predict the affine transformation between the atrial model represented by each feature point in the point cloud data and the atrial template model, and further map each feature point in the input point cloud data to the atrial template model.

[0024] In one embodiment, using the affine transformation network to map the partial feature points in the atrial model to be modeled to the atrial template model, the step of obtaining the point cloud data mapped to the atrial template model is as follows: The step of performing normalization processing on the point cloud data of the atrial model to be modeled; Inputting the point cloud data after normalization processing into a fully-connected neural network to obtain an affine transformation matrix between the atrial model to be modeled and the atrial template model by prediction; Based on the affine transformation matrix obtained by prediction, mapping some feature points in the atrial model to be modeled to the atrial template model, and obtaining the point cloud data mapped to the atrial template model.

[0025] Here, the affine transformation network includes a ReLU activation function and a batch normalization layer. Let the function represented by this affine transformation network be f, then f is from R (m×3) to R q , and σ = f θ (Q, V init ). Here, m represents the number of feature points in the point cloud data input into the affine transformation network, and R (m×3)represents the point cloud data set input to the affine transformation network, q represents the number of parameters required for the described affine transformation, and R q σ represents the set of parameters required for the affine transformation, where σ is the parameter of the affine transformation and θ is the parameter of the affine transformation network. Typically, any one affine transformation needs to be represented by 12 parameters, but the number of parameters may be reduced as needed. For example, if the atrium corresponding to the collected point cloud data is approximately unidirectional with the atrial template model, then to roughly align the atrial template model, only three parameters representing scaling in the x, y, and z directions are needed, as only the size of the atrium represented by the point cloud data needs to be adjusted, and no other parameters representing rotation are needed. Finally, the parameters output by the affine transformation network are converted into the corresponding affine transformation matrix to map the feature points in the input point cloud data to the atrial template model.

[0026] Step 104: Input the point cloud data mapped to the atrial template model into the prediction network to predict the displacement of control feature points in the atrial template model. Based on the displacement of the control feature points, deform the atrial template model to obtain a 3D anatomical model of the atrium being modeled.

[0027] In this embodiment, point cloud data is mapped to an atrial template model and then input into a prediction network. The prediction network is used to predict the displacement of control feature points in the atrial template model. By adding the displacement of the control feature points to the initial positions of the control feature points in the atrial template model, the atrial template model can be deformed. Based on the deformed atrial template model and the initially input point cloud data, a three-dimensional anatomical model of the atrium to be modeled is obtained. That is, V p =(T1 -1 (WP init +ΔP) T ) T And here, T1 -1V is the inverse transform of the affine transform output by the affine transform network. p ΔP represents the mesh vertex coordinates of the 3D anatomical model of the atrium being modeled, and P represents the displacement of the control feature points. init This represents the initial position of the control feature point.

[0028] In one embodiment, a neural network similar to the PointNet++ structure (Reference 4) can be employed as the prediction network, and the PointNet++ neural network includes multiple set abstraction modules. The step of inputting point cloud data mapped to an atrial template model into the prediction network and obtaining the displacement of control feature points in the atrial template model by prediction is as follows: The point cloud data mapped to the atrial template model is input to the first point set abstraction module, the data output by the first point set abstraction module is then input to the second point set abstraction module, and this process is repeated until the data output by the final point set abstraction module is used as the displacement of the control feature points in the atrial template model obtained by prediction.

[0029] Here, in each point set abstraction module, the following operations are performed.

[0030] k points are randomly sampled from the input data, and m points are selected within a sphere of a specified radius around each sampled point. Each m point corresponding to a sampled point is considered a set of points. Feature information is extracted from each set of points. In this embodiment, feature information may be extracted according to the method provided in Reference 5, and global feature information corresponding to each sampled point is obtained using a maximum pooling operation. The k sampled points and the global feature information corresponding to each sampled point are output.

[0031] Unlike typical PointNet++ neural networks that randomly sample points from input point cloud data by sampling the furthest points, the prediction network in this embodiment groups input feature points by selecting a specified number of center points in the atrial template model by sampling the furthest points. Since the center points can be pre-selected in the atrial template model, the calculation of sampling the furthest points is not required during the training and testing processes, saving time. Furthermore, by directly selecting center points from the atrial template model rather than from the input data, it is possible to ensure that the final displacement is predicted around the control feature points, thus improving the accuracy of the prediction.

[0032] As described above, the method for constructing a three-dimensional anatomical model of the heart according to the embodiment of the present invention requires the use of an affine transformation network and a predictive network. Therefore, before using the affine transformation network and the predictive network, they must first be trained. The training process is as follows.

[0033] (1) Construct training data. The training data consists of two parts: one is the ground truth of the 3D atrial anatomy model, and the other is the point cloud data for training. To obtain the ground truth of the 3D atrial anatomy model, multiple cardiac CT images are acquired, and then, using an existing machine learning method (Reference 3), the 3D atrial anatomy model represented by a 3D mesh is extracted from the multiple cardiac CT images, and these 3D atrial anatomy models extracted from the cardiac CT images are used as the ground truth for training. For the point cloud data for training, point cloud data can be randomly generated as training point cloud data by simulating the process by which electrodes acquire points in the atrium. Here, the random generation of point cloud data by simulating the process by which electrodes acquire points in the atrium may be as follows: Since the shapes of electrodes of different patterns (linear, mesh) can both be represented by analysis, the coordinates of the electrodes can be recorded in a timely manner, i.e., point cloud data can be obtained, by simulating the motion trajectory of the electrodes in the atrium. The electrode's motion trajectory can include translation and rotation to reach the entrance of the pulmonary veins and other important locations in the atria, and by adding a certain degree of randomness to the electrode's motion trajectory, it can also simulate the uncertainties in actual operation.

[0034] All training data is divided into training datasets, validation datasets, and test datasets in predetermined proportions. Here, the training dataset is used to train the neural network and optimize its parameters, the validation dataset is used to tune the hyperparameters of this deep learning method, and the test dataset is used to verify the accuracy of the point cloud meshing. Furthermore, the training data in the embodiment of the present invention may include actually collected atrial point cloud data to further test the trained neural network.

[0035] (2) Parameter optimization is performed on the affine transformation network and the prediction network using the training data.

[0036] Regarding the optimization of the affine transformation network, the affine transformation network is a fully connected neural network, which is obtained by training it using backpropagation based on a first loss function. The first loss function is determined based on the following equation. L1=||V init -(T1V T groundtruth ) T || 2 2 Here, L1 represents the value of the first loss function, V init represents a matrix corresponding to the position coordinate information of each mesh vertex in the atrial template model, T1 represents the affine transformation matrix, and V groundtruth This represents a matrix corresponding to the true coordinate values ​​of each mesh vertex in the 3D network model corresponding to the atrium being modeled, || || 2 2 represents the square of the Euclidean norm.

[0037] Regarding the optimization of the prediction network, the prediction network was obtained by training it using backpropagation based on a second loss function. The second loss function is determined based on the following equation. L2=||V p -V groundtruth || 2 2 Here, L2 represents the value of the second loss function, V p This represents a matrix corresponding to the predicted coordinate values ​​of each mesh vertex in the 3D mesh model of the atrium being modeled, V groundtruth This represents a matrix corresponding to the true coordinate values ​​of each mesh vertex in the 3D anatomical model corresponding to the atrium being modeled, || || 2 2 represents the square of the Euclidean norm.

[0038] Based on the first and second loss functions, the total loss function is L = L1 + L2, and the model can be trained using the training method of stochastic gradient descent.

[0039] Referring to Figure 2, in this embodiment of the present invention, the collected point cloud data of the to-model atrium within the heart and the atrial template model (STN) are input to an affine transformation network (f) to perform an affine transformation. The affine transformation network (f) outputs affine-transformed point cloud data, which is the point cloud data mapped to the atrial template model. This affine-transformed point cloud data is input to a prediction network (PointNet++), which outputs the displacement (△P) of the control feature points obtained by prediction. Next, based on the initial position of the control feature points in the atrial template model, the position of the moved control feature points is determined, and the deformed atrial template model is obtained. Finally, the deformed atrial template model and the point cloud data of the to-model atrium within the heart initially collected (the affine-transformed point cloud data) are subjected to the inverse transformation T1 -1 Based on the above, a 3D anatomical model of the target atrium is obtained.

[0040] Referring to Figure 3, embodiments of the present invention further provide a three-dimensional mapping system for the heart. The three-dimensional mapping system for the heart comprises an acquisition device 201, a processor 202, and a display device 203. The acquisition device 201 is used to acquire point cloud data of the atrium to be modeled within the heart. The processor 202 may be used to perform a method for constructing a three-dimensional anatomical model of the heart based on the point cloud data, according to any one embodiment described above. The display device 203 is used to display the three-dimensional anatomical model of the atrium.

[0041] The data collection device 201 may be an electrode catheter and / or electrode pads. Positional coordinate information of feature points within the atria may be collected using the electrode catheter alone, or the electrode catheter and electrode pads may be used in combination to collect positional coordinate information of feature points within the atria. The electrode catheter is a catheter having one or more electrodes, and positional coordinate information of feature points within the atria is collected by inserting the catheter into the cardiac chambers and moving the electrodes. Electrode pads are attached to the chest or back of the subject to position the electrodes mounted on the catheter, thereby enabling the electrode catheter to collect more accurate positional coordinate information of feature points. The processor 202 may be an integrated circuit chip with signal processing capabilities. The processor 202 may be a general-purpose processor such as a Central Processing Unit (CPU) or a Network Processor (NP). Each method, step, and logical block diagram disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor may be a microprocessor, or it may be any conventional processor. The display device 203 may be a device capable of realizing human-computer interaction functions, and may be capable of displaying a three-dimensional anatomical model of the atrium, or of realizing functions such as receiving commands entered by the user. For example, it may be a device such as a touch panel display, and the display device 203 may also display only the three-dimensional anatomical model of the atrium independently.

[0042] It should be understood that the structure in Figure 3 is illustrative only and may include more or fewer components than those shown in Figure 3, or may have a different configuration than that shown in Figure 3. Each component shown in Figure 3 may be implemented in hardware and / or software. For example, a three-dimensional mapping system of the heart may further include memory (not shown) for storing structured three-dimensional image data of a target object that has been collected, and for storing a computer program that is executed by processor 202 to realize a technical solution in a method for constructing a three-dimensional anatomical model of the heart according to any embodiment of the above method. The memory may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electric Erasable Programmable Read-Only Memory (EEPROM), etc.

[0043] This specification has been described with reference to various exemplary embodiments. However, those skilled in the art will recognize that changes and modifications can be made to the exemplary embodiments without departing from the scope of this disclosure. For example, various operational steps and the components used to perform the operational steps may be implemented in different ways depending on the particular application or taking into account any number of cost functions related to the operation of the system (for example, one or more steps may be deleted, modified, or incorporated into other steps).

[0044] Furthermore, as will be understood by those skilled in the art, the principles herein can be reflected in a computer program product on a computer-readable storage medium preloaded with computer-readable program code. Any tangible, non-temporary, computer-readable storage medium may be used, including magnetic storage devices (such as hard disks and floppy disks), optical storage devices (such as CD-ROMs, DVDs, and Blu-ray discs), flash memory, and / or similar. These computer program instructions can be loaded into a general-purpose computer, a dedicated computer, or other programmable data processing device to form a machine, and as a result, these instructions executed on the computer or other programmable data processing device can generate means for implementing a specified function. These computer program instructions may also be stored in computer-readable memory, which can instruct the computer or other programmable data processing device to function in a particular way, and thus, instructions stored in computer-readable memory can form a product including means for implementing a specified function. Computer program instructions can also be loaded into a computer or other programmable data processing device to execute a series of operational steps on the computer or other programmable device to generate a computer implementation process, thereby enabling the instructions executed on the computer or other programmable device to provide steps for implementing a specified function.

[0045] While the principles of this specification are illustrated in various embodiments, many modifications of structure, arrangement, proportions, elements, materials, and components are used to suit specific environmental and operating requirements, provided they do not deviate from the principles and scope of this disclosure. Such modifications and other variations or modifications are intended to be included within the scope of this specification.

[0046] The detailed description above has been given with reference to various embodiments. However, those skilled in the art will recognize that various modifications and changes can be made without departing from the scope of this disclosure. Therefore, this disclosure should be considered illustrative and not restrictive, and all such changes are intended to be within its scope. Similarly, advantages, other benefits, and solutions to problems relating to various embodiments have been described above. However, advantages, benefits, solutions to problems, and the elements that give rise to them, or solutions that make them clearer, should not be construed as important, necessary, or essential. The term “including” as used herein and any other variation thereof means non-exclusive inclusion, and such process, method, article, or apparatus including a list of elements also includes not only those elements but also other elements that are not expressly enumerated or that do not belong to this process, method, system, article, or apparatus. Furthermore, the term “connected” as used herein and any other variation thereof means physical, electrical, magnetic, optical, communication, functional, and / or any other connection.

[0047] Those skilled in the art will recognize that many modifications can be made to the details of the embodiments described above without departing from the basic principles of the present invention. Accordingly, the scope of the present invention should be determined according to the following claims.

Claims

1. A method for constructing a three-dimensional anatomical model of the heart based on point cloud data, A step of obtaining an atrial template model, wherein the atrial template model is a three-dimensional mesh model having a predetermined atrial boundary, the atrial template model is provided with a plurality of predetermined control feature points, and the atrial template model is deformed according to the displacement of the control feature points. A step of acquiring point cloud data of a to-model atrium within the heart, wherein the point cloud data of the to-model atrium is position coordinate information of some feature points in the to-model atrium; The steps include: mapping some of the feature points in the atrium to be modeled to the atrial template model using an affine transformation network, and obtaining point cloud data mapped to the atrial template model; The steps include inputting the point cloud data mapped to the atrial template model into a prediction network and obtaining the displacement of control feature points in the atrial template model by prediction, A method characterized by comprising the step of deforming the atrial template model based on the displacement of the control feature points to obtain a three-dimensional anatomical model of the atrium to be modeled.

2. The method according to claim 1, characterized in that the displacement of the control feature point and the displacement of each mesh vertex in the atrial template model have a linear mapping relationship.

3. The step of acquiring point cloud data of the atrium to be modeled within the heart is: The method according to claim 1, characterized by comprising the step of collecting position coordinate information of some feature points in the atrium to be modeled using electrodes, and obtaining point cloud data of the atrium to be modeled.

4. The steps of using an affine transformation network to map some of the feature points in the atrium to be modeled to the atrial template model and obtaining point cloud data mapped to the atrial template model are: The steps include: applying normalization to the point cloud data of the atrial to be modeled; The process involves inputting the normalized point cloud data into a fully connected neural network to predict and obtain the affine transformation matrix between the atrium to be modeled and the atrial template model, and The method according to claim 1, comprising the steps of mapping some feature points in the atrium to be modeled to the atrial template model based on the affine transformation matrix obtained by prediction, and obtaining point cloud data mapped to the atrial template model.

5. The aforementioned fully connected neural network was obtained by training it using backpropagation based on a first loss function, the first loss function being determined based on the following equation: L １ =||V init -(T １ V Ｔ groundtruth ) Ｔ || 2 2 Here, L １ represents the value of the first loss function, V init represents a matrix corresponding to the position coordinate information of each mesh vertex in the atrial template model, T １ represents an affine transformation matrix, V groundtruth represents a matrix corresponding to the true coordinate values of each mesh vertex in the three-dimensional network model corresponding to the atrial model to be modeled, || || 2 2 represents the square of the Euclidean norm. The method according to claim 4, characterized in that

6. The prediction network includes a plurality of point set abstraction modules, and the step of inputting the point cloud data mapped to the atrial template model into the prediction network and obtaining the displacement of control feature points in the atrial template model by prediction is: The point cloud data mapped to the atrial template model is input to the first point set abstraction module, the data output by the first point set abstraction module is then input to the second point set abstraction module, and this process is repeated until an inference is made. The data output by the final point set abstraction module is taken as the displacement of the control feature points in the atrial template model obtained by prediction. Here, in each point set abstraction module, The method according to claim 1, characterized by performing the following operations: randomly sampling k points from input data, selecting m points within a sphere of a specified radius from each sampled point as the center point, forming a set of m points corresponding to each sampled point, extracting feature information from each set of points, obtaining global feature information corresponding to each sampled point using a maximum pooling operation, and outputting the k sampled points and the global feature information corresponding to each sampled point.

7. The prediction network was obtained by training it using backpropagation based on a second loss function, which is determined based on the following equation: L ２ =||V ｐ -V groundtruth || 2 2 Here, L ２ This represents the value of the second loss function, V ｐ V represents a matrix corresponding to the predicted coordinate values ​​of each mesh vertex in the three-dimensional mesh model of the atrium being modeled, groundtruth This represents a matrix corresponding to the true coordinate values ​​of each mesh vertex in the three-dimensional anatomical model corresponding to the atrium being modeled, || || 2 2 The method according to 6, characterized in that represents the square of the Euclidean norm.

8. It is a three-dimensional mapping system for the heart. A data acquisition device for obtaining point cloud data of the atrium to be modeled within the heart, A processor for performing the method for constructing a three-dimensional anatomical model of the heart according to any one of claims 1 to 7 and outputting the three-dimensional anatomical model of the atrium to be modeled, A three-dimensional mapping system for the heart, comprising a display device for displaying a three-dimensional anatomical model of the atrium to be modeled.

9. The cardiac three-dimensional mapping system according to claim 8, characterized in that the collection device includes at least one electrode catheter and an electrode pad.

10. A computer-readable storage medium characterized by storing a program that, when executed by a processor, can implement the method described in any one of claims 1 to 7.

Images