A blood vessel skeleton overall feature vector extraction method, system, device and storage medium based on a Transformer model

Abstract

The application relates to the technical field of medical image processing, and particularly provides a blood vessel skeleton overall feature vector extraction method, system and device based on a Transformer model and a storage medium. The method can extract a blood vessel skeleton from a three-dimensional blood vessel surface model, take the blood vessel skeleton as a center line structure of the blood vessel, accurately describe a complex bifurcation structure of a blood vessel tree, and accurately fit the center line shape of the blood vessel without relying on extremely high point density. Secondly, the application constructs a graph data structure, naturally retains the connection relationship between vertices, and can completely carry the connection relationship and hierarchical structure of the blood vessel branches. The blood vessel skeleton overall feature vector obtained by the application can provide complete and accurate overall shape descriptors for complex blood vessel-shaped three-dimensional structures, provide standardized feature inputs for downstream medical tasks based on blood vessel shape features, and provide a brand-new scheme for disease diagnosis and treatment.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, specifically to a method, system, device, and storage medium for extracting the overall feature vector of a vascular skeleton based on the Transformer model. Background Technology

[0002] The role of angiomics based on vascular shape modeling in medical diagnosis and treatment is receiving increasing attention. The geometry and topology of blood vessels are important biomarkers reflecting physiological states and pathological changes in the human body, and their quantitative characteristics can provide objective evidence for the diagnosis and treatment of various diseases. For example, the geometric and topological features of renal vessels are significant for the classification and metastasis risk assessment of renal cell carcinoma; the shape features of the portal vein in the liver can provide a reference for determining the edge of liver segments in anatomical liver resection surgery, helping to preserve as much healthy liver volume as possible while completely removing the tumor; in addition, vascular shape features can also be applied to clinical scenarios such as vascular malformation diagnosis, arteriosclerosis screening, stroke risk prediction, and precise surgical planning. All of these applications rely on feature vectors that can comprehensively describe the three-dimensional shape of blood vessels.

[0003] Currently, commonly used methods for extracting vascular feature vectors include point cloud methods and polygon mesh methods. These two types of methods can convert three-dimensional vascular structures into numerical feature vectors for downstream machine learning tasks such as classification, prediction, and matching.

[0004] However, existing methods for extracting blood vessel feature vectors have the following drawbacks: (1) Point cloud is a collection of discrete three-dimensional points. When describing sharp structures with drastic curvature changes, extremely high point density is required to ensure the accuracy of the shape. However, the point cloud density of blood vessels obtained from medical image reconstruction is limited, and point cloud methods are difficult to accurately describe complex shapes with sharp bifurcations, such as blood vessel trees.

[0005] (2) Since point clouds are unordered three-dimensional point sets, they do not naturally contain the connection relationship between points and cannot preserve the connection relationship and hierarchical structure between blood vessel branches, making it difficult for the extracted features to fully reflect the topological information of blood vessels.

[0006] (3) The size range of human vascular tree is large. Polygonal mesh needs to take into account both the overall shape of large blood vessels and the detailed features of small branches. If a uniform high-resolution mesh is used, the amount of data and computation will increase sharply. If the resolution is reduced, the structure of small branches cannot be accurately represented. Therefore, the polygonal mesh method is difficult to adapt to complex vascular trees with large scale span. (4) Feature extraction of polygonal meshes is mainly based on convolution operations of local neighborhoods, which makes it difficult to effectively capture long-distance topological connections in the vascular tree and easily causes loss of topological information during feature extraction.

[0007] Therefore, how to design a method, system, device, and storage medium for extracting the overall feature vector of the vascular skeleton based on the Transformer model that can accurately describe the complex branching structure of the vascular tree and completely extract the vascular topology information is a technical problem that has not yet been solved in the prior art. Summary of the Invention

[0008] Therefore, this application proposes a method, system, device, and storage medium for extracting the overall feature vector of the vascular skeleton based on the Transformer model, which can accurately describe the complex branching structure of the vascular tree and completely extract the vascular topology information, so as to improve or solve at least one of the above-mentioned technical problems.

[0009] This application mainly includes the following aspects: In a first aspect, embodiments of this application provide a method for extracting the overall feature vector of a vascular skeleton based on a Transformer model, the method comprising: Obtain a three-dimensional blood vessel surface model; The vascular skeleton was extracted from the three-dimensional vascular surface model using the average curvature flow algorithm. Construct a graph data structure for the vascular skeleton and calculate the vascular radius at each skeleton vertex of the vascular skeleton; the vertices and edges of the graph data structure correspond one-to-one with the skeleton vertices and skeleton edges of the vascular skeleton, and the attributes of the vertex include the spatial coordinates of the vertex and the list of surface point coordinates of the corresponding three-dimensional vascular surface model before contraction. Repairing the vascular skeleton includes bridging the broken skeleton, deleting skeleton noise debris, and removing skeleton noise burrs. By using a recursive backtracking algorithm, each vertex in the graph data structure is assigned a level number to complete the level classification of the vascular skeleton; The vascular skeleton is divided into a predetermined number of skeleton segments of equal length; Extract the segment feature vector and position code of each skeleton segment, wherein the segment feature vector is... ; The segment feature vector and the position encoding are input into the trained Transformer model to obtain the encoded feature vector of each skeleton segment; By merging all the encoded feature vectors using an unrelated symmetric function, the overall feature vector of the vascular skeleton is obtained.

[0010] According to one embodiment of the present invention, the training steps of the Transformer model include: S1: Input the segment feature vector and the position encoding into the random mask layer of the untrained Transformer model; S2: Randomly select 50% of the skeleton segments, hide the corresponding segment feature vectors, and generate masked input features; S3: The masked input features are fed into the encoder of the Transformer model. The encoder processes all the unmasked segment feature vectors and outputs the encoded feature vector. S4: Input the encoded feature vector into the decoder of the Transformer model. Based on the encoded feature vector, the decoder reconstructs the segment feature vectors of all the skeleton segments, thereby outputting the complete reconstructed segment feature vectors. S5: Calculate the loss function, which is defined as the reconstruction error between the reconstructed fragment feature vector and the original input fragment feature vector; S6: Optimize the model parameters using the backpropagation algorithm to minimize the reconstruction error; S7: Repeat steps S2-S6 until the reconstructed segment feature vector output by the Transformer model can stably and accurately restore the original input segment feature vector, indicating that the Transformer model has mastered the shape rules of complex blood vessels in the human body and can stably and accurately extract the encoded feature vector.

[0011] According to one embodiment of the present invention, the step of bridging the fractured skeleton includes: Receive the first mouse click position input by the user, the first mouse click position corresponding to the rupture area to be bridged on the vascular skeleton displayed on the screen; Traverse all vertices of the graph data structure and calculate the Euclidean distance between the vertex and the first mouse click position, and take the vertex with the smallest Euclidean distance as the first bridge endpoint; Receive the second mouse click position input by the user; Traverse all vertices of the graph data structure and calculate the Euclidean distance between the vertex and the second mouse click position, and take the vertex with the smallest Euclidean distance as the second bridge endpoint; A connecting edge is added between the first bridging endpoint and the second bridging endpoint, so that the broken skeleton is connected through the connecting edge, thereby completing the bridging of the broken skeleton.

[0012] According to one embodiment of the present invention, the step of deleting skeleton noise fragments includes: The connected component set is obtained by calculating the connected component to which each vertex belongs using the boost:connected_components function; Count the number of vertices contained in each of the connected regions; Minimum number of user inputs required threshold ; Traverse the set of connected regions; if the number of vertices contained in it is less than the minimum threshold number... Then all vertices of the connected region are marked as noise fragments; Traverse all vertices of the graph data structure, delete all vertices marked as noise fragments, and the edges they form, thereby completing the deletion of skeleton noise fragments.

[0013] According to one embodiment of the present invention, the step of removing skeleton noise burrs includes: Receive the root node selection command input by the user and determine the root node; Starting from the root node, all vertices in the graph data structure are traversed using a breadth-first search algorithm, and the number of neighboring vertices of the vertex is obtained using the boost:degree function, which is used as the degree of the vertex. The vertices with a degree of 1 are marked as leaf nodes, the vertices with a degree of 2 are marked as ordinary nodes, and the vertices with a degree greater than 2 are marked as branching points. Traverse all the leaf nodes, and perform a reverse breadth-first traversal with each leaf node as the initial vertex until the first branch point is reached, thereby obtaining the leaf branch corresponding to the leaf node, and calculating the total length of the leaf branch. Minimum length threshold for receiving user input ; Traverse all the leaf branches; if their total length is less than the minimum length threshold... Then all vertices in the leaf branch except for the bifurcation point are marked as noise spikes; Traverse all vertices of the graph data structure, delete all vertices marked as noise spikes, and the edges they form, thereby completing the removal of skeleton noise spikes.

[0014] According to one embodiment of the present invention, the step of assigning level numbers to each vertex in the graph data structure and completing the level division of the vascular skeleton by means of a recursive backtracking algorithm includes: Set the level number of the root node to 0; Using the root node as the current vertex, initiate the recursive backtracking algorithm: Obtain all adjacent vertices of the current vertex, exclude adjacent vertices that have already been assigned level numbers, and obtain the set of adjacent vertices to be processed; Traverse each vertex to be processed in the set of adjacent vertices to be processed. If the vertex to be processed is a normal vertex, assign the current level number to the vertex to be processed and use the vertex to be processed as the new current vertex. Recursively execute the recursive backtracking algorithm. If the vertex to be processed is the branch point, then the current level number is incremented by 1, the updated level number is assigned to the vertex to be processed, and the vertex to be processed is used as the new current vertex. The recursive backtracking algorithm is then executed recursively. If the vertex to be processed is a leaf node, then the current level number is assigned to the vertex to be processed, and the recursive backtracking algorithm of this branch is terminated. Once all vertices in the graph data structure have been assigned level numbers, the recursive backtracking algorithm is terminated, thereby completing the level classification of the vascular skeleton.

[0015] According to one embodiment of the present invention, the step of extracting the segment feature vector and positional encoding of each of the skeleton segments includes: Traverse all the skeleton segments and obtain all the skeleton vertices contained in each skeleton segment; Calculate the average of the spatial coordinates of all the skeleton vertices within the skeleton segment to obtain the centroid coordinates of the skeleton segment, and use the centroid coordinates as the position code of the skeleton segment; The spatial coordinates of all the skeleton vertices within the skeleton segment are converted into a coordinate matrix. Principal component analysis is performed on the coordinate matrix to extract the eigenvector corresponding to the largest eigenvalue, which is used as the orientation of the skeleton segment. The average inner diameter of the skeleton segment is obtained by averaging the blood vessel radii corresponding to all skeleton vertices within the skeleton segment. The skeleton vertex connected to the previous skeleton segment in the skeleton segment is selected as the segment starting point, and the spatial coordinates of the segment starting point are used as the geodesic coordinates of the previous skeleton segment. Obtain the level number corresponding to the skeleton vertex within the skeleton segment, and use it as the level of the skeleton segment; The orientation, the average inner diameter, the geodesic upper coordinates, and the level are integrated to obtain the segment feature vector of the skeleton segment.

[0016] Secondly, embodiments of this application also provide a system for extracting the overall feature vector of a vascular skeleton based on a Transformer model, the system comprising: The acquisition module is used to acquire a three-dimensional blood vessel surface model; A skeleton generation module is used to extract the vascular skeleton from the three-dimensional vascular surface model using an average curvature flow algorithm. The data construction and radius calculation module is used to construct the graph data structure of the vascular skeleton and calculate the vascular radius at each skeleton vertex of the vascular skeleton; the vertices and edges of the graph data structure correspond one-to-one with the skeleton vertices and skeleton edges of the vascular skeleton, and the attributes of the vertex include the spatial coordinates of the vertex and the list of surface point coordinates of the corresponding three-dimensional vascular surface model before contraction. The skeleton repair module is used to repair the vascular skeleton, including bridging the broken skeleton, deleting skeleton noise debris, and removing skeleton noise burrs. The grading module is used to assign a grading number to each vertex in the graph data structure through a recursive backtracking algorithm, thereby completing the grading of the vascular skeleton. The skeleton segmentation module is used to segment the vascular skeleton into a set number of skeleton segments of equal length; The feature extraction module is used to extract the segment feature vectors and position codes of each skeleton segment; The encoding feature vector extraction module is used to input the segment feature vector and the position encoding into the trained Transformer model to obtain the encoding feature vector of each skeleton segment; The overall feature vector extraction module is used to fuse all the encoded feature vectors by arranging independent symmetric functions to obtain the overall feature vector of the vascular skeleton.

[0017] Thirdly, embodiments of this application also provide a computer device, the computer device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the steps of the method described above when executing the instructions.

[0018] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method described above.

[0019] The technical solution provided in this application has the following advantages: 1. The vascular skeleton feature vector extraction method based on the Transformer model in this application can extract the vascular skeleton from the three-dimensional vascular surface model and use the vascular skeleton as the centerline structure of the vascular system, thereby accurately describing the complex branching structure of the vascular tree. It can accurately fit the centerline shape of the vascular system without relying on extremely high point density.

[0020] 2. The method for extracting the overall feature vector of the vascular skeleton based on the Transformer model in this application naturally preserves the connection relationship between vertices by constructing a graph data structure that corresponds one-to-one with the skeleton vertices and edges of the vascular skeleton. It can fully carry the connection relationship and hierarchical structure of the vascular branches, so that the extracted features can fully reflect the topological information of the blood vessels.

[0021] 3. The method for extracting the overall feature vector of the vascular skeleton based on the Transformer model in this application uses vascular skeleton data as the basic data structure. It is not affected by the difference in the scale of vascular diameter and can accurately represent the morphological features of the main trunk at the centimeter level and the terminal at the millimeter level at the millimeter level. It will not cause the problem of a sharp increase in data volume and computational load, and can be adapted to complex vascular trees with large scale span.

[0022] 4. The method for extracting the overall feature vector of the vascular skeleton based on the Transformer model in this application adopts the Transformer encoder architecture for feature extraction, which can effectively capture the long-distance topological connection relationship in the vascular tree; at the same time, the features output by the encoder are globally fused by the permutation-independent symmetric function, and the generated overall feature vector of the vascular skeleton can completely reflect the global topological information of the blood vessels.

[0023] 5. Compared to the traditional radius calculation method that requires calculating the normal direction and searching for the nearest point on the blood vessel surface point by point, this application pre-constructs a graph data structure of the blood vessel skeleton and performs Euclidean distance calculation based on the list of surface point coordinates of the three-dimensional blood vessel surface model before contraction corresponding to the skeleton vertices. This transforms the complex blood vessel radius calculation into a fast retrieval of fixed data and simple geometric operations, significantly simplifying the blood vessel radius calculation process and improving the overall processing efficiency of blood vessel skeleton feature extraction. It provides efficient and accurate technical support for feature extraction and downstream analysis tasks of batch blood vessel samples.

[0024] 6. The overall feature vector of the vascular skeleton obtained in this application can provide a complete and accurate overall shape descriptor for complex vascular three-dimensional structures, provide standardized feature input for downstream medical tasks based on vascular shape features, and provide a brand-new solution for disease diagnosis and treatment.

[0025] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is the overall flowchart of the method for extracting the overall feature vector of the vascular skeleton based on the Transformer model in this application.

[0028] Figure 2 This is an architecture diagram of the system for extracting the overall feature vector of the vascular skeleton based on the Transformer model in this application.

[0029] Figure 3 This is the vascular skeleton extracted in this application, taking the portal vein of the liver as an example.

[0030] Figure 4 This is a schematic diagram of the bridging fracture skeleton of this application, wherein (a) is a schematic diagram before bridging, (b) is a schematic diagram during bridging, and (c) is a schematic diagram after bridging.

[0031] Figure 5 This is a schematic diagram of the removal of skeleton noise fragments in this application, where (d) is a schematic diagram before removal and (e) is a schematic diagram after removal.

[0032] Figure 6 This is a schematic diagram of removing skeleton noise burrs in this application, where (f) is a schematic diagram before removal and (g) is a schematic diagram after removal.

[0033] Figure 7 This is a diagram illustrating the classification of levels in this application.

[0034] Figure 8 This is a schematic diagram of the feature vector of the fragment in this application.

[0035] Figure 9 This is a schematic diagram of the binary mask data.

[0036] Figure 10 This is a schematic diagram of a voxel cube unit. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0038] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0039] The present application will now be described in detail with reference to the accompanying drawings and embodiments. Example

[0040] like Figure 1 As shown, this embodiment provides a method for extracting the overall feature vector of a vascular skeleton based on the Transformer model, including the following steps: S101: Obtain a three-dimensional blood vessel surface model; S102: Extract the vascular skeleton from the three-dimensional vascular surface model using the average curvature flow algorithm; S103: Construct the graph data structure of the vascular skeleton and calculate the vascular radius at each skeleton vertex of the vascular skeleton; the vertices and edges of the graph data structure correspond one-to-one with the skeleton vertices and skeleton edges of the vascular skeleton, and the attributes of the vertex include the spatial coordinates of the vertex and the list of surface point coordinates of the corresponding three-dimensional vascular surface model before contraction. S104: Repair the vascular skeleton, including bridging the broken skeleton, deleting skeleton noise fragments, and removing skeleton noise burrs. S105: By using a recursive backtracking algorithm, assign level numbers to each vertex in the graph data structure to complete the level classification of the vascular skeleton; S106: Divide the vascular skeleton into a predetermined number of skeleton segments of equal length; S107: Extract the segment feature vector and position code of each skeleton segment, wherein the segment feature vector is... ; S108: Input the segment feature vector and the position encoding into the trained Transformer model to obtain the encoded feature vector of each skeleton segment; S109: By merging all the encoded feature vectors using an unrelated symmetric function, the overall feature vector of the vascular skeleton is obtained.

[0041] The following details each step of the method for extracting the overall feature vector of the vascular skeleton based on the Transformer model: S101: Obtain a three-dimensional blood vessel surface model; In one possible implementation, the step of obtaining a three-dimensional blood vessel surface model includes the following steps: S1011: Acquire thin-section CT scan sequence of the target area; Here, the target site thin-slice CT scan sequence refers to the tomographic scan data stored in nii.gz format obtained by performing thin-slice scans on the target vascular region of the patient using clinical CT equipment. These data retain metadata such as voxel coordinates and spatial spacing of the original images, and can accurately present the subtle anatomical features of the vascular structure in the target area.

[0042] S1012: Convert thin-slice CT scan sequences of the target area into vascular mask data; In one possible implementation, step S1012 involves converting the thin-slice CT scan sequence of the target area into vascular mask data using a deep learning image segmentation model; the deep learning image segmentation model is a three-dimensional segmentation neural network built based on the 3DU-Net network.

[0043] This step utilizes a deep learning image segmentation model to accurately identify vascular regions from thin-slice CT scan sequences of the target area and converts them into binary mask data (such as...). Figure 9 As shown in the figure, this is a matrix where the target area is 1 and the background is 0. This mask data can intuitively mark the distribution of blood vessels in three-dimensional space, providing a clear outline boundary for the subsequent construction of a three-dimensional blood vessel surface model using the moving cube algorithm, thus avoiding the subjectivity and error of manual segmentation.

[0044] It should be noted that this deep learning image segmentation model uses a three-dimensional segmentation neural network built on the 3DU-Net network. The 3DU-Net network is a classic architecture in the field of medical image segmentation. Through the combination of an encoder-decoder structure and skip connections, it can extract deep features of the image while retaining shallow spatial information.

[0045] S1013: Convert vascular mask data into a three-dimensional vascular surface model using the moving cube algorithm; Specifically, first, the vascular mask data is input, and then the vascular mask data is divided into multiple 2*2*2 voxel cube units (e.g., Figure 10 As shown in the figure, each cell contains 8 voxels with values ​​of 0 or 1. For each cubic cell, if its 8 voxels contain both 0 and 1, i.e., there is a blood vessel boundary, the intersection coordinates are determined by linear interpolation, and the intersection of the voxel edge and the blood vessel surface is calculated. Then, based on the 0 and 1 distribution of the voxels in the cell, a predefined configuration table is queried to determine the connection method of the intersection points, generate triangular patches, and fit the contour of the blood vessel surface. Finally, all cubic cells are traversed, and the generated triangular patches are stitched together to form a complete three-dimensional blood vessel surface model.

[0046] Compared to traditional manual modeling, the moving cube algorithm can automatically generate a complete vascular tree model from mask data, avoiding manual intervention and improving modeling efficiency. In addition, the algorithm can capture the fine structure of blood vessels through voxel-level intersection calculation, ensuring the consistency between the 3D model and the patient's real anatomical structure, and providing a reliable basis for subsequent vascular skeleton extraction and feature analysis.

[0047] S102: Extract the vascular skeleton from the three-dimensional vascular surface model using the average curvature flow algorithm; Here, a three-dimensional vascular surface model (i.e., a surface triangular mesh) is first input. The average curvature flow algorithm is then used to shrink the three-dimensional model. Through iterative evolution of the model surface, surface points move along the direction of their average curvature, ultimately converging into the vascular skeleton of the model (e.g., ...). Figure 3 (As shown in the figure). This process can be implemented using the extract_mean_curvature_flow_skeleton function in the CGAL library.

[0048] Specifically, during the iteration process, the surface mesh in the intermediate shrinkage stage is called the mid-skeleton. After each iteration, the system performs local remeshing of the mid-skeleton through angle segmentation and edge shrinkage. When the change in the mid-skeleton between two iterations is small enough, the process ends. Each edge shrinkage merges two vertices into one vertex. During this process, the algorithm records the vertices of the triangular facets of the original surface mesh and assigns them to the skeleton vertices on the finally merged vascular skeleton.

[0049] The average curvature flow algorithm can accurately preserve the branch point positions and branch angles of the vascular tree, avoiding the offset of the skeletal bifurcation points caused by algorithm errors, and providing a real anatomical reference for the subsequent construction of graph data structures.

[0050] The method for extracting the overall feature vector of the vascular skeleton based on the Transformer model in this application can extract the vascular skeleton from the three-dimensional vascular surface model and use the vascular skeleton as the centerline structure of the vascular system, thereby accurately describing the complex branching structure of the vascular tree and accurately fitting the centerline shape of the vascular system without relying on extremely high point density.

[0051] S103: Construct the graph data structure of the vascular skeleton and calculate the vascular radius at each skeleton vertex of the vascular skeleton; the vertices and edges of the graph data structure correspond one-to-one with the skeleton vertices and skeleton edges of the vascular skeleton, and the attributes of the vertex include the spatial coordinates of the vertex and the list of surface point coordinates of the corresponding three-dimensional vascular surface model before contraction. Here, we first input a 3D blood vessel surface model and a blood vessel skeleton, and then use the Boost library's adjacency matrix data structure, specifically boost:adjacency_list, to construct a graph data structure. Specifically, each skeleton vertex in the blood vessel skeleton is used as a vertex in the graph data structure, and by traversing the topology of the blood vessel skeleton, a one-to-one correspondence is established between vertices and skeleton vertices.

[0052] Each vertex stores the following attributes: 1) Spatial coordinates: Records the position of the skeleton vertex in three-dimensional space, used to locate the vascular anatomy structure; 2) Surface point coordinate list: Stores the coordinates of the vascular surface points of the original three-dimensional vascular surface model before contraction corresponding to the skeleton vertex; It should be noted that this list is retained during contraction by the average curvature flow algorithm, and the vascular surface geometry information can be obtained without additional calculation.

[0053] In addition, edges of the graph data structure are generated based on the connection relationship of the skeleton vertices. The length of the edge is defined as the Euclidean distance between two adjacent skeleton vertices, which is used to reflect the actual length of the blood vessel. The blood vessel radius corresponding to the current skeleton vertex is obtained by calculating the average distance between each skeleton vertex and the coordinates of all its corresponding surface points.

[0054] Compared to the traditional radius calculation method that requires calculating the normal direction and searching for the nearest point on the blood vessel surface point by point, this application pre-constructs a graph data structure of the blood vessel skeleton and performs Euclidean distance calculation based on the list of surface point coordinates of the three-dimensional blood vessel surface model before contraction corresponding to the skeleton vertices. This transforms the complex blood vessel radius calculation into a fast retrieval of fixed data and simple geometric operations, significantly simplifying the blood vessel radius calculation process and improving the overall processing efficiency of blood vessel skeleton feature extraction. It provides efficient and accurate technical support for feature extraction and downstream analysis tasks of batch blood vessel samples.

[0055] The method for extracting the overall feature vector of the vascular skeleton based on the Transformer model in this application naturally preserves the connection relationship between vertices by constructing a graph data structure that corresponds one-to-one with the skeleton vertices and edges of the vascular skeleton. It can fully carry the connection relationship and hierarchical structure of vascular branches, so that the extracted features can fully reflect the topological information of the blood vessels.

[0056] The method for extracting the overall feature vector of the vascular skeleton based on the Transformer model in this application uses vascular skeleton data as the basic data structure. It is not affected by the difference in the scale of vascular diameter and can accurately represent the morphological features of the main trunk at the centimeter level and the terminal end at the millimeter level at the millimeter level. It does not have the problem of a sharp increase in data volume and computational load and can be adapted to complex vascular trees with large scale span.

[0057] S104: Repair the vascular skeleton, including bridging the broken skeleton, deleting skeleton noise fragments, and removing skeleton noise burrs. like Figure 4 As shown in (a), (b), and (c) above, in one possible implementation, the step of bridging the fractured skeleton includes: S1041a: Receive the first mouse click position input by the user, the first mouse click position corresponding to the rupture area to be bridged on the vascular skeleton displayed on the screen; S1042a: Traverse all vertices of the graph data structure and calculate the Euclidean distance between the vertex and the first mouse click position, and take the vertex with the smallest Euclidean distance as the first bridge endpoint; S1043a: Receives the second mouse click position input by the user; S1044a: Traverse all vertices of the graph data structure and calculate the Euclidean distance between the vertex and the second mouse click position, and take the vertex with the smallest Euclidean distance as the second bridge endpoint; S1045a: Add a connecting edge between the first bridging endpoint and the second bridging endpoint so that the broken skeleton is connected through the connecting edge, thereby completing the bridging of the broken skeleton.

[0058] Specifically, when a user clicks on one end of the skeleton to be bridged on the screen with the mouse, the Euclidean distance between the mouse click position and the vertex is calculated, and the vertex with the smallest Euclidean distance is taken as the first bridging endpoint (v1). The second bridging endpoint (v2) is found in the same way. Then, the boost:add_edge(v1,v2,g) function is run, where g is the graph data structure of the vascular skeleton, which adds a connecting edge between the first and second bridging endpoints, completing the bridging of the fractured skeleton.

[0059] like Figure 5As shown in (d) and (e) above, in one possible implementation, the step of removing skeleton noise fragments includes: S1041b: The connected component to which each vertex belongs is calculated using the boost:connected_components function to obtain a set of connected components; S1042b: Count the number of vertices contained in each of the connected regions; S1043b: Minimum number of user inputs received threshold ; S1044b: Traverse the set of connected regions. If the number of vertices contained in the set is less than the minimum number threshold... Then all vertices of the connected region are marked as noise fragments; S1045b: Traverse all vertices of the graph data structure, delete all vertices marked as noise fragments, and the edges they form, thereby completing the deletion of skeleton noise fragments.

[0060] like Figure 6 As shown in (f) and (g) above, in one possible implementation, the step of removing skeleton noise burrs includes: S1041c: Receives the root node selection command input by the user and determines the root node; S1042c: Starting from the root node, traverse all vertices in the graph data structure using the breadth-first search algorithm, and obtain the number of neighboring vertices of the vertex using the boost:degree function, which is used as the degree of the vertex. S1043c: Mark the vertex with a degree of 1 as a leaf node, mark the vertex with a degree of 2 as a normal node, and mark the vertex with a degree greater than 2 as a branching point; S1044c: Traverse all the leaf nodes, and perform a reverse breadth-first traversal with each leaf node as the initial vertex until the first branch point is reached, thereby obtaining the leaf branch corresponding to the leaf node, and calculating the total length of the leaf branch. S1045c: Minimum length threshold for receiving user input ; S1046c: Traverse all the leaf branches; if their total length is less than the minimum length threshold... Then all vertices in the leaf branch except for the bifurcation point are marked as noise spikes; S1047c: Traverse all vertices of the graph data structure, delete all vertices marked as noise spikes, and the edges they form, thereby completing the removal of skeleton noise spikes.

[0061] S105: By using a recursive backtracking algorithm, assign level numbers to each vertex in the graph data structure to complete the level classification of the vascular skeleton; like Figure 7 As shown, in one possible implementation, the step of assigning level numbers to each vertex in the graph data structure through a recursive backtracking algorithm to complete the level division of the vascular skeleton includes: Set the level number of the root node to 0; Using the root node as the current vertex, initiate the following recursive backtracking algorithm: Obtain all adjacent vertices of the current vertex, exclude adjacent vertices that have already been assigned level numbers, and obtain the set of adjacent vertices to be processed; Traverse each vertex to be processed in the set of adjacent vertices to be processed. If the vertex to be processed is a normal vertex, assign the current level number to the vertex to be processed and use the vertex to be processed as the new current vertex. Recursively execute the recursive backtracking algorithm. If the vertex to be processed is the branch point, then the current level number is incremented by 1, the updated level number is assigned to the vertex to be processed, and the vertex to be processed is used as the new current vertex. The recursive backtracking algorithm is then executed recursively. If the vertex to be processed is a leaf node, then the current level number is assigned to the vertex to be processed, and the recursive backtracking algorithm of this branch is terminated. Once all vertices in the graph data structure have been assigned level numbers, the recursive backtracking algorithm is terminated, thereby completing the level classification of the vascular skeleton.

[0062] S106: Divide the vascular skeleton into a predetermined number of skeleton segments of equal length; Specifically, users input the number of segments to be divided into on the vascular skeleton through the interface. Typical values ​​include 64 and 256. Traverse all edges in the vascular skeleton graph data structure and sum the lengths of all edges to obtain the skeleton length. ,use Get the fragment length ; Traverse all topological paths in the vascular skeleton graph data structure. Each topological path includes independent vascular segments consisting of a root vertex to a branch point, branch points to branch points, and branch points to leaf nodes. The segment length is calculated as a single topological path. Based on this, the segments are sequentially divided to obtain skeleton segments with matching lengths; For single topology paths with insufficient endpoints The remaining part of the length, if its length is less than If it is, it will be merged into the previous skeleton segment; otherwise, it will form a separate skeleton segment.

[0063] S107: Extract the segment feature vector and position code of each skeleton segment, wherein the segment feature vector is... ; In one possible implementation, the step of extracting the segment feature vectors and position codes of each of the skeleton segments includes: Traverse all the skeleton segments and obtain all the skeleton vertices contained in each skeleton segment; Calculate the average of the spatial coordinates of all the skeleton vertices within the skeleton segment to obtain the centroid coordinates of the skeleton segment, and use the centroid coordinates as the position code of the skeleton segment; The spatial coordinates of all the skeleton vertices within the skeleton segment are converted into a coordinate matrix. Principal component analysis is performed on the coordinate matrix to extract the eigenvector corresponding to the largest eigenvalue, which is used as the orientation of the skeleton segment. Specifically, when calculating the orientation of a vascular skeletal segment, the spatial coordinates of the m vertices of the segment are integrated into an m x 3 coordinate matrix. The covariance matrix of this matrix is ​​calculated, and eigenvalues ​​and eigenvectors are derived from it. The eigenvector corresponding to the largest eigenvalue is extracted as the principal component direction of the vascular segment, i.e., its orientation. Since vascular skeletal segments may have some curvature, the orientation, represented by a single vector, indicates the overall main extension direction of the segment and corresponds to the direction of blood flow within the blood vessel.

[0064] The average inner diameter of the skeleton segment is obtained by averaging the blood vessel radii corresponding to all skeleton vertices within the skeleton segment. The skeleton vertex connected to the previous skeleton segment in the skeleton segment is selected as the segment starting point, and the spatial coordinates of the segment starting point are used as the geodesic coordinates of the previous skeleton segment. It should be noted that a geodesic refers to the shortest path distance between two points along the edges of the vascular skeleton, which is different from the Euclidean distance, which is directly calculated to determine the spatial position of the two points. The starting point of a segment is the first skeleton vertex reached during the traversal of the vascular skeleton from the root node, corresponding to the vertex that connects the segment to the previous level skeleton segment. By introducing the coordinates of the previous level geodesic, the relative position of the segment on the vascular skeleton can be characterized from the perspective of the topological path. This allows for accurate differentiation of parallel vascular branches that are close in spatial Euclidean distance but far apart in topological path, thus more accurately reflecting the branching topology of the vascular tree.

[0065] Obtain the level number corresponding to the skeleton vertex within the skeleton segment, and use it as the level of the skeleton segment; By integrating the orientation, the average inner diameter, the geodesic upper-level coordinates, and the level, the segment feature vector of the skeleton segment is obtained (e.g., Figure 8 (As shown).

[0066] S108: Input the segment feature vector and the position encoding into the trained Transformer model to obtain the encoded feature vector of each skeleton segment; In one possible implementation, the training steps of the Transformer model include: S1: Input the segment feature vector and the position encoding into the random mask layer of the untrained Transformer model; S2: Randomly select 50% of the skeleton segments, hide the corresponding segment feature vectors, and generate masked input features; S3: The masked input features are fed into the encoder of the Transformer model. The encoder processes all the unmasked segment feature vectors and outputs the encoded feature vector. S4: Input the encoded feature vector into the decoder of the Transformer model. Based on the encoded feature vector, the decoder reconstructs the segment feature vectors of all the skeleton segments, thereby outputting the complete reconstructed segment feature vectors. S5: Calculate the loss function, which is defined as the reconstruction error between the reconstructed fragment feature vector and the original input fragment feature vector; S6: Optimize the model parameters using the backpropagation algorithm to minimize the reconstruction error; S7: Repeat steps S2-S6 until the reconstructed segment feature vector output by the Transformer model can stably and accurately restore the original input segment feature vector, indicating that the Transformer model has mastered the shape rules of complex blood vessels in the human body and can stably and accurately extract the encoded feature vector.

[0067] It should be noted that the ability to "accurately reconstruct the complete shape of a blood vessel by looking at only 50% of the vascular skeleton fragments" is only used to verify whether the model truly understands and masters the shape patterns of complex human blood vessels, and is part of the training process. In actual use, simply inputting the fragment feature vectors and the position codes into the encoder of the trained Transformer model will yield the encoded feature vectors of each skeleton fragment.

[0068] S109: By merging all the encoded feature vectors using an unrelated symmetric function, the overall feature vector of the vascular skeleton is obtained.

[0069] Specifically, the symmetric function used is the max pooling function. First, it iterates through all feature components of each encoded feature vector output by the encoder, then extracts the maximum feature component at the same position in all encoded feature vectors, and finally integrates these extracted maximum feature components to generate the overall feature vector of the vascular skeleton.

[0070] The Transformer encoder outputs encoded feature vectors of multiple independent skeleton segments, which are essentially scattered local feature representations and cannot directly represent the overall shape of the entire blood vessel. By using max pooling fusion, any number of local segment features can be integrated into a single feature vector that can completely describe the global shape of the blood vessel, thus realizing the transformation from local segment description to overall blood vessel description.

[0071] The proposed method for extracting the overall feature vector of the vascular skeleton based on the Transformer model employs a Transformer encoder architecture for feature extraction, which can effectively capture long-distance topological connections in the vascular tree. At the same time, it uses a permutation-independent symmetric function to globally fuse the features output by the encoder, and the generated overall feature vector of the vascular skeleton can completely reflect the global topological information of the blood vessels.

[0072] The resulting overall feature vector of the vascular skeleton can provide a complete and accurate overall shape descriptor for complex vascular 3D structures, provide standardized feature input for downstream medical tasks based on vascular shape features, and provide a new solution for disease diagnosis and treatment.

[0073] For example, it can be used for prognostic risk prediction, such as predicting the risk of renal cell carcinoma metastasis, the risk of liver failure after hepatectomy, and the risk of stroke recurrence. Specifically, it can collect a large amount of vascular data and long-term follow-up results from patients; analyze the correlation between the overall feature vector of the vascular skeleton and the prognostic results; and establish a prognostic prediction model. By inputting the vector of a new patient, it is possible to predict their disease progression and survival probability in the next few years.

[0074] Secondly, it can also be used for precise surgical planning, such as anatomical liver resection planning, partial nephrectomy planning, and vascular interventional surgery planning. Specifically, it generates the overall feature vector of the patient's vascular skeleton; compares it with the vectors of a large number of successful surgical cases in the database; finds several cases that are most similar to the patient's vascular shape, and refers to the surgical plans of these cases; and formulates a personalized and precise surgical plan based on the patient's specific situation. This can help doctors more accurately determine the surgical resection range, completely remove the tumor, and preserve healthy tissue to the maximum extent, thereby improving the success rate of surgery and the patient's prognosis.

[0075] like Figure 2 As shown, based on the same application concept, this application embodiment also provides a system for extracting the overall feature vector of a vascular skeleton based on the Transformer model, including: The acquisition module is used to acquire a three-dimensional blood vessel surface model; A skeleton generation module is used to extract the vascular skeleton from the three-dimensional vascular surface model using an average curvature flow algorithm. The data construction and radius calculation module is used to construct the graph data structure of the vascular skeleton and calculate the vascular radius at each skeleton vertex of the vascular skeleton; the vertices and edges of the graph data structure correspond one-to-one with the skeleton vertices and skeleton edges of the vascular skeleton, and the attributes of the vertex include the spatial coordinates of the vertex and the list of surface point coordinates of the corresponding three-dimensional vascular surface model before contraction. The skeleton repair module is used to repair the vascular skeleton, including bridging the broken skeleton, deleting skeleton noise debris, and removing skeleton noise burrs. The grading module is used to assign a grading number to each vertex in the graph data structure through a recursive backtracking algorithm, thereby completing the grading of the vascular skeleton. The skeleton segmentation module is used to segment the vascular skeleton into a set number of skeleton segments of equal length; The feature extraction module is used to extract the segment feature vectors and position codes of each skeleton segment; The encoding feature vector extraction module is used to input the segment feature vector and the position encoding into the trained Transformer model to obtain the encoding feature vector of each skeleton segment; The overall feature vector extraction module is used to fuse all the encoded feature vectors by arranging independent symmetric functions to obtain the overall feature vector of the vascular skeleton.

[0076] Based on the same concept, embodiments of this application also provide a computer device, including at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by a processor, the instructions being executed by the at least one processor to cause the at least one processor to implement the above-described method when executing the instructions.

[0077] Based on the same concept, embodiments of this application also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method.

[0078] In this embodiment, the computer program, when run by the processor, can also execute other machine-readable instructions to perform other methods as described in the embodiments. For details on the specific execution steps and principles, please refer to the description of the embodiments, which will not be repeated here.

[0079] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0080] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0081] In addition, the functional units in the embodiments provided in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0082] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0084] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for extracting the overall feature vector of a vascular skeleton based on the Transformer model, characterized in that, include: Obtain a three-dimensional blood vessel surface model; The vascular skeleton was extracted from the three-dimensional vascular surface model using the average curvature flow algorithm. Construct a graph data structure for the vascular skeleton and calculate the vascular radius at each skeleton vertex of the vascular skeleton; the vertices and edges of the graph data structure correspond one-to-one with the skeleton vertices and skeleton edges of the vascular skeleton, and the attributes of the vertex include the spatial coordinates of the vertex and the list of surface point coordinates of the corresponding three-dimensional vascular surface model before contraction. Repairing the vascular skeleton includes bridging the broken skeleton, deleting skeleton noise debris, and removing skeleton noise burrs. By using a recursive backtracking algorithm, each vertex in the graph data structure is assigned a level number to complete the level classification of the vascular skeleton; The vascular skeleton is divided into a predetermined number of skeleton segments of equal length; Extract the segment feature vector and position code of each skeleton segment, wherein the segment feature vector is... ; The segment feature vector and the position encoding are input into the trained Transformer model to obtain the encoded feature vector of each skeleton segment; By merging all the encoded feature vectors using an unrelated symmetric function, the overall feature vector of the vascular skeleton is obtained.

2. The method for extracting the overall feature vector of a vascular skeleton based on the Transformer model according to claim 1, characterized in that, The training steps for the Transformer model include: S1: Input the segment feature vector and the position encoding into the random mask layer of the untrained Transformer model; S2: Randomly select 50% of the skeleton segments, hide the corresponding segment feature vectors, and generate masked input features; S3: The masked input features are fed into the encoder of the Transformer model. The encoder processes all the unmasked segment feature vectors and outputs the encoded feature vector. S4: Input the encoded feature vector into the decoder of the Transformer model. Based on the encoded feature vector, the decoder reconstructs the segment feature vectors of all the skeleton segments, thereby outputting the complete reconstructed segment feature vectors. S5: Calculate the loss function, which is defined as the reconstruction error between the reconstructed fragment feature vector and the original input fragment feature vector; S6: Optimize the model parameters using the backpropagation algorithm to minimize the reconstruction error; S7: Repeat steps S2-S6 until the reconstructed segment feature vector output by the Transformer model can stably and accurately restore the original input segment feature vector, indicating that the Transformer model has mastered the shape rules of complex blood vessels in the human body and can stably and accurately extract the encoded feature vector.

3. The method for extracting the overall feature vector of the vascular skeleton based on the Transformer model according to claim 1, characterized in that, The steps for bridging the fractured skeleton include: Receive the first mouse click position input by the user, the first mouse click position corresponding to the rupture area to be bridged on the vascular skeleton displayed on the screen; Traverse all vertices of the graph data structure and calculate the Euclidean distance between the vertex and the first mouse click position, and take the vertex with the smallest Euclidean distance as the first bridge endpoint; Receive the second mouse click position input by the user; Traverse all vertices of the graph data structure and calculate the Euclidean distance between the vertex and the second mouse click position, and take the vertex with the smallest Euclidean distance as the second bridge endpoint; A connecting edge is added between the first bridging endpoint and the second bridging endpoint, so that the broken skeleton is connected through the connecting edge, thereby completing the bridging of the broken skeleton.

4. The method for extracting the overall feature vector of the vascular skeleton based on the Transformer model according to claim 3, characterized in that, The step of removing skeleton noise fragments includes: The connected component set is obtained by calculating the connected component to which each vertex belongs using the boost:connected_components function; Count the number of vertices contained in each of the connected regions; Minimum number of user inputs required threshold ; Traverse the set of connected regions; if the number of vertices contained in it is less than the minimum threshold number... Then all vertices of the connected region are marked as noise fragments; Traverse all vertices of the graph data structure, delete all vertices marked as noise fragments, and the edges they form, thereby completing the deletion of skeleton noise fragments.

5. The method for extracting the overall feature vector of the vascular skeleton based on the Transformer model according to claim 4, characterized in that, The step of removing skeleton noise burrs includes: Receive the root node selection command input by the user and determine the root node; Starting from the root node, all vertices in the graph data structure are traversed using a breadth-first search algorithm, and the number of neighboring vertices of the vertex is obtained using the boost:degree function, which is used as the degree of the vertex. The vertices with a degree of 1 are marked as leaf nodes, the vertices with a degree of 2 are marked as ordinary nodes, and the vertices with a degree greater than 2 are marked as branching points. Traverse all the leaf nodes, and perform a reverse breadth-first traversal with each leaf node as the initial vertex until the first branch point is reached, thereby obtaining the leaf branch corresponding to the leaf node, and calculating the total length of the leaf branch. Minimum length threshold for receiving user input ; Traverse all the leaf branches; if their total length is less than the minimum length threshold... Then all vertices in the leaf branch except for the bifurcation point are marked as noise spikes; Traverse all vertices of the graph data structure, delete all vertices marked as noise spikes, and the edges they form, thereby completing the removal of skeleton noise spikes.

6. The method for extracting the overall feature vector of a vascular skeleton based on the Transformer model according to claim 5, characterized in that, The steps for assigning level numbers to each vertex in the graph data structure using a recursive backtracking algorithm to complete the level classification of the vascular skeleton include: Set the level number of the root node to 0; Using the root node as the current vertex, initiate the following recursive backtracking algorithm: Obtain all adjacent vertices of the current vertex, exclude adjacent vertices that have already been assigned level numbers, and obtain the set of adjacent vertices to be processed; Traverse each vertex to be processed in the set of adjacent vertices to be processed. If the vertex to be processed is a normal vertex, assign the current level number to the vertex to be processed and use the vertex to be processed as the new current vertex. Recursively execute the recursive backtracking algorithm. If the vertex to be processed is the branch point, then the current level number is incremented by 1, the updated level number is assigned to the vertex to be processed, and the vertex to be processed is used as the new current vertex. The recursive backtracking algorithm is then executed recursively. If the vertex to be processed is a leaf node, then the current level number is assigned to the vertex to be processed, and the recursive backtracking algorithm of this branch is terminated. Once all vertices in the graph data structure have been assigned level numbers, the recursive backtracking algorithm is terminated, thereby completing the level classification of the vascular skeleton.

7. The method for extracting the overall feature vector of a vascular skeleton based on the Transformer model according to claim 6, characterized in that, The steps for extracting the segment feature vectors and positional codes of each skeleton segment include: Traverse all the skeleton segments and obtain all the skeleton vertices contained in each skeleton segment; Calculate the average of the spatial coordinates of all the skeleton vertices within the skeleton segment to obtain the centroid coordinates of the skeleton segment, and use the centroid coordinates as the position code of the skeleton segment; The spatial coordinates of all the skeleton vertices within the skeleton segment are converted into a coordinate matrix. Principal component analysis is performed on the coordinate matrix to extract the eigenvector corresponding to the largest eigenvalue, which is used as the orientation of the skeleton segment. The average inner diameter of the skeleton segment is obtained by averaging the blood vessel radii corresponding to all skeleton vertices within the skeleton segment. The skeleton vertex connected to the previous skeleton segment in the skeleton segment is selected as the segment starting point, and the spatial coordinates of the segment starting point are used as the geodesic coordinates of the previous skeleton segment. Obtain the level number corresponding to the skeleton vertex within the skeleton segment, and use it as the level of the skeleton segment; The orientation, the average inner diameter, the geodesic upper coordinates, and the level are integrated to obtain the segment feature vector of the skeleton segment.

8. A system for extracting the overall feature vector of a vascular skeleton based on the Transformer model, characterized in that, include: The acquisition module is used to acquire a three-dimensional blood vessel surface model; A skeleton generation module is used to extract the vascular skeleton from the three-dimensional vascular surface model using an average curvature flow algorithm. The data construction and radius calculation module is used to construct the graph data structure of the vascular skeleton and calculate the vascular radius at each skeleton vertex of the vascular skeleton; the vertices and edges of the graph data structure correspond one-to-one with the skeleton vertices and skeleton edges of the vascular skeleton, and the attributes of the vertex include the spatial coordinates of the vertex and the list of surface point coordinates of the corresponding three-dimensional vascular surface model before contraction. The skeleton repair module is used to repair the vascular skeleton, including bridging the broken skeleton, deleting skeleton noise debris, and removing skeleton noise burrs. The grading module is used to assign a grading number to each vertex in the graph data structure through a recursive backtracking algorithm, thereby completing the grading of the vascular skeleton. The skeleton segmentation module is used to segment the vascular skeleton into a set number of skeleton segments of equal length; The feature extraction module is used to extract the segment feature vectors and position codes of each skeleton segment; The encoding feature vector extraction module is used to input the segment feature vector and the position encoding into the trained Transformer model to obtain the encoding feature vector of each skeleton segment; The overall feature vector extraction module is used to fuse all the encoded feature vectors by arranging independent symmetric functions to obtain the overall feature vector of the vascular skeleton.

9. A computer device, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the steps of the method according to any one of claims 1 to 7 when executing the instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images