Three-dimensional reconstruction method and device for medical image sequence

Abstract

The invention discloses a three-dimensional reconstruction method and device for a medical image sequence, which can improve the reconstruction quality and efficiency under the condition of ensuring high precision, provide a reliable scheme for accurate anatomical modeling of medical imaging, have obvious advantages in processing a complex structure, and can generate a better three-dimensional grid model. The method comprises the following steps: (1) reading an original medical image sequence of CT scanning, and carrying out data preprocessing operation; (2) extracting features from the CT image volume by using a graph convolutional network, generating coordinate offset values of corresponding grid vertexes, and learning spatial position adjustment parameters of each vertex by the graph convolutional network through a neighborhood topological relation so as to establish a local topological structure; (3) performing density optimization on initial grid vertexes generated by the CT image; (4) adjusting the topological structure of the optimized grid; and (5) defining and calculating a loss function to optimize the precision of the reconstruction model.

Description

Technical Field

[0001] The present invention relates to the technical field of computer graphics and deep learning, and in particular to a three-dimensional reconstruction method of a medical image sequence, and a three-dimensional reconstruction device of a medical image sequence. Background Art

[0002] CT (Computed Tomography) is a computer tomography scan. It uses precisely collimated X-ray beams, Y-rays, ultrasound waves, etc., together with highly sensitive detectors to scan sections of a certain part of the human body one by one. It has the characteristics of fast scanning time and clear images, and can be used to examine a variety of diseases. CT images are layer images, and cross-sections are commonly used. In order to display the entire organ, the result of CT scanning is a series of CT column images.

[0003] In recent years, deep learning has achieved great success in medical image segmentation, especially in the segmentation of liver, pancreas and other abdominal structures. Despite these advances, automatic reconstruction of CT or MRI medical images remains an important challenge. Traditional methods (such as U-Net), while effective on many tasks, suffer from poor segmentation when the input is downsampled due to the high memory footprint of full-resolution medical images. In addition, there is a growing need for models that can generate arbitrary resolution outputs from coarse inputs to enhance the flexibility and applicability of these systems.

[0004] Many bioimaging tools are designed for studying the morphology of structures such as cells, organs, and tissues. Researchers and clinicians often prefer to visualize these structures as 3D surfaces at the desired level of detail, without being limited by the voxel resolution of the original data. Although volumes and distances can be estimated using voxels, surface analysis is best done using meshes. This is because meshes are able to represent complex geometries more naturally and flexibly, allowing for detailed and accurate visualization and analysis.

[0005] However, state-of-the-art methods usually generate volumes that need to be converted into surface meshes using algorithms such as Marching Cubes, followed by mesh smoothing. These algorithms introduce artifacts and hinder end-to-end training, thereby reducing the quality of the final 3D model. Artifacts during the meshing process can lead to reduced accuracy in the representation of anatomical structures, making high-fidelity reconstruction challenging. Summary of the invention

[0006] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to provide a three-dimensional reconstruction method for a medical image sequence, which can improve the reconstruction quality and efficiency while ensuring high precision, provide a reliable solution for precise anatomical modeling of medical imaging, have significant advantages in processing complex structures, and can generate better three-dimensional mesh models.

[0007] The technical solution of the present invention is: the three-dimensional reconstruction method of the medical image sequence comprises the following steps:

[0008] (1) Read the original medical image sequence of CT scan, perform data preprocessing, normalize the grayscale of each layer of image, and generate voxel data for subsequent feature extraction and analysis;

[0009] (2) Using graph convolutional networks to extract features from the CT image volume and generate coordinate offset values ​​of the corresponding mesh vertices, the graph convolutional network learns the spatial position adjustment parameters of each vertex through the neighborhood topological relationship to establish a local topological structure;

[0010] (3) Density optimization is performed on the initial mesh vertices generated from the CT image. According to the different density characteristics of the image regions, the distribution density of the vertices is adjusted to ensure that a higher vertex density is used in high-detail areas and a lower vertex density is used in low-detail areas.

[0011] (4) Adjust the topological structure of the optimized mesh, identify the topologically discontinuous or inconsistent areas in the mesh, and reconnect or adjust the vertices of these areas according to the neighborhood topological relationship to ensure that the topological structure of the mesh conforms to the actual boundaries and contours of the medical image;

[0012] (5) Define and calculate the loss function to optimize the accuracy of the reconstructed model, and optimize the model parameters by minimizing the loss functions such as chamfer loss, intersection-over-union loss, and Dice coefficient loss.

[0013] The present invention can improve the quality and efficiency of reconstruction while ensuring high precision, provide a reliable solution for precise anatomical modeling in medical imaging, have significant advantages in processing complex structures, and can generate better three-dimensional mesh models.

[0014] A three-dimensional reconstruction device for a medical image sequence is also provided, the device comprising:

[0015] The data acquisition module reads the original medical image sequence of the CT scan, performs data preprocessing operations, performs grayscale normalization processing on each layer of the image, and generates voxel data for subsequent feature extraction and analysis;

[0016] The feature extraction module uses a graph convolutional network to extract features from the CT image volume and generate coordinate offset values ​​of the corresponding mesh vertices. The graph convolutional network learns the spatial position adjustment parameters of each vertex through the neighborhood topological relationship to establish a local topological structure;

[0017] Density optimization module, which optimizes the density of the initial mesh vertices generated by the CT image, and adjusts the distribution density of the vertices according to the different density characteristics of the image area to ensure that higher vertex density is used in high-detail areas and reduced vertex density in low-detail areas;

[0018] A topology adjustment module adjusts the topology of the optimized mesh, identifies topologically discontinuous or inconsistent areas in the mesh, and reconnects or adjusts the vertices of these areas according to neighborhood topological relationships, ensuring that the topology of the mesh conforms to the actual boundaries and contours of the medical image;

[0019] The loss evaluation module defines and calculates the loss function to optimize the accuracy of the reconstructed model and optimizes the model parameters by minimizing the loss functions such as chamfer loss, intersection-over-union loss, and Dice coefficient loss. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flow chart of a method for three-dimensional reconstruction of a medical image sequence according to the present invention is shown.

[0021] Figure 2 A schematic diagram of topology modification in an embodiment of a method for three-dimensional reconstruction of a medical image sequence according to the present invention is shown. DETAILED DESCRIPTION

[0022] like Figure 1 , Figure 2 As shown, the three-dimensional reconstruction method of the medical image sequence includes the following steps:

[0023] (1) Read the original medical image sequence of CT scan, perform data preprocessing, normalize the grayscale of each layer of image, and generate voxel data for subsequent feature extraction and analysis;

[0024] (2) Using graph convolutional networks to extract features from the CT image volume and generate coordinate offset values ​​of the corresponding mesh vertices, the graph convolutional network learns the spatial position adjustment parameters of each vertex through the neighborhood topological relationship to establish a local topological structure;

[0025] (3) Density optimization is performed on the initial mesh vertices generated from the CT image. According to the different density characteristics of the image regions, the distribution density of the vertices is adjusted to ensure that a higher vertex density is used in high-detail areas and a lower vertex density is used in low-detail areas.

[0026] (4) Adjust the topological structure of the optimized mesh, identify the topologically discontinuous or inconsistent areas in the mesh, and reconnect or adjust the vertices of these areas according to the neighborhood topological relationship to ensure that the topological structure of the mesh conforms to the actual boundaries and contours of the medical image;

[0027] (5) Define and calculate the loss function to optimize the accuracy of the reconstructed model, and optimize the model parameters by minimizing the loss functions such as chamfer loss, intersection-over-union loss, and Dice coefficient loss.

[0028] The present invention can improve the quality and efficiency of reconstruction while ensuring high precision, provide a reliable solution for precise anatomical modeling in medical imaging, have significant advantages in processing complex structures, and can generate better three-dimensional mesh models.

[0029] Preferably, the step (2) comprises the following sub-steps:

[0030] (2.1) Apply the graph convolutional network to the preprocessed CT image sequence to extract the feature information of each layer of the image. The role of the graph convolutional network at this stage is to process the preprocessed CT image data and capture the structural features in the image through convolution operations. These features will be used for subsequent mesh deformation;

[0031] (2.2) Use graph convolutional networks to learn local and global features of meshes to generate high-fidelity 3D models that approximate the desired object. A 3D mesh is a collection of vertices, edges, and faces that defines the shape of a 3D object and is represented by a graph. in, represents a set of N vertices in the mesh, represents a set of E edges, each edge connecting two vertices, To attach feature vectors to vertices, the graph convolution layer defined on the irregular graph is described as:

[0032]

[0033] in, Represents the characteristics of node r in the lth layer, which is composed of the neighbor node set of node r Indicates that ω 0 and ω 1 is the weight parameter;

[0034] (2.3) The extracted features are used to adjust the coordinates of the mesh vertices to generate

[0035] 3D mesh model for CT image matching. The final position of each vertex is obtained by iteratively applying the deformation steps of the graph convolutional network. In each step, the network predicts vertex-specific displacement vectors, which are predicted relative to the mesh produced by the previous step. By iteratively applying these displacements, the template mesh is gradually adjusted to match the target object.

[0036] (2.4) In order to enhance the model’s ability to focus on key areas in a larger context, the attention mechanism is integrated into graph convolution. The attention mechanism helps the model capture global geometric relationships and identify key deformation areas. For a typical multi-head self-attention module, the input and output feature maps are represented as and At a specific pixel location (i,j)

[0037] The tensors at are represented as input and output The attention weight calculation formula is:

[0038]

[0039] Where s represents The feature dimension of the input feature f ij and f ab Through linear transformation and

[0040] Calculate the similarity and normalize it through the Softmax function to get the attention weight

[0041] Heavy, including

[0042] Preferably, in the step (2.2), the mesh decoder takes the template mesh as input. The input of the mesh decoder is a spherical mesh, in which all three-dimensional vertices form countless small planes, and the edges are used to perform graph convolution. The mesh decoder gradually refines the spherical mesh through learning to match the target object. r = 0 represents the input of the decoder, and 1 ≤ r ≤ R represents the output of the subsequent module. Each mesh vertex is defined as:

[0043] z r =h r (x r ,z r-1 ,v r-1 )

[0044] v r =v r-1 +△ t (z r )

[0045] Among them, v r represents the three-dimensional vertex coordinates after passing through the rth module; x r and z r-1 are the feature vectors generated by the rth and r-1th modules in the voxel decoder and grid decoder, respectively; h r and z rThese are two functions implemented by 4 graph convolutional layers, whose weights are learned during the training process.

[0046] Preferably, in step (2.4), in order to integrate the adjacency matrix into the self-attention layer of the graph convolution, the following steps are taken:

[0047] Q i+1 ,K i+1 ,V i+1 =h i W q ,h i W k ,h i W v

[0048]

[0049] h i+1 =S i+1 V i+1

[0050] Among them, h i represents the hidden state of the i-th layer in the graph convolution, W q , W k , W v and W A Yes

[0051] The learned parameters, s k is the feature dimension and ⊙ represents element-wise multiplication.

[0052] Preferably, the step (3) comprises the following sub-steps:

[0053] (3.1) Based on the initial mesh generated from the CT image, unpooling is applied to increase the number of vertices;

[0054] (3.2) Vertices are added using a unified solution pooling strategy, and then the mesh is deformed and the shortest distance from each candidate vertex to its parent edge is calculated; if the distance exceeds the specified threshold, the vertex is retained; otherwise, the vertex is discarded;

[0055] (3.3) Utilize the continuous mapping learned by the mesh decoder to associate the surface points on the input sphere with the surface points on the object surface, restore the edge connections by computing the convex hull on the sphere, and then transfer it directly to the target mesh;

[0056] (3.4) Feature aggregation is performed using Monte Carlo sampling implemented in the mesh network. These variant density coefficients are defined per node after pooling or unpooling, ensuring that vertex density is adaptively managed throughout the network;

[0057] (3.5) The mesh is further optimized to adapt its density to the needs of local details, including applying a local residual connection operation. In order to control the vertex density, a weighted edge length regularization term is defined, which forces shorter edge lengths in areas with higher weights, focusing more vertices where more details are needed.

[0058] Preferably, in the step (3.4), the aggregation functions of the pooling layer and the de-pooling layer are:

[0059]

[0060] in, is a learnable parameter, Represents the density value, which controls the distribution of vertices during the adaptive optimal variable density mesh unpooling process.

[0061] Preferably, in step (3.5), the weighted edge length regularization term is defined as:

[0062]

[0063] For a vertex r in the predicted grid, the weight of the nearest vertex is used to represent its own weight, and all edges connected to r share the same weight when calculating the edge length.

[0064] Preferably, the step (4) comprises the following sub-steps:

[0065] (4.1) To prevent this artifact and further improve the visual quality of the reconstructed mesh, a boundary optimization module is designed to only predict the displacement relative to each input boundary vertex. Each boundary vertex can only move on the two-dimensional plane formed by the two boundary edges intersecting the vertex; the regularization term penalizes boundary markings by forcing the boundary curve to remain smooth and consistent.

[0066] The definition is as follows:

[0067]

[0068] in, represents the set of vertices on the open boundary, Indicates the boundary

[0069] The set of neighbor vertices of vertex x;

[0070] (4.2) In order to evaluate the quality of each triangular surface, the surface quality index SQI is introduced:

[0071] The SQI is defined as:

[0072] SQI=α·E pos +β·E smooth

[0073] Among them, E pos represents the position error of the vertex, E smooth represents the smoothing error of the surface, α and β are weight parameters used to balance the importance of these two factors;

[0074] (4.3) In order to further enhance the effect of the boundary optimization module, the graph convolutional network is used to predict the displacement of boundary vertices. The displacement prediction formula of boundary vertices is:

[0075]

[0076] Among them, d(r) represents the displacement vector of vertex r, x r represents the eigenvector of vertex r, A is the adjacency matrix describing the connection relationship between boundary vertices, represents the set of neighbor vertices of vertex r, W is the weight matrix, and σ is the activation function.

[0077] Preferably, in step (5), the chamfer loss measures the average closest distance between two point sets, and evaluates the closeness between the generated 3D mesh and the target CT scan data; the IoU loss reflects the reconstruction accuracy by calculating the ratio of the intersection and union of the predicted mesh and the real mesh, and is used to evaluate the surface integrity of the reconstruction; the Dice coefficient loss evaluates the overall morphological similarity of the reconstructed surface, and determines the performance of the model by calculating the overlap between the predicted and actual meshes.

[0078] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes the steps of the above-mentioned embodiment method, and the storage medium can be: ROM / RAM, disk, CD, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a three-dimensional reconstruction device for medical image sequences, which is usually represented in the form of functional modules corresponding to the steps of the method. The device includes:

[0079] The data acquisition module reads the original medical image sequence of the CT scan, performs data preprocessing operations, performs grayscale normalization processing on each layer of the image, and generates voxel data for subsequent feature extraction and analysis;

[0080] The feature extraction module uses a graph convolutional network to extract features from the CT image volume and generate coordinate offset values ​​of the corresponding mesh vertices. The graph convolutional network learns the spatial position adjustment parameters of each vertex through the neighborhood topological relationship to establish a local topological structure;

[0081] Density optimization module, which optimizes the density of the initial mesh vertices generated by the CT image, and adjusts the distribution density of the vertices according to the different density characteristics of the image area to ensure that higher vertex density is used in high-detail areas and reduced vertex density in low-detail areas;

[0082] A topology adjustment module adjusts the topology of the optimized mesh, identifies topologically discontinuous or inconsistent areas in the mesh, and reconnects or adjusts the vertices of these areas according to neighborhood topological relationships, ensuring that the topology of the mesh conforms to the actual boundaries and contours of the medical image;

[0083] The loss evaluation module defines and calculates the loss function to optimize the accuracy of the reconstructed model and optimizes the model parameters by minimizing the loss functions such as chamfer loss, intersection-over-union loss, and Dice coefficient loss.

[0084] The present invention counts the chamfer loss distances of networks with and without topology modification in three different organ data sets. In the three organ data sets of liver, hippocampus and pancreas, the network with topology modification has a significant advantage in chamfer loss distance, and its chamfer distance is smaller than the chamfer distance of the network without topology modification. As shown in Table 1, the chamfer distance of the network with topology modification in the liver data set is 0.0020, and that of the network without topology is 0.0038, etc.

[0085] Table 1 Chamfer distance (CD) of different data sets

[0086] Dataset liver Hippocampus pancreas No topology 0.0038 0.0020 0.0040 With topology 0.0020 0.0015 0.0035

[0087] In general, this shows that the network with topology modification can reduce the chamfer loss and improve the model accuracy. On the contrary, the disadvantage of the network without topology modification is that its chamfer distance is larger in all three organ datasets, resulting in relatively low model accuracy.

[0088] During the training process, the present invention designs a This provides a feedback mechanism for model training. By calculating the loss value, we can know how the model performs in the current state. The total loss function is as follows:

[0089]

[0090] in, Indicates the chamfer distance, represents the surface normal loss, represents weighted edge length regularization, represents the vertex error estimate, represents Laplace regularization, Represents boundary regularization; the parameter weights set in the experiment are: 1 is 0.1, λ 2is 0.1, λ 3 is 0.1, λ 4 is 1, λ 5 It is 0.3.

[0091] The present invention verifies the effectiveness of the residual graph convolution, topology modification network and adaptive de-pooling strategy in the grid decoder by conducting evaluation studies.

[0092] When evaluating residual graph convolution, we systematically remove key components from the model to analyze their impact on 3D reconstruction performance. First, we remove the graph convolution residual block in the grid decoder (G-ResNet), and find that the network cannot effectively reuse features at different levels, resulting in the loss of local geometric details and partial shape loss in the final reconstruction. Next, we remove the self-attention mechanism, and the results show that the 3D shape in high-concavity areas has unexpected convexities or concavities, because graph convolution mainly relies on local information and lacks the integration of distant nodes.

[0093] For the topology modification network, after removing this module (topology modification), the retrained model has self-intersecting connections in thin and complex areas (such as liver lobes), which is due to error prediction and insufficient surface pruning. Although the surface error is small without topology modification, the key reconstruction accuracy is still affected.

[0094] In terms of the adaptive unpooling strategy, after removing the adaptive unpooling, although the boundary regularization minimizes the error difference, the surface appears jagged near thin structures.

[0095] Through these experiments, it was found that complex surfaces (such as the liver) were more affected by the removal of modules than simple geometries, indicating that the method is well suited for the reconstruction of complex anatomical structures. Overall, the complete model containing all components provided the best results, thus validating the effectiveness of these components in the model, which together ensure that the model can achieve high-precision 3D reconstruction when dealing with complex anatomical structures.

[0096] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the protection scope of the technical solution of the present invention.

Claims

1. A three-dimensional reconstruction method for a medical image sequence, characterized in that: The method comprises the following steps: (1) Read the original medical image sequence of CT scan, perform data preprocessing, normalize the grayscale of each layer of image, and generate voxel data for subsequent feature extraction and analysis; (2) Using graph convolutional networks to extract features from the CT image volume and generate coordinate offset values ​​of the corresponding mesh vertices, the graph convolutional network learns the spatial position adjustment parameters of each vertex through the neighborhood topological relationship to establish a local topological structure; (3) Density optimization is performed on the initial mesh vertices generated from the CT image. According to the different density characteristics of the image regions, the distribution density of the vertices is adjusted to ensure that a higher vertex density is used in high-detail areas and a lower vertex density is used in low-detail areas. (4) Adjust the topological structure of the optimized mesh, identify the topologically discontinuous or inconsistent areas in the mesh, and reconnect or adjust the vertices of these areas according to the neighborhood topological relationship to ensure that the topological structure of the mesh conforms to the actual boundaries and contours of the medical image; (5) Define and calculate the loss function to optimize the accuracy of the reconstructed model, and optimize the model parameters by minimizing the loss functions such as chamfer loss, intersection-over-union loss, and Dice coefficient loss.

2. The three-dimensional reconstruction method of medical image sequence according to claim 1, characterized in that: The step (2) comprises the following sub-steps: (2.1) Apply the graph convolutional network to the preprocessed CT image sequence to extract the feature information of each layer of the image. The role of the graph convolutional network at this stage is to process the preprocessed CT image data and capture the structural features in the image through convolution operations. These features will be used for subsequent mesh deformation; (2.2) Use graph convolutional networks to learn local and global features of meshes to generate high-fidelity 3D models that approximate the desired object. A 3D mesh is a collection of vertices, edges, and faces that defines the shape of a 3D object and is represented by a graph. in, represents a set of N vertices in the mesh, represents a set of E edges, each edge connecting two vertices, To attach feature vectors to vertices, the graph convolution layer defined on the irregular graph is described as: in, Represents the characteristics of node r in the lth layer, which is composed of the neighbor node set of node r Indicates that ω0 and ω1 are weight parameters; (2.3) The extracted features are used to adjust the coordinates of the mesh vertices to generate a 3D mesh model that matches the original CT image. The final position of each vertex is obtained by iteratively applying the deformation steps of the graph convolutional network. In each step, the network predicts vertex-specific displacement vectors, which are predicted relative to the mesh produced by the previous step. By iteratively applying these displacements, the template mesh is gradually adjusted to match the target object. (2.4) In order to enhance the model’s ability to focus on key areas in a larger context, the attention mechanism is integrated into graph convolution. The attention mechanism helps the model capture global geometric relationships and identify key deformation areas. For a typical multi-head self-attention module, the input and output feature maps are represented as and The tensors at a specific pixel location (i, j) are represented as input and output The attention weight calculation formula is: Where s represents The feature dimension of the input feature f ij and f ab Through linear transformation and Calculate the similarity and normalize it through the Softmax function to get the attention weight, where 3. The three-dimensional reconstruction method of medical image sequence according to claim 2, characterized in that: In the step (2.2), the mesh decoder takes the template mesh as input. The input of the mesh decoder is a spherical mesh, in which all three-dimensional vertices form countless small planes, and its edges are used to perform graph convolution. The mesh decoder gradually refines the spherical mesh through learning to match the target object. r = 0 represents the input of the decoder, and 1 ≤ r ≤ R represents the output of the subsequent module. Each mesh vertex is defined as: z r =h r (x r ,z r-1 ,v r-1 ) in r =in r-1 +△ t (with r ) Among them, v r represents the three-dimensional vertex coordinates after passing through the rth module; x r and z r-1 are the feature vectors generated by the rth and r-1th modules in the voxel decoder and grid decoder, respectively; h r and z r These are two functions implemented by 4 graph convolutional layers, whose weights are learned during the training process.

4. The three-dimensional reconstruction method of medical image sequence according to claim 3, characterized in that: In step (2.4), in order to integrate the adjacency matrix into the self-attention layer of the graph convolution, the following steps are taken: Q i+1 ,K i+1 ,V i+1 =h i W q ,h i W k ,h i W v h i+1 =S i+1 V i+1 Among them, h i represents the hidden state of the i-th layer in the graph convolution, W q , W k , W v and W A is a learnable parameter, s k is the feature dimension and ⊙ represents element-wise multiplication.

5. The three-dimensional reconstruction method of medical image sequence according to claim 4, characterized in that: The step (3) comprises the following sub-steps: (3.1) Based on the initial mesh generated from the CT image, unpooling is applied to increase the number of vertices; (3.2) Vertices are added using a unified solution pooling strategy, and then the mesh is deformed and the shortest distance from each candidate vertex to its parent edge is calculated; if the distance exceeds the specified threshold, the vertex is retained; otherwise, the vertex is discarded; (3.3) Utilize the continuous mapping learned by the mesh decoder to associate the surface points on the input sphere with the surface points on the object surface, restore the edge connections by computing the convex hull on the sphere, and then transfer it directly to the target mesh; (3.4) Feature aggregation is performed using Monte Carlo sampling implemented in the mesh network. These variant density coefficients are defined per node after pooling or unpooling, ensuring that vertex density is adaptively managed throughout the network; (3.5) The mesh is further optimized to adapt its density to the needs of local details, including applying a local residual connection operation. In order to control the vertex density, a weighted edge length regularization term is defined, which forces shorter edge lengths in areas with higher weights, focusing more vertices where more details are needed.

6. The three-dimensional reconstruction method of medical image sequence according to claim 5, characterized in that: In the step (3.4), the aggregation functions of the pooling layer and the de-pooling layer are: in, is a learnable parameter, Represents the density value, which controls the distribution of vertices during the adaptive optimal variable density mesh unpooling process.

7. The method for three-dimensional reconstruction of a medical image sequence according to claim 6, characterized in that: In step (3.5), the weighted edge length regularization term is defined as: For a vertex r in the predicted grid, the weight of the nearest vertex is used to represent its own weight, and all edges connected to r share the same weight when calculating the edge length.

8. The method for three-dimensional reconstruction of a medical image sequence according to claim 7, characterized in that: The step (4) comprises the following sub-steps: (4.1) To prevent such artifacts and further improve the visual quality of the reconstructed mesh, a boundary optimization module is designed to only predict the displacement relative to each input boundary vertex. Each boundary vertex can only move on the two-dimensional plane formed by the two boundary edges intersecting the vertex; the regularization term penalizes boundary markings by forcing the boundary curve to remain smooth and consistent. The boundary regularization term is defined as follows: in, represents the set of vertices on the open boundary, Represents the set of neighbor vertices of vertex x on the boundary; (4.2) In order to evaluate the quality of each triangular surface, the surface quality index SQI is introduced: The SQI is defined as: SQI=α·E pos +β·E smooth Among them, E pos represents the position error of the vertex, E smooth represents the smoothing error of the surface, α and β are weight parameters used to balance the importance of these two factors; (4.3) In order to further enhance the effect of the boundary optimization module, the graph convolutional network is used to predict the displacement of boundary vertices. The displacement prediction formula of boundary vertices is: Among them, d(r) represents the displacement vector of vertex r, x r represents the eigenvector of vertex r, A is the adjacency matrix describing the connection relationship between boundary vertices, represents the set of neighbor vertices of vertex r, W is the weight matrix, and σ is the activation function.

9. The method for three-dimensional reconstruction of a medical image sequence according to claim 8, characterized in that: In step (5), the chamfer loss measures the average closest distance between two point sets, and evaluates the closeness between the generated 3D mesh and the target CT scan data; the IoU loss reflects the reconstruction accuracy by calculating the ratio of the intersection and union of the predicted mesh and the real mesh, and is used to evaluate the surface integrity of the reconstruction; the Dice coefficient loss evaluates the overall morphological similarity of the reconstructed surface, and determines the performance of the model by calculating the overlap between the predicted and actual meshes.

10. A three-dimensional reconstruction device for a medical image sequence, characterized in that: The device includes: The data acquisition module reads the original medical image sequence of the CT scan, performs data preprocessing operations, performs grayscale normalization processing on each layer of the image, and generates voxel data for subsequent feature extraction and analysis; The feature extraction module uses a graph convolutional network to extract features from the CT image volume and generate coordinate offset values ​​of the corresponding mesh vertices. The graph convolutional network learns the spatial position adjustment parameters of each vertex through the neighborhood topological relationship to establish a local topological structure; Density optimization module, which optimizes the density of the initial mesh vertices generated by the CT image, and adjusts the distribution density of the vertices according to the different density characteristics of the image area to ensure that higher vertex density is used in high-detail areas and reduced vertex density in low-detail areas; A topology adjustment module adjusts the topology of the optimized mesh, identifies topologically discontinuous or inconsistent areas in the mesh, and reconnects or adjusts the vertices of these areas according to neighborhood topological relationships, ensuring that the topology of the mesh conforms to the actual boundaries and contours of the medical image; The loss evaluation module defines and calculates the loss function to optimize the accuracy of the reconstructed model and optimizes the model parameters by minimizing the loss functions such as chamfer loss, intersection-over-union loss, and Dice coefficient loss.

Images