A high-precision vision-guided endoscopic intelligent minimally invasive surgery robot system

Abstract

This invention discloses a high-precision visual-guided laparoscopic intelligent minimally invasive surgical robot system, belonging to the field of intelligent surgical robot systems. It involves multiple modules including image acquisition, 3D reconstruction, semantic recognition, path planning, obstacle avoidance correction, execution control, force feedback and visual linkage, and human-machine collaborative interaction. The system constructs a real-time 3D anatomical model of the surgical area based on multispectral and depth images. By fusing force sensing data and image features to identify tissue types, it achieves adaptive adjustment of resection parameters. During path planning, it integrates multiple spatial and kinematic constraints to generate an optimal operational trajectory that can avoid obstacles. The human-machine interface displays the structural distribution and tool status in real time, supporting surgeon interaction and path adjustment. This invention improves surgical accuracy, stability, and safety, and is suitable for complex minimally invasive surgical scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent surgical robot systems, and more specifically to a high-precision visually guided laparoscopic intelligent minimally invasive surgical robot system. Background Technology

[0002] With the development of minimally invasive surgical techniques, laparoscopic surgery has been widely used in various departments such as thoracic surgery, urology, and gynecology due to its advantages such as less trauma, faster recovery, and less postoperative pain. However, traditional laparoscopic surgery still faces many challenges in scenarios with complex anatomical structures, limited organ space, unclear lesion localization, or interference from physiological movements (such as heartbeat and respiratory movements).

[0003] Current laparoscopic surgery primarily relies on manual operation by the surgeon, supplemented by two-dimensional video image navigation. However, two-dimensional images suffer from limitations in perspective, lack of depth information, and image distortion, making it difficult to accurately identify and avoid target tissues, and easily leading to accidental damage to important neurovascular structures. Meanwhile, traditional surgical robot systems rely heavily on preoperative image registration or static markers for target recognition, lacking the ability to respond in real-time to dynamic changes during surgery (such as tissue deformation, fluid obstruction, and surgical smoke), affecting their operational stability and accuracy in complex environments.

[0004] Especially when performing high-risk minimally invasive procedures such as abdominal aortic aneurysm repair (EVAR) and adrenal tumor resection, the target anatomical structures are small in size, deep in location, and close to important organs. Any slight error may cause irreversible damage, which places higher demands on the precise positioning and path planning of surgical instruments in clinical practice. Summary of the Invention

[0005] The purpose of this invention is to provide a high-precision visually guided laparoscopic intelligent minimally invasive surgical robot system to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a high-precision visually guided laparoscopic intelligent minimally invasive surgical robot system, comprising: The image acquisition module is used to acquire raw image data within the endoscopic field of view, including multispectral images and depth images; The 3D reconstruction module is used to construct a real-time 3D anatomical model of the target surgical area based on the image data obtained by the image acquisition module, and extract key tissue boundary features from it; The target recognition module is used to perform semantic recognition and segmentation of key surgical structures based on a preset training model and a three-dimensional anatomical model, and to obtain a structural distribution map. The surgical path planning module is used to dynamically calculate the optimal operation path under multiple constraints based on the structural distribution map and the resection path priority set preoperatively. The obstacle avoidance and correction module is used to monitor the positional changes of key structures in the 3D anatomical model in real time, and dynamically correct the operation path and generate an obstacle avoidance path in combination with the current pose of the surgical tools. The robot control and execution module is used to control the robotic arm to perform multi-degree-of-freedom operations on surgical tools according to the obstacle avoidance path, so as to achieve positioning and resection. The force feedback and vision linkage module is used to jointly calculate the force sensing data and image data of the surgical tool during tissue contact, to help determine the tissue type and optimize the cutting force and angle. The human-machine interface module is used to display the structure distribution diagram, path planning diagram and surgical tool status to the surgeon in real time, and allows the surgeon to adjust the path.

[0007] Preferably, the 3D reconstruction module includes: Multi-channel image registration processing is performed on the acquired multispectral image and depth image. Using the registered depth image data, the corresponding three-dimensional point cloud data is constructed by pixel back projection method. Each pixel is back projected to three-dimensional space according to its image coordinates and depth value, and reconstructed into a set of spatial coordinate points according to the intrinsic parameter matrix to form the original point cloud model. The original point cloud data is subjected to noise removal and homogenization processing. Point cloud resampling is performed using a voxel mesh-based method, and background interference and edge anomalies are removed to obtain the target anatomical point cloud model. Based on the point cloud surface reconstruction results, a boundary detection algorithm based on the rate of change of normal vectors is used to extract the edge features of key anatomical structures, and curved regions and abrupt change nodes are identified through spatial curvature analysis.

[0008] Preferably, the target recognition module includes: The three-dimensional anatomical model is spatially normalized, and the point cloud data is projected onto the two-dimensional image plane to generate a multi-channel input image that integrates depth information. An intraoperative semantic recognition model is constructed based on a deep convolutional neural network. The model adopts an encoder-decoder structure. The model is trained by transfer learning on a preoperative labeled dataset and outputs semantic label maps including lesion areas, important blood vessels and neural structures. The identified semantic label map is back-mapped to the 3D point cloud space. A multi-view label fusion algorithm based on spatial correspondence is used to assign semantic labels to each point cloud data point and construct a 3D structure distribution map. Based on semantic labels in the 3D structural distribution map, the spatial positional relationship between target structures is modeled, and a graph structure is used to represent the topological connections, adjacency relationships and spatial constraints between organizations.

[0009] Preferably, the surgical path planning module includes: Based on the three-dimensional structural distribution map output by the target recognition module, the spatial boundary information of the lesion area in the surgical area is extracted. Combined with the preoperative medical images and the lesion resection priority setting, a planned spatial area containing the target tissue and the safety buffer is automatically constructed. A multi-constraint model is constructed in the planned spatial area, including spatial obstacle avoidance constraints, robotic arm kinematic constraints, and field of view continuity constraints, and the constraints are transformed into cost function parameters in the path planning algorithm; Based on the constructed multi-constraint model, an improved heuristic search algorithm is used to generate the optimal surgical path. The algorithm combines the three-dimensional D*Lite search framework with the Euclidean distance heuristic function, integrates the safety cost graph and the minimum path curvature optimization strategy, and outputs a multi-degree-of-freedom operation trajectory. A dynamic feasibility analysis is performed on the generated path. The action sequence of the surgical tool under the planned path is simulated using time-series simulation. Conflict or over-limit positions are marked and adjusted through a local path reconstruction algorithm.

[0010] Preferably, the obstacle avoidance and correction module includes: Real-time acquisition of the latest spatial location information of key anatomical structures in the 3D anatomical model; Combining the current pose and velocity information of the surgical tool during the operation, the tool's motion trajectory is constructed based on a spatial temporal prediction model. By comparing the real-time positions of key structures, an ellipsoidal approximation model is used to make collision warning judgments, and the judgment results are output as a conflict risk matrix. After identifying potential conflict areas, local routes are replanned based on existing paths.

[0011] Preferably, the force feedback and vision linkage module includes: Real-time force data during tissue contact are collected, the force vector magnitude, direction, and rate of change of force in the contact area are calculated, and a set of force characteristic parameters for tissue identification is extracted. Based on the multispectral and depth image data provided by the image acquisition module, image texture features are extracted; The extracted force features and image texture features are used as multimodal inputs to construct a joint classification model based on decision trees or shallow neural networks, which can identify the tissue type of the current contact area in real time and output the tissue type label and confidence level. Based on the identified tissue type and its corresponding mechanical properties, the surgical tool's cutting force and angle are dynamically adjusted, including cutting rate, cutting angle, and maximum permissible contact force.

[0012] Preferably, the human-computer collaborative interface module includes: Based on the output three-dimensional structural distribution map, a voxel-based rendering and boundary highlighting algorithm is used to visualize the lesion area, important anatomical structures and their spatial relationships in real time in the graphical user interface. The generated multi-degree-of-freedom path trajectory is superimposed on the 3D structural map in a segmented color-coded form to display the distance, angle and curvature information between the path and the anatomical structure, and the path risk level is indicated by the color change of the trajectory nodes. The acquired tool pose, execution status, and motion trajectory are dynamically updated in the interface to reflect the real-time position, posture changes, and operation mode of the tool's 3D model.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs an intelligent minimally invasive laparoscopic surgical robot system that integrates high-precision visual guidance, 3D reconstruction, semantic recognition, path planning, obstacle avoidance correction, execution control, force linkage, and human-computer interaction. It breaks through the technical bottlenecks in traditional laparoscopic surgery, such as limited field of vision, path dependence on surgeon experience, fuzzy tissue recognition, and high operational risks. It achieves automatic recognition, dynamic perception, and precise intervention of multiple anatomical structures in complex surgical scenarios, significantly improving the intelligence level and operational safety of the surgical process.

[0014] 2. The technical solution provided by this invention has advantages such as multimodal data fusion decision-making, autonomous path optimization and obstacle avoidance, tool state closed-loop control and surgeon interactive intervention visualization. It can maintain path stability and tissue recognition accuracy under complex changes in the surgical environment, support personalized intervention and real-time system feedback collaboration, and is particularly suitable for minimally invasive surgical scenarios with extremely high requirements for anatomical precision and operational error tolerance, and has broad clinical application prospects. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] For examples, please refer to Figure 1 As shown in the figure, the high-precision visual-guided laparoscopic intelligent minimally invasive surgical robot system described in this embodiment includes: The image acquisition module is used to acquire raw image data within the endoscopic field of view, including multispectral images and depth images; The 3D reconstruction module is used to construct a real-time 3D anatomical model of the target surgical area based on the image data obtained by the image acquisition module, and extract key tissue boundary features from it; The target recognition module is used to perform semantic recognition and segmentation of key surgical structures based on a preset training model and a three-dimensional anatomical model, and to obtain a structural distribution map. The surgical path planning module is used to dynamically calculate the optimal operation path under multiple constraints based on the structural distribution map and the resection path priority set preoperatively. The obstacle avoidance and correction module is used to monitor the positional changes of key structures in the 3D anatomical model in real time, and dynamically correct the operation path and generate an obstacle avoidance path in combination with the current pose of the surgical tools. The robot control and execution module is used to control the robotic arm to perform multi-degree-of-freedom operations on surgical tools according to the obstacle avoidance path, so as to achieve positioning and resection. The force feedback and vision linkage module is used to jointly calculate the force sensing data and image data of the surgical tool during tissue contact, to help determine the tissue type and optimize the cutting force and angle. The human-machine interface module is used to display the structure distribution diagram, path planning diagram and surgical tool status to the surgeon in real time, and allows the surgeon to adjust the path.

[0019] The image acquisition module is one of the core front-end modules of the system of this invention. Its main function is to acquire raw image data within the laparoscopic field of view and provide high-quality, multi-dimensional input information for subsequent modules such as 3D reconstruction and target recognition. This module comprehensively utilizes multispectral imaging and structured light depth sensing technologies to achieve precise identification and spatial positioning of different tissue types in the surgical scene, overcoming the problem of insufficient dimensionality of traditional two-dimensional images.

[0020] Specifically, the image acquisition module mainly includes a multispectral image acquisition unit, a depth image acquisition unit, an image synchronization timing controller, and an anti-interference image enhancement unit. The multispectral image acquisition unit, configured with a multi-channel optical sensor array encompassing visible light (red, green, and blue) and near-infrared bands, can acquire multi-band images covering a wavelength range from 400 nm to 900 nm. Employing a narrowband filter and time-division excitation control strategy, images of different bands are acquired sequentially within different time windows and uniformly aligned to the principal coordinate system, avoiding spectral interference and crosstalk.

[0021] To obtain spatial depth information of tissue structures within the laparoscopic field of view, the depth image acquisition unit employs a fusion method of structured light projection and binocular stereo vision. The structured light module projects a high-density coded light spot pattern onto the tissue surface by integrating a dot-matrix near-infrared laser projector into the inner wall of the endoscope. The binocular camera unit acquires the deformation features of the pattern on different tissue surfaces and, combined with a pre-calibrated camera parameter model (including intrinsic matrix, extrinsic matrix, and distortion correction model), uses a 3D reconstruction algorithm to recover the spatial 3D coordinates of each pixel in the surgical scene, generating a high-precision depth image.

[0022] Specifically, the coding spacing of the structured light projection dot matrix was set to 1 mm, the image resolution to 1920×1080 pixels, the calibrated baseline length to 30 mm, and the projection angle to 30 degrees. The pixel depth value was calculated using triangulation, with the error controlled within ±0.5 mm. The 3D reconstruction algorithm employed a block matching optimization algorithm based on spatial correlation, combined with a constrained back-projection correction strategy, significantly improving the accuracy of reconstructing tissue edges and curved surface regions.

[0023] To ensure the stability and robustness of image data, an anti-interference image enhancement unit is integrated into the image acquisition module. This unit can perform adaptive brightness compensation, color equalization, and image dehazing on the original image, ensuring the acquisition of clear and continuous image frames even in complex environments such as surgical smoke, blood obscuring the image, and reflective interference. This unit uses a multi-scale image enhancement algorithm based on Retinex theory to automatically identify areas of decreased contrast in the image and perform local enhancement processing.

[0024] Furthermore, data from all image acquisition units are coordinated through a unified image synchronization timing controller to ensure that multispectral and depth images are acquired synchronously at the same timestamp, avoiding spatial misalignment caused by tissue movement or endoscope jitter. Image synchronization accuracy is controlled within 10 milliseconds, and all image data is stored and transmitted to subsequent modules in a unified format (such as OpenEXR or 16-bit TIFF).

[0025] The 3D reconstruction module first acquires multispectral image data and depth image data output from the image acquisition module. Since different spectral channels have different optical distortion characteristics and shooting angles, directly using them will lead to displacement, scale errors, and rotation differences between the images, affecting the accuracy of the subsequent 3D structure. Therefore, this invention employs a multi-channel image registration method based on joint matching of feature points and edge contours.

[0026] In the specific implementation, firstly, the improved Scale Invariant Feature Transform (SIFT) algorithm is used to extract corner and edge points from the multispectral image as the key feature point set A. Then, the Canny edge detection algorithm is used to extract edge contours from the visible light image as the auxiliary constraint feature set B. By minimizing the Euclidean distance between feature point sets A and B in the image space, and combining this with the image pyramid multi-resolution registration strategy, unified spatial alignment of the multispectral image is achieved. The registration accuracy is controlled to within one pixel by calculating the feature point pairing error.

[0027] After registration, the input depth image is used to reconstruct the 3D point cloud model of the target region. The spatial position of each pixel consists of its image coordinates (u, v) and depth value d. The 3D coordinates (x, y, z) are calculated using the following projection relationships: x equals the depth value d multiplied by the image point's horizontal coordinate minus the optical center's horizontal coordinate, then divided by the focal length fx; y equals the depth value d multiplied by the image point's vertical coordinate minus the optical center's vertical coordinate, then divided by the focal length fy; z equals the depth value d.

[0028] The above calculations rely on the camera's intrinsic parameter matrix, where fx and fy represent the focal length, and cx and cy represent the principal point coordinates. All parameters are pre-calibrated using a calibration board method and processed by an inverse distortion correction algorithm to ensure the accuracy of the depth mapping. The resulting point cloud set contains the spatial three-dimensional positions of all pixels in each frame, constructing the original point cloud model M0.

[0029] To enhance point cloud quality, this invention selects pixels with an effective depth range of 20 mm to 120 mm for projection, ignoring poorly reflected edge areas and distorted corners to eliminate invalid information and improve modeling speed.

[0030] Raw point cloud models typically contain a large number of noisy points and sparse regions, and directly using them in modeling can lead to model instability. This invention employs a combination of voxel grid filtering and statistical outlier removal algorithms to optimize point cloud data.

[0031] The basic idea of ​​the voxel mesh filtering algorithm is to divide the point cloud into several fixed-size three-dimensional cubes (i.e., voxels), and retain only one representative point within each voxel, whose coordinates are the geometric center of all points within the voxel, thereby achieving sparsity and homogenization of the point cloud. The voxel side length is set to 2 mm in the filtering parameters, which can be adjusted appropriately according to the complexity of the target structure.

[0032] Building upon this, the statistical outlier removal algorithm further analyzes the average distance between each point and its K nearest neighbors (K is set to 50 by default). If the average distance of a point exceeds three times the standard deviation of the overall mean of its point cloud, it is identified as an outlier and deleted. This algorithm effectively suppresses noise interference caused by reflections, occlusions, or discontinuities at tissue interfaces.

[0033] The filtered point cloud data M1 has the characteristics of uniform point distribution, consistent density, and clear boundaries, making it suitable for subsequent detailed modeling.

[0034] Based on optimized point cloud data, this invention employs a normal variation analysis algorithm to extract key tissue boundary features. Specific steps include: For each point in the point cloud model M1, calculate its normal vector direction, and use the PCA principal component analysis algorithm to select the principal axis direction of the neighboring points as the normal vector direction of that point; Calculate the angle between the normal vector of each point and its neighboring points. If the angle change exceeds a set threshold (set to 25 degrees), mark it as a boundary point. Further, curvature estimation algorithms are used to analyze the spatial curvature of boundary points, identify curvature abrupt change points, and weight them as "important structural points"; All boundary points are constructed into an edge feature map F1 and stored in the form of a KD tree for real-time retrieval by subsequent modules.

[0035] To improve processing speed, the above feature extraction process is implemented on a GPU parallel architecture, and the time taken for boundary point extraction is controlled within 50 milliseconds per frame.

[0036] The target recognition module first performs data preprocessing on the three-dimensional anatomical model output by the three-dimensional reconstruction module, mainly including spatial normalization processing, image projection fusion processing, and image quality enhancement processing.

[0037] In the spatial normalization stage, the input 3D point cloud model is first standardized in terms of coordinate system, unifying all spatial data to the robot's operational workspace coordinate system. The normalization process includes the following steps: Translate the centroid of the point cloud data to the origin of the coordinate system; Normalize the coordinates of all points so that the length of the point cloud in each axis is normalized to the interval [−1, 1]. A spatial rotation matrix is ​​applied to align the main direction to a fixed pose, ensuring the consistency of the model input.

[0038] Next, to facilitate semantic recognition processing by the convolutional neural network, the 3D point cloud model is converted into a multi-channel 2D image using a projection method. Specifically, a view-based synthetic projection algorithm is employed, projecting the point cloud onto a 2D image plane from three directions: front view, top view, and side view. RGB color information, depth information, and normal vector information are retained for each pixel, generating a fused input image containing multiple channels. The depth information is normalized to the [0, 1] interval, and the normal vector information is encoded as a grayscale image with three directional components.

[0039] For image enhancement, a histogram equalization algorithm is used to enhance image contrast, a median filtering algorithm is used to suppress high-frequency noise, and a local contrast enhancement algorithm based on Retinex theory is applied to improve image detail quality. The final output is a standardized multi-channel image sample with a size of 512×512 pixels and 6 channels (including RGB, depth, and normal vectors).

[0040] This invention employs a deep convolutional neural network (CNN)-based structure to construct a semantic recognition model. The model uses an encoder-decoder architecture to achieve pixel-level semantic segmentation of key tissues in images. The model incorporates multi-scale feature fusion and attention mechanisms to enhance its ability to recognize the boundaries of complex tissue structures.

[0041] The encoder section uses a five-layer convolutional module, each layer consisting of a convolutional layer, a batch normalization layer, and a ReLU activation function. Semantic features at different scales are extracted after downsampling at each layer. Decoder section: Uses transposed convolution (i.e., deconvolution) to restore spatial resolution layer by layer, and fuses intermediate features from the encoder, enhancing boundary preservation through skip connections; Attention Module: A spatial attention mechanism is inserted in the intermediate layer to enhance the feature representation of salient regions in space.

[0042] The model output is a semantic label map with a size of 512×512 pixels. The category of each pixel is determined by the maximum probability output by the Softmax function. The label categories include: lesion area (such as tumor or diseased tissue), important blood vessels (such as aorta and renal artery), nerve bundle, normal tissue background, etc., for a total of 8 categories.

[0043] Model training was performed using transfer learning on a preoperative labeled dataset. Initial network parameters were loaded with pre-trained weights from a public medical image segmentation dataset (such as MICCAI2020), and then fine-tuned using a clinically labeled laparoscopic image dataset (with at least 2000 samples). The loss function employed a hybrid of class-weighted cross-entropy and Dice coefficients, with weights set according to the annotation ratio of various tissue types. The Adam optimizer was used during training, with an initial learning rate of 0.0001 and a batch size of 16.

[0044] The two-dimensional label map output by the semantic recognition model needs to be mapped back to a three-dimensional point cloud space to achieve three-dimensional semantic modeling. This invention employs a strategy based on spatial projection inverse mapping and multi-view label fusion for the mapping operation.

[0045] First, the spatial correspondence between the pixels of the 2D image and the points in the original point cloud is recorded during the projection process at each viewpoint (i.e., the projection index matrix). After recognition is completed, the category information of each pixel in the 2D label image is assigned to the corresponding 3D point cloud point using this index, thus initially constructing a 3D semantic point cloud.

[0046] Since multiple perspectives may overlap, the same spatial point may be assigned different labels by multiple perspectives. This invention designs a label fusion algorithm based on confidence weighting: For each spatial point, the label category and corresponding confidence level obtained from each viewpoint are statistically analyzed; If the cumulative confidence score of a certain category exceeds 70% of the total confidence score, then that category is determined as the final label; If the confidence difference is less than 10%, the neighborhood voting method is used to determine the final category by majority voting based on the labels of spatially nearest points.

[0047] This label fusion algorithm can effectively improve the consistency and spatial coherence of the recognition results.

[0048] To provide more efficient support for path planning and obstacle avoidance, after obtaining the structural distribution map, this invention further models the spatial relationships between key organizational structures in the 3D semantic point cloud and expresses them using a graph structure modeling approach.

[0049] Specifically, each key tissue structure is treated as a node. If the spatial distance between two tissues is less than a set threshold (e.g., 5 mm), an edge is established between them. The weight of the edge can be represented by the reciprocal of the spatial distance, forming a weighted undirected graph. This graph structure not only represents the proximity between tissues but also describes their topological connections, such as clinical information like "tumor is adjacent to the left wall of the aorta," which assists in subsequent path detour planning and high-risk area prediction.

[0050] To further improve modeling efficiency, the graph structure is stored using a KD-Tree index, supporting real-time queries and dynamic updates. The spatial distance threshold can be dynamically set according to specific surgical types, offering strong adaptability.

[0051] The first step in path planning is to determine the effective operating space, that is, to construct the actual planned area available for the movement of surgical instruments based on the three-dimensional structural distribution map. This invention uses the following method for construction: Lesion region boundary extraction: Using the 3D structural distribution map output by the target recognition module, the 3D boundary of the lesion region is extracted. A voxel-based boundary fitting algorithm is used to mesh the lesion surface to form a structural contour model. If the lesion is irregular in shape, the boundary can be reconstructed using the point cloud concave shell fitting method (Alpha Shape).

[0052] Preoperative image fusion: Preoperative CT / MRI image data is introduced, and after registration (such as rigid registration method based on mutual information), it is fused with the intraoperative 3D model. The lesion tissue depth information and the resection priority marked by the surgeon are superimposed (such as priority sequence list based on risk level, with the level from 1 to 5 from high to low).

[0053] Buffer zone construction: A safety buffer zone of 3 mm to 10 mm is generated around the lesion area. The buffer distance is automatically set according to the tissue type, such as 10 mm around neural structures and 3 mm for soft tissue areas. The buffer zone and the lesion together form a "planning space area" as a constraint boundary for path search.

[0054] The spatial region is constructed in the form of a voxel cube with a side length of 2 mm, and uses an octree structure for compressed storage to improve search efficiency.

[0055] Within the aforementioned planned spatial area, various clinical and physical constraints need to be considered to ensure the feasibility, safety, and continuity of the pathway. This invention constructs the following three types of core constraints and integrates them into the pathway search algorithm in the form of a cost function: Spatial obstacle avoidance constraint: A minimum distance of 5 millimeters is set between the target path and high-risk structures (such as blood vessels and nerves). In the cost map, an exponential distance penalty function is constructed centered on this structure, with higher costs for closer proximity. Specifically, this is expressed as follows: The cost C equals the base cost plus the risk coefficient multiplied by an exponential function, where the exponential function uses the reciprocal of the distance as the exponent.

[0056] Robotic arm kinematic constraints: Considering the degrees of freedom, joint limits, and redundancy of the surgical robotic arm. Only motion trajectories within the reachable range of the robotic arm are allowed during path generation. The reachability of each pose point is evaluated using an inverse kinematics solution function (such as a Jacobian pseudo-inverse solution). If a joint angle exceeds a set limit (±120 degrees), the path point will be marked as invalid.

[0057] Visual field continuity constraint: To ensure the surgeon keeps the target area within the laparoscopic field of view throughout the procedure, visibility analysis is performed on each key point along the path. The visibility model is simulated using a ray tracing algorithm. If a point is obscured by tissue, the path cost is automatically weighted, prompting the path to prioritize points with good visibility.

[0058] All constraints are transformed into the cost graph input of the path search algorithm. The total cost of each path candidate node is the weighted sum of the above three types of constraints, where the weights can be adjusted and optimized in actual deployment (e.g., obstacle avoidance weight is 0.5, robotic arm constraint is 0.3, and field of view constraint is 0.2).

[0059] After constructing the complete cost model, the path generation stage begins. This invention employs an improved 3D D*Lite search algorithm as the main path generation framework and integrates the following key optimization strategies: Heuristic function selection: Euclidean distance is used as the basic heuristic function to quickly approximate the target region in three-dimensional space. At the same time, it is combined with the cost graph function to form a comprehensive heuristic valuation, so as to balance the shortest path and risk avoidance.

[0060] Path smoothing: To avoid sharp angles or abrupt changes in the path, this invention introduces a path curvature optimization algorithm. For continuous path points, their curvature values ​​(i.e., the rate of change of the normal vector per unit length of the trajectory) are calculated, and a maximum curvature threshold (set to no more than 0.2 radians per millimeter) is applied. Nodes in areas exceeding the limit are adjusted or replanned.

[0061] Multi-degree-of-freedom trajectory output: Combining the robotic arm motion model, the path points are converted from Cartesian coordinates to robotic arm joint space coordinates (i.e., six-axis or seven-axis joint angle values), forming a sequence of actions that can be directly executed. The output trajectory is represented by a sequence of timestamps, with each trajectory point containing spatial position, tool posture, and corresponding execution time, forming a "path-time-posture" triplet.

[0062] The path search process is executed on a GPU-accelerated platform, with the average planning time controlled within 200 milliseconds, meeting the real-time requirements during surgery.

[0063] To further ensure the stable execution of the pathway in a real surgical environment, this invention designs a dynamic feasibility assessment mechanism for the pathway and combines it with a local pathway reconstruction method for intelligent adjustment: Action sequence simulation: Using a surgical robot control model, the entire process of the robotic arm executing along the path is simulated in a simulation environment, and each point is checked for problems such as conflicts, exceeding kinematic limits, or visibility failures.

[0064] Conflict marker: If a node causes the robotic arm to come into contact with the surrounding tissue during the simulation, or if the maximum acceleration limit is exceeded due to too fast a pose change (set to 3 radians per second), the node is marked as a conflict point.

[0065] Local path reconstruction: For trajectory segments 2 to 3 nodes before and after the conflict point, a local RRT replanning algorithm is used for replacement generation. The RRT algorithm is limited to fast convergence within a local region, while ensuring a smooth transition with the preceding and following path segments.

[0066] Path verification and confirmation: After adjustment, simulation verification is performed again. If no new conflict nodes are found, the path is uploaded to the robot control module for execution. The system supports repeated path updates during surgery to cope with dynamic situations such as changes in tissue position and changes in surgeon's intent.

[0067] To achieve continuous spatial perception of key tissue structures (such as blood vessels, nerves, and tumor boundaries) within the surgical area, this module employs a boundary tracking algorithm based on multi-frame point cloud fusion to update the 3D structural model in real time.

[0068] Input data: Depth images and 3D point cloud sequences continuously acquired by the image acquisition module.

[0069] Fusion algorithm implementation: Temporal alignment of point cloud data in consecutive frames is performed, and coordinate consistency is achieved through a sparse point cloud registration algorithm based on iterative nearest point (ICP). Use voxel filtering (voxel side length set to 1 mm) to unify point density and eliminate redundancy; A point cloud trajectory estimation algorithm based on Kalman filtering is applied to predict the state of the time series of organizational boundary points and output the current frame spatial boundary model of the key structure. For tissues that undergo significant deformation (such as lung tissue under the influence of respiration), non-rigid point cloud registration algorithms (such as CPD algorithm) can be introduced to ensure dynamic tracking accuracy and control the boundary error within 2 mm.

[0070] Through the above processing, the system can maintain a real-time 3D model of key organizational structures and provide the latest spatial reference for subsequent collision detection and path correction.

[0071] To evaluate the spatial behavior of surgical tools during execution and the spatial interaction between them and key structures, this module establishes a spatial trajectory prediction model based on temporal information and uses a geometrically simplified model for collision warning.

[0072] Tool trajectory prediction model construction: The tool pose is output in real time by the robot control module, including the position, orientation (Euler angles or quaternions) and instantaneous velocity of the end effector; The pose state of the tool is recorded in the past 10 frames using a sliding window. Combined with the tool's current acceleration, a state transition equation based on Kalman prediction is used to predict the motion trajectory in the next 0.5 seconds, with a time step of 50 milliseconds. The prediction results are constructed as a discrete sequence of path points and stored as a set of points in three-dimensional space.

[0073] Collision warning judgment model: The surgical tool is modeled as an equivalent ellipsoid approximation model, with the principal axis direction consistent with the tool's principal axis direction, and the ellipsoid radius is set according to the tool size (e.g., major axis is 60 mm, minor axis is 5 mm). The boundary points of key organizational structures are simplified into point sets or grid elements to construct a spatial distance detection model; The judgment rule is based on the minimum distance between the ellipsoid and the point set. When the distance is less than a set threshold (e.g., 3 mm), it is marked as a potential conflict. All detection points are mapped to risk values, and the output is a conflict risk matrix. The matrix dimension is consistent with the number of path points predicted by the tool, and each element corresponds to the degree of conflict of a path point (0 means safe, 1 means inevitable collision, and 0.5 means moderate risk).

[0074] The risk matrix is ​​input into the path control module in real time to determine whether a path segment needs to be corrected.

[0075] Once a high-risk segment is identified in the predicted path, this invention designs a path replanning method based on local graph optimization to complete local obstacle avoidance correction while maintaining path continuity.

[0076] Affected area extraction: Traverse the conflict risk matrix to locate path segments with consecutive risk values ​​greater than 0.7; extract 3 path points before and after the segment to form a local path segment with a window length of 7; set the three-dimensional spatial region corresponding to the path segment as the "correction region".

[0077] Obstacle avoidance path reconstruction algorithm: Within the correction area, a three-dimensional search graph is constructed using the start and end points of the affected segments as path endpoints; The nodes in the diagram are discrete spatial points (divided with a resolution of 2 millimeters), and the edges represent feasible movements between adjacent points. Introduce a path optimization cost function, including path length, rate of curvature change, and risk penalty term: The path length cost is the Euclidean distance between nodes; The curvature cost is calculated using a three-point fitting method to determine the curvature rate, with a weighted penalty applied if the curvature exceeds the maximum allowable value (0.3 radians / mm). Risk penalties are weighted based on previous risk matrix values.

[0078] Using gradient descent-based path optimization algorithms (such as CHOMP or STOMP), the initial path is iteratively optimized to obtain a smooth and obstacle-avoiding shortest path.

[0079] Path continuation and verification: Transition interpolation is performed at the connection between the new path segment and the original path to ensure continuous attitude and velocity changes; The final path is converted into a tool-executable trajectory format, and the accessibility and collision-free nature of the path in the surgical scenario are verified through a kinematic simulation engine. If conflicts still exist after the correction path is corrected, the correction area will be expanded and the path will be replanned until it passes verification.

[0080] The obstacle avoidance and correction module outputs a discretized 3D spatial trajectory point (including position and attitude information), in the following format: Each trajectory point is represented as a triple (P, R, t), where P is the three-dimensional position vector, R is the attitude quaternion, and t is the timestamp.

[0081] This trajectory needs to be converted into the desired angle values ​​of each joint of the robotic arm. To this end, this module introduces a kinematic model based on an Analytical Inverse Kinematics (AIK) solver or a numerical optimization inverse solver (such as the Jacobian pseudo-inverse method), to establish the kinematic model according to the robot arm type (six-axis or seven-axis redundancy): If a closed analytical solution exists (such as a standard six-DOF joint structure), the analytical solution should be used first, as it is faster and more accurate. For redundant degrees of freedom (such as a seven-axis structure), an optimization method based on redundancy resolution is adopted, and secondary optimization objectives such as "elbow height" and "joint center distance" are added to ensure that the joint configuration meets physical constraints. The calculation frequency is above 1000 Hz to ensure high responsiveness control.

[0082] The original path points are represented discretely and require time-series interpolation to generate continuous control commands. This invention uses a fifth-order polynomial interpolation algorithm to generate joint positions, velocities, and accelerations within each time slice, ensuring a smooth path without abrupt changes, with an interpolation interval of 1 millisecond.

[0083] The control mode adopts a composite servo control strategy, including: Position servo control: used for main path tracking to ensure arrival accuracy (control error less than 0.3 mm); Velocity feedforward compensation: Pre-sets joint motion trends based on interpolation results to improve response speed; Impedance control mode: Activated during tool-tissue contact, it simulates a virtual spring-damping system to adjust compliance, reduce stress on biological tissue, and ensure tissue protection. Safety limit control: When the joint angle, speed or acceleration exceeds the threshold (e.g., more than 3 radians per second), the control system automatically switches to protection mode.

[0084] The robotic end effector needs to perform different operational functions depending on the current surgical procedure, such as tissue grasping, cutting, suction, and electrocoagulation. This module achieves unified control of various surgical tools through a multi-channel execution management mechanism.

[0085] Each tool corresponds to a control channel, and the control signals include: Action pattern signals: such as opening / closing, rotation, triggering; The intensity signals, such as clamping force and electrocoagulation intensity, can be adjusted. Status feedback signals: tool open / closed status, load changes, current changes, etc., used for closed-loop control.

[0086] All end-effector movements are synchronized with the path trajectory to ensure that the tool's actions are precisely applied to the surgical target area and to avoid accidental contact with important structures.

[0087] To ensure the safety and stability of path execution, the robot control execution module introduces a multi-level feedback mechanism: Position and attitude feedback: Current pose data is provided in real time by the encoder and inertial measurement unit (IMU); Force feedback monitoring: The contact force is monitored by a six-dimensional force sensor integrated into the tool base. If the force exceeds the set threshold (e.g., 5 Newtons), compliance control or emergency stop protection is triggered immediately. Current and temperature monitoring: detects changes in drive unit current and temperature, and identifies abnormal conditions such as overload and stall; Troubleshooting process: Triggering safety threshold: Enter deceleration mode; Continuous over-limit: Enter protection shutdown mode; A pop-up window on the human-computer interface prompts the operator to intervene.

[0088] This step uses a six-dimensional force / torque sensor integrated into the proximal end of the surgical tool to acquire real-time force data when the surgical tool comes into contact with the tissue. The sampling frequency is set to 1000 Hz to meet the requirements of dynamic and rapid response.

[0089] The sensor output is a 6-dimensional vector set, including force (in Newtons) and torque (in Newton-meters) along the X, Y, and Z directions, represented as: F=(fx,fy,fz), T=(τx,τy,τz); the following feature parameter set for tissue discrimination is extracted from the sensor data (each contact cycle is 100 milliseconds): Force vector magnitude: equal to the square root of fx² plus fy² plus fz²; Contact direction angle: The angle between the tool spindle and the force vector, used to determine whether it is a vertical compression or shear contact; Force change rate (first derivative): The difference in magnitude between the current frame and the previous frame divided by the time difference, reflecting the impedance characteristics of the tissue; Average peak impact: The maximum instantaneous force value recorded over multiple contact cycles, reflecting tissue rigidity; Contact stability index: the variance of force signal during contact; the softer the tissue, the more obvious the fluctuation.

[0090] The above parameter combination forms a force feature vector, which serves as part of the input model.

[0091] This section uses multispectral and depth images acquired by the image acquisition module, combined with tissue contact areas, to perform image segmentation and texture analysis, and extract visually relevant tissue discrimination features.

[0092] Based on the real-time pose of the surgical tool, the image ROI of the tool contact area is determined by back projection. Select an image block with an area of ​​approximately 5 mm x 5 mm at the tip of the tool for processing; Multispectral images (RGB + near-infrared) and depth images (depth accuracy ≤ 1 mm) are acquired simultaneously.

[0093] The following algorithm is used to extract image texture and structural features: Gray-level co-occurrence matrix (GLCM) features: extracting statistical texture metrics such as energy, contrast, entropy, and correlation; Local Binary Pattern (LBP): Analyzes changes in microscopic texture structure to distinguish between muscle, fat, and tumor tissue; Depth edge gradient features: Calculate the first-order gradient in the Z-direction of the contact area in the depth map, which reflects the depth deformation response of soft and hard tissues.

[0094] The aforementioned image processing features constitute the image feature vector.

[0095] To achieve real-time identification of organization types, this invention constructs a lightweight, embeddable multimodal joint classification model, which can choose one of the following two model structures: The input is the merged feature vector; The output is a tissue type label (such as muscle, blood vessel, connective tissue, nerve, tumor, etc., a total of 5 categories); A node splitting strategy based on information gain is adopted; The model training set consists of preoperative labeled data and intraoperative labeled data, with a sample size of more than 3,000 groups. The output includes classification labels and confidence scores (a confidence score greater than 0.8 is considered a valid identification).

[0096] 2. Shallow neural networks (such as 3-layer MLPs): The number of nodes in the input layer depends on the feature dimension (approximately 50 dimensions). The hidden layer has two layers, with 64 and 32 nodes in each layer; The activation function used is ReLU, and the output layer uses Softmax; The training was performed using the Adam optimizer with a learning rate of 0.001. Network inference latency is controlled within 5 milliseconds to meet real-time requirements.

[0097] The model is deployed in an embedded GPU chip (such as NVIDIA Jetson TX2) or an edge computing module, supporting intraoperative recognition at more than 100 frames per second.

[0098] After identifying the current contact tissue type, the system adjusts the following operating parameters based on a pre-built database of tissue mechanical property parameters for that type: Resection force adjustment: Set the maximum allowable contact force threshold (in Newtons); for example: 2 Newtons for adipose tissue, 1.5 Newtons for nerves, and 3 Newtons for tumors; if the real-time force sensor exceeds the threshold, the system will automatically trigger the force limiting mechanism to reduce the advance speed or cutting force.

[0099] Cutting angle and speed adjustment: Adjust the cutting angle according to the directionality of the tissue structure; for example, the cutting angle of muscle tissue is optimized to within 45 degrees according to the direction of muscle fibers; combine the trajectory output by the path planning module and add the angle fine-tuning function to each trajectory point.

[0100] Cutting frequency / energy output adjustment: For electrocoagulation shears or ultrasonic scalpels, adjust the excitation frequency and power according to the tissue hardness; for example, connective tissue requires a higher frequency band (e.g., 60 kHz), while soft tissue can be treated with a lower frequency (40 kHz).

[0101] The structural distribution maps output by the target recognition module and the 3D reconstruction module are visualized. These maps are 3D point clouds or voxel models containing multiple types of tissues, and their data includes semantic labels and spatial coordinates for different tissue structures such as lesion areas, nerves, blood vessels, fat, and muscles. To facilitate the surgeon's intuitive understanding of the anatomical structure of the surgical area, the system renders the structural distribution maps as highly interactive 3D models within the interface.

[0102] The GPU-based voxel rendering algorithm is employed, and the specific methods include: The three-dimensional structural data was voxelized, and the voxel size was set to 1 mm. Rendering is performed using the volumetric ray casting method; Introduce lighting models (such as the Phong model) to enhance the sense of depth and shadow representation; Color coding is used to distinguish tissue types, such as red for arteries, blue for veins, green for nerves, and purple for tumor areas; A boundary highlighting algorithm is used for the boundaries of each tissue, and a color high-pass filter and stroke line are added at the interface between tissues to make the structural edges clearer; The rendering results allow the operator to perform rotation, scaling, translation, and sectioning operations. The sectioning function is implemented through a draggable section control tool provided by the interface (such as a slider to move the XY plane), which can cut the model in any direction to observe the internal structure distribution.

[0103] To highlight the lesion area and key tissues, this module uses the following display logic: The lesion area is set to a blinking border mode, which occurs once per second. If the lesion is located near an important anatomical structure (less than 5 mm), add a semi-transparent danger zone marker at the lesion boundary; All structures support information hover functionality, which displays meta-information such as structure name, label confidence level, and spatial coordinates when the mouse hovers over it.

[0104] To clearly demonstrate the surgical operation path generated by the system and its relationship with the structure to the surgeon, the system overlays the path trajectory as a color-coded path line onto the three-dimensional structural model, realizing spatial understanding of the path and visualization of risk assessment.

[0105] The path trajectory is output by the path planning module, and is typically a sequence of 3D position and attitude information points arranged in chronological order, represented as: Path point set Each p contains a position (x, y, z), an orientation (quaternion or Euler angles), and a timestamp t, where n is the total number of points in the set.

[0106] The trajectory is drawn as a continuous curve in the interface and displayed according to the following principles: The color of the curve reflects the risk level of the trajectory (green for low risk, yellow for medium risk, and red for high risk); Each trajectory point is a visual node, with a node radius set to 2 millimeters; The path playback animation function can be enabled to simulate the dynamic advancement process of the surgical procedure path, and the speed is adjustable.

[0107] The distance, angle, and curvature information between each path and the surrounding anatomical structures can be viewed directly in the interface: Distance display: The minimum Euclidean distance between the path and the nearest important structure is displayed in floating form as a number. If it is less than the warning threshold (e.g., 3 mm), the path segment is highlighted in red. Angle display: The tangential angle between the current point and the next point on the path is calculated as the angle of change of the path direction, which is used to help determine whether the angle of entry is reasonable; Curvature display: The path curvature is calculated using a three-point fitting method. Excessive curvature (such as exceeding 0.3 radians per millimeter) will prompt a risk of "discontinuous operation". The above information and path are updated in the interface view, and the operator can choose to browse segment by segment, observe globally, or zoom in on a specific area.

[0108] By monitoring the tool status information output by the robot control execution module, the 3D model status of the surgical tool is updated in real time on the interface, including its position, posture, motion trajectory, and type of execution action.

[0109] Each type of surgical tool (such as electrocautery scissors, ultrasonic scalpel, and grasping forceps) is bound to a visual 3D model in the system, and the system renders the state in the following way: The tool model performs attitude synchronization on the path trajectory, that is, the current tool position is equal to the coordinates of the current trajectory point, and the attitude is directly driven by the attitude quaternion; The model's motion trajectory is displayed as a "wake line" in the interface, and its length can be customized (e.g., displaying the motion trajectory of the last 5 seconds). The current tool action status (such as "cutting", "electrocoagulation", "at rest") is marked on the model with color; for example, the model turns bright blue when the action is active.

[0110] The system supports simultaneous rendering by multiple tools and uses a screen partitioning layout for display: The main viewpoint displays the 3D model and path trajectory of the main operating tool; The auxiliary window displays the pose of the second tool or observer; All tool data, including tool name, current mode, and contact force (from the force feedback module), is displayed in real time in the parameter bar on the right. If a tool malfunctions or exceeds its limits (such as abnormal posture or abnormal current), the system will issue a graphical warning, such as a flashing red box and a pop-up warning box.

[0111] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A high-precision visual-guided laparoscopic intelligent minimally invasive surgical robot system, characterized in that: include: The image acquisition module is used to acquire raw image data within the endoscopic field of view, including multispectral images and depth images; The 3D reconstruction module is used to construct a real-time 3D anatomical model of the target surgical area based on the image data obtained by the image acquisition module, and extract key tissue boundary features from it; The target recognition module is used to perform semantic recognition and segmentation of key surgical structures based on a preset training model and a three-dimensional anatomical model, and to obtain a structural distribution map. The surgical path planning module is used to dynamically calculate the optimal operation path under multiple constraints based on the structural distribution map and the resection path priority set preoperatively. The obstacle avoidance and correction module is used to monitor the positional changes of key structures in the 3D anatomical model in real time, and dynamically correct the operation path and generate obstacle avoidance paths based on the current pose of the surgical tools, including: Real-time acquisition of the latest spatial location information of key anatomical structures in the 3D anatomical model; Combining the current pose and velocity information of the surgical tool during the operation, the tool's motion trajectory is constructed based on a spatial temporal prediction model. By comparing the real-time positions of key structures, an ellipsoidal approximation model is used to make collision warning judgments, and the judgment results are output as a conflict risk matrix. After identifying potential conflict areas, local routes are replanned based on existing routes; The robot control and execution module is used to control the robotic arm to perform multi-degree-of-freedom operations on surgical tools according to the obstacle avoidance path, so as to achieve positioning and resection. The force feedback and vision linkage module is used to jointly calculate the force sensing data and image data of the surgical tool during tissue contact, assisting in determining the tissue type and optimizing the resection force and angle, including: Real-time force data during tissue contact are collected, the force vector magnitude, direction, and rate of change of force in the contact area are calculated, and a set of force characteristic parameters for tissue identification is extracted. Based on the multispectral and depth image data provided by the image acquisition module, image texture features are extracted; The extracted force features and image texture features are used as multimodal inputs to construct a joint classification model based on decision trees or shallow neural networks, which can identify the tissue type of the current contact area in real time and output the tissue type label and confidence level. Based on the identified tissue type and its corresponding mechanical properties, the surgical tool's cutting force and angle are dynamically adjusted, including cutting rate, cutting angle, and maximum permissible contact force. The human-machine interface module is used to display the surgeon's structural distribution diagram, path planning diagram, and surgical tool status in real time, and allows the surgeon to adjust the path, including: Based on the output three-dimensional structural distribution map, a voxel-based rendering and boundary highlighting algorithm is used to visualize the lesion area, important anatomical structures and their spatial relationships in real time in the graphical user interface. The generated multi-degree-of-freedom path trajectory is superimposed on the 3D structural map in a segmented color-coded form to display the distance, angle and curvature information between the path and the anatomical structure, and the path risk level is indicated by the color change of the trajectory nodes. The acquired tool pose, execution status, and motion trajectory are dynamically updated in the interface to reflect the real-time position, posture changes, and operation mode of the tool's 3D model.

2. The high-precision vision-guided endoscopic intelligent minimally invasive surgical robotic system according to claim 1, characterized in that: The 3D reconstruction module includes: Multi-channel image registration processing is performed on the acquired multispectral image and depth image. Using the registered depth image data, the corresponding three-dimensional point cloud data is constructed by pixel back projection method. Each pixel is back projected to three-dimensional space according to its image coordinates and depth value, and reconstructed into a set of spatial coordinate points according to the intrinsic parameter matrix to form the original point cloud model. The original point cloud data is subjected to noise removal and homogenization processing. Point cloud resampling is performed using a voxel mesh-based method, and background interference and edge anomalies are removed to obtain the target anatomical point cloud model. Based on the point cloud surface reconstruction results, a boundary detection algorithm based on the rate of change of normal vectors is used to extract the edge features of key anatomical structures, and curved regions and abrupt change nodes are identified through spatial curvature analysis.

3. The high-precision vision-guided endoscopic intelligent minimally invasive surgical robotic system according to claim 2, characterized in that: The target recognition module includes: The three-dimensional anatomical model is spatially normalized, and the point cloud data is projected onto the two-dimensional image plane to generate a multi-channel input image that integrates depth information. An intraoperative semantic recognition model is constructed based on a deep convolutional neural network. The model adopts an encoder-decoder structure. The model is trained by transfer learning on a preoperative labeled dataset and outputs semantic label maps including lesion areas, important blood vessels and neural structures. The identified semantic label map is back-mapped to the 3D point cloud space. A multi-view label fusion algorithm based on spatial correspondence is used to assign semantic labels to each point cloud data point and construct a 3D structure distribution map. Based on semantic labels in the 3D structural distribution map, the spatial positional relationship between target structures is modeled, and a graph structure is used to represent the topological connections, adjacency relationships and spatial constraints between organizations.

4. The high-precision visual-guided laparoscopic intelligent minimally invasive surgical robot system according to claim 3, characterized in that: The surgical pathway planning module includes: Based on the three-dimensional structural distribution map output by the target recognition module, the spatial boundary information of the lesion area in the surgical area is extracted. Combined with the preoperative medical images and the lesion resection priority setting, a planned spatial area containing the target tissue and the safety buffer is automatically constructed. A multi-constraint model is constructed in the planned spatial area, including spatial obstacle avoidance constraints, robotic arm kinematic constraints, and field of view continuity constraints, and the constraints are transformed into cost function parameters in the path planning algorithm; Based on the constructed multi-constraint model, an improved heuristic search algorithm is used to generate the optimal surgical path. The algorithm combines the three-dimensional D*Lite search framework with the Euclidean distance heuristic function, integrates the safety cost graph and the minimum path curvature optimization strategy, and outputs a multi-degree-of-freedom operation trajectory. A dynamic feasibility analysis is performed on the generated path. The action sequence of the surgical tool under the planned path is simulated using time-series simulation. Conflict or over-limit positions are marked and adjusted through a local path reconstruction algorithm.

Images