Search

CN-121685880-B - Virtual reconstruction system before thoracic surgery operation

CN121685880BCN 121685880 BCN121685880 BCN 121685880BCN-121685880-B

Abstract

The invention relates to the technical field of three-dimensional reconstruction and intelligent analysis of medical images, and particularly discloses a pre-thoracic surgery virtual reconstruction system which comprises the steps of constructing a space diagram structure by taking voxels as nodes, introducing respiratory motion displacement tracks as dynamic attributes, extracting and fusing static structural features and dynamic motion features to generate context feature vectors rich in anatomical and functional information, sequencing the nodes of the diagram according to topological geodesic distances and feature significance in a sequencing topological constraint segmentation module, realizing layering segmentation from trunk to terminal by combining an anatomical topological constraint and sequence prediction model based on a rule base, carrying out connectivity verification, fracture repair and geometric rationality verification, and outputting a high-quality pre-operative virtual reconstruction three-dimensional model which is accurate, free of errors and accords with anatomical rules.

Inventors

  • Duan Wanshi
  • JING XIN

Assignees

  • 中国人民解放军空军军医大学

Dates

Publication Date
20260508
Application Date
20260210

Claims (9)

  1. 1. A pre-thoracic surgical virtual reconstruction system, comprising: The system comprises a multi-mode data fusion and dynamic graph construction module, a graph node and a graph node, wherein the multi-mode data fusion and dynamic graph construction module is used for synchronously acquiring CT image data and respiratory CT image data of the chest; The cross-modal context perception feature coding module extracts a local three-dimensional image block and a corresponding displacement track sequence in CT image data aiming at each graph node, respectively processes the image block and the track sequence to obtain a static structure feature and a dynamic motion feature; The sequence prediction model is used for carrying out label prediction by integrating the current graph node characteristics, the processed neighbor graph node states and a predefined anatomical topological rule base, wherein the rule base prescribes the morphology and connectivity constraint of a key anatomical structure; the method comprises the steps of generating a sequence prediction label, generating a sequence prediction model, carrying out global anatomical rationality test and correction on the sequence prediction label, and outputting a preoperative virtual reconstruction three-dimensional model for clinical use before thoracic surgery.
  2. 2. The pre-thoracic surgical virtual reconstruction system of claim 1, wherein the chest-based CT image data constructs an initial map structure, comprising: Traversing all voxel units based on a preset 26 adjacency rule, and if two voxel units are connected in a three-dimensional space at a plane, an edge or a point, establishing an undirected edge between the map nodes corresponding to the two voxel units; Calculating the image gray value gradient modular length between adjacent voxel units, and endowing each edge with initial weight according to the image gray value gradient modular length; and fusing the image block features with the initial weights to generate enhanced edge attributes, thereby completing the optimized construction of the initial graph structure.
  3. 3. The pre-thoracic virtual reconstruction system according to claim 1, wherein the extracting the displacement track of each voxel unit in respiratory motion from respiratory CT image data as a dynamic attribute is added to a corresponding map node, specifically comprising: carrying out respiratory phase division on respiratory CT image data, and selecting the end-inspiration or end-expiration phase as a reference frame; For each voxel unit, according to the spatial position of each voxel unit in each phase image, a space-time displacement track which spans the complete respiratory cycle is formed by interpolating displacement vectors of each voxel unit in a continuous phase motion vector field; Extracting three characteristic quantities of track length, average displacement amplitude and main motion direction from the displacement track to form a track characteristic vector, and adding the track characteristic vector as dynamic attribute to the corresponding graph node in the initial graph structure.
  4. 4. The pre-thoracic surgical virtual reconstruction system of claim 1, wherein the processing of the image block and the sequence of trajectories to obtain static structural features and dynamic motion features comprises: For a local three-dimensional image block, performing feature extraction by adopting a group of three-dimensional convolution kernels with separable directions, wherein the three-dimensional convolution kernels respectively have different weights along the axial direction, the sagittal direction and the coronal direction so as to anisotropically sense the structural information of different anatomical planes and output a multi-directional feature map; for a displacement track sequence, decomposing a displacement vector of the displacement track sequence into three motion primitive components of translation, rotation and scaling; calculating the amplitude and phase variation spectrum of each motion element component in each respiratory period to form a dynamic motion characteristic vector; And respectively carrying out principal component analysis dimension reduction treatment on the multi-direction feature map and the dynamic motion feature vector, and mapping the dimension reduced features to a unified numerical range to obtain standardized static structural features and dynamic motion features.
  5. 5. The pre-thoracic virtual reconstruction system of claim 1, wherein the integrating the static structural features with the dynamic motion features generates a fused contextual feature vector for each graph node, comprising: Constructing a full-connection temporary graph by taking all graph nodes in the current batch process as vertexes, wherein static structural features and dynamic motion features of each node are respectively used as two attributes of the corresponding vertexes; The dynamic motion characteristics of the target vertexes are weighted and aggregated by using the attention weights, and a dynamic characteristic enhanced by the structural context is generated for each vertex; And the gating fusion unit learns a dynamic weight to control the fusion proportion of the two types of features, and finally outputs the unique fusion context feature vector of each graph node.
  6. 6. The pre-thoracic virtual reconstruction system of claim 1, wherein the ordering of the graph nodes according to the fused context feature vector forms a processing sequence, specifically comprising: Based on the fused context feature vector, identifying graph nodes belonging to the main starting point of the key anatomical structure as an initial root node set; Calculating the topological geodesic distance from all other graph nodes to the nearest root node on the initial graph structure by taking the root node set as a starting point, wherein the topological geodesic distance refers to the shortest hop count of a connection path along a graph edge; And sequencing all the graph nodes according to the sequence from small to large of the topological geodesic distance, and performing secondary sequencing on the graph nodes with the same distance according to the modular length of the fusion context feature vector of the graph nodes, so as to finally generate a processing sequence radiating from the structural core to the edge.
  7. 7. The pre-thoracic virtual reconstruction system of claim 1, wherein the sequential processing of each of the map nodes using the sequence prediction model comprises: before predicting the label of the current graph node, inquiring the label state of the processed neighbor graph node, and judging whether the label state triggers a rule in a predefined anatomical topology rule base or not; If the rule is triggered, adjusting the attention weight in the sequence prediction model according to the rule type; the sequence prediction model integrates the current node characteristics and the regulated neighborhood node state context, and finally decides the label of the current node through a gating circulation unit.
  8. 8. The pre-thoracic virtual reconstruction system of claim 1, wherein the global anatomic rationality test and correction of the initial segmentation mask comprises: Calculating the number of connected components of the topological skeleton network, comparing the number with the number of normal connected components specified in a predefined anatomical rule base, and identifying whether abnormal disconnection exists or not; Determining a position fracture endpoint for the identified abnormal disconnection area, and calculating the geometric characteristics of the three-dimensional space gap at the fracture; and generating a smooth transition restoration curved surface in the fracture gap by taking the fracture end points and the neighborhood contour points thereof as constraint conditions, and recovering the correct topological connectivity of the anatomical structure.
  9. 9. The thoracic surgical pre-operative virtual reconstruction system of claim 1, wherein the outputting the pre-operative virtual reconstructed three-dimensional model specifically comprises: calculating the total volume and the surface area of the repaired three-dimensional segmentation mask, comparing the total volume and the surface area with the range of typical organ geometric parameters recorded in a predefined anatomical topology rule base, and verifying the geometric rationality of the typical organ geometric parameters; Performing multi-resolution analysis on the initial surface mesh model obtained by the mask conversion after repair, smoothing unnatural surface fluctuation introduced by interpolation repair under a coarse resolution level, and maintaining real anatomical details derived from original image data under a fine resolution level; And performing water tightness treatment and error-free surface patch inspection on the optimized curved surface grid, and finally outputting a preoperative virtual reconstruction three-dimensional model which can be directly used for three-dimensional visualization, printing or surgical navigation.

Description

Virtual reconstruction system before thoracic surgery operation Technical Field The invention relates to the technical field of three-dimensional reconstruction and intelligent analysis of medical images, in particular to a pre-thoracic surgery virtual reconstruction system. Background In pre-thoracic planning, accurate three-dimensional reconstruction of organs and lesions based on Computed Tomography (CT) images is of paramount importance. The prior art generally employs segmentation algorithms that achieve tissue separation by mapping the images into graph structures and minimizing energy functions. However, when dealing with complex regions with highly overlapping gray values and tightly adhered structures, these methods rely heavily on the constraints of image gray gradients and spatial smoothness, and it is often difficult to accurately distinguish between different tissue boundaries, resulting in limited segmentation accuracy. The prior art has the following defects: When the graph-cut algorithm based on energy minimization processes adhesion tissues with similar gray scales and complex topological structures, the smooth term constraint of the graph-cut algorithm can preferentially select a segmentation path which passes through a few high-contrast areas instead of a real tortuous boundary due to a mathematical short-circuit effect, so that non-physical topological fracture of tubular structures such as blood vessels and the like is caused in a three-dimensional reconstruction model. Such errors are less in the conventional simple cases, but are particularly prominent in extreme scenes such as 'dead lock' adhesion of lung gate areas, the generated topological distortion model can seriously mislead surgical planning, even cause surgical selection errors, and the traditional method lacks an explicit maintenance mechanism for anatomical topological connectivity, so that the risk cannot be effectively avoided. Disclosure of Invention The present invention aims to provide a thoracic surgery pre-operative virtual reconstruction system which solves the above-mentioned problems. The aim of the invention can be achieved by the following technical scheme: a thoracic surgical pre-operative virtual reconstruction system, comprising: The system comprises a multi-mode data fusion and dynamic graph construction module, a graph node and a graph node, wherein the multi-mode data fusion and dynamic graph construction module is used for synchronously acquiring CT image data and respiratory CT image data of the chest; The cross-modal context perception feature coding module extracts a local three-dimensional image block and a corresponding displacement track sequence in CT image data aiming at each graph node, respectively processes the image block and the track sequence to obtain a static structure feature and a dynamic motion feature; The sequence prediction model is used for carrying out label prediction by integrating the current graph node characteristics, the processed neighbor graph node states and a predefined anatomical topological rule base, wherein the rule base prescribes the morphology and connectivity constraint of a key anatomical structure; the method comprises the steps of generating a sequence prediction label, generating a sequence prediction model, carrying out global anatomical rationality test and correction on the sequence prediction label, and outputting a preoperative virtual reconstruction three-dimensional model for clinical use before thoracic surgery. As a further scheme of the invention, the CT image data based on the chest is used for constructing an initial graph structure, which specifically comprises the following steps: Traversing all voxel units based on a preset 26 adjacency rule, and if two voxel units are connected in a three-dimensional space at a plane, an edge or a point, establishing an undirected edge between the map nodes corresponding to the two voxel units; Calculating the image gray value gradient modular length between adjacent voxel units, and endowing each edge with initial weight according to the image gray value gradient modular length; and fusing the image block features with the initial weights to generate enhanced edge attributes, thereby completing the optimized construction of the initial graph structure. The invention further provides a method for extracting displacement tracks of each voxel unit in respiratory motion from respiratory CT image data, and adding the displacement tracks as dynamic attributes to corresponding graph nodes, wherein the method specifically comprises the following steps: carrying out respiratory phase division on respiratory CT image data, and selecting the end-inspiration or end-expiration phase as a reference frame; For each voxel unit, according to the spatial position of each voxel unit in each phase image, a space-time displacement track which spans the complete respiratory cycle is formed by interpolating displacement vectors of e