Intelligent optimal path planning system and method for complex three-dimensional space

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an optimal path intelligent planning system and method for a complex three-dimensional space, and the method comprises the steps: obtaining the preoperative CT image data of a patient, extracting the spatial relation between a tumor and a blood vessel in a three-dimensional anatomical model through a GNN graph neural network, generating a dynamic adjacency matrix, and carrying out the optimal path intelligent planning through the dynamic adjacency matrix; outputting a plurality of operation candidate paths from the thoracic cavity inlet to the tumor core; an improved NSGA-II algorithm is adopted to optimize the multiple operation candidate paths, and an initial operation path control instruction is obtained; a rapid collision detection model based on an AABB hierarchical bounding box and a GJK algorithm is utilized to calculate the position and the invasion depth of an end effector of a surgical robot in real time to obtain a collision detection result, a target surgical path control instruction is generated according to the collision detection result and net contact force, the surgical path planning time is effectively shortened to the minute level, and the surgical path planning efficiency is improved. And compared with the traditional method, the path safety margin is greatly improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an optimal path intelligent planning system and method for complex three-dimensional spaces. Background Technology

[0002] Current three-dimensional path planning in lung cancer surgery relies on traditional methods, which have significant limitations. Preoperative planning mainly depends on doctors manually annotating anatomical structures in CT images, which is time-consuming and easily influenced by subjective experience, resulting in low path design efficiency. Intraoperative navigation often uses static images and fixed algorithms, making it difficult to dynamically avoid unexpected situations such as intraoperative bleeding and tissue displacement, resulting in a high risk of misoperation. In addition, existing technologies do not fully integrate multiple medical indicators such as tumor boundary integrity, lung function preservation rate, and postoperative recovery, and a single optimization goal may easily overlook the balance between safety and functionality. Traditional algorithms lack adaptive capabilities and cannot adjust the path in real time to cope with complex intraoperative environmental changes, making three-dimensional path planning inefficient. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an intelligent optimal path planning system and method for complex three-dimensional spaces.

[0004] In a first aspect, embodiments of this application provide an intelligent optimal path planning system for complex three-dimensional spaces, the intelligent optimal path planning system comprising the following modules: A module for building a three-dimensional model of the CT image is used to acquire the patient's preoperative CT image data and perform real-time semantic segmentation of the lung structure in the CT image data using the U-Net+Transformer hybrid model to identify lesion areas and blood vessel walls, thereby obtaining a three-dimensional anatomical model. The surgical path generation module is used to extract the spatial relationship between the tumor and blood vessels in the three-dimensional anatomical model using a GNN graph neural network, generate a dynamic adjacency matrix, design an Actor-Critic framework, and define the state space as the current position, tumor boundary distance, and avoidance structure distribution, and the action space as a six-degree-of-freedom surgical instrument operation command, and output multiple surgical candidate paths from the thoracic cavity entrance to the tumor core. The initial path selection module is used to optimize multiple surgical candidate paths using an improved NSGA-II algorithm to obtain initial surgical path control instructions; The net contact force calculation module is used to acquire contact force signals in real time through the torque sensor on the end effector of the surgical robot based on the initial surgical path control command, and to calculate the net contact force between the instrument and the tissue using a signal fusion and noise reduction algorithm based on Kalman filtering. The target path planning module is used to calculate the position and intrusion depth of the surgical robot end effector in real time using a fast collision detection model based on AABB hierarchical bounding boxes and GJK algorithm, obtain collision detection results, and generate target surgical path control commands based on the collision detection results and net contact force.

[0005] In one possible implementation, the 3D model building module includes the following units: The image data acquisition unit is used to acquire the patient's preoperative CT image data. It reads the CT image data through the Pydicom library, converts the data into a 32-bit floating-point image matrix, and obtains the first image data. The image data filtering unit is used to remove Gaussian noise from the first image data using Gaussian filtering, eliminate salt-and-pepper noise by combining median filtering, automatically crop invalid background areas based on the patient's body contour, retain the lung region of interest, and obtain the second image data. The image data processing unit is used to normalize the gray values ​​of the second image data to the [0,1] interval, and use the CLAHE algorithm to enhance the contrast of lung structures to obtain preprocessed image data.

[0006] In one possible implementation, the 3D model building module includes the following units: The feature extraction unit is used to extract local spatial features based on the U-Net architecture adopted by the encoder of the U-Net+Transformer hybrid model. The first 4 layers are convolutional modules, each containing 2 3×3 convolutions + BN + ReLU, with a stride of 2 for downsampling. The feature fusion unit is used to connect the Transformer encoder after the 4th layer of the encoder. It contains 4 attention heads and has a hidden layer dimension of 512. It is used to capture global context information. The decoding end of the hybrid model uses transposed convolution for upsampling and fuses with the corresponding layer features of the encoder. The feature segmentation unit is used to input preprocessed image data into the U-Net+Transformer hybrid model, and output the segmentation result through convolution to obtain the image feature vector; The model generation unit is used to extract lesion regions and blood vessel walls from the segmentation mask of image feature vectors, filter out missegmented small regions, obtain semantic label regions, stack the semantic label regions into a three-dimensional voxel model based on CT spatial coordinate information, and generate a three-dimensional anatomical model of lesions and blood vessels through the MarchingCubes surface reconstruction algorithm.

[0007] In one possible implementation, the surgical path generation module includes the following sub-modules: The graph node definition submodule is used to discretize key structures in a 3D anatomical model into graph nodes, including tumor surface sampling points, vascular branch points and endpoints. Each node is characterized by 3D coordinates, structure type, radius and curvature. The adjacency weight calculation submodule is used to calculate the adjacency weight based on the spatial GCN graph convolutional network, using the Euclidean distance and direction vector between nodes, and to learn the spatial constraints of tumors and blood vessels. The adjacency matrix generation submodule is used to construct an initial adjacency matrix based on the spatial relationships output by the spatial GCN graph convolutional network. When the current position is close to the structure in the path planning, the adjacency matrix is ​​updated to obtain a dynamic adjacency matrix.

[0008] In one possible implementation, the surgical path generation module includes the following sub-modules: The State Definition submodule is used to define the continuous states in the state space of the Actor-Critic framework, including the current position, the minimum distance to the tumor boundary, and the spatial distribution vector of the avoidance structure. The discrete state is whether it is close to the danger zone. The candidate path generation submodule is used to define the action space as a 6DoF six-degree-of-freedom surgical instrument operation command, including translation, range and rotation. The action step size is dynamically adjusted according to real-time accuracy. The Actor outputs the action probability distribution and the Critic evaluates the state value through the PPO algorithm. During inference, based on different initial exploration noise, 10-20 candidate paths from the thoracic cavity entrance to the tumor core are output, resulting in multiple surgical candidate paths.

[0009] In one possible implementation, the initial path filtering module includes the following units: The algorithm improvement unit is used to improve the NSGA-II algorithm by adopting an elite preservation and adaptive weight strategy, giving higher selection probability to paths close to the Pareto optimal front, while increasing the weight of safety targets in the near-tumor region; retaining the safe segments of the two paths during crossover, and randomly fine-tuning the action instructions in the non-dangerous region during mutation, thus obtaining the improved NSGA-II algorithm. The mapping and association unit is used to set the population size to 0 and iterate 30 times. The optimal solution is selected by non-dominated sorting and crowding calculation. The forward and inverse kinematics models of the lung intervention surgery robot are established based on the optimal solution. The joint coordinate system is defined by the DH parameter method. The inverse kinematics is solved by numerical solution to obtain the mapping relationship between joint angles and end effector pose. Path resolution unit, used for employing RRT The algorithm plans the motion path nodes of the end effector from the initial position to the edge of the lesion, converts the motion path nodes into the pose commands of the surgical robot end effector, converts them into the angle commands of each joint through inverse kinematics, and uses a trapezoidal velocity curve to plan the joint motion. The path control unit is used to add joint limit constraints and obstacle avoidance constraints, delete instructions that exceed the constraint range, and generate initial surgical path control instructions.

[0010] In one possible implementation, the net contact force calculation module includes the following units: The signal preprocessing unit is used to install a 6-dimensional torque sensor between the surgical robot's end effector and the instrument clamping mechanism. It acquires the raw force signal and torque signal output by the torque sensor in real time through a data acquisition card, and simultaneously acquires the robot's end effector pose data. The acquired raw force signal and torque signal are sequentially subjected to amplitude limiting filtering and mean filtering to obtain the preprocessed signal. Model building units are used to construct Kalman filter models, with net contact force as the state variable in the state equations. The calculation formula is as follows: ;in, The process noise follows a Gaussian distribution with a mean of 0 and a variance of Q = 0.001. , represents the state transition matrix, express Net contact force state quantity at any given moment; Observation equation: using the pre-processed sensor-measured force as the observed value. The calculation formula is as follows: ;in, Represents the observation matrix. The observation noise is represented by a variance R = 0.01, determined through static sensor testing. The gain calculation unit is used to perform Kalman filtering on the acquired force signal to predict the state and calculate the predicted value. ; Calculate variance prediction: Update step, calculate Kalman gain: Status Update: Covariance update: ; in, express The covariance matrix of the predicted state values ​​at any given time represents the uncertainty of the predicted values; express The covariance matrix of the optimal estimate of net contact force. Represents the state transition matrix The transpose of the matrix; Indicates process noise The variance; express The time-major Kalman gain is used to balance the reliability of predicted values ​​and observed values; Represents the observation matrix The transpose of the matrix; Indicates observation noise The variance; Indicates fusion The optimal estimate of net contact force after the time-based observation; Represents the observation residual; express The covariance matrix of the optimal estimate at time step 1 is used for the filtering iteration at the next time step. This represents the identity matrix, with the same dimensions as the covariance matrix. Indicates Kalman gain; The contact force calculation unit is used to decompose the filtered force signal into a normal force along the axis of the instrument and a lateral force perpendicular to the axis, and synthesize them into the net contact force between the instrument and the tissue.

[0011] In one possible implementation, the target path planning module includes the following units: Layered bounding units are used to enclose the surgical robot end effector and virtual constraint boundaries in layers. A 3-5 layer AABB tree is constructed from the root node to the leaf node. The leaf node corresponds to the basic triangular facet of the model. The collision detection range is simplified by axis-aligned cuboid bounding boxes, and the minimum coordinates of each bounding box are calculated. The collision detection unit is used to set the iteration accuracy and maximum number of iterations of the GJK algorithm, define the collision threshold of the end effector, obtain the actual pose data of the end effector, update the global coordinates of the end effector's AABB tree, and perform layer-by-layer collision detection with the AABB tree of the virtual constraint boundary. If there is no collision, the child node is skipped; if there is a collision, the child node is recursively detected until the leaf node. The GJK algorithm is used to calculate the shortest distance and intrusion depth of the end effector for the triangular facets corresponding to the leaf nodes to obtain the collision detection results. The path planning unit is used to generate target surgical path control commands based on collision detection results and net contact force, including at least maintenance commands, deceleration commands, and emergency stop commands.

[0012] Secondly, this application provides an intelligent optimal path planning method for complex three-dimensional spaces, the intelligent optimal path planning method comprising the following steps: Preoperative CT image data of patients is acquired, and the lung structure of the CT image data is semantically segmented in real time using the U-Net+Transformer hybrid model to identify lesion areas and blood vessel walls, thereby obtaining a three-dimensional anatomical model. The spatial relationship between tumor and blood vessels in the three-dimensional anatomical model is extracted using a graph neural network (GNN) to generate a dynamic adjacency matrix. An Actor-Critic framework is designed, with the state space including the current position, tumor boundary distance, and avoidance structure distribution. The action space is defined as six-degree-of-freedom surgical instrument operation instructions, and multiple surgical candidate paths from the thoracic cavity entrance to the tumor core are output. An improved NSGA-II algorithm was used to optimize multiple surgical candidate paths to obtain initial surgical path control instructions; Based on the initial surgical path control command, the contact force signal is acquired in real time by the torque sensor on the end effector of the surgical robot, and the net contact force between the instrument and the tissue is calculated by using a signal fusion and noise reduction algorithm based on Kalman filtering. A fast collision detection model based on AABB hierarchical bounding boxes and GJK algorithm is used to calculate the position and intrusion depth of the surgical robot end effector in real time, obtain the collision detection results, and generate target surgical path control commands based on the collision detection results and net contact force.

[0013] In one possible implementation, the improved NSGA-II algorithm is used to optimize multiple surgical candidate paths to obtain initial surgical path control instructions, including: Based on the NSGA-II algorithm, an elite preservation and adaptive weighting strategy is adopted to give higher selection probability to paths close to the Pareto optimal front, while increasing the weight of safety objectives in the near-tumor region; when crossing, the safe segments of the two paths are preserved, and when mutating, the action instructions are randomly fine-tuned in the non-dangerous region, resulting in the improved NSGA-II algorithm. With a population size of 0 and 30 iterations, the optimal solution is selected by non-dominated sorting and crowding calculation. Based on the optimal solution, the forward and inverse kinematics models of the lung interventional surgery robot are established. The DH parameter method is used to define the joint coordinate system. The inverse kinematics is solved by numerical solution to obtain the mapping relationship between joint angles and end effector pose. Using RRT The algorithm plans the motion path nodes of the end effector from the initial position to the edge of the lesion, converts the motion path nodes into the pose commands of the surgical robot end effector, converts them into the angle commands of each joint through inverse kinematics, and uses a trapezoidal velocity curve to plan the joint motion. Add joint limit constraints and obstacle avoidance constraints, delete instructions that exceed the constraint range, and generate initial surgical path control instructions.

[0014] Based on the method provided in this application, its beneficial effects lie in the fact that by integrating graph neural networks and deep reinforcement learning, a three-dimensional intelligent path planning system for lung cancer surgery is constructed, significantly improving the efficiency and safety of preoperative planning and intraoperative navigation. The hybrid model can automatically extract the spatial topological relationships of lung anatomy and generate multiple candidate paths through training in a virtual surgical environment, effectively shortening the surgical path planning time to the minute level, and significantly improving the path safety margin compared to traditional methods. The multi-objective optimization module comprehensively evaluates the integrity of tumor resection, lung function preservation rate, and postoperative recovery indicators, outputting a Pareto optimal solution set, enabling surgeons to make flexible decisions among different priorities. The intraoperative dynamic adjustment mechanism updates the path through real-time image feedback, avoiding the risk of bleeding or tissue displacement and reducing the incidence of intraoperative complications. In addition, the three-dimensional visualization and navigation interface intuitively presents key parameters, including instrument angles and depths, assisting doctors in precise operation, not only improving the success rate of surgery but also reducing the use of intraoperative consumables and postoperative hospitalization time, providing an intelligent solution for minimally invasive treatment of lung cancer. Attached Figure Description

[0015] 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.

[0016] Figure 1 This is a schematic diagram of the structure of the intelligent optimal path planning system for complex three-dimensional space provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the three-dimensional model building module provided in the embodiments of this application; Figure 3 This is a flowchart of an embodiment of the intelligent optimal path planning method for complex three-dimensional space provided in this application. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] The specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 , Figure 1 This is a schematic diagram of the structure of the intelligent optimal path planning system for complex three-dimensional space provided in the embodiments of this application, including: The 3D model building module is used to acquire the patient's preoperative CT image data. It uses the U-Net+Transformer hybrid model to perform real-time semantic segmentation of the lung structure in the CT image data, identify lesion areas and blood vessel walls, and obtain a 3D anatomical model. Understandably, the 3D model building module also includes: an image data acquisition unit, which is used to acquire the patient's preoperative CT image data, read the CT image data through the Pydicom library, convert the data into a 32-bit floating-point image matrix, and obtain the first image data; The image data filtering unit is used to remove Gaussian noise from the first image data using Gaussian filtering, eliminate salt-and-pepper noise by combining median filtering, automatically crop invalid background areas based on the patient's body contour, retain the lung region of interest, and obtain the second image data. The image data processing unit is used to normalize the gray values ​​of the second image data to the [0,1] interval, and use the CLAHE algorithm to enhance the contrast of lung structures to obtain preprocessed image data.

[0019] The feature extraction unit is used to extract local spatial features based on the U-Net architecture adopted by the encoder of the U-Net+Transformer hybrid model. The first 4 layers are convolutional modules, each containing 2 3×3 convolutions + BN + ReLU, with a stride of 2 for downsampling. The feature fusion unit is used to connect the Transformer encoder after the 4th layer of the encoder. It contains 4 attention heads and has a hidden layer dimension of 512. It is used to capture global context information. The decoding end of the hybrid model uses transposed convolution for upsampling and fuses with the corresponding layer features of the encoder. The feature segmentation unit is used to input preprocessed image data into the U-Net+Transformer hybrid model, and output the segmentation result through convolution to obtain the image feature vector; The model generation unit is used to extract lesion regions and blood vessel walls from the segmentation mask of image feature vectors, filter out missegmented small regions, obtain semantic label regions, stack the semantic label regions into a three-dimensional voxel model based on CT spatial coordinate information, and generate a three-dimensional anatomical model of lesions and blood vessels through the MarchingCubes surface reconstruction algorithm.

[0020] Preoperative CT image data acquisition and preprocessing: 1. Data acquisition: The patient's lungs were scanned using a 64-slice or higher spiral CT scanner. Scanning parameters were set as follows: tube voltage 120kV, tube current 150-200mA, slice thickness 0.625-1mm, and pitch 1.0, ensuring that the image covered the entire lung area. Detailed data of the lung lobe containing the lesion was acquired and stored in DICOM standard format. 2. Data preprocessing: Format conversion: DICOM files were read using the Pydicom library and converted to a 32-bit floating-point image matrix; Denoising: Gaussian filtering was used with σ=0.8-1.2 to remove Gaussian noise from the image, combined with median filtering and a 3×3 window size to eliminate salt-and-pepper noise; Region cropping: Invalid background areas were automatically cropped based on the patient's body contour, retaining the lung region of interest (ROI); Normalization: Image grayscale values ​​were normalized to the [0,1] interval, and the CLAHE algorithm was used to enhance the contrast of intrapulmonary structures.

[0021] U-Net+Transformer Hybrid Model Construction and Deployment: 1. Model Structure Construction: Encoder End: U-Net basic architecture is adopted. The first 4 layers are convolutional modules, each containing 2 3×3 convolutions + BN + ReLU, with a stride of 2 for downsampling to extract local spatial features; after the 4th layer of the encoder end, a Transformer encoder is connected, containing 4 attention heads and a hidden layer dimension of 512 to capture global context information; Decoder End: Transposed convolution is used, with a stride of 2 to achieve upsampling, which is fused with the corresponding layer features of the encoder end. Finally, the segmentation result is output through 1×1 convolution, with 3 categories: lesion, blood vessel wall, and background. 2. Model Training and Optimization: Dataset Preparation: A public dataset, LIDC-IDRI combined with clinical labeled data, was used to construct a sample set of 1000 cases: 700 cases for training, 200 cases for validation, and 100 cases for testing. The samples were randomly rotated (±15°), flipped, and scaled, with data augmentation of 0.8-1.2 times. Training Parameters: The optimizer used was AdamW, with an initial learning rate of 1e-4, decaying by 10% every 50 epochs. The loss function was Dice loss + cross-entropy loss, with a batch size of 8. The training lasted for 100 epochs, and training was stopped when the Dice coefficient on the validation set did not improve for 10 consecutive epochs. Model Deployment: The trained model was converted to ONNX format and deployed to the edge computing unit of the surgical robot control system. GPU: NVIDIA Jetson AGXXavier.

[0022] Real-time semantic segmentation and feature vector extraction: 1. Real-time segmentation inference: The preprocessed CT image is input into the deployed model, and the inference time is controlled within 50ms. The output is a semantic segmentation mask of the lung structure. False lesion areas with an area of ​​less than 50 pixels are removed by connected component analysis to correct the segmentation results. 2. 3D anatomical model extraction: Features are extracted from the segmented lesion area and blood vessel wall area respectively: shape features (area, perimeter, roundness, principal axis direction); texture features (energy, entropy, and contrast of the gray-level co-occurrence matrix); position features (center point and bounding box parameters in 3D coordinates). Based on the spatial coordinate information of CT, the semantic labels of 2D slices are stacked into a 3D voxel model. The MarchingCubes surface reconstruction algorithm is used to generate triangular mesh surface models of lesions and blood vessels, preserving accurate spatial position and morphological parameters, volume, center coordinates, and surface normal vectors, to generate 3D anatomical models of lesions and blood vessels.

[0023] The surgical path generation module is used to extract the spatial relationship between tumor and blood vessels in the three-dimensional anatomical model using a GNN graph neural network, generate a dynamic adjacency matrix, design an Actor-Critic framework, and define the state space as the current position, tumor boundary distance, and avoidance structure distribution, while the action space is defined as a six-degree-of-freedom surgical instrument operation command. It outputs multiple surgical candidate paths from the thoracic cavity entrance to the tumor core. Understandably, the surgical path generation module also includes a graph node definition submodule, which is used to discretize key structures in the three-dimensional anatomical model into graph nodes, including tumor surface sampling points, vascular branch points and endpoints. Each node is characterized by three-dimensional coordinates, structure type, radius and curvature. The adjacency weight calculation submodule is used to calculate the adjacency weight based on the spatial GCN graph convolutional network, using the Euclidean distance and direction vector between nodes, and to learn the spatial constraints of tumors and blood vessels. The adjacency matrix generation submodule is used to construct an initial adjacency matrix based on the spatial relationships output by the spatial GCN graph convolutional network. When the current position is close to the structure in the path planning, the adjacency matrix is ​​updated to obtain a dynamic adjacency matrix.

[0024] The State Definition submodule is used to define the continuous states in the state space of the Actor-Critic framework, including the current position, the minimum distance to the tumor boundary, and the spatial distribution vector of the avoidance structure. The discrete state is whether it is close to the danger zone. The candidate path generation submodule is used to define the action space as a 6DoF six-degree-of-freedom surgical instrument operation command, including translation, range and rotation. The action step size is dynamically adjusted according to real-time accuracy. The Actor outputs the action probability distribution and the Critic evaluates the state value through the PPO algorithm. During inference, based on different initial exploration noise, 10-20 candidate paths from the thoracic cavity entrance to the tumor core are output, resulting in multiple surgical candidate paths.

[0025] Graph structure construction of the 3D anatomical model: 1. Node feature definition: Key structures in the 3D anatomical model are discretized into graph nodes, including tumor surface sampling points (one point every 5mm²), vascular branch points and endpoints. Each node feature includes 3D coordinates, structure type, tumor / vascular type, radius, vascular / curvature type, and tumor surface. 2. Spatial relationship extraction using GNN: A spatial GNN is used to calculate adjacency weights through Euclidean distance and direction vectors between nodes to learn the spatial constraints between the tumor and blood vessels.

[0026] Dynamic adjacency matrix generation: 1. Initial adjacency matrix: Based on the spatial relationships output by the spatial GCN graph convolutional network, an initial adjacency matrix is ​​constructed. The element values ​​are the safe passage probabilities between two nodes, which are positively correlated with distance and negatively correlated with structural hazard. 2. Dynamic update mechanism: When the current position is close to a structure in the path planning, the adjacency matrix is ​​updated in real time. If the distance is less than the safety threshold of 1mm, the weight of the corresponding position is set to 0, and passage is prohibited; if the distance increases, the weight dynamically increases.

[0027] The Actor-Critic reinforcement learning framework is designed as follows: 1. State Space Definition: Continuous State: Current position, 3D coordinates (x, y, z), minimum distance to the tumor boundary, avoidance structure, spatial distribution vector of blood vessels, distance and direction of the k nearest blood vessels, k=5. Discrete State: Whether it is close to the danger zone, 0 / 1, 1 when the distance is < the safety threshold. 2. Action Space Definition: Six-degree-of-freedom (6DoF) surgical instrument operation commands, including translation, Δx, Δy, Δz, range [-0.5mm, 0.5mm] and rotation, Δα, Δβ, Δγ, range [-5°, 5°]. The action step size is dynamically adjusted according to real-time accuracy, and the step size is reduced when approaching the tumor. 3. Reward Function Design: Positive Reward: Close to the tumor core, far from blood vessels, smooth path. Penalty: Touching blood vessels, distance < 0.5mm, deviation from the approximate direction from the thoracic inlet to the tumor, angle deviation > 30°. 4. Training and Inference: Train the Actor using the PPO algorithm, output the action probability distribution and Critic, evaluate the state value, and generate 10-20 candidate paths based on different initial exploration noise during inference.

[0028] The initial path selection module is used to optimize multiple surgical candidate paths using an improved NSGA-II algorithm to obtain initial surgical path control instructions; Understandably, the initial path selection module also includes: an algorithm improvement unit, which is used to assign higher selection probability to paths close to the Pareto optimal front based on the NSGA-II algorithm, using an elite retention and adaptive weight strategy, while increasing the weight of safety targets in the near-tumor region; retaining the safe segments of the two paths during crossover, and randomly fine-tuning the action instructions in the non-dangerous region during mutation, to obtain the improved NSGA-II algorithm. The mapping and association unit is used to set the population size to 0 and iterate 30 times. The optimal solution is selected by non-dominated sorting and crowding calculation. The forward and inverse kinematics models of the lung intervention surgery robot are established based on the optimal solution. The joint coordinate system is defined by the DH parameter method. The inverse kinematics is solved by numerical solution to obtain the mapping relationship between joint angles and end effector pose. Path resolution unit, used for employing RRT The algorithm plans the motion path nodes of the end effector from the initial position to the edge of the lesion, converts the motion path nodes into the pose commands of the surgical robot end effector, converts them into the angle commands of each joint through inverse kinematics, and uses a trapezoidal velocity curve to plan the joint motion. The path control unit is used to add joint limit constraints and obstacle avoidance constraints, delete instructions that exceed the constraint range, and generate initial surgical path control instructions.

[0029] The multi-objective optimization objectives are defined as follows: Objective 1: Minimize the total path length to reduce operation time; Objective 2: Maximize the minimum distance to blood vessels to improve safety; Objective 3: Minimize the change in path curvature to improve instrument operation stability.

[0030] The NSGA-II algorithm has been improved in the following ways: 1. Selection operator optimization: An elite retention + adaptive weight strategy is adopted, giving higher selection probabilities to paths closer to the Pareto optimal front. Simultaneously, for the safety objective (Objective 2), the weight is increased in the near-tumor region, with weight = 1 / d_tumor; the smaller d_tumor, the greater the weight. 2. Crossover and mutation operators: During crossover, the safe segments of both paths are retained (those >2mm from blood vessels). During mutation, the action instructions are randomly fine-tuned only in non-dangerous areas (>1mm from blood vessels).

[0031] Initial surgical path control command generation: 1. Optimization iteration: Set the population size to 50, iterate for 30 generations, and select the optimal solution through non-dominated sorting and crowding calculation. 2. Path discretization: Discretize the optimized continuous path into a series of 6DoF pose commands at time steps of 10ms / step, which serve as the initial control sequence for the surgical robot.

[0032] Robot Kinematic Modeling and Path Planning: 1. Kinematic Modeling: Establish the forward and inverse kinematics models of the lung intervention surgery robot, consisting of a 6-DOF robotic arm and an end effector. The DH parameter method is used to define the joint coordinate system. The inverse kinematics is solved numerically using the Newton-Raphson method to obtain the mapping relationship between joint angles and end effector pose. 2. Initial Path Planning: Using the guiding boundary generated in step 2 as constraints, the RRT (Real-Time Tracking) method is employed. The algorithm plans the initial motion path of the end effector from its initial position, the surgical entry point, to the edge of the lesion. The spacing between path nodes is set to 0.1 mm to ensure path smoothness, and the rate of change of joint angle between adjacent nodes is ≤5° / s.

[0033] Initial control command generation: 1. Convert the planned path nodes into pose commands for the end effector, with position coordinates X / Y / Z and attitude angles α / β / γ. Use inverse kinematics to convert these into angle commands for each joint. Plan joint motion using trapezoidal velocity curves to ensure smoothness during start-up and shutdown, with acceleration ≤10° / s². 2. Command optimization: Add joint limit constraints. The angle range of each joint is set according to the robot's mechanical structure and obstacle avoidance constraints. Delete commands exceeding the constraint range and re-interpolate to generate compliant commands.

[0034] Initial Real-Time Control Execution: 1. The optimized initial control commands are sent to the robot joint actuators via the EtherCAT bus, with the control frequency set to 1000Hz, to achieve the movement of the end effector along the planned path. 2. Real-Time Monitoring: The robot joint encoders collect the angles of each joint in real time at a sampling frequency of 1000Hz, calculate the actual pose of the end effector, and compare it with the commanded pose. If the error exceeds 0.1mm, command correction is triggered. A PID controller is used, with a proportional coefficient Kp=5.0, an integral coefficient Ki=0.1, and a derivative coefficient Kd=0.5.

[0035] The net contact force calculation module is used to acquire contact force signals in real time through the torque sensor on the end effector of the surgical robot based on the initial surgical path control command, and to calculate the net contact force between the instrument and the tissue using a signal fusion and noise reduction algorithm based on Kalman filtering. Understandably, the net contact force calculation module also includes: a signal preprocessing unit, which is used to install a 6-dimensional torque sensor between the surgical robot's end effector and the instrument clamping mechanism, and to collect the raw force signal and torque signal output by the torque sensor in real time through a data acquisition card, while simultaneously collecting the robot's end effector pose data; the collected raw force signal and torque signal are sequentially subjected to amplitude limiting filtering and mean filtering to obtain the preprocessed signal; Model building units are used to construct Kalman filter models, with net contact force as the state variable in the state equations. The calculation formula is as follows: ;in, The process noise follows a Gaussian distribution with a mean of 0 and a variance of Q = 0.001. , represents the state transition matrix, express Net contact force state quantity at any given moment; Observation equation: using the pre-processed sensor-measured force as the observed value. The calculation formula is as follows: ;in, Represents the observation matrix. The observation noise is represented by a variance R = 0.01, determined through static sensor testing. The gain calculation unit is used to perform Kalman filtering on the acquired force signal to predict the state and calculate the predicted value. ; Calculate variance prediction: Update step, calculate Kalman gain: Status Update: Covariance update: ; in, express The covariance matrix of the predicted state values ​​at any given time represents the uncertainty of the predicted values; express The covariance matrix of the optimal estimate of net contact force. Represents the state transition matrix The transpose of the matrix; Indicates process noise The variance; express The time-major Kalman gain is used to balance the reliability of predicted values ​​and observed values; Represents the observation matrix The transpose of the matrix; Indicates observation noise The variance; Indicates fusion The optimal estimate of net contact force after the time-based observation; Represents the observation residual; express The covariance matrix of the optimal estimate at time step 1 is used for the filtering iteration at the next time step. This represents the identity matrix, with the same dimensions as the covariance matrix. Indicates Kalman gain; The contact force calculation unit is used to decompose the filtered force signal into a normal force along the axis of the instrument and a lateral force perpendicular to the axis, and synthesize them into the net contact force between the instrument and the tissue.

[0036] The target path planning module is used to calculate the position and invasion depth of the surgical robot end effector in real time using a fast collision detection model based on AABB hierarchical bounding boxes and GJK algorithm, obtain collision detection results, and generate target surgical path control commands based on collision detection results and net contact force. Understandably, the target path planning module also includes: a hierarchical bounding unit, which is used to hierarchically bound the surgical robot end effector and virtual constraint boundary, constructing a 3-5 layer AABB tree from the root node to the leaf node, with the leaf node corresponding to the basic triangular facet of the model, simplifying the collision detection range through axis-aligned cuboid bounding boxes, and calculating the minimum coordinates of each bounding box; The collision detection unit is used to set the iteration accuracy and maximum number of iterations of the GJK algorithm, define the collision threshold of the end effector, obtain the actual pose data of the end effector, update the global coordinates of the end effector's AABB tree, and perform layer-by-layer collision detection with the AABB tree of the virtual constraint boundary. If there is no collision, the child node is skipped; if there is a collision, the child node is recursively detected until the leaf node. The GJK algorithm is used to calculate the shortest distance and intrusion depth of the end effector for the triangular facets corresponding to the leaf nodes to obtain the collision detection results. The path planning unit is used to generate target surgical path control commands based on collision detection results and net contact force, including at least maintenance commands, deceleration commands, and emergency stop commands.

[0037] Rapid Collision Detection Model Construction: 1. AABB Hierarchical Bounding Box Construction: The 3D model of the end effector, in STL format, and the virtual constraint boundary model generated in step 2 are layered and bounded. A 3-5 layer AABB tree is constructed from the root node to the leaf node, with the leaf nodes corresponding to the basic triangular faces of the model. The collision detection range is simplified by using axis-aligned cuboid bounding boxes. The minimum / maximum coordinates (xmin / xmax, ymin / ymax, zmin / zmax) of each bounding box are calculated. 2. GJK Algorithm Parameter Configuration: The iteration accuracy of the GJK algorithm is set to ≤0.01mm (distance error ≤ 0.01mm), and the maximum number of iterations is set to 50. A collision threshold is defined for the end effector and the constraint boundary; an intrusion depth ≥ 0.5mm is considered a collision risk.

[0038] Real-time collision detection execution: 1. Real-time acquisition of the actual pose data of the end effector from step 3, updating the global coordinates of the end effector's AABB tree, and performing layer-by-layer collision detection with the AABB tree of the virtual constraint boundary: root node collision detection, skipping child nodes if no collision occurs; if a collision occurs, recursively detecting child nodes until leaf nodes; for the triangular facets corresponding to leaf nodes, the GJK algorithm is used to calculate the shortest distance and intrusion depth between the end effector and the constraint boundary, obtaining the collision detection result: no collision, close to constraint, or intrusion into constraint. 2. Detection frequency optimization: AABB tree traversal and GJK algorithm iteration are accelerated through GPU parallel computing to ensure a collision detection frequency ≥ 500Hz, meeting real-time control requirements.

[0039] Target Surgical Path Control Command Generation and Execution: 1. Control Strategy Formulation: For collision-free conditions and net contact force < safety threshold, settings are based on tissue type: 2N for lung soft tissue, 1.5N for blood vessel walls: maintain initial control command and normal movement; For proximity constraints, distance < 1mm or net contact force ≥ safety threshold: generate deceleration command, reducing the end effector's movement speed to 50% of its original speed, and simultaneously issuing an audible and visual warning; For intrusion constraints, intrusion depth ≥ 0.5mm or net contact force ≥ danger threshold, 3N for soft tissue, 2.5N for blood vessel walls: generate emergency stop command, controlling the end effector to immediately stop movement, simultaneously cutting off the instrument drive power, and issuing a red alarm signal. 2. Command Execution and Feedback: The target surgical path control command is sent to the robot driver in real time via the bus. After execution, the end effector status is collected through joint encoders and force sensors to confirm that the movement status is consistent with the command; alarm signals are synchronously transmitted to the surgical console, displaying the alarm type (proximity alarm / intrusion alarm / force overload alarm) and specific location information, allowing the surgeon to decide whether to adjust the surgical plan.

[0040] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of the 3D model building module in the intelligent optimal path planning system for complex 3D space provided in the embodiments of this application, including: 201. Feature extraction unit, which is used to extract local spatial features based on the encoding end of the U-Net+Transformer hybrid model, which adopts the U-Net basic architecture. The first 4 layers are convolutional modules, each group contains 2 3×3 convolutions + BN + ReLU, with a stride of 2 for downsampling. 202. Feature fusion unit, used to connect the Transformer encoder after the 4th layer of the encoder, contains 4 attention heads, the hidden layer dimension is 512, used to capture global context information. The decoding end of the hybrid model uses transposed convolution for upsampling and fuses with the corresponding layer features of the encoder. 203. Feature segmentation unit, used to input preprocessed image data into the U-Net+Transformer hybrid model, and output segmentation results through convolution to obtain image feature vectors; 204. Model generation unit, used to extract lesion region and blood vessel wall from the segmentation mask of image feature vector, filter out missegmented small regions, obtain semantic label region, stack the semantic label region into a three-dimensional voxel model based on CT spatial coordinate information, and generate a three-dimensional anatomical model of lesion and blood vessel through MarchingCubes surface reconstruction algorithm.

[0041] Based on the modules provided in this application embodiment, CT data noise is effectively purified through format conversion, artifact removal, and HU value normalization, providing high-quality input for subsequent segmentation; the data augmentation strategy during the training phase significantly improves the model's generalization ability, ensuring adaptability to images of different patients; the U-Net+Transformer hybrid model combines the local detail capture capability of U-Net with the global spatial dependency modeling advantage of Transformer, accurately identifying lesion areas and blood vessel walls, especially optimizing the segmentation boundaries of complex structures; after lightweighting and GPU acceleration, the segmentation time for 300-500 layers of single CT data is controlled within 5 seconds, meeting real-time requirements; post-processing further improves the accuracy of semantic labels by filtering out small mis-segmented areas <5mm³; and the three-dimensional reconstruction technology based on spatial coordinates generates a triangular mesh model containing accurate morphological parameters, volume, center coordinates, etc.

[0042] The above describes the intelligent optimal path planning system for complex three-dimensional space in the embodiments of this application. The following describes the intelligent optimal path planning method for complex three-dimensional space in the embodiments of this application. Please refer to [link / reference]. Figure 3 , Figure 3 A flowchart illustrating the intelligent optimal path planning method for complex three-dimensional space provided in this application embodiment includes: 301. Obtain the patient's preoperative CT image data, and use the U-Net+Transformer hybrid model to perform real-time semantic segmentation of the lung structure in the CT image data to identify lesion areas and blood vessel walls, thereby obtaining a three-dimensional anatomical model. 302. Using GNN graph neural network to extract the spatial relationship between tumor and blood vessels in 3D anatomical model, generate dynamic adjacency matrix, design Actor-Critic framework, state space includes current position, tumor boundary distance, and avoidance structure distribution, action space is defined as six degrees of freedom surgical instrument operation instructions, output multiple surgical candidate paths from thoracic cavity entrance to tumor core; 303. An improved NSGA-II algorithm is used to optimize multiple surgical candidate paths to obtain initial surgical path control instructions; 304. Based on the initial surgical path control command, the contact force signal is acquired in real time by the torque sensor on the end effector of the surgical robot, and the net contact force between the instrument and the tissue is calculated by using a signal fusion and noise reduction algorithm based on Kalman filtering. 305. Using a fast collision detection model based on AABB hierarchical bounding boxes and GJK algorithm, the position and invasion depth of the surgical robot end effector are calculated in real time to obtain the collision detection results. Based on the collision detection results and net contact force, the target surgical path control command is generated.

[0043] In this embodiment of the invention, a three-dimensional intelligent path planning system for lung cancer surgery is constructed by fusing graph neural networks and deep reinforcement learning, significantly improving the efficiency and safety of preoperative planning and intraoperative navigation. The hybrid model can automatically extract the spatial topological relationships of lung anatomy and generate multiple candidate paths through training in a virtual surgical environment, effectively shortening the surgical path planning time to the minute level, and significantly improving the path safety margin compared to traditional methods. It comprehensively evaluates the integrity of tumor resection, lung function preservation rate, and postoperative recovery indicators, outputting a Pareto optimal solution set, enabling surgeons to make flexible decisions among different priorities. The intraoperative dynamic adjustment mechanism updates the path through real-time image feedback, avoiding the risk of bleeding or tissue displacement and reducing the incidence of intraoperative complications. The three-dimensional visualization and navigation interface intuitively presents key parameters, including instrument angles and depths, assisting doctors in precise operation, which not only improves the success rate of surgery but also reduces the use of intraoperative consumables and postoperative hospitalization time, providing an intelligent solution for minimally invasive treatment of lung cancer.

[0044] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent optimal path planning system for complex three-dimensional spaces, characterized in that, The optimal path intelligent planning system includes the following modules: The 3D model building module is used to acquire the patient's preoperative CT image data, and to perform real-time semantic segmentation of the lung structure of the CT image data through the U-Net+Transformer hybrid model to identify lesion areas and blood vessel walls, thereby obtaining a 3D anatomical model. The surgical path generation module is used to extract the spatial relationship between the tumor and blood vessels in the three-dimensional anatomical model using a GNN graph neural network, generate a dynamic adjacency matrix, design an Actor-Critic framework, and define the state space as the current position, tumor boundary distance, and avoidance structure distribution, and the action space as a six-degree-of-freedom surgical instrument operation command, and output multiple surgical candidate paths from the thoracic cavity entrance to the tumor core. The initial path selection module is used to optimize multiple surgical candidate paths using an improved NSGA-II algorithm to obtain initial surgical path control instructions; The net contact force calculation module is used to acquire contact force signals in real time through the torque sensor on the end effector of the surgical robot based on the initial surgical path control command, and to calculate the net contact force between the instrument and the tissue using a signal fusion and noise reduction algorithm based on Kalman filtering. The target path planning module is used to calculate the position and intrusion depth of the surgical robot end effector in real time using a fast collision detection model based on AABB hierarchical bounding boxes and GJK algorithm, obtain collision detection results, and generate target surgical path control commands based on the collision detection results and net contact force.

2. The intelligent optimal path planning system for complex three-dimensional space as described in claim 1, characterized in that, The 3D model creation module includes the following units: The image data acquisition unit is used to acquire the patient's preoperative CT image data. It reads the CT image data through the Pydicom library, converts the data into a 32-bit floating-point image matrix, and obtains the first image data. The image data filtering unit is used to remove Gaussian noise from the first image data using Gaussian filtering, eliminate salt-and-pepper noise by combining median filtering, automatically crop invalid background areas based on the patient's body contour, retain the lung region of interest, and obtain the second image data. The image data processing unit is used to normalize the gray values ​​of the second image data to the [0,1] interval, and use the CLAHE algorithm to enhance the contrast of lung structures to obtain preprocessed image data.

3. The intelligent optimal path planning system for complex three-dimensional space as described in claim 2, characterized in that, The 3D model creation module includes the following units: The feature extraction unit is used to extract local spatial features based on the U-Net architecture adopted by the encoder of the U-Net+Transformer hybrid model. The first 4 layers are convolutional modules, each containing 2 3×3 convolutions + BN + ReLU, with a stride of 2 for downsampling. The feature fusion unit is used to connect the Transformer encoder after the 4th layer of the encoder. It contains 4 attention heads and has a hidden layer dimension of 512. It is used to capture global context information. The decoding end of the hybrid model uses transposed convolution for upsampling and fuses with the corresponding layer features of the encoder. The feature segmentation unit is used to input preprocessed image data into the U-Net+Transformer hybrid model, and output the segmentation result through convolution to obtain the image feature vector; The model generation unit is used to extract lesion regions and blood vessel walls from the segmentation mask of image feature vectors, filter out missegmented small regions, obtain semantic label regions, stack the semantic label regions into a three-dimensional voxel model based on CT spatial coordinate information, and generate a three-dimensional anatomical model of lesions and blood vessels through the MarchingCubes surface reconstruction algorithm.

4. The intelligent optimal path planning system for complex three-dimensional space as described in claim 1, characterized in that, The surgical path generation module includes the following sub-modules: The graph node definition submodule is used to discretize key structures in a 3D anatomical model into graph nodes, including tumor surface sampling points, vascular branch points and endpoints. Each node is characterized by 3D coordinates, structure type, radius and curvature. The adjacency weight calculation submodule is used to calculate the adjacency weight based on the spatial GCN graph convolutional network, using the Euclidean distance and direction vector between nodes, and to learn the spatial constraints of tumors and blood vessels. The adjacency matrix generation submodule is used to construct an initial adjacency matrix based on the spatial relationships output by the spatial GCN graph convolutional network. When the current position is close to the structure in the path planning, the adjacency matrix is ​​updated to obtain a dynamic adjacency matrix.

5. The intelligent optimal path planning system for complex three-dimensional space as described in claim 1, characterized in that, The surgical path generation module includes the following sub-modules: The State Definition submodule is used to define the continuous states in the state space of the Actor-Critic framework, including the current position, the minimum distance to the tumor boundary, and the spatial distribution vector of the avoidance structure. The discrete state is whether it is close to the danger zone. The candidate path generation submodule is used to define the action space as a 6DoF six-degree-of-freedom surgical instrument operation command, including translation, range and rotation. The action step size is dynamically adjusted according to real-time accuracy. The Actor outputs the action probability distribution and the Critic evaluates the state value through the PPO algorithm. During inference, based on different initial exploration noise, 10-20 candidate paths from the thoracic cavity entrance to the tumor core are output, resulting in multiple surgical candidate paths.

6. The intelligent optimal path planning system for complex three-dimensional space as described in claim 1, characterized in that, The initial path filtering module includes the following units: The algorithm improvement unit is used to improve the NSGA-II algorithm by adopting an elite preservation and adaptive weight strategy, giving higher selection probability to paths close to the Pareto optimal front, while increasing the weight of safety targets in the near-tumor region; retaining the safe segments of the two paths during crossover, and randomly fine-tuning the action instructions in the non-dangerous region during mutation, thus obtaining the improved NSGA-II algorithm. The mapping and association unit is used to set the population size to 0 and iterate 30 times. The optimal solution is selected by non-dominated sorting and crowding calculation. The forward and inverse kinematics models of the lung intervention surgery robot are established based on the optimal solution. The joint coordinate system is defined by the DH parameter method. The inverse kinematics is solved by numerical solution to obtain the mapping relationship between joint angles and end effector pose. Path resolution unit, used for employing RRT The algorithm plans the motion path nodes of the end effector from the initial position to the edge of the lesion, converts the motion path nodes into the pose commands of the surgical robot end effector, converts them into the angle commands of each joint through inverse kinematics, and uses a trapezoidal velocity curve to plan the joint motion. The path control unit is used to add joint limit constraints and obstacle avoidance constraints, delete instructions that exceed the constraint range, and generate initial surgical path control instructions.

7. The intelligent optimal path planning system for complex three-dimensional space as described in claim 1, characterized in that, The net contact force calculation module includes the following units: The signal preprocessing unit is used to install a 6-dimensional torque sensor between the surgical robot's end effector and the instrument clamping mechanism. It acquires the raw force signal and torque signal output by the torque sensor in real time through a data acquisition card, and simultaneously acquires the robot's end effector pose data. The acquired raw force signal and torque signal are sequentially subjected to amplitude limiting filtering and mean filtering to obtain the preprocessed signal. Model building units are used to construct Kalman filter models, with net contact force as the state variable in the state equations. The calculation formula is as follows: ;in, The process noise follows a Gaussian distribution with a mean of 0 and a variance of Q = 0.

001. , represents the state transition matrix, express Net contact force state quantity at any given moment; Observation equation: using the pre-processed sensor-measured force as the observed value. The calculation formula is as follows: ;in, Represents the observation matrix. The variance represents the observation noise and is determined through static testing of the sensor. R=0.01 ; The gain calculation unit is used to perform Kalman filtering on the acquired force signal to predict the state and calculate the predicted value. ; Calculate variance prediction: Update step, calculate Kalman gain: Status Update: Covariance update: ; in, express The covariance matrix of the predicted state values ​​at any given time represents the uncertainty of the predicted values; express The covariance matrix of the optimal estimate of net contact force. Represents the state transition matrix The transpose of the matrix; Indicates process noise The variance; express The time-major Kalman gain is used to balance the reliability of predicted values ​​and observed values; Represents the observation matrix The transpose of the matrix; Indicates observation noise The variance; Indicates fusion The optimal estimate of net contact force after the time-based observation; Represents the observation residual; express The covariance matrix of the optimal estimate at time step 1 is used for the filtering iteration at the next time step. This represents the identity matrix, with the same dimensions as the covariance matrix. Indicates Kalman gain; The contact force calculation unit is used to decompose the filtered force signal into a normal force along the axis of the instrument and a lateral force perpendicular to the axis, and synthesize them into the net contact force between the instrument and the tissue.

8. The intelligent optimal path planning system for complex three-dimensional space as described in claim 7, characterized in that, The target path planning module includes the following units: Layered bounding units are used to enclose the surgical robot end effector and virtual constraint boundaries in layers. A 3-5 layer AABB tree is constructed from the root node to the leaf node. The leaf node corresponds to the basic triangular facet of the model. The collision detection range is simplified by axis-aligned cuboid bounding boxes, and the minimum coordinates of each bounding box are calculated. The collision detection unit is used to set the iteration accuracy and maximum number of iterations of the GJK algorithm, define the collision threshold of the end effector, obtain the actual pose data of the end effector, update the global coordinates of the end effector's AABB tree, and perform layer-by-layer collision detection with the AABB tree of the virtual constraint boundary. If there is no collision, the child node is skipped; if there is a collision, the child node is recursively detected until the leaf node. The GJK algorithm is used to calculate the shortest distance and intrusion depth of the end effector for the triangular facets corresponding to the leaf nodes to obtain the collision detection results. The path planning unit is used to generate target surgical path control commands based on collision detection results and net contact force, including at least maintenance commands, deceleration commands, and emergency stop commands.

9. An intelligent optimal path planning method for complex three-dimensional spaces, characterized in that, The optimal path intelligent planning method includes the following steps: Preoperative CT image data of patients is acquired, and the lung structure of the CT image data is semantically segmented in real time using the U-Net+Transformer hybrid model to identify lesion areas and blood vessel walls, thereby obtaining a three-dimensional anatomical model. The spatial relationship between tumor and blood vessels in the three-dimensional anatomical model is extracted using a graph neural network (GNN) to generate a dynamic adjacency matrix. An Actor-Critic framework is designed, with the state space including the current position, tumor boundary distance, and avoidance structure distribution. The action space is defined as six-degree-of-freedom surgical instrument operation instructions, and multiple surgical candidate paths from the thoracic cavity entrance to the tumor core are output. An improved NSGA-II algorithm was used to optimize multiple surgical candidate paths to obtain initial surgical path control instructions; Based on the initial surgical path control command, the contact force signal is acquired in real time by the torque sensor on the end effector of the surgical robot, and the net contact force between the instrument and the tissue is calculated by using a signal fusion and noise reduction algorithm based on Kalman filtering. A fast collision detection model based on AABB hierarchical bounding boxes and GJK algorithm is used to calculate the position and intrusion depth of the surgical robot end effector in real time, obtain the collision detection results, and generate target surgical path control commands based on the collision detection results and net contact force.

10. The intelligent optimal path planning method for complex three-dimensional space as described in claim 9, characterized in that, The improved NSGA-II algorithm is used to optimize multiple surgical candidate paths to obtain initial surgical path control instructions, including: Based on the NSGA-II algorithm, an elite preservation and adaptive weighting strategy is adopted to give higher selection probability to paths close to the Pareto optimal front, while increasing the weight of safety objectives in the near-tumor region; when crossing, the safe segments of the two paths are preserved, and when mutating, the action instructions are randomly fine-tuned in the non-dangerous region, resulting in the improved NSGA-II algorithm. With a population size of 0 and 30 iterations, the optimal solution is selected by non-dominated sorting and crowding calculation. Based on the optimal solution, the forward and inverse kinematics models of the lung interventional surgery robot are established. The DH parameter method is used to define the joint coordinate system. The inverse kinematics is solved by numerical solution to obtain the mapping relationship between joint angles and end effector pose. Using RRT The algorithm plans the motion path nodes of the end effector from the initial position to the edge of the lesion, converts the motion path nodes into the pose commands of the surgical robot end effector, converts them into the angle commands of each joint through inverse kinematics, and uses a trapezoidal velocity curve to plan the joint motion. Add joint limit constraints and obstacle avoidance constraints, delete instructions that exceed the constraint range, and generate initial surgical path control instructions.

Images