3D heart shape reconstruction method based on deep neural network and model prior

Abstract

The application provides a 3D heart shape reconstruction method based on a deep neural network and model prior, comprising the following steps: constructing a three-dimensional model set of a heart, performing principal component analysis to generate a low-dimensional parameter hidden space; data enhancement, sampling from the obtained parameter space to generate a training set of a parameter regression network; constructing the parameter regression network, training the parameter regression network by using the generated training set; generating a training set of a segmentation network, constructing the segmentation network, training the segmentation network by using the generated data set; connecting the trained segmentation network and the trained parameter network to form a joint network, training the joint network by using a training set formed by real data, and performing a three-dimensional reconstruction task by using the trained joint network. The application makes the inference process efficient and fast, the result more reasonable, and the segmentation and reconstruction effects improved, and a data enhancement method is proposed to improve the generalization performance of the parameter estimation network from a large amount of synthetic data training.

Description

Technical Field

[0001] This invention relates to the field of medical 3D reconstruction technology, specifically to a 3D heart shape reconstruction method based on end-to-end deep neural networks and model priors. Background Technology

[0002] The heart is one of the body's most vital organs, and cardiovascular disease is a leading cause of death and morbidity worldwide. Therefore, modeling the heart to obtain crucial clinical parameters such as volume and ejection fraction helps physicians assess cardiovascular function and diagnose diseases. Currently, clinicians often face significant challenges in interventions guided by limited 2D images; thus, instantiating 3D shapes from limited 2D medical images is crucial to alleviating this challenge. More accurate and reliable shape models will further assist research in cardiac electrophysiology and mechanical simulation. Furthermore, the visualization of 3D heart models lowers the barrier to learning and interpreting medical images, facilitating an intuitive understanding of the heart's overall structure and function, and promoting knowledge transfer and learning.

[0003] To date, medical 3D reconstruction technology has achieved some research results. These mainly include:

[0004] 1) Segmentation-based 3D medical reconstruction methods often require fine segmentation masks, primarily focusing on the preliminary task of medical image segmentation. The process begins by outlining the volume of each image layer, then using the Marching Cubes algorithm to generate a mesh for the segmented regions, followed by manual surface post-processing or smoothing. However, this method is highly dependent on segmentation quality and cannot be easily trained end-to-end to obtain fine 3D shapes. Furthermore, due to the low resolution of medical images, the segmentation results often produce surfaces with step-like artifacts.

[0005] 2) Model-based methods are generally two-stage: first, a coarse segment is generated, and then the segment contour is fitted by a constructed statistical shape model (SSM) through least squares regression to complete shape prediction. However, these two stages are often isolated from each other. Deformable model-based methods, on the other hand, use registration to minimize the energy function to make the reference mesh and the target contour consistent, but the computation is very complex.

[0006] 3) Methods that directly reconstruct surfaces are dedicated to building the geometric structure of point clouds or meshes directly from images. Some extend this to modeling irregular graphs and using graph convolutional neural networks, but they can often only reconstruct some simple structures.

[0007] In conclusion, existing methods have shortcomings, and there is no satisfactory solution yet. Summary of the Invention

[0008] To address the aforementioned technical issues, such as the need for a large amount of labeled data during training, this invention aims to provide a 3D heart shape reconstruction method based on end-to-end deep neural networks and model priors.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A 3D heart shape reconstruction method based on end-to-end deep neural networks and model priors includes the following steps:

[0011] Step S1: Construct a three-dimensional model set of the heart from the original CT scan, and perform principal component analysis to generate a low-dimensional parametric latent space;

[0012] Step S2: Generate a training set for the parameter regression network from the 3D model set, and enhance the training set of the parameter regression network by sampling the latent space of the parameters obtained in step S1.

[0013] Step S3: Construct a parametric regression network by training the parametric regression network using the training set generated in step S2.

[0014] Step S4: Generate a training set for the segmentation network from the 3D model set, construct the segmentation network, and train the segmentation network using the training set.

[0015] Step S5: Connect the segmentation network trained in step S4 with the parameter regression network trained in step S3 to form a joint network, and train the joint network.

[0016] Step S6: Perform the 3D reconstruction task using the trained joint network.

[0017] Furthermore, the specific steps of step S1 are as follows:

[0018] 1-1. Slice each original CT three-dimensional volume along the cross-sectional normal direction to obtain a series of two-dimensional images, and segment the left ventricle in each two-dimensional image to obtain a mask for the two-dimensional image.

[0019] 1-2. Stack a series of 2D image masks to form a mask volume, use the Marching Cubes algorithm to generate surface meshes for segmented regions, and perform smoothing operations to obtain a 3D model of the heart.

[0020] 1-3. Perform rigid registration and non-rigid registration on all the obtained 3D models in sequence to make the spatial distribution and topological structure of all 3D models similar.

[0021] 1-4. Apply principal component analysis to all 3D models to obtain the parameters of each 3D model, thereby generating a low-dimensional parameter latent space;

[0022] The 3D model set consists of the original CT 3D volume, mask volume, 3D model, and parameters of the 3D model.

[0023] Furthermore, the specific steps of step S2 are as follows:

[0024] 2-1. Manually determine 5 key points on the mean model to determine the position of the tangent plane, thereby obtaining the corresponding 5 vertex indices. Since the vertices of the 3D models obtained by the parameters are in one-to-one correspondence, the key points of any 3D model can be directly determined based on the vertex indices.

[0025] 2-2. The training set of the parametric regression network consists of 3D model slices in the 3D model set obtained from the keypoint pairs in step 2-1. The resulting set of slice masks of the 3D model has corresponding labels as the parameters of the 3D model.

[0026] 2-3. Analyze the parameter distribution of the 3D model set and perform Gaussian sampling centered on the parameters of each sample to obtain parameters different from the samples; generate a 3D model from the obtained parameters, slice the 3D model using the key points obtained in step 2-1, and obtain a set of slice masks for the 3D model; add the slice masks and the parameters of the 3D model to the training set of the parameter regression network obtained in step 2-2, thereby enhancing the training set of the parameter regression network.

[0027] Furthermore, the specific steps are as follows:

[0028] 3-1. For a known 3D model, label the 5 key points obtained as A, B, C, D, and E respectively;

[0029] 3-2. The short-axis view plane sequence of the heart consists of 10 planes with the line connecting A and B as the normal and passing through points with intervals on the line connecting A and B. These 10 planes are used to generate 10 slice masks for the cross-section of the 3D model.

[0030] 3-3. The long axis two-chamber view plane of the heart is defined as the plane formed by the three points A, B, and C. A slice mask is generated for the cross section of the three-dimensional model using this plane.

[0031] 3-4. The long axis three-chamber view plane of the heart is defined as the plane formed by the three points A, B, and D. A slice mask is generated for the cross section of the three-dimensional model using this plane.

[0032] 3-5. The long axis four-chamber view plane of the heart is defined as the plane formed by the three points A, B, and E. A slice mask is generated for the cross section of the three-dimensional model using this plane.

[0033] Using the above method, each 3D model can obtain a set of 13 slice masks.

[0034] Furthermore, the parameter regression network described in step S3 consists of 4 downsampling modules and 3 fully connected layers. Each downsampling module consists of two 3×3 convolutional layers, two ReLU activation functions, and one 2×2 pooling layer. The network input is a set of 13 slice masks with a size of 192×192, and the output is a K-dimensional parameter vector.

[0035] Furthermore, the loss function of the parameter regression network described in step S3 is:

[0036]

[0037] weight w i Positively correlated with the variance ratio of the eigenvectors, α i It is the true value of the model parameters in the i-th dimension. It is the predicted value of the model parameters in the i-th dimension.

[0038] Furthermore, the specific method for generating the training set of the segmentation network in step S4 is as follows: for each original CT 3D volume and its corresponding mask volume in the 3D model set, the original CT 3D volume is sliced ​​to obtain view images and their corresponding slice masks; the training set of the generated segmentation network consists of 13 view images obtained from an original CT 3D volume, and the corresponding labels are 13 slice masks.

[0039] Furthermore, the specific steps for obtaining the view image and its corresponding slice mask are as follows:

[0040] 4-1. Locate the midpoint between the apex of the left ventricle and the mitral valve on the cross-sectional view of the original CT three-dimensional volume, and then scan along this line to obtain the long axis two-chamber view image. Then perform the same operation on the mask volume to obtain the slice mask.

[0041] 4-2. On the long-axis two-chamber view plane, scan the line connecting the midpoint of the mitral valve and the apex of the left ventricle to obtain the long-axis four-chamber view image and its slice mask. Then perform the same operation on the mask volume to obtain the slice mask.

[0042] 4-3. Using the long axis four-chamber view plane as the reference plane, continuously scan vertically along the long axis of the left ventricle to obtain a series of images to form a view sequence image. Then, perform the same operation on the mask volume to obtain the slice mask.

[0043] 4-4. On the basal short-axis view plane, scan the line connecting the midpoint of the aortic valve and the midpoint of the left ventricle to obtain the long-axis three-chamber view image and its slice mask. Then perform the same operation on the mask volume to obtain the slice mask.

[0044] Using the above method, a set of 13 view images and corresponding slice masks can be obtained for each original CT 3D volume.

[0045] Furthermore, the total loss of the joint network described in step S5 is:

[0046] L joint =λ·L seg +μ·L seg

[0047] Where L seg It is the segmentation loss, L seg It is the parametric regression loss, where λ and μ are their respective weights.

[0048] The training set for the joint network consists of 13 view images obtained from an original CT 3D volume in the 3D model set, with the corresponding labels being 13 slice masks and the parameters of the 3D model.

[0049] According to claim 1, a 3D heart shape reconstruction method based on end-to-end deep neural network and model prior is characterized in that, in step S6, 13 view images obtained from any original CT 3D volume are fed into a segmentation network to obtain 13 slice masks, which are then input into a parameter regression network to predict the parameters of the 3D model, and finally the 3D model is generated from the parameters.

[0050] Compared with the closest existing technology, the technical solution provided in this aspect has the following beneficial effects:

[0051] This invention proposes a novel end-to-end deep learning network that leverages the advantages of end-to-end deep learning methods to make the inference process efficient and fast.

[0052] This invention incorporates model priors by constructing a parametric model, making the results more consistent with reality and more reasonable.

[0053] This invention proposes a joint optimization architecture consisting of a segmentation network and a parametric regression network. With the assistance of the segmentation task, personalized model parameters are obtained, while the addition of parametric regression loss is equivalent to adding the model's anatomical prior information to the segmentation, thereby making the segmentation results more reasonable.

[0054] To address the issue of small training data size, this invention proposes a data augmentation method to improve the generalization performance of parameter estimation networks by training with large amounts of synthetic data. Attached Figure Description

[0055] Figure 1 A flowchart of a 3D heart shape reconstruction method based on end-to-end deep neural network and model prior provided by the present invention;

[0056] Figure 2 The present invention provides a framework diagram of a 3D heart shape reconstruction method based on end-to-end deep neural network and model prior, wherein (a) is the parametric model construction, (b) is the pre-training stage of parametric regression network, and (c) is the joint network training stage;

[0057] Figure 3 This is a schematic diagram showing the relative positional relationship between each slice location and the three-dimensional model of the left ventricle in this invention;

[0058] Figure 4 This is a partial segmentation result image obtained using the present invention. Figure 4 In the image, (a) and (b) represent an example. In each example, the first row is the original image, the second row is the ground truth mask, and the third row is the mask generated by this method.

[0059] Figure 5 These are partial 3D reconstruction results obtained using this method. Figure 5 In the diagram, (a)(b)(c)(d)(e)(f) represent an example. In each example, the first row is the 3D model and the second row is the error plot. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Example 1

[0062] Figure 1 A flowchart of a 3D heart shape reconstruction method based on end-to-end deep neural networks and model priors provided by this invention is shown below. Figure 1 Please explain each step in detail.

[0063] Step S110: Construct a three-dimensional model set of the heart from the original CT scan, and perform principal component analysis to generate a low-dimensional parametric latent space.

[0064] In this embodiment, the sample set of the acquired original cardiac CT images contains multiple three-dimensional volumes. Each three-dimensional volume is sliced ​​along the cross-sectional normal direction to obtain a series of two-dimensional images, and the left ventricle in each two-dimensional image is segmented.

[0065] The mask volume is obtained by stacking masks of two-dimensional images, and the surface mesh of the segmented region is generated using the Marching Cubes algorithm. After smoothing, a three-dimensional model of the left ventricle is obtained.

[0066] Assuming a set of three-dimensional models of the left ventricle There are M models, each with a different number of vertices. The parameterized model mainly consists of parameters. Decision made. Once the parameters are obtained, the heart model can be reconstructed.

[0067] First, the Iterative Closest Point (ICP) algorithm is used to rigidly register each 3D model with the template. Then, the template is non-rigidly registered and deformed onto each 3D model, ensuring that all 3D models have similar spatial distribution and consistent topology. This results in a set of 3D models. and Where N is the number of vertices in the template.

[0068] After the 3D model set is constructed, its parameter latent space can be obtained through statistical analysis. We use Principal Component Analysis (PCA) on the set obtained through registration. The parameters for constructing the model. The shape S i Mapping to vector Thus obtaining S map =[s1,s2,…,s M and average shape The singular value decomposition (SVD) of the covariance matrix is ​​used to obtain... The formula for obtaining the 3D shape from the parameters is as follows:

[0069]

[0070] Where α represents the parameter belonging to the parametric latent space, and K is the dimension of the parametric latent space.

[0071] In this embodiment, M is 150, M is 8000, and K is 50.

[0072] The constructed 3D model set consists of the original CT 3D volume, mask volume, 3D model, and parameters of the 3D model.

[0073] Step S120: Generate a training set for the parameter regression network from the 3D model set, and enhance the training set of the parameter regression network by sampling the latent space of the parameters obtained in step S110.

[0074] The training set used in this method for the parametric regression network is generated as follows:

[0075] Based on experience, radiologists manually determined five key points on the mean model to determine the position of the slicing plane. Since the vertices of the 3D models obtained from the parameters are in one-to-one correspondence, the key points of any 3D model can be directly determined based on the vertex indices. For each 3D model, slices are made on a specific plane using the obtained key points (labeled A, B, C, D, and E), resulting in a set of slice masks for the 3D model. A schematic diagram illustrating the relative positional relationship between each slice location and the 3D model of the left ventricle is shown below. Figure 3 As shown. The method for obtaining a set of slice masks for a 3D model is as follows:

[0076] 1) The short-axis view plane sequence of the heart consists of 10 planes with the line connecting A and B as the normal and passing through points with intervals on the line connecting A and B. These 10 planes are used to generate 10 slice masks for the cross section of the three-dimensional model.

[0077] 2) The long axis two-chamber view plane of the heart is defined as the plane formed by the three points A, B, and C. A slice mask is generated for the cross section of the three-dimensional model using this plane.

[0078] 3) The long axis three-chamber view plane of the heart is defined as the plane formed by the three points A, B, and D. A slice mask is generated for the cross section of the three-dimensional model using this plane.

[0079] 4) The long axis four-chamber view plane of the heart is defined as the plane formed by the three points A, B, and E. A slice mask is generated for the cross section of the three-dimensional model using this plane.

[0080] Using the above method, each 3D model can obtain a set of 13 slice masks.

[0081] Through the above implementation method, we can obtain the training set of the parameter regression network. The data consists of 13 slice masks obtained from a slice of a 3D model in the 3D model set, and the corresponding labels are the parameters that generated this 3D model.

[0082] In this method, the training set of the parametric regression network is enhanced in the following way:

[0083] The parameter distribution of the 3D model set is analyzed, and Gaussian sampling is performed with the parameters of each sample as the center to obtain parameters different from those of the samples. A 3D model is generated from the obtained parameters, and the 3D model is sliced ​​using the key points obtained in step 2-1 to obtain a set of slice masks for the 3D model. The slice masks and the parameters of the 3D model are added to the training set of the parameter regression network obtained in step 2-2 to enhance the training set of the parameter regression network.

[0084] Step S130: Construct a parametric regression network. Train the parametric regression network using the training set of the parametric regression network generated in step S120.

[0085] The parameter regression network consists of four downsampling modules and three fully connected layers. Each downsampling module comprises two 3×3 convolutional layers, two ReLU activation functions, and one 2×2 pooling layer. The network input consists of 13 slice masks of size 192×192, and the output is a K-dimensional parameter vector.

[0086] Furthermore, considering that the model parameters are generated through PCA operations, the coefficients corresponding to feature vectors with larger variance contributions are more important and better reflect the model's topological structure. Therefore, we use the weighted L2 function as the loss function for the regression network:

[0087]

[0088] weight w i Positively correlated with the variance ratio of the eigenvectors, α i It is the true value of the model parameters in the i-th dimension. It is the predicted value of the model parameters in the i-th dimension.

[0089] Step S140: Generate a training set for the segmentation network from the 3D model set, construct the segmentation network, and train the segmentation network using the training set.

[0090] The training set used by the segmentation network in this method is generated through the following steps:

[0091] 1) Locate the midpoint between the apex of the left ventricle and the mitral valve on the cross-sectional view of the original CT three-dimensional volume, and then scan along this line to obtain the long axis two-chamber view image. Then perform the same operation on the mask volume to obtain the slice mask.

[0092] 2) On the long-axis two-chamber view plane, scan the line connecting the midpoint of the mitral valve and the apex of the left ventricle to obtain the long-axis four-chamber view image and its slice mask. Then perform the same operation on the mask volume to obtain the slice mask.

[0093] 3) Using the long axis four-chamber view plane as the reference plane, continuously scan vertically along the long axis of the left ventricle to obtain a series of images to form a view sequence image. Then, perform the same operation on the mask volume to obtain the slice mask.

[0094] 4) On the basal short-axis view plane, scan the line connecting the midpoint of the aortic valve and the midpoint of the left ventricle to obtain the long-axis three-chamber view image and its slice mask. Then perform the same operation on the mask volume to obtain the slice mask.

[0095] Through the above implementation method, a total of 13 view images and their slice masks can be obtained. The pixel value in each view image is obtained by trilinear interpolation of the original CT volume.

[0096] Through the above implementation method, a training set for the segmentation network can be obtained. The data consists of 13 view images obtained from an original CT 3D volume in the 3D model set, and the corresponding labels are 13 slice masks.

[0097] Preferably, the segmentation network is a UNet network, consisting of a downsampling path, an upsampling path, and skip connections. The downsampling path consists of four downsampling modules, each composed of two 3x3 convolutional layers, two ReLU activation functions, and one 2x2 pooling layer. The upsampling path consists of four upsampling modules, each composed of two 3x3 convolutional layers, two ReLU activation functions, and one bilinear interpolation function. The feature maps generated in the downsampling stage are combined with the feature maps generated in the upsampling stage of the same layer through skip connections. The network input consists of 13 images of size 192×192, and the output consists of 13 slice masks of size 192×192. The loss function is:

[0098]

[0099] Among them, y i p(y) represents the label of pixel i. i ) is the probability of being predicted as positive.

[0100] Step S150: Connect the segmentation network trained in step S140 and the parameter regression network trained in step S130 to form a joint network, and train the joint network.

[0101] The training set used in this method for the joint network consists of 13 view images obtained from an original CT 3D volume in the 3D model set, with the corresponding labels being 13 slice masks and the parameters of the 3D model.

[0102] The total loss of the joint network is:

[0103] L joint =λ·L seg +μ·L seg

[0104] Where L seg It is the segmentation loss, L seg It is a parametric regression loss, where λ and μ are their respective weights. Step S160: Perform the 3D reconstruction task using the trained joint network.

[0105] Thirteen view images obtained from any original CT 3D volume are processed by a segmentation network to obtain 13 slice masks. These masks are then input into a parametric regression network to predict the parameters of the 3D model. Finally, the 3D model is generated from these parameters. The process is as follows: Figure 2 As shown in (c).

[0106] Through the above implementation methods, we can obtain the corresponding three-dimensional model from the original CT volume. The partial segmentation output results of this invention are as follows: Figure 4 As shown, some of the reconstruction output results are as follows: Figure 5 As shown.

[0107] Any aspects of this invention not described in detail are well-known to those skilled in the art.

[0108] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for 3D heart shape reconstruction based on deep neural networks and model priors, characterized in that, Includes the following steps: Step S1: Construct a three-dimensional model set of the heart from the original CT scan, and perform principal component analysis to generate a low-dimensional parametric latent space; Step S2: Generate a training set for the parameter regression network from the 3D model set, and enhance the training set of the parameter regression network by sampling the latent space of the parameters obtained in step S1. The specific steps of step S2 are as follows: 2-1. Manually determine 5 key points on the mean model to determine the position of the tangent plane, thereby obtaining the corresponding 5 vertex indices. Since the vertices of the 3D models obtained by the parameters are in one-to-one correspondence, the key points of any 3D model can be directly determined based on the vertex indices. 2-2. The training set of the parametric regression network consists of 3D model slices in the 3D model set obtained from the keypoint pairs in step 2-1. The resulting set of slice masks of the 3D model has corresponding labels as the parameters of the 3D model. 2-3. Analyze the parameter distribution of the 3D model set, and perform Gaussian sampling centered on the parameters of each sample to obtain parameters different from the samples; generate a 3D model from the obtained parameters, slice the 3D model using the key points obtained in step 2-1, and obtain a set of slice masks for the 3D model; add the slice masks and the parameters of the 3D model to the training set of the parameter regression network obtained in step 2-2, thereby enhancing the training set of the parameter regression network; Step S3: Construct a parametric regression network by training the parametric regression network using the training set generated in step S2. Step S4: Generate a training set for the segmentation network from the 3D model set, construct the segmentation network, and train the segmentation network using the training set. Step S5: Connect the segmentation network trained in step S4 with the parameter regression network trained in step S3 to form a joint network, and train the joint network. Step S6: Perform the 3D reconstruction task using the trained joint network.

2. The 3D heart shape reconstruction method based on deep neural networks and model priors according to claim 1, characterized in that, The specific steps of step S1 are as follows: 1-1. Slice each original CT three-dimensional volume along the cross-sectional normal direction to obtain a series of two-dimensional images, and segment the left ventricle in each two-dimensional image to obtain a mask for the two-dimensional image. 1-2. Stack a series of 2D image masks to form a mask volume, use the Marching Cubes algorithm to generate surface meshes for segmented regions, and perform smoothing operations to obtain a 3D model of the heart. 1-3. Perform rigid registration and non-rigid registration on all the obtained 3D models in sequence to make the spatial distribution and topological structure of all 3D models similar. 1-4. Apply principal component analysis to all 3D models to obtain the parameters of each 3D model, thereby generating a low-dimensional parameter latent space; The 3D model set consists of the original CT 3D volume, mask volume, 3D model, and parameters of the 3D model.

3. The 3D heart shape reconstruction method based on deep neural networks and model priors according to claim 1, characterized in that, The method for obtaining a set of slice masks for a 3D model comprises the following steps: 3-1. For a known 3D model, label the 5 key points obtained as A, B, C, D, and E respectively; 3-2. The short-axis view plane sequence of the heart consists of 10 planes with the line connecting A and B as the normal and passing through points with intervals on the line connecting A and B. These 10 planes are used to generate 10 slice masks for the cross-section of the 3D model. 3-3. The long axis two-chamber view plane of the heart is defined as the plane formed by the three points A, B, and C. A slice mask is generated for the cross section of the three-dimensional model using this plane. 3-4. The long axis three-chamber view plane of the heart is defined as the plane formed by the three points A, B, and D. A slice mask is generated for the cross section of the three-dimensional model using this plane. 3-5. The long axis four-chamber view plane of the heart is defined as the plane formed by the three points A, B, and E. A slice mask is generated for the cross section of the three-dimensional model using this plane. Using the above method, each 3D model can obtain a set of 13 slice masks.

4. The 3D heart shape reconstruction method based on deep neural networks and model priors according to claim 1, characterized in that, The parameter regression network described in step S3 consists of 4 downsampling modules and 3 fully connected layers. Each downsampling module consists of two 3×3 convolutional layers, two ReLU activation functions, and one 2×2 pooling layer. The network input is a set of 13 slice masks with a size of 192×192, and the output is a K-dimensional parameter vector.

5. The 3D heart shape reconstruction method based on deep neural networks and model priors according to claim 1, characterized in that, The loss function of the parametric regression network described in step S3: ； Weight Positively correlated with the variance ratio of the eigenvectors, It is the true value of the model parameters in the i-th dimension. It is the predicted value of the model parameters in the i-th dimension.

6. The 3D heart shape reconstruction method based on deep neural networks and model priors according to claim 1, characterized in that, The specific method for generating the training set of the segmentation network in step S4 is to slice the original CT three-dimensional volume and its corresponding mask volume for each original CT three-dimensional volume in the three-dimensional model set to obtain the view image and its corresponding slice mask. The training set for the generated segmentation network consists of 13 view images obtained from an original CT 3D volume, with the corresponding labels being 13 slice masks.

7. The 3D heart shape reconstruction method based on deep neural networks and model priors according to claim 6, characterized in that, The specific steps for obtaining the view image and its corresponding slice mask are as follows: 4-1. Locate the midpoint between the apex of the left ventricle and the mitral valve on the cross-sectional view of the original CT three-dimensional volume, and then scan along this line to obtain the long-axis two-chamber view image. Then perform the same operation on the mask volume to obtain the slice mask; 4-2. On the long-axis two-chamber view plane, scan the line connecting the midpoint of the mitral valve and the apex of the left ventricle to obtain the long-axis four-chamber view image and its slice mask. Then perform the same operation on the mask volume to obtain the slice mask. 4-3. Using the long axis four-chamber view plane as the reference plane, continuously scan vertically along the long axis of the left ventricle to obtain a series of images to form a view sequence image. Then, perform the same operation on the mask volume to obtain the slice mask. 4-4. On the basal short-axis view plane, scan the line connecting the midpoint of the aortic valve and the midpoint of the left ventricle to obtain the long-axis three-chamber view image and its slice mask. Then perform the same operation on the mask volume to obtain the slice mask. Using the above method, a set of 13 view images and corresponding slice masks can be obtained for each original CT 3D volume.

8. The 3D heart shape reconstruction method based on deep neural networks and model priors according to claim 1, characterized in that, The total loss of the joint network described in step S5 is: ； in It is a segmentation loss. It is parametric regression loss. and These are their respective weights; The training set for the joint network consists of 13 view images obtained from an original CT 3D volume in the 3D model set, with the corresponding labels being 13 slice masks and the parameters of the 3D model.

9. The 3D heart shape reconstruction method based on deep neural networks and model priors according to claim 1, characterized in that, In step S6, 13 view images obtained from any original CT 3D volume are fed into a segmentation network to obtain 13 slice masks, which are then input into a parametric regression network to predict the parameters of the 3D model. Finally, the 3D model is generated from the parameters.

Images