A percutaneous lung puncture path planning method based on pulmonary segment level blood vessel structure constraint

Abstract

The application discloses a percutaneous lung puncture path planning method and device based on lung segment level blood vessel structure constraint, and the method comprises the following steps: performing isotropic resampling, lung lobe segmentation and bronchial tree extraction on CT image data to generate a three-dimensional lung dissection model; inputting the three-dimensional lung dissection model into a lung segment semantic segmentation module to output a lung segment boundary; reconstructing the topological connection relationship between lung segment level blood supply arteries and drainage veins from the lung segment boundary to obtain a structured lung segment level blood vessel atlas; constructing a dynamic puncture risk field voxel map according to the structured lung segment level blood vessel atlas; sampling a plurality of candidate needle entry points in a surface feasible needle entry region to generate an initial path set in the three-dimensional lung dissection model; and weighting the path length, cumulative risk value, puncture angle penalty and high-density lung tissue volume according to a multi-objective optimization function to output a final puncture path. The application significantly improves the safety boundary and clinical applicability of percutaneous lung puncture.

Description

Technical Field

[0001] This invention relates to the field of computer-aided interventional therapy technology, specifically to a percutaneous lung puncture path planning method and device, and computer equipment based on lung segmental vascular structure constraints. Background Technology

[0002] Percutaneous lung biopsy is the gold standard for pathological diagnosis of pulmonary space-occupying lesions and plays an irreplaceable role in the early diagnosis, molecular subtyping, and personalized treatment of lung cancer. With the widespread use of low-dose spiral CT screening, the detection rate of pulmonary nodules has increased significantly, leading to a continuous rise in the number of lung biopsies. However, pneumothorax is the most common postoperative complication, and in severe cases, it can progress to tension pneumothorax, endangering life. Furthermore, pulmonary vascular injury can lead to hemorrhage or even air embolism; although the incidence is low, the mortality rate is high, not only prolonging hospital stays and increasing medical costs but also potentially delaying subsequent treatment.

[0003] Traditional lung biopsy path planning heavily relies on the surgeon's clinical experience. Physicians must mentally visualize the three-dimensional anatomical structure on two-dimensional CT images, selecting the entry point and puncture path while avoiding multiple constraints such as ribs, interlobar fissures, major blood vessels, and pulmonary bullae. The length of air-bearing lung tissue traversed along the puncture path is the primary predictor of pneumothorax risk, and the pleural puncture angle also has a significant impact. However, physicians struggle to quantitatively weigh these multivariate constraints within a limited timeframe, leading to significant differences in surgical quality among surgeons with varying experience levels.

[0004] To address the above issues, this invention proposes a percutaneous lung puncture path planning method based on lung segmental vascular structure constraints, in order to improve the safety boundaries and clinical applicability of percutaneous lung puncture. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method and device for percutaneous lung puncture path planning based on lung segmental vascular structure constraints, as well as computer equipment.

[0006] According to one aspect of the present invention, a percutaneous lung puncture path planning method based on lung segmental vascular structure constraints is provided, comprising:

[0007] Preoperative chest CT images of patients are acquired and isotropically resampled, segmented into lung lobes, and extracted into bronchial trees to generate a three-dimensional lung anatomy model. The three-dimensional lung anatomy model is then input into a lung segment semantic segmentation module, which outputs lung segment boundaries. Based on a threshold centerline tracing algorithm, the topological connections between the supplying arteries and draining veins at the lung segment boundaries are reconstructed to obtain a structured lung segment vascular atlas containing information on vessel diameter, branch hierarchy, and spatial orientation.

[0008] The puncture target point is marked in the three-dimensional lung anatomy model of the target lesion, and a dynamic puncture risk field voxel map is constructed based on the structured lung segmental vascular atlas. Among them, the risk attenuation radius is dynamically set according to the diameter of each blood vessel centerline as the source, and a Gaussian kernel function is used to generate a continuously differentiable risk potential energy distribution to form a three-dimensional puncture risk field voxel map covering the entire lung parenchyma.

[0009] Multiple candidate needle insertion points are sampled in the feasible needle insertion area on the body surface. An initial path set is generated in the three-dimensional lung anatomy model using the fast travel method. Each path in the initial path set is mapped to the three-dimensional puncture risk field voxel map to obtain the cumulative risk value. A multi-objective optimization function is constructed based on the cumulative risk value, the lung tissue density weighting factor, and the puncture angle penalty term.

[0010] The multi-objective optimization function weighs path length, cumulative risk value, puncture angle penalty, and volume of high-density lung tissue traversed, and outputs a set of non-dominated optimal paths as the final puncture path; wherein, the final puncture path includes a spatial coordinate sequence, real-time risk thermal projection, and avoidance suggestions.

[0011] In one alternative approach, reconstructing the topological connectivity between the pulmonary segmental feeding arteries and draining veins from the pulmonary segment boundary using a threshold centerline tracing algorithm to obtain a structured pulmonary segmental vascular atlas containing information on vessel diameter, branching hierarchy, and spatial orientation further includes:

[0012] The lung segment boundaries are spatially registered with the bronchial tree skeleton extracted from the bronchial tree, and the anatomical labels of each bronchial segment are used as prior constraints to serve as the tracking starting regions for the accompanying arteries within the lung segment and the intersegmental veins, respectively. Specifically, for the accompanying arteries within the lung segment, the lateral bronchial wall neighborhood is used as the seed point set; for the draining veins, the adjacent area of ​​the lung segmental septum and the subpleural reflux area are used as the seed point set.

[0013] An adaptive morphological top-hat transformation method is used to enhance the low-contrast small vessel signal in the three-dimensional lung anatomy model, resulting in an enhanced vessel probability map. A threshold centerline tracing algorithm is then executed on this enhanced vessel probability map, starting from the seed point set. This algorithm iteratively searches for regions in adjacent voxels that satisfy the circularity and continuity constraints of the vessel cross-section, reconstructing the vessel centerline path pixel by pixel. During this pixel-by-pixel reconstruction, the equivalent circle diameter is calculated as the vessel diameter based on the set of cross-sectional pixels perpendicular to the local direction within the centerline's neighborhood. Branch levels are automatically assigned based on this vessel diameter, its relationship to the corresponding bronchus, its cross-segmental course, and its order of entry into the portal vein. The three-dimensional tangential vector of each point on the centerline is accumulated and recorded as the spatial direction of the vessel, forming the vessel's vector growth trajectory.

[0014] The central lines of all accompanying arteries within the same lung segment are connected hierarchically to form an arterial tree topology, and the draining veins spanning multiple lung segments are connected according to their confluence relationship to form a venous network topology. The arterial tree topology and the venous network topology are then combined to generate a structured lung segmental vascular atlas covering the entire lung.

[0015] In one alternative approach, constructing a dynamic puncture risk field voxel map based on the structured lung segmental vascular atlas further includes:

[0016] The specific location, branching level and spatial orientation of each feeding artery and draining vein are obtained from the structured lung segmental vascular atlas, resulting in a refined set of vascular centerlines;

[0017] For each refined vessel centerline in the refined vessel centerline set, a dynamic risk attenuation radius is set according to the vessel diameter to obtain a weighted risk attenuation radius set.

[0018] The risk decay radius set is generated into a continuously differentiable risk potential energy distribution by using a Gaussian kernel function, forming a dynamic puncture risk field voxel map covering the entire lung parenchyma.

[0019] In one alternative approach, the lung segment semantic segmentation module includes:

[0020] A 3D convolutional encoder-decoder network outputs an initial lung segment probability map. The encoder network consists of 5 downsampling stages, each containing 2 3D convolutional layers and 1 3D max-pooling layer, used to progressively compress spatial resolution and expand the number of channels to 512 dimensions to extract cross-scale lung parenchyma texture features and boundary blurring features. The decoder network consists of 5 upsampling stages, each containing 1 3D transposed convolutional layer and 2 3D convolutional layers, used to restore spatial resolution and fuse multi-level semantic information.

[0021] The bronchial tree anatomy prior injection module is connected to the bronchial tree skeleton and its corresponding segmental bronchial labels, respectively, and receives the intermediate layer feature map output by the fourth layer encoder; the bronchial tree skeleton is encoded using a graph convolutional network, which contains three graph convolutional layers. Each layer uses Chebyshev polynomial approximation to perform spectral domain convolution, mapping the spatial coordinates of bronchial branch points and bronchial diameter features into a 128-dimensional bronchial spatial embedding vector;

[0022] The lung fissure enhancement detection subnetwork includes three sets of parallel branches, corresponding to 2D convolution kernels in the coronal, sagittal, and transverse directions, respectively. Each set of branches contains three convolutional layers, which are used to extract the linear high-density response of the lung fissure on the corresponding plane. The outputs of each branch are rearranged in three dimensions and then stitched along the channel dimension to generate a lung fissure probability map.

[0023] A conditional random field post-processing optimization network is used to output a 3D lung segment semantic label map with spatial resolution aligned with the original CT. The lung segment boundary confidence map is used as a univariate potential term, and a binary potential term is constructed using voxel spatial proximity and CT value similarity. The mean field approximation algorithm is used for 10 iterations of message passing to perform probabilistic inference across the entire map, forcing the lung segment boundary to coincide with the high-confidence lung fissure location and maintaining label consistency within the lung segment.

[0024] The segment-level anatomical compliance verification module is used to receive the three-dimensional lung segment semantic label map and compare it with the preset lung segment anatomical knowledge base. It transmits the coordinates of abnormal areas back to increase the weight of the binary potential term of the abnormal area and re-executes the mean field inference until the final lung segment boundary that conforms to the anatomical rules is output.

[0025] In one alternative approach, a set of non-dominated optimal paths is output as the final puncture path based on a multi-objective optimization function that balances path length, cumulative risk value, puncture angle penalty, and volume of high-density lung tissue traversed, further including:

[0026] The initial path set is divided into Pareto levels according to the fast non-dominated sorting algorithm. Paths that are not dominated by any other path in all four objective dimensions are marked as the first frontier and form a seed set of candidate non-dominated paths.

[0027] Using the candidate non-dominated path seed set as the parent population, the spatial control point sequences of the two parent paths are interpolated and recombined using a simulated binary crossover operator to generate new offspring paths that simultaneously inherit the local geometric features of both parents.

[0028] A polynomial mutation operator is used to apply random perturbation to the path control point in the low potential energy region of the three-dimensional puncture risk field voxel map, driving the path to adaptively drift towards the sparse vascular region, ensuring that the offspring path is still a continuous and passable physical puncture path in the lung parenchyma and is remapped to the three-dimensional puncture risk field voxel map to update its multi-target feature vector.

[0029] The multi-objective optimization function is iteratively executed until a preset population generation threshold is reached. All individuals on the first frontier in the last generation population are output as the Pareto optimal solution set. Based on the three-dimensional spatial coordinate sequence of each path in the Pareto optimal solution set, the corresponding cumulative risk value thermal profile, and the avoidance suggestion text, a non-dominated optimal path is constructed as the final puncture path.

[0030] In an alternative approach, the cumulative risk value is obtained based on a weighted Gaussian potential field integral model of the stress concentration effect on the blood vessel wall and the probability of rupture, and its calculation formula is as follows:

[0031]

[0032] in, , The overall rupture tendency of the i-th blood vessel; The critical stress threshold corresponding to a 50% probability of fracture; The risk decay radius scaling factor recommended by the Fleischner Society; , The velocity component of the puncture needle perpendicular to the local normal of the blood vessel at path point s; Candidate puncture paths Total arc length; The total number of vascular centerlines included in the structured lung segmental vascular atlas; For path In arc length parameter The three-dimensional spatial coordinates at [0, L(p)]; Let x(s) be the projection point on the center line of the i-th blood vessel that is closest to the path point x(s). Let be the equivalent diameter of the i-th blood vessel; Let be the volumetric flow rate of the i-th blood vessel; This refers to pulmonary artery systolic blood pressure. Let be the wall thickness of the i-th blood vessel; For the nearest blood vessel point corresponding to path point x(s) The local normal vector of the normal plane at the centerline of the blood vessel.

[0033] In one alternative approach, the puncture angle penalty term is calculated based on a needle tract offset error propagation model under chest wall curvature constraints, and its calculation formula is as follows:

[0034]

[0035] in, For the angle of needle insertion; The average curvature at the needle insertion point on the body surface; The coefficient for sensitivity of curvature and inter-rib spacing; The pneumothorax index; This is the normalization coefficient.

[0036] In one alternative approach, the formula for calculating the volume traversed through high-density lung tissue is:

[0037]

[0038] in, For local Young's modulus, ; To calibrate the compressibility coefficient of isolated pig lungs; Here is the Hounsfield unit for the CT image at spatial location x; As a benchmark for healthy lung parenchyma modulus; This is an empirical coefficient; As a continuous indicator of the degree of fibrosis, ; The additional resistance coefficient for fibrosis; Let s be the three-dimensional coordinates corresponding to the arc length parameter s on path p; Candidate puncture paths The total arc length.

[0039] According to another aspect of the present invention, a percutaneous lung puncture path planning device based on lung segmental vascular structure constraints is provided, comprising:

[0040] A lung segmental vascular atlas construction module is used to acquire preoperative chest CT image data of patients and perform isotropic resampling, lobe segmentation and bronchial tree extraction on the CT image data to generate a three-dimensional lung anatomy model; the three-dimensional lung anatomy model is input into the lung segment semantic segmentation module to output the lung segment boundaries; the topological connection relationship between the lung segment feeding arteries and draining veins is reconstructed from the lung segment boundaries according to the threshold centerline tracing algorithm to obtain a structured lung segmental vascular atlas containing information on vessel diameter, branch level and spatial orientation.

[0041] The dynamic puncture risk field generation module is used to mark the puncture target point in the three-dimensional lung anatomy model of the target lesion and construct a dynamic puncture risk field voxel map based on the structured lung segmental vascular atlas. Specifically, the risk attenuation radius is dynamically set according to the diameter of each blood vessel centerline as the source, and a Gaussian kernel function is used to generate a continuously differentiable risk potential energy distribution to form a three-dimensional puncture risk field voxel map covering the entire lung parenchyma.

[0042] The path set generation and risk mapping module is used to sample multiple candidate needle insertion points in the feasible needle insertion area on the body surface, generate an initial path set in the three-dimensional lung anatomy model using the fast travel method, map each path in the initial path set to the three-dimensional puncture risk field voxel map to obtain a cumulative risk value, and construct a multi-objective optimization function based on the cumulative risk value, lung tissue density weighting factor and puncture angle penalty term.

[0043] The non-dominated path generation module is used to weigh path length, cumulative risk value, puncture angle penalty and high-density lung tissue volume traversed according to a multi-objective optimization function, and output a set of non-dominated optimal paths as the final puncture path; wherein, the final puncture path includes a spatial coordinate sequence, real-time risk thermal projection and avoidance suggestions.

[0044] According to another aspect of the present invention, a computer device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus;

[0045] The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the percutaneous lung puncture path planning method based on lung segment-level vascular structure constraints described above.

[0046] According to the solution provided by the present invention, preoperative chest CT image data of the patient is acquired, and isotropic resampling, lobe segmentation, and bronchial tree extraction are performed on the CT image data to generate a three-dimensional lung anatomy model; the three-dimensional lung anatomy model is input into a lung segment semantic segmentation module to output lung segment boundaries; the topological connection relationship between the lung segment-level feeding arteries and draining veins is reconstructed from the lung segment boundaries according to a threshold centerline tracing algorithm to obtain a structured lung segment-level vascular atlas containing information on vessel diameter, branch level, and spatial orientation; puncture target points are marked in the three-dimensional lung anatomy model of the target lesion, and a dynamic puncture risk field voxel map is constructed based on the structured lung segment-level vascular atlas; wherein, the risk attenuation radius is dynamically set according to the diameter of each vessel centerline as the source. A continuously differentiable risk potential energy distribution is generated using a Gaussian kernel function to form a three-dimensional puncture risk field voxel map covering the entire lung parenchyma. Multiple candidate needle insertion points are sampled in the feasible needle insertion area on the body surface. An initial path set is generated in the three-dimensional lung anatomy model using a fast-moving method. Each path in the initial path set is mapped to the three-dimensional puncture risk field voxel map to obtain a cumulative risk value. A multi-objective optimization function is constructed based on the cumulative risk value, lung tissue density weighting factor, and puncture angle penalty term. The multi-objective optimization function balances path length, cumulative risk value, puncture angle penalty, and the volume of high-density lung tissue traversed, outputting a set of non-dominated optimal paths as the final puncture path. The final puncture path includes a spatial coordinate sequence, real-time risk thermal projection, and avoidance suggestions. This invention significantly improves the safety boundaries and clinical applicability of percutaneous lung puncture.

[0047] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0048] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0049] Figure 1 A flowchart illustrating the percutaneous lung puncture path planning method based on lung segment-level vascular structure constraints according to an embodiment of the present invention is shown.

[0050] Figure 2A A schematic diagram illustrating the process of generating cumulative risk according to an embodiment of the present invention is shown;

[0051] Figure 2B This diagram illustrates the process of generating the puncture path according to an embodiment of the present invention.

[0052] Figure 3A A schematic diagram of a structured lung segmental vascular atlas according to an embodiment of the present invention is shown;

[0053] Figure 3B A schematic diagram of the final puncture path in an embodiment of the present invention is shown;

[0054] Figure 4 This invention illustrates a schematic diagram of the percutaneous lung puncture path planning device based on lung segment-level vascular structure constraints according to an embodiment of the present invention.

[0055] Figure 5 A schematic diagram of the structure of a computer device according to an embodiment of the present invention is shown. Detailed Implementation

[0056] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0057] Figure 1 A flowchart illustrating the percutaneous lung puncture path planning method based on lung segmental vascular structure constraints according to an embodiment of the present invention is shown. Specifically, as... Figure 1 As shown, it includes the following steps:

[0058] Step S101: Acquire the patient's preoperative chest CT image data and perform isotropic resampling, lobe segmentation, and bronchial tree extraction on the CT image data to generate a three-dimensional lung anatomy model; input the three-dimensional lung anatomy model into the lung segment semantic segmentation module and output the lung segment boundaries; reconstruct the topological connection relationship between the lung segment-level feeding arteries and draining veins from the lung segment boundaries according to the threshold centerline tracing algorithm to obtain a structured lung segment-level vascular atlas containing information on vessel diameter, branch level, and spatial orientation.

[0059] In this embodiment, the generated three-dimensional lung anatomy model is segmented using a lung segment semantic segmentation module to accurately identify lung segment boundaries, especially key areas that need to be avoided when planning the puncture path. The threshold centerline tracing algorithm reconstructs the topological connections between the supplying arteries and draining veins at the lung segment boundaries, obtaining a structured atlas containing information on vessel diameter, branching hierarchy, and spatial orientation. This helps reduce the risk of damage to important vessels during puncture and improves surgical safety.

[0060] In one alternative approach, reconstructing the topological connectivity between the pulmonary segmental feeding arteries and draining veins from the pulmonary segment boundary using a threshold centerline tracing algorithm to obtain a structured pulmonary segmental vascular atlas containing information on vessel diameter, branching hierarchy, and spatial orientation further includes:

[0061] The lung segment boundaries are spatially registered with the bronchial tree skeleton extracted from the bronchial tree, and the anatomical labels of each bronchial segment are used as prior constraints to serve as the tracking starting regions for the accompanying arteries within the lung segment and the intersegmental veins, respectively. Specifically, for the accompanying arteries within the lung segment, the lateral bronchial wall neighborhood is used as the seed point set; for the draining veins, the adjacent area of ​​the lung segmental septum and the subpleural reflux area are used as the seed point set.

[0062] An adaptive morphological top-hat transformation method is used to enhance the low-contrast small vessel signal in the three-dimensional lung anatomy model, resulting in an enhanced vessel probability map. A threshold centerline tracing algorithm is then executed on this enhanced vessel probability map, starting from the seed point set. This algorithm iteratively searches for regions in adjacent voxels that satisfy the circularity and continuity constraints of the vessel cross-section, reconstructing the vessel centerline path pixel by pixel. During this pixel-by-pixel reconstruction, the equivalent circle diameter is calculated as the vessel diameter based on the set of cross-sectional pixels perpendicular to the local direction within the centerline's neighborhood. Branch levels are automatically assigned based on this vessel diameter, its relationship to the corresponding bronchus, its cross-segmental course, and its order of entry into the portal vein. The three-dimensional tangential vector of each point on the centerline is accumulated and recorded as the spatial direction of the vessel, forming the vessel's vector growth trajectory.

[0063] The central lines of all accompanying arteries within the same lung segment are connected hierarchically to form an arterial tree topology, and the draining veins spanning multiple lung segments are connected according to their confluence relationship to form a venous network topology. The arterial tree topology and the venous network topology are then combined to generate a structured lung segmental vascular atlas covering the entire lung.

[0064] In this embodiment, the lung segment boundaries are spatially registered with the bronchial tree framework, and the anatomical tags of each bronchial segment are used as geofences for arterial and venous tracking. For accompanying arteries, tracking is initiated within a 1.5-3.0 mm annular neighborhood on the outer side of the bronchial wall to exclude interference from non-accompanying vessels. For draining veins, the adjacent areas of the lung segmental septa and the subpleural reflux area are anchored. The capture of intersegmental veins and subpleural veins can compress the search space for vessel identification by more than 85%, reducing the false positive rate to 1 / 4 of that of traditional global tracking methods. Arteries are stored as a root-trunk-branch tree structure, and veins are stored as a tributary-converging trunk-main trunk network structure. The two topologies can be queried independently or fused for analysis (e.g., "puncture path distance from A3a 2.1 mm, distance from V3c 3.4 mm"), allowing physicians to obtain a complete lung segmental vascular map preoperatively. Figure 3A As shown, a thin line (intra-segmental vessels, level 1.1) originates from the outer wall of the bronchus in the basal lobe of the lung, with its diameter decreasing to 0.5 mm; an arc-shaped line (inter-segmental vessels, level 1.2) originates from the subpleural reflux area of ​​the upper lobe of the lung and merges with the thin line at the hilum; the level number is marked at the end of each line, and the transparency varies with the diameter (larger diameter = higher transparency).

[0065] Specifically, using the tracheal carina and the bifurcation points of the bilateral main bronchi as control points, the translation and rotation parameters are optimized using the mutual information maximization criterion to achieve sub-voxel-level alignment (target registration error <1.2mm). After registration, a lung segment-bronchus mapping table is established; for example, the anterior segment of the right upper lobe (S2) is associated with the B1 bronchus, and the lingular segment of the left upper lobe (S1) is associated with the B2 bronchus. This mapping table serves as the anatomical index for subsequent vascular tracking. Traversing the bronchial centerline of each segment, an annular region 1.5-3.0mm away from the bronchial centerline is extracted as a candidate region on a plane perpendicular to the local bronchial orientation. Within this region, voxels meeting the following criteria are selected: CT value 150-450 HU, vascular probability map response >0.7, and local grayscale peak value; these are marked as arterial seed points. If no candidate point is detected in the annular region lateral to a certain bronchial segment, the search radius is automatically expanded to 4.0mm and the probability threshold is reduced to 0.6 to address situations where small vessels are poorly visualized. Voxels at lung segment boundaries are extracted, and a distance transform field is generated. Regions within 2 mm of the segment boundary are selected. Simultaneously, voxels on the pleural surface are extracted and expanded inwards by 3-8 mm to generate a subpleural region. The union of these two regions is taken as the vein candidate region. Within this region, voxels satisfying a CT value of -100-150 HU, a vessel probability map response >0.65, and a local gray-scale mean are selected and marked as vein seed points. For large veins near the hilum, a constraint of no accompanying relationship with the bronchi is applied to avoid mistakenly including arteries in vein tracking. A spherical structural element sequence is constructed, with a diameter set of {1.5 mm, 2.5 mm, 4.0 mm, 6.0 mm}. For each voxel, a four-scale top-cap transform result is calculated, and the maximum response is taken as the enhancement value for that voxel. Simultaneously, the scale parameter at which the maximum response is obtained is recorded as a reference for the estimated vessel diameter of that voxel. The enhanced image is then linearly normalized to 0-1, and an enhanced vessel probability map is output. The voxel with the highest probability value is selected from the seed point set as the starting point. An empty queue is established, and this point is enqueued and marked as visited. Within a circular neighborhood with a radius of 3 mm in the normal plane, all pixels with a vessel probability > 0.5 are extracted, and an ellipse is fitted using the least squares method. If the ratio of the major and minor axes of the ellipse is ≤ 1.33 (corresponding to a circularity > 0.75), the equivalent circle diameter is taken as the vessel diameter at that location; if the ratio is > 1.33 and ≤ 1.54 (corresponding to a circularity of 0.65-0.75), it is still accepted as a vein; if the ratio is > 1.54, it is judged as a non-vascular structure, and the branch tracking is terminated. The tracking of the current seed point is terminated when the queue is empty, or the vessel diameter at the current point is < 1.0 mm, or no new points are added for 5 consecutive iterations, or the tracking reaches the preset mask boundary of the hilum. The hilum is used as the root node, and the traversal is performed outward along the centerline. At the branch point, the rate of decrease in the diameter of the child vessel relative to the parent vessel is calculated. If the descent rate is ≥30% and the progeny vessel clearly accompanies a known subsegmental bronchus, it is assigned to level 2 (subsegmental level); if the descent rate is <30% or there is no clear accompanies the progeny vessel, it is assigned to a continuation branch of level 1 (segmental level).Similarly, level 3 corresponds to subsegments, and level 4 and below are collectively referred to as terminal branches. Tracing is performed retrogradely along the central line, with the main portal vein as the converging endpoint. Levels are assigned based on the distance across lung segments: those spanning ≥3 lung segments are marked as main branch veins (level 1), those spanning 2 lung segments are intersegmental veins (level 2), and those located only within a single lung segment are intrasegmental veins (level 3). During central line tracing, a three-dimensional tangential vector is stored for each point, taking the unit vector pointing from that point to its next neighboring point in the tracing path. For endpoints or cases without subsequent points, the opposite direction of the previous unit vector pointing to the current point is used. Using lung segments as units, the vessel with the thickest diameter extending to the hilum within that segment is selected as the main root node. Branch points are treated as internal nodes, and terminal points as leaf nodes, connected to form a directed acyclic graph. The storage structure includes node ID, three-dimensional coordinates, diameter, tangential vector, parent node ID, list of child node IDs, branch level, and anatomical name. All venous segments are identified, and their spatial proximity and directional consistency are calculated. If two venous segments are less than 3 mm apart near the hilum and their directional angle is less than 45°, they are considered to be confluenced. A tributary-confluence trunk connection is established, allowing a confluence trunk to receive multiple tributaries and allowing a vein to successively merge with multiple veins to form a network structure. The arterial topology map and the venous topology map are stored in the same data structure, distinguished by a vessel type field.

[0066] In one alternative approach, the lung segment semantic segmentation module includes:

[0067] A 3D convolutional encoder-decoder network outputs an initial lung segment probability map. The encoder network consists of 5 downsampling stages, each containing 2 3D convolutional layers and 1 3D max-pooling layer, used to progressively compress spatial resolution and expand the number of channels to 512 dimensions to extract cross-scale lung parenchyma texture features and boundary blurring features. The decoder network consists of 5 upsampling stages, each containing 1 3D transposed convolutional layer and 2 3D convolutional layers, used to restore spatial resolution and fuse multi-level semantic information.

[0068] The bronchial tree anatomy prior injection module is connected to the bronchial tree skeleton and its corresponding segmental bronchial labels, respectively, and receives the intermediate layer feature map output by the fourth layer encoder; the bronchial tree skeleton is encoded using a graph convolutional network, which contains three graph convolutional layers. Each layer uses Chebyshev polynomial approximation to perform spectral domain convolution, mapping the spatial coordinates of bronchial branch points and bronchial diameter features into a 128-dimensional bronchial spatial embedding vector;

[0069] The lung fissure enhancement detection subnetwork includes three sets of parallel branches, corresponding to 2D convolution kernels in the coronal, sagittal, and transverse directions, respectively. Each set of branches contains three convolutional layers, which are used to extract the linear high-density response of the lung fissure on the corresponding plane. The outputs of each branch are rearranged in three dimensions and then stitched along the channel dimension to generate a lung fissure probability map.

[0070] A conditional random field post-processing optimization network is used to output a 3D lung segment semantic label map with spatial resolution aligned with the original CT. The lung segment boundary confidence map is used as a univariate potential term, and a binary potential term is constructed using voxel spatial proximity and CT value similarity. The mean field approximation algorithm is used for 10 iterations of message passing to perform probabilistic inference across the entire map, forcing the lung segment boundary to coincide with the high-confidence lung fissure location and maintaining label consistency within the lung segment.

[0071] The segment-level anatomical compliance verification module is used to receive the three-dimensional lung segment semantic label map and compare it with the preset lung segment anatomical knowledge base. It transmits the coordinates of abnormal areas back to increase the weight of the binary potential term of the abnormal area and re-executes the mean field inference until the final lung segment boundary that conforms to the anatomical rules is output.

[0072] In this embodiment, the spatial coordinates of bronchial branch points and their diameter features are mapped to a 128-dimensional spatial embedding vector. This vector is then concatenated with the intermediate layer feature map output from the fourth layer encoder along the channel dimension. This allows the network to understand anatomical rules such as the B3 bronchus being located within the S3 segment and the B4 bronchus being located within the S4 segment. In incomplete lung fissure regions, the segment boundaries are rationally divided along the bronchiovascular bundle pathway. Considering the curved surface of the lung fissure in three-dimensional space and its linear shape in a single axial section, three sets of parallel 2D convolutional branches are used to process the coronal, sagittal, and transverse planes respectively. Each branch specifically learns the linear high-density features of the lung fissure within its own plane. The outputs are then concatenated and fused after three-dimensional rearrangement, providing high-confidence boundary constraints for the conditional random field. The encoder employs a 5-stage downsampling process, progressively increasing the number of channels to 512 dimensions. This ensures the network simultaneously captures the global geometric orientation of the lung fissures (receptive field > 128 mm) and subtle intra-segment texture differences (receptive field < 8 mm). The decoder fuses multi-level features through skip connections, preventing deep semantic features from overshadowing shallow boundary details. The bronchial tree is not a planar structure, and conventional convolutions struggle to model its three-dimensional branch topology. Chebyshev graph convolution is used to construct bronchial branch points as graph nodes and connections as edges. Convolution operations are performed in the spectral domain to explicitly learn the global spatial configuration of the bronchial tree, effectively suppressing abnormal segmental boundary drift in pathological conditions such as atelectasis and pleural effusion.

[0073] Step S102: Mark the puncture target point in the three-dimensional lung anatomy model of the target lesion, and construct a dynamic puncture risk field voxel map based on the structured lung segmental vascular atlas; wherein, taking the center line of each blood vessel as the source, the risk attenuation radius is dynamically set according to its diameter, and a Gaussian kernel function is used to generate a continuously differentiable risk potential energy distribution to form a three-dimensional puncture risk field voxel map covering the entire lung parenchyma.

[0074] In this embodiment, a continuously differentiable risk potential energy distribution is generated using a Gaussian kernel function. This transforms the discrete vessel centerline into a risk field with continuous derivatives at any spatial location. This makes risk calculation in path planning no longer a simple collision detection, but rather capable of sensing gradual risk at different distances from the vessel. The risk attenuation radius is dynamically set based on the actual diameter of each vessel. Larger vessels have a wider risk impact range, while the risk field of smaller capillaries is relatively concentrated. The risk of massive bleeding or air embolism due to needle damage to larger vessels is much higher than that of damage to smaller vessels. Because the risk field is continuous and differentiable, the cumulative risk value is sensitive to and quantifiable to minor path deformations. This allows for a precise quantitative trade-off between path length (shortening the puncture path) and risk avoidance (bypassing large vessels), avoiding path jitter or local optimum traps caused by discrete risk assessment, thus generating a smoother and safer puncture path.

[0075] In one alternative approach, constructing a dynamic puncture risk field voxel map based on the structured lung segmental vascular atlas further includes:

[0076] The specific location, branching level and spatial orientation of each feeding artery and draining vein are obtained from the structured lung segmental vascular atlas, resulting in a refined set of vascular centerlines;

[0077] For each refined vessel centerline in the refined vessel centerline set, a dynamic risk attenuation radius is set according to the vessel diameter to obtain a weighted risk attenuation radius set.

[0078] The risk decay radius set is generated into a continuously differentiable risk potential energy distribution by using a Gaussian kernel function, forming a dynamic puncture risk field voxel map covering the entire lung parenchyma.

[0079] In this embodiment, by superimposing the Gaussian kernel contributions from all vascular sources, a dynamic puncture risk field voxel map covering every voxel in the entire lung is ultimately generated. Essentially, this transforms the complex vascular network constraints into an intuitive "risk terrain," with low-risk areas corresponding to valleys (sparse vascular zones) and high-risk areas corresponding to peaks (concentrated zones of large vessels). This allows physicians to assess puncture safety from a global perspective, avoiding the possibility of missing optimal pathways from a localized viewpoint. The structured vascular atlas does not rely on any external atlas libraries or statistical templates. Once the dynamic attenuation radius is set, it can be uniformly applied to all patients, but the actual attenuation range of each vessel is adaptively determined by its own true diameter, ensuring a stable output of an individualized risk field in populations with different body types and varying degrees of vascular variability, demonstrating good clinical generalization ability.

[0080] Step S103: Sample multiple candidate needle insertion points in the feasible needle insertion area on the body surface, generate an initial path set in the three-dimensional lung anatomy model using the fast travel method, map each path in the initial path set to the three-dimensional puncture risk field voxel map to obtain a cumulative risk value, and construct a multi-objective optimization function based on the cumulative risk value, lung tissue density weighting factor and puncture angle penalty term.

[0081] In this embodiment, the Fast Marching Method is used to generate an initial path set in the three-dimensional lung anatomy model, which can efficiently solve for the shortest path between candidate needle insertion points on the body surface and puncture target points inside the body. For example... Figure 2A As shown, compared to traditional random sampling or grid search methods, this approach ensures the generation of a diverse population of geometrically continuous, physically feasible initial paths covering different spatial orientations within complex lung parenchyma. By mapping the spatial coordinate sequence of each initial path to a constructed 3D puncture risk field voxel map, the accumulated vascular injury risk value along the path can be accurately calculated, achieving a shift from qualitative judgment (whether the path touches a blood vessel) to quantitative assessment (what is the accumulated risk value of this path). The accumulated risk value not only reflects whether a blood vessel has been touched but also the proximity of the path to blood vessels of different diameters and the length of the path traversing high-risk areas. Differences in CT values ​​in different regions (such as the near-pleural region, near-hilar region, emphysema region, and fibrosis region) reflect differences in tissue density, elastic modulus, and bleeding tendency. A density-weighted optimization function can actively guide the path to avoid high-density sclerotic or inflammatory consolidation areas, reducing puncture resistance, needle deviation risk, and pneumothorax incidence.

[0082] Step S104: Based on the multi-objective optimization function, a set of non-dominated optimal paths is output as the final puncture path by weighing the path length, cumulative risk value, puncture angle penalty, and the volume of high-density lung tissue traversed. The final puncture path includes a spatial coordinate sequence, real-time risk thermal projection, and avoidance suggestions.

[0083] In this embodiment, as Figure 2B As shown, the cumulative risk value mitigates the risk of damage to feeding arteries or draining veins during puncture, exhibiting high sensitivity, especially for small-diameter but functionally critical segmental vessels. The puncture angle penalty term constrains the needle insertion angle, avoiding operational difficulties or pleural injury caused by excessively steep or flat approaches, while limiting the volume of high-density lung tissue (such as fibrotic or consolidated areas) traversed, reducing puncture resistance and the probability of complications. The non-dominated optimal path set not only includes a spatial coordinate sequence but also includes real-time risk thermal projection and structured avoidance suggestions, providing operators with intuitive and interpretable multi-option comparison criteria. For example, a patient has a 15mm ground-glass nodule in the posterior segment (S6) of the right lower lobe, requiring a puncture biopsy. After path planning, three optimal non-dominated paths are output: Path A: Shortest (85mm), but requires traversing a small section of high-density fibrotic area (volume 12mm³), and its end is close to a segmental vein (medium cumulative risk); Path B: Slightly longer (92mm), completely avoiding the high-density area, lowest risk, but the needle insertion angle is relatively flat (58°), close to the upper limit of the angle; Path C: Moderate length (88mm), low risk, ideal angle (45°), but requires slight detours, traversing a small amount of normal lung tissue. The choice can be made based on the patient's specific situation. If the patient has poor lung function and low tolerance, Path B should be prioritized for safety; if efficiency is prioritized and the patient is young, Path A can be selected; if multiple factors need to be considered, Path C should be selected. Each path is accompanied by a heatmap display and text suggestions such as "It is recommended to avoid the venous branches between segments S6 and S10," improving the accuracy of preoperative planning and the ability to adapt during surgery.

[0084] In one alternative approach, a set of non-dominated optimal paths is output as the final puncture path based on a multi-objective optimization function that balances path length, cumulative risk value, puncture angle penalty, and volume of high-density lung tissue traversed, further including:

[0085] The initial path set is divided into Pareto levels according to the fast non-dominated sorting algorithm. Paths that are not dominated by any other path in all four objective dimensions are marked as the first frontier and form a seed set of candidate non-dominated paths.

[0086] Using the candidate non-dominated path seed set as the parent population, the spatial control point sequences of the two parent paths are interpolated and recombined using a simulated binary crossover operator to generate new offspring paths that simultaneously inherit the local geometric features of both parents.

[0087] A polynomial mutation operator is used to apply random perturbation to the path control point in the low potential energy region of the three-dimensional puncture risk field voxel map, driving the path to adaptively drift towards the sparse vascular region, ensuring that the offspring path is still a continuous and passable physical puncture path in the lung parenchyma and is remapped to the three-dimensional puncture risk field voxel map to update its multi-target feature vector.

[0088] The multi-objective optimization function is iteratively executed until a preset population generation threshold is reached. All individuals on the first frontier in the last generation population are output as the Pareto optimal solution set. Based on the three-dimensional spatial coordinate sequence of each path in the Pareto optimal solution set, the corresponding cumulative risk value thermal profile, and the avoidance suggestion text, a non-dominated optimal path is constructed as the final puncture path.

[0089] In this embodiment, the path set is divided into Pareto levels to avoid the risk of optimization deviating from actual clinical needs due to improper weight allocation. The first frontier path set ensures that path length, vascular risk, puncture angle, and tissue density have reached an equilibrium state where further unidirectional improvement is impossible without worsening other objectives. By simulating a binary crossover operator, the spatial control point sequences of the two parent paths are interpolated and recombined to achieve sexual reproduction of path morphological features. The offspring path can simultaneously inherit a certain segment of the bypass arc of parent A and the needle insertion angle preference of parent B, avoiding getting trapped in local optima. A polynomial mutation operator is used to apply random perturbation to the path control points in the low-potential region of the three-dimensional puncture risk field voxel map, making the evolutionary process both exploratory (random perturbation) and exploitative (low-potential guidance). While maintaining the excellent framework of the parent path, the offspring path actively and adaptively migrates to a safer anatomical space. After each crossover and mutation operation, a path validity check is enforced to ensure that the offspring path remains a continuous and passable physical puncture path within the lung parenchyma, preventing infeasible solutions that might arise from simple numerical optimization, such as crossing large blood vessels, breaching lung boundaries, or exhibiting non-physiological sharp bends or reversals. The multi-objective optimization engine is iteratively executed until a preset population generation threshold is reached. Each newly generated offspring path is remapped to a 3D puncture risk field voxel map to update its multi-objective feature vector, ensuring that every selection, crossover, and mutation during the evolutionary process is a precise feedback based on the current optimal risk field information. As the population progresses through generations, the first frontier continuously approaches the true Pareto front. Figure 3B As shown, the path enters the lesion from the lateral side of the rib at the entry point Pin → passes through key fulcrums S1 and S2 → terminates at the point Pout. The cumulative risk value is 0.27, and avoidance suggestions include "avoiding radial vessels ≥8mm", "keeping the puncture angle ≤20°", and "avoiding entry into high-density lung tissue with a volume >1.8cm³".

[0090] In an alternative approach, the cumulative risk value is obtained based on a weighted Gaussian potential field integral model of the stress concentration effect on the blood vessel wall and the probability of rupture, and its calculation formula is as follows:

[0091]

[0092] in, , The overall rupture tendency of the i-th blood vessel; The critical stress threshold corresponding to a 50% probability of fracture; The risk decay radius scaling factor recommended by the Fleischner Society; , The velocity component of the puncture needle perpendicular to the local normal of the blood vessel at path point s; Candidate puncture paths Total arc length; The total number of vascular centerlines included in the structured lung segmental vascular atlas; For path In arc length parameter The three-dimensional spatial coordinates at [0, L(p)]; Let x(s) be the projection point on the center line of the i-th blood vessel that is closest to the path point x(s). Let be the equivalent diameter of the i-th blood vessel; Let be the volumetric flow rate of the i-th blood vessel; This refers to pulmonary artery systolic blood pressure. Let be the wall thickness of the i-th blood vessel; For the nearest blood vessel point corresponding to path point x(s) The local normal vector of the normal plane at the centerline of the blood vessel.

[0093] In this embodiment, the velocity component of the puncture needle perpendicular to the local normal of the blood vessel introduces operational dynamics into static path planning, accurately characterizing the impact effect of the needle tip on the blood vessel wall at the moment of puncture (when the needle tip moves directly towards the blood vessel wall, even at the same distance, the risk of vascular tearing caused by the instantaneous impact force is much higher than that of tangential glancing), enabling the cumulative risk value to distinguish between two distinct clinical scenarios: "slow approach" and "rapid impact." The risk attenuation radius scaling factor directly anchors the construction of the risk field to authoritative clinical guidelines, establishing a mathematical mapping between the blood vessel diameter observed in imaging and the clinically recommended safe buffer distance, ensuring that the risk value output by the model is highly consistent with clinical consensus, thus enhancing the clinical credibility and acceptability of the algorithm. The contribution of the blood vessel along the entire candidate path is integrated and accumulated, ensuring that the path risk assessment comprehensively considers the superposition effect of all vascular networks encountered along the way. For example, although a path is far from a large artery, it continuously crosses multiple small venous plexuses, and its integral value can still accurately reflect the cumulative damage risk.

[0094] In one alternative approach, the puncture angle penalty term is calculated based on a needle tract offset error propagation model under chest wall curvature constraints, and its calculation formula is as follows:

[0095]

[0096] in, For the angle of needle insertion; The average curvature at the needle insertion point on the body surface; The coefficient for sensitivity of curvature and inter-rib spacing; The pneumothorax index; This is the normalization coefficient.

[0097] In this embodiment, the pneumothorax index penalty term quantifies the risk of excessive needle path length leading to pleural or lung tissue perforation when the needle insertion angle is too small (approaching parallel to the pleura). A higher pneumothorax index results in a more severe penalty for excessively flat puncture angles, thus guiding path planning to avoid high-risk, small-angle needle insertions. This transforms implicit clinical experience ("don't insert the needle close to the bone," "inserting the needle too flat easily perforates the pleura") into a mathematical model. By adjusting the pneumothorax index and the curvature and intercostal space sensitivity coefficient, doctors can adjust the penalty term based on the patient's specific situation (e.g., patients with emphysema have a high risk of pneumothorax, so the penalty term is increased). Value; osteoporosis patients have brittle ribs, thus increasing The penalty intensity is dynamically adjusted based on the value.

[0098] In one alternative approach, the formula for calculating the volume traversed through high-density lung tissue is:

[0099]

[0100] in, For local Young's modulus, ; To calibrate the compressibility coefficient of isolated pig lungs; Here is the Hounsfield unit for the CT image at spatial location x; As a benchmark for healthy lung parenchyma modulus; This is an empirical coefficient; As a continuous indicator of the degree of fibrosis, ; The additional resistance coefficient for fibrosis; Let s be the three-dimensional coordinates corresponding to the arc length parameter s on path p; Candidate puncture paths The total arc length.

[0101] In this embodiment, Young's modulus converts the CT value (HU) into a stiffness index of the tissue. The puncture path will actively avoid hard tissues (such as fibrous lesions and calcified lymph nodes) because the puncture needle encounters great resistance, is prone to deviation, and increases patient pain when passing through hard tissues.

[0102] According to the solution provided by the present invention, preoperative chest CT image data of the patient is acquired, and isotropic resampling, lobe segmentation, and bronchial tree extraction are performed on the CT image data to generate a three-dimensional lung anatomy model; the three-dimensional lung anatomy model is input into a lung segment semantic segmentation module to output lung segment boundaries; the topological connection relationship between the lung segment-level feeding arteries and draining veins is reconstructed from the lung segment boundaries according to a threshold centerline tracing algorithm to obtain a structured lung segment-level vascular atlas containing information on vessel diameter, branch level, and spatial orientation; puncture target points are marked in the three-dimensional lung anatomy model of the target lesion, and a dynamic puncture risk field voxel map is constructed based on the structured lung segment-level vascular atlas; wherein, the risk attenuation radius is dynamically set according to the diameter of each vessel centerline as the source. A continuously differentiable risk potential energy distribution is generated using a Gaussian kernel function to form a three-dimensional puncture risk field voxel map covering the entire lung parenchyma. Multiple candidate needle insertion points are sampled in the feasible needle insertion area on the body surface. An initial path set is generated in the three-dimensional lung anatomy model using a fast-moving method. Each path in the initial path set is mapped to the three-dimensional puncture risk field voxel map to obtain a cumulative risk value. A multi-objective optimization function is constructed based on the cumulative risk value, lung tissue density weighting factor, and puncture angle penalty term. The multi-objective optimization function balances path length, cumulative risk value, puncture angle penalty, and the volume of high-density lung tissue traversed, outputting a set of non-dominated optimal paths as the final puncture path. The final puncture path includes a spatial coordinate sequence, real-time risk thermal projection, and avoidance suggestions. This invention significantly improves the safety boundaries and clinical applicability of percutaneous lung puncture.

[0103] Figure 4 A schematic diagram of the framework of a percutaneous lung puncture path planning device based on lung segmental vascular structure constraints according to an embodiment of the present invention is shown. The percutaneous lung puncture path planning device based on lung segmental vascular structure constraints includes:

[0104] The lung segmental vascular atlas construction module 410 is used to acquire the patient's preoperative chest CT image data and perform isotropic resampling, lung lobe segmentation, and bronchial tree extraction on the CT image data to generate a three-dimensional lung anatomical model; input the three-dimensional lung anatomical model into the lung segment semantic segmentation module to output the lung segment boundary; and reconstruct the topological connection relationship between the lung segmental feeding arteries and draining veins from the lung segment boundary according to the threshold centerline tracing algorithm to obtain a structured lung segmental vascular atlas containing information on vessel diameter, branch level, and spatial orientation.

[0105] The dynamic puncture risk field generation module 420 is used to mark the puncture target point in the three-dimensional lung anatomy model of the target lesion and construct a dynamic puncture risk field voxel map based on the structured lung segmental vascular atlas. Specifically, the risk attenuation radius is dynamically set according to the diameter of each blood vessel centerline as the source, and a Gaussian kernel function is used to generate a continuously differentiable risk potential energy distribution to form a three-dimensional puncture risk field voxel map covering the entire lung parenchyma.

[0106] The path set generation and risk mapping module 430 is used to sample multiple candidate needle insertion points in the feasible needle insertion area on the body surface, generate an initial path set in the three-dimensional lung anatomy model using the fast travel method, map each path in the initial path set to the three-dimensional puncture risk field voxel map to obtain a cumulative risk value, and construct a multi-objective optimization function based on the cumulative risk value, lung tissue density weighting factor and puncture angle penalty term.

[0107] The non-dominated path generation module 440 is used to weigh path length, cumulative risk value, puncture angle penalty and volume of high-density lung tissue traversed according to a multi-objective optimization function, and output a set of non-dominated optimal paths as the final puncture path; wherein, the final puncture path includes a spatial coordinate sequence, real-time risk thermal projection and avoidance suggestions.

[0108] Figure 5 The diagram shows a structural schematic of an embodiment of the computer device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computer device.

[0109] like Figure 5 As shown, the computer device may include: a processor 502, a communications interface 504, a memory 506, and a communications bus 508.

[0110] The processor 502, communication interface 504, and memory 506 communicate with each other via communication bus 508. Communication interface 504 is used to communicate with other network elements such as clients or other servers. Processor 502 executes program 510, specifically performing the relevant steps in the above-described embodiment of the percutaneous lung puncture path planning method based on lung segmental vascular structure constraints.

[0111] Specifically, program 510 may include program code that includes computer operation instructions.

[0112] Processor 502 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0113] Memory 506 is used to store program 510. Memory 506 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0114] According to the solution provided by the present invention, preoperative chest CT image data of the patient is acquired, and isotropic resampling, lobe segmentation, and bronchial tree extraction are performed on the CT image data to generate a three-dimensional lung anatomy model; the three-dimensional lung anatomy model is input into a lung segment semantic segmentation module to output lung segment boundaries; the topological connection relationship between the lung segment-level feeding arteries and draining veins is reconstructed from the lung segment boundaries according to a threshold centerline tracing algorithm to obtain a structured lung segment-level vascular atlas containing information on vessel diameter, branch level, and spatial orientation; puncture target points are marked in the three-dimensional lung anatomy model of the target lesion, and a dynamic puncture risk field voxel map is constructed based on the structured lung segment-level vascular atlas; wherein, the risk attenuation radius is dynamically set according to the diameter of each vessel centerline as the source. A continuously differentiable risk potential energy distribution is generated using a Gaussian kernel function to form a three-dimensional puncture risk field voxel map covering the entire lung parenchyma. Multiple candidate needle insertion points are sampled in the feasible needle insertion area on the body surface. An initial path set is generated in the three-dimensional lung anatomy model using a fast-moving method. Each path in the initial path set is mapped to the three-dimensional puncture risk field voxel map to obtain a cumulative risk value. A multi-objective optimization function is constructed based on the cumulative risk value, lung tissue density weighting factor, and puncture angle penalty term. The multi-objective optimization function balances path length, cumulative risk value, puncture angle penalty, and the volume of high-density lung tissue traversed, outputting a set of non-dominated optimal paths as the final puncture path. The final puncture path includes a spatial coordinate sequence, real-time risk thermal projection, and avoidance suggestions. This invention significantly improves the safety boundaries and clinical applicability of percutaneous lung puncture.

[0115] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination of all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed can be employed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose. Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

Claims

1. A percutaneous lung puncture path planning method based on lung segmental vascular structure constraints, characterized in that, include: Preoperative chest CT images of patients are acquired and isotropically resampled, segmented into lung lobes, and extracted into bronchial trees to generate a three-dimensional lung anatomy model. The three-dimensional lung anatomy model is then input into a lung segment semantic segmentation module, which outputs lung segment boundaries. Based on a threshold centerline tracing algorithm, the topological connections between the supplying arteries and draining veins at the lung segment boundaries are reconstructed to obtain a structured lung segment vascular atlas containing information on vessel diameter, branch hierarchy, and spatial orientation. The puncture target point is marked in the three-dimensional lung anatomy model of the target lesion, and a dynamic puncture risk field voxel map is constructed based on the structured lung segmental vascular atlas. Among them, the risk attenuation radius is dynamically set according to the diameter of each blood vessel centerline as the source, and a Gaussian kernel function is used to generate a continuously differentiable risk potential energy distribution to form a three-dimensional puncture risk field voxel map covering the entire lung parenchyma. Multiple candidate needle insertion points are sampled in the feasible needle insertion area on the body surface. An initial path set is generated in the three-dimensional lung anatomy model using the fast travel method. Each path in the initial path set is mapped to the three-dimensional puncture risk field voxel map to obtain the cumulative risk value. A multi-objective optimization function is constructed based on the cumulative risk value, the lung tissue density weighting factor, and the puncture angle penalty term. The multi-objective optimization function weighs path length, cumulative risk value, puncture angle penalty, and volume of high-density lung tissue traversed, and outputs a set of non-dominated optimal paths as the final puncture path; wherein, the final puncture path includes a spatial coordinate sequence, real-time risk thermal projection, and avoidance suggestions.

2. The percutaneous lung puncture path planning method based on lung segmental vascular structure constraints according to claim 1, characterized in that, The topological connectivity between the pulmonary segmental feeding arteries and draining veins is reconstructed from the pulmonary segment boundary using a threshold centerline tracing algorithm, resulting in a structured pulmonary segmental vascular atlas containing information on vessel diameter, branch hierarchy, and spatial orientation. This atlas further includes: The lung segment boundaries are spatially registered with the bronchial tree skeleton extracted from the bronchial tree, and the anatomical labels of each bronchial segment are used as prior constraints to serve as the tracking starting regions for the accompanying arteries within the lung segment and the intersegmental veins, respectively. Specifically, for the accompanying arteries within the lung segment, the lateral bronchial wall neighborhood is used as the seed point set; for the draining veins, the adjacent area of ​​the lung segmental septum and the subpleural reflux area are used as the seed point set. An adaptive morphological top-hat transformation method is used to enhance the low-contrast small vessel signal in the three-dimensional lung anatomy model, resulting in an enhanced vessel probability map. A threshold centerline tracing algorithm is then executed on this enhanced vessel probability map, starting from the seed point set. This algorithm iteratively searches for regions in adjacent voxels that satisfy the circularity and continuity constraints of the vessel cross-section, reconstructing the vessel centerline path pixel by pixel. During this pixel-by-pixel reconstruction, the equivalent circle diameter is calculated as the vessel diameter based on the set of cross-sectional pixels perpendicular to the local direction within the centerline's neighborhood. Branch levels are automatically assigned based on this vessel diameter, its relationship to the corresponding bronchus, its cross-segmental course, and its order of entry into the portal vein. The three-dimensional tangential vector of each point on the centerline is accumulated and recorded as the spatial direction of the vessel, forming the vessel's vector growth trajectory. The central lines of all accompanying arteries within the same lung segment are connected hierarchically to form an arterial tree topology, and the draining veins spanning multiple lung segments are connected according to their confluence relationship to form a venous network topology. The arterial tree topology and the venous network topology are then combined to generate a structured lung segmental vascular atlas covering the entire lung.

3. The percutaneous lung puncture path planning method based on lung segmental vascular structure constraints according to claim 1, characterized in that, The construction of a dynamic puncture risk field voxel map based on the structured lung segmental vascular atlas further includes: The specific location, branching level and spatial orientation of each feeding artery and draining vein are obtained from the structured lung segmental vascular atlas, resulting in a refined set of vascular centerlines; For each refined vessel centerline in the refined vessel centerline set, a dynamic risk attenuation radius is set according to the vessel diameter to obtain a weighted risk attenuation radius set. The risk decay radius set is generated by using a Gaussian kernel function to produce a continuously differentiable risk potential energy distribution, forming a dynamic puncture risk field voxel map covering the entire lung parenchyma.

4. The percutaneous lung puncture path planning method based on lung segmental vascular structure constraints according to claim 1, characterized in that, The lung segment semantic segmentation module includes: A 3D convolutional encoder-decoder network outputs an initial lung segment probability map. The encoder network consists of 5 downsampling stages, each containing 2 3D convolutional layers and 1 3D max-pooling layer, used to progressively compress spatial resolution and expand the number of channels to 512 dimensions to extract cross-scale lung parenchyma texture features and boundary blurring features. The decoder network consists of 5 upsampling stages, each containing 1 3D transposed convolutional layer and 2 3D convolutional layers, used to restore spatial resolution and fuse multi-level semantic information. The bronchial tree anatomy prior injection module is connected to the bronchial tree skeleton and its corresponding segmental bronchial labels, respectively, and receives the intermediate layer feature map output by the fourth layer encoder; the bronchial tree skeleton is encoded using a graph convolutional network, which contains three graph convolutional layers. Each layer uses Chebyshev polynomial approximation to perform spectral domain convolution, mapping the spatial coordinates of bronchial branch points and bronchial diameter features into a 128-dimensional bronchial spatial embedding vector; The lung fissure enhancement detection subnetwork includes three sets of parallel branches, corresponding to 2D convolution kernels in the coronal, sagittal, and transverse directions, respectively. Each set of branches contains three convolutional layers, which are used to extract the linear high-density response of the lung fissure on the corresponding plane. The outputs of each branch are rearranged in three dimensions and then stitched along the channel dimension to generate a lung fissure probability map. A conditional random field post-processing optimization network is used to output a 3D lung segment semantic label map with spatial resolution aligned with the original CT. The lung segment boundary confidence map is used as a univariate potential term, and a binary potential term is constructed using voxel spatial proximity and CT value similarity. The mean field approximation algorithm is used for 10 iterations of message passing to perform probabilistic inference across the entire map, forcing the lung segment boundary to coincide with the high-confidence lung fissure location and maintaining label consistency within the lung segment. The segment-level anatomical compliance verification module is used to receive the three-dimensional lung segment semantic label map and compare it with the preset lung segment anatomical knowledge base. It transmits the coordinates of abnormal areas back to increase the weight of the binary potential term of the abnormal area and re-executes the mean field inference until the final lung segment boundary that conforms to the anatomical rules is output.

5. The percutaneous lung puncture path planning method based on lung segmental vascular structure constraints according to claim 1, characterized in that, Based on a multi-objective optimization function, a set of non-dominated optimal paths is output as the final puncture path, balancing path length, cumulative risk value, puncture angle penalty, and the volume of high-density lung tissue traversed. Further components include: The initial path set is divided into Pareto levels according to the fast non-dominated sorting algorithm. Paths that are not dominated by any other path in all four objective dimensions are marked as the first frontier and form a seed set of candidate non-dominated paths. Using the candidate non-dominated path seed set as the parent population, the spatial control point sequences of the two parent paths are interpolated and recombined using a simulated binary crossover operator to generate new offspring paths that simultaneously inherit the local geometric features of both parents. A polynomial mutation operator is used to apply random perturbation to the path control point in the low potential energy region of the three-dimensional puncture risk field voxel map, driving the path to adaptively drift towards the sparse vascular region, ensuring that the offspring path is still a continuous and passable physical puncture path in the lung parenchyma and is remapped to the three-dimensional puncture risk field voxel map to update its multi-target feature vector. The multi-objective optimization function is iteratively executed until a preset population generation threshold is reached. All individuals on the first frontier in the last generation population are output as the Pareto optimal solution set. Based on the three-dimensional spatial coordinate sequence of each path in the Pareto optimal solution set, the corresponding cumulative risk value thermal profile, and the avoidance suggestion text, a non-dominated optimal path is constructed as the final puncture path.

6. The percutaneous lung puncture path planning method based on lung segmental vascular structure constraints according to claim 1, characterized in that, The cumulative risk value is obtained based on a weighted Gaussian potential field integral model of the stress concentration effect on the blood vessel wall and the probability of rupture, and its calculation formula is as follows: ； in, , The overall rupture tendency of the i-th blood vessel; The critical stress threshold corresponding to a 50% probability of fracture; The risk decay radius scaling factor recommended by the Fleischner Society; , The velocity component of the puncture needle perpendicular to the local normal of the blood vessel at path point s; Candidate puncture paths Total arc length; The total number of vascular centerlines included in the structured lung segmental vascular atlas; For path In arc length parameter The three-dimensional spatial coordinates at [0, L(p)]; Let x(s) be the projection point on the center line of the i-th blood vessel that is closest to the path point x(s). Let be the equivalent diameter of the i-th blood vessel; Let be the volumetric flow rate of the i-th blood vessel; This refers to pulmonary artery systolic blood pressure. Let be the wall thickness of the i-th blood vessel; For the nearest blood vessel point corresponding to path point x(s) The local normal vector of the normal plane at the centerline of the blood vessel.

7. The percutaneous lung puncture path planning method based on lung segmental vascular structure constraints according to claim 1, characterized in that, The puncture angle penalty term is calculated based on the needle tract offset error propagation model under chest wall curvature constraints, and its calculation formula is as follows: ； in, For the needle angle; The average curvature at the needle insertion point on the body surface; The coefficient of sensitivity between curvature and rib spacing; The pneumothorax index; This is the normalization coefficient.

8. The percutaneous lung puncture path planning method based on lung segmental vascular structure constraints according to any one of claims 1-7, characterized in that, The formula for calculating the volume of high-density lung tissue traversed is: ； in, For local Young's modulus, ; To calibrate the compressibility coefficient of isolated pig lungs; Here are Hounsfield units for the CT image at spatial location x; As a benchmark for healthy lung parenchyma modulus; This is an empirical coefficient; As a continuous indicator of the degree of fibrosis, ; The additional resistance coefficient for fibrosis; Let s be the three-dimensional coordinates corresponding to the arc length parameter s on path p; Candidate puncture paths The total arc length.

9. A percutaneous lung puncture path planning device based on lung segmental vascular structure constraints, characterized in that, include: A lung segmental vascular atlas construction module is used to acquire preoperative chest CT image data of patients and perform isotropic resampling, lobe segmentation and bronchial tree extraction on the CT image data to generate a three-dimensional lung anatomy model; the three-dimensional lung anatomy model is input into the lung segment semantic segmentation module to output the lung segment boundaries; the topological connection relationship between the lung segment feeding arteries and draining veins is reconstructed from the lung segment boundaries according to the threshold centerline tracing algorithm to obtain a structured lung segmental vascular atlas containing information on vessel diameter, branch level and spatial orientation. The dynamic puncture risk field generation module is used to mark the puncture target point in the three-dimensional lung anatomy model of the target lesion and construct a dynamic puncture risk field voxel map based on the structured lung segmental vascular atlas. Specifically, the risk attenuation radius is dynamically set according to the diameter of each blood vessel centerline as the source, and a Gaussian kernel function is used to generate a continuously differentiable risk potential energy distribution to form a three-dimensional puncture risk field voxel map covering the entire lung parenchyma. The path set generation and risk mapping module is used to sample multiple candidate needle insertion points in the feasible needle insertion area on the body surface, generate an initial path set in the three-dimensional lung anatomy model using the fast travel method, map each path in the initial path set to the three-dimensional puncture risk field voxel map to obtain a cumulative risk value, and construct a multi-objective optimization function based on the cumulative risk value, lung tissue density weighting factor and puncture angle penalty term. The non-dominated path generation module is used to weigh path length, cumulative risk value, puncture angle penalty and high-density lung tissue volume traversed according to a multi-objective optimization function, and output a set of non-dominated optimal paths as the final puncture path; wherein, the final puncture path includes a spatial coordinate sequence, real-time risk thermal projection and avoidance suggestions.

10. A computer device, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the percutaneous lung puncture path planning method based on lung segmental vascular structure constraints as described in any one of claims 1-8.

Images