Image registration algorithm method and system in transforaminal endoscopic surgery

Abstract

This invention discloses an image registration algorithm and system for percutaneous endoscopic discectomy (PED) surgery, belonging to the field of image processing technology. It includes: reconstructing a preoperative three-dimensional skeletal model; acquiring intraoperative fluoroscopic images and reconstructing a three-dimensional point cloud; sequentially performing initial registration and fine registration to obtain the fine registration transformation relationship; constructing a dynamic error compensation model, using respiratory phase and surgical operation parameters as input, predicting the three-dimensional displacement compensation amount through a single mapping function and correcting the registration transformation to obtain the final transformation matrix; and aligning and fusing the model with the intraoperative images for display. This invention employs multi-scale feature fusion and an anatomical prior fine registration mechanism to reduce registration errors; uses a hybrid kernel regression model with heterogeneous parameter fusion to uniformly compensate for respiratory and mechanical deformation errors; utilizes GPU parallel computing to meet real-time requirements; and constructs a multi-index closed-loop monitoring system based on feature correspondence residuals, feature distribution uniformity, and post-compensation displacement to avoid transient noise false alarms and ensure surgical safety.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically to an image registration algorithm and system for percutaneous endoscopic discectomy. Background Technology

[0002] Percutaneous endoscopic discectomy (PED) is an important branch of minimally invasive spinal surgery. Its core advantage lies in establishing a working channel through percutaneous puncture, allowing for procedures such as discectomy and spinal canal decompression under direct endoscopic visualization, thus avoiding the significant trauma and prolonged recovery period associated with traditional open surgery. This technique, with its significant advantages of minimal trauma, less bleeding, and faster postoperative recovery, has been widely used in the clinical treatment of common spinal diseases such as lumbar disc herniation and lumbar spinal stenosis, and has become one of the important technical platforms for modern minimally invasive spinal surgery. With the continuous expansion of clinical applications and the gradual extension of surgical indications, the need for precise intraoperative localization and real-time tracking of target anatomical structures is becoming increasingly prominent, directly driving the rapid development and widespread application of intraoperative image registration technology in this field.

[0003] The basic principle of intraoperative image registration technology lies in spatially aligning preoperative three-dimensional image data (such as CT and MRI) with intraoperative real-time two-dimensional images (such as C-arm X-ray fluoroscopy images) or three-dimensional scene information. This allows the rich anatomical information contained in the preoperative images to be superimposed and displayed in the surgical field, providing the surgeon with intuitive and accurate spatial navigation reference. In the special application scenario of percutaneous endoscopic discectomy (PED), due to the limited surgical space and the target being located in deep tissues, it is often difficult to accurately determine the three-dimensional spatial relationship between instruments and key anatomical structures by visual identification under direct endoscopic visualization alone. Therefore, the introduction of image registration technology has significant clinical value in improving surgical safety and reducing serious complications such as nerve damage.

[0004] Existing intraoperative registration methods mainly include several technical approaches such as feature-point-based registration, surface-matching-based registration, and voxel-based registration. Feature-point-based registration methods identify and extract corresponding anatomical landmarks or artificial markers from preoperative and intraoperative images, and then use the least squares method or iterative nearest-point algorithm to solve the spatial transformation relationship between the two point sets. This method can achieve high registration accuracy provided the markers are accurately selected, but it is highly dependent on the markers, and the precise calibration of the markers themselves presents certain technical challenges. Surface-matching-based registration methods use the 3D bone surface reconstructed from preoperative images and the digital surface point cloud acquired intraoperatively for iterative registration. Its advantage is that it does not require deliberate setting of markers, but the registration accuracy is limited by factors such as surface reconstruction quality and point cloud density. Voxel-based registration methods use the grayscale or gradient information of images as similarity measures. By optimizing algorithms to search for spatial transformation relationships that maximize the similarity measure, automatic image registration is achieved. The advantage of this method is that it has a high degree of automation and can make full use of the global information of the image. However, when dealing with deformable structures such as the spine that have physiological movements, its registration accuracy is often affected to some extent.

[0005] However, with the increasing application of percutaneous endoscopic discectomy (PED) in complex cases and the ever-increasing clinical demands for surgical precision, the existing registration techniques have gradually revealed several deep-seated contradictions and inherent limitations in practical applications. First, the target area for PED is typically located near the intervertebral disc and the posterior margin of adjacent vertebral bodies. This area has complex anatomical structures and significant individual differences. Existing registration methods based on feature points or surface matching struggle to guarantee registration accuracy in cases with significant anatomical variations or obvious osteophyte formation at the vertebral body margins. Second, although PED falls under the category of minimally invasive surgery, patients inevitably experience periodic positional changes due to respiratory movements during the procedure. Furthermore, the surgical manipulation itself may cause slight relative displacements between vertebral bodies. This dynamic change introduces a risk of intraoperative registration drift over time, and existing static registration methods are inadequate in addressing this dynamic error. Furthermore, the working channel for percutaneous endoscopic discectomy (PED) is established based on percutaneous puncture. The puncture needle and working cannula generate mechanical stress during entry into the intervertebral foramen. This stress, transmitted to the target vertebral body, can induce significant mechanical deformation. Existing registration algorithms typically assume the spinal structure is a rigid body when establishing registration relationships, an assumption that deviates significantly from actual surgical conditions. In addition, existing registration methods often fail to meet the high real-time requirements of PED, especially in applications requiring multiple dynamic registrations to track tissue displacement, where the computational time of the registration algorithms becomes even more pronounced.

[0006] In summary, achieving high-precision, robust, and real-time registration of intraoperative images during percutaneous endoscopic discectomy (PED) while fully considering multiple factors such as dynamic physiological movement of the spine, mechanical deformation caused by surgical procedures, and complex individual anatomical variations has become a key technical challenge for those skilled in the art. Fundamental improvements and innovations to existing registration techniques at the algorithmic level are urgently needed. Summary of the Invention

[0007] The purpose of this invention is to provide an image registration algorithm method and system for percutaneous endoscopic discectomy, in order to solve the problems of existing registration techniques mentioned in the background art in terms of complex anatomical variations, dynamic physiological movements, mechanical deformation and real-time performance.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The image registration algorithm method for percutaneous endoscopic discectomy includes the following steps: Step S1: Acquire preoperative three-dimensional image data and reconstruct a three-dimensional skeletal model of the target vertebra; Step S2: Acquire intraoperative fluoroscopic images and reconstruct the intraoperative three-dimensional point cloud structure; Step S3: Perform initial registration on the 3D skeleton model and the 3D point cloud structure to obtain the initial spatial transformation relationship, and extract the registration feature point set based on the initial spatial transformation relationship to perform fine registration and obtain the fine registration spatial transformation relationship; Step S4: Construct a dynamic error compensation model. Combine the real-time respiratory phase parameters with the surgical operation parameters as the input vector. Predict the three-dimensional displacement compensation amount through a single unified mapping function. Then, use the three-dimensional displacement compensation amount to correct the fine registration spatial transformation relationship and obtain the final registration transformation matrix. Step S5: Use the final registration transformation matrix to spatially align and fuse the 3D skeletal model with the intraoperative images for display.

[0009] According to the above technical solution, in step S2, reconstructing the intraoperative three-dimensional point cloud structure refers to extracting the pure skeletal edge contour of the target vertebra from at least two intraoperative fluoroscopic images at mutual angles using a deep convolutional neural network without external hardware spatial markers, and directly reconstructing the three-dimensional point cloud structure representing the real spatial pose of the bone during the operation through multi-view stereo matching under epipolar geometric constraints.

[0010] According to the above technical solution, the initial registration in step S3 specifically includes: Simultaneously, Gaussian smoothing filtering is applied to the 3D skeletal model and 3D point cloud structure at multiple different spatial scales to simultaneously capture the local geometric undulations and overall anatomical orientation of the vertebral surface. Calculate the geometric curvature characteristics of grid nodes at each spatial scale. The geometric curvature characteristics include Gaussian curvature, average curvature, and a high-dimensional local shape histogram descriptor based on the statistical distribution of the angle between the local neighborhood normal vectors. Calculate the similarity matrix between the 3D skeleton model and the 3D point cloud structure in the feature space, match feature descriptors, and use a bidirectional matching strategy and random sampling consistency algorithm to remove abnormal corresponding point pairs; Based on the preserved set of corresponding interior points, the initial rigid space transformation relationship that minimizes the sum of squared Euclidean distances between corresponding point pairs is solved using the singular value decomposition algorithm.

[0011] According to the above technical solution, in step S3, the registration feature point set is extracted to perform fine registration, specifically including: Based on prior anatomical knowledge of the spinal structure, control constraint rules are selected, and a set of feature control points that cover the entire spatial circumference and are repeatable across patients is automatically determined on the surface of the coarsely aligned 3D skeletal model. The feature control point set is distributed in the posterior margin region of the target vertebral body, the bilateral pedicle stenosis apex region, the spinous process tip region, the lateral apex region of the transverse process, and the central region of the articular facet. An iterative nearest-point algorithm based on geometric feature distance weighting is executed, using the cosine similarity of point pairs on multi-scale feature vectors as the matching weight. A dual screening mechanism is used to eliminate erroneous correspondences caused by local point cloud missingness or soft tissue occlusion. The cumulative spatial transformation matrix is ​​iteratively updated using the calculated incremental rigid body transformation parameters to obtain the finely registered spatial transformation relationship.

[0012] According to the above technical solution, in step S4, the input vector includes at least the instantaneous respiratory phase, the needle insertion angle, the puncture depth, the outer diameter of the working cannula, the cannula advancement rate, and the cumulative parameters representing the duration of the operation; the unified mapping function is a variant structure of the radial basis function network based on data-driven and physical-mechanical prior training, which is used to implicitly establish a cross-coupling model between different respiratory phases and different mechanical compression states of surgical instruments.

[0013] Based on the above technical solution, the regression equation of the unified mapping function and the expression of the hybrid kernel function are as follows: In the formula, This refers to the three-dimensional displacement compensation amount; The input vector is composed of real-time respiratory phase sub-vectors. Surgical operation biomechanical parameter subvector Composed of splicing elements; Representing the The center points of the support vector kernels are determined by pre-updating and training based on an offline high-fidelity finite element simulation database. This is the regression weight vector; This is a balancing weight adjustment factor; and The kernel width parameter is used to adjust the smoothness of the regression space; The total number of core central nodes.

[0014] According to the above technical solution, in step S4, the respiratory phase parameters are optimized and corrected in real time by a respiratory pattern recognition unit, specifically including: A sliding window analysis strategy is used to extract continuous respiratory waveform signals, and a multidimensional respiratory feature vector including the period variation coefficient, the proportion of inspiratory phase time, the waveform peak amplitude variation coefficient, the peak frequency of power spectral density, and the number of waveform slope abrupt changes is calculated online. The offline-trained support vector machine classifier performs pattern determination on the multidimensional respiratory feature vector and outputs the pattern assignment weight coefficients for different typical respiratory patterns. Adaptive interpolation updates are performed on multiple pre-set kernel regression parameter sets based on pattern attribution weight coefficients to correct instantaneous registration drift of the dynamic error compensation model during non-stationary breathing switching.

[0015] According to the above technical solution, step S4 also includes introducing a non-rigid spatial penalty term based on the elastic connection relationship between adjacent vertebrae, in the registration transformation matrix of the locked target surgical vertebra. Under constraints, construct a joint multi-constraint optimization objective function: In the formula, For the perspective camera's two-dimensional projection function; This is the set of three-dimensional coordinates of anatomical auxiliary feature points on adjacent non-target vertebrae; These are the corresponding intraoperative two-dimensional feature observation point plane coordinates; It is a set of equivalent biomechanical elastic connections between adjacent vertebral bodies; and Let be the rigid body space transformation matrix of the adjacent vertebrae to be corrected; and These are the coordinates of the geometric center of the cone mesh model; This is the original relative position reference vector before surgery; The weighting coefficient of the elastic constraint penalty term is used to equivalently simulate the physiological stiffness of the intervertebral disc and ligament complex.

[0016] According to the above technical solution, in step S5, while merging the display, the registration reliability closed-loop continuous monitoring mechanism is started to calculate the feature correspondence residual index, the spatial distribution uniformity index of feature control points in the dissected area, and the residual displacement index between consecutive frames after dynamic compensation. When at least two of the three indices exceed their respective preset safety and reliability thresholds, the system determines that the registration has failed, immediately triggers a registration failure warning, and forcibly suspends the fusion display of the augmented reality image.

[0017] A percutaneous endoscopic discectomy image registration system includes: The preoperative image acquisition and reconstruction module is used to receive preoperative three-dimensional image data and extract a smooth, pure bony triangular mesh model of the target vertebral body for reconstruction. The intraoperative image acquisition and point cloud construction module is used to acquire multi-angle fluoroscopic image sequences in surgical scenarios without hardware marker constraints, and reconstruct the intraoperative three-dimensional pure bone point cloud structure based on the depth segmentation matching algorithm. The multi-scale coarse and fine registration module is used to match and solve the initial spatial transformation relationship under high-dimensional shape descriptors at multiple spatial scales, and perform fine registration with feature distance weighting based on anatomical prior feature control points to output the fine registration spatial transformation relationship. The high-dimensional coupled dynamic compensation module is used to construct a radial basis hybrid kernel regression model. It combines the real-time respiratory phase vector with multiple surgical instrument operation mechanical parameters to construct a high-dimensional input vector. It implicitly predicts the three-dimensional displacement compensation amount of coupled deformation with a single mapping function, thereby correcting the output final registration transformation matrix. The spatial alignment and closed-loop monitoring module is used to perform rigid body transformation and deformation rendering to achieve augmented reality 3D stereo fusion guidance, and to perform safe closed-loop disconnection control based on the parallel decision logic of feature correspondence residuals, distribution uniformity and remaining displacement to ensure the accuracy and reliability of the algorithm.

[0018] Compared with the prior art, the present invention has the following beneficial effects: In terms of registration accuracy, by constructing an initial registration strategy based on multi-scale feature fusion, geometric features are extracted and matched at multiple spatial scales, from local cortical bone undulations to the overall curvature of the vertebral body. Combined with an iterative fine registration mechanism based on feature point optimization and feature distance weighting based on anatomical prior knowledge, the mean registration error is reduced to the sub-millimeter level, and the accuracy is significantly improved. The complementarity of multi-scale features effectively enhances the robustness to local occlusion, point cloud sparsity, and noise interference. The anatomically guided feature control point selection strategy ensures the representativeness of the matching points in spatial distribution and cross-patient repeatability.

[0019] In terms of dynamic error compensation, this paper breaks through the conventional approach of separating respiratory motion compensation and mechanical deformation correction in existing technologies. It proposes a hybrid kernel regression dynamic compensation model based on heterogeneous parameter fusion. The model combines real-time respiratory phase parameters and surgical operation parameters to form a high-dimensional input vector. The three-dimensional displacement compensation amount is directly predicted through a single unified mapping function. The central node automatically learns and encodes the interaction mode between different respiratory phases and different mechanical states in the high-dimensional space, implicitly realizing the unified modeling of the coupling effect between the two. This significantly reduces the periodic displacement error caused by respiratory motion and the deformation error caused by the mechanical action of the working cannula.

[0020] In terms of real-time performance, by migrating all computationally intensive tasks such as multi-scale feature extraction, high-dimensional feature distance matrix calculation, iterative nearest point search and transformation solution to the graphics processing unit for parallel execution, the execution time of the complete single registration process, including dynamic error compensation, meets the real-time requirements of the operation, significantly shortening the time of a single registration. In terms of surgical safety, a closed-loop monitoring mechanism for registration reliability is constructed, which comprehensively examines three dimensions of indicators: feature correspondence residual, feature distribution uniformity, and residual displacement after dynamic compensation. A failure warning is only triggered when at least two of these indicators exceed their respective thresholds, effectively avoiding false alarms caused by non-systematic factors such as transient noise or local occlusion. When registration fails, the augmented reality navigation screen is automatically paused and the surgeon is prompted to reacquire fluoroscopic images to refresh the registration, forming a complete perception-evaluation-intervention safety closed loop. Attached Figure Description

[0021] Figure 1 This is a flowchart of the image registration method of the present invention. Detailed Implementation

[0022] 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, and 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.

[0023] Example 1 The system of this invention comprises a workstation host, a graphics processing unit (GPU), an image acquisition card, a video encoder, and a medical display device at the hardware level. The workstation host is equipped with a multi-core CPU and large-capacity memory to meet the requirements of concurrent multimodal data processing; the GPU uses a dedicated graphics card supporting a unified computing architecture, with sufficient video memory and computing cores to support high-speed parallel computation of intraoperative 3D volumetric data; the image acquisition card supports real-time input of at least two channels of high-definition video signals, ensuring zero-latency transmission of C-arm fluoroscopic images; the video encoder outputs the fused image, including augmented reality navigation information, to the medical display at a high frame rate, providing the surgeon with continuous and smooth visual guidance. Functionally, the system mainly includes a preoperative image acquisition module, a 3D model reconstruction module, an intraoperative image acquisition module, a 3D structure restoration module, a multi-scale feature extraction module, an initial registration module, a feature point optimization module, a dynamic error compensation module, a registration fusion module, a computational acceleration module, and a registration accuracy evaluation module. These modules collaboratively execute the following complete method pipeline.

[0024] This embodiment mainly addresses the issue of eliminating dependence on external hardware spatial markers in minimally invasive surgery, achieving pure image-driven, markerless, high-precision registration, and solving the dynamic registration drift problem caused by the superposition of spontaneous breathing and surgical mechanical stress through high-dimensional implicit coupling modeling.

[0025] In the preoperative phase, the system's preoperative image acquisition module first receives continuous tomographic CT data of the target vertebral segment via the DICOM protocol interface. The CT slice thickness is preferably set to 0.5mm to 1.0mm, and the scanning range should cover the target vertebra and at least two adjacent vertebral segments. Specifically, the received raw data undergoes bone-specific window width and level adjustment in the data preprocessing unit. The window width is typically set to 1500HU to 2000HU, and the window level to 300HU to 500HU, thereby maximizing the contrast between bone tissue and surrounding soft tissue. Furthermore, the system performs image grayscale normalization processing, standardizing the grayscale value ranges under different scanning protocols to a unified interval, and employs a noise reduction method based on a nonlocal mean algorithm to filter out quantum noise, aiming to effectively preserve the edge details of the bone cortex while smoothing out noise. After the above preprocessing is completed, the segmentation unit in the 3D model reconstruction module then executes a bone tissue segmentation algorithm based on thresholding and morphological operations. First, regions with voxel values ​​between 200 HU and 1000 HU are marked as candidate bone tissues. Then, a spherical structural element with a radius of 1 mm is used to perform a mathematical morphological opening operation to eliminate false pores caused by noise inside the cancellous bone. Next, a spherical structural element with a radius of 1.5 mm is used to perform a closing operation to fill the tiny serrated burrs at the edge of the bone cortex, finally obtaining a smooth and complete binary bone mask. Preferably, the meshing unit extracts isosurfaces from the binary voxel data based on the Moving Cubes algorithm, and the threshold of the isosurface is adaptively selected according to the bone density distribution histogram of the target vertebral region to ensure the topological integrity of the mesh patches (triangular meshes) at the junction of the bone cortex and cancellous bone. This allows the reconstructed 3D bone model to be stored in the form of a set of vertex coordinate vectors and their triangular patch topological connections, providing a high-precision pure bony geometric reference for subsequent registration.

[0026] During the intraoperative image acquisition phase, this invention, without any constraints from the body surface or surgical instrument hardware space markers, establishes bidirectional communication between the X-ray machine control unit in the intraoperative image acquisition module and the C-arm X-ray machine via a protocol, precisely controlling the X-ray machine's projection angle, exposure parameters, and image acquisition trigger signal. Specifically, the system acquires fluoroscopic images of the target vertebral region from two mutually angular projection positions, anteroposterior and lateral, with the angle difference between the two projection positions preferably set to 60 to 120 degrees. The exposure parameters are automatically adjusted according to the patient's body size within a range of tube voltage 70kVp to 120kVp, tube current 2mA to 10mA, and exposure time 10ms to 100ms. Furthermore, the image acquisition unit acquires fluoroscopic image sequences at a continuous acquisition rate of 15 to 30 frames per second and caches the acquired image data in real time into a circular buffer queue with a preset depth of 32 frames. This effectively smooths out instantaneous jitter during data transmission, ensuring the continuity of the image stream consumed by the downstream processing module in the time dimension. After acquiring each frame of perspective image, the 3D structure reconstruction module immediately initiates bone edge detection and 3D point cloud reconstruction. Specifically, this module incorporates a deep convolutional neural network (CNN) based on an encoder-decoder architecture, specifically designed to generate pixel-by-pixel bone edge probability maps. The encoder portion of this network consists of four cascaded convolutional downsampling stages, while the decoder portion gradually recovers the spatial resolution of the feature map through symmetrical four-stage upsampling convolutions. Finally, a sigmoid activation function maps the output features to the edge probability map. Preferably, the stereo matching unit, under the epipolar geometric constraint framework established by the C-arm X-ray machine's geometric calibration parameters, scans and searches for corresponding image points in the edge maps of the two viewpoints line by line along the corresponding epipolar lines. Based on triangulation, it directly reconstructs the 3D point cloud data representing the true spatial pose of the bone during the current operation by inverse projection of the 2D disparity information. The sampling density of the point cloud is controlled to have an adjacent point spacing of 0.5mm to 1.0mm to achieve the best balance between point cloud density and computational overhead.

[0027] Once the preoperative triangular mesh model and the intraoperative sparse point cloud are constructed, the registration process enters the coarse-fine integrated alignment stage to address the robust coarse alignment problem between heterogeneous 3D representations. Specifically, the multi-scale feature extraction module first performs a unified multi-scale geometric feature extraction pipeline on both the preoperative 3D skeletal model and the intraoperative 3D point cloud. For the preoperative mesh model, it is first subjected to Gaussian smoothing at three different spatial scales, with the standard deviation of each scale uniquely set to... , , This constructs a scale-space representation capable of simultaneously capturing local high-frequency undulations on the cortical bone surface and the overall anatomical orientation of the vertebral body. On each smoothed model version, the curvature calculation unit uses a local quadratic surface fitting method to calculate the Gaussian curvature and mean curvature at each vertex, generating dual-channel curvature features at that scale. Based on this, the local descriptor generation unit further constructs a high-dimensional local shape histogram for each vertex as its spatial context feature descriptor. This histogram, through the neighborhood of the vertex… The curvature values ​​of the nearest neighbor vertices are statistically distributed. Further, the initial registration module concatenates the multi-scale feature vectors of the preoperative model and the intraoperative point cloud and maps them to a unified high-dimensional feature space. In this space, the Euclidean distance matrix between the two feature sets is calculated to measure the feature similarity between point pairs. Preferably, the system employs a bidirectional matching strategy to confirm candidate correspondences and further eliminates abnormal correspondences that do not conform to the consistent spatial transformation model using the Random Sample Consensus Algorithm (RANSAC). Based on the retained set of interior point correspondences, its transformation solution unit calculates the initial rigid spatial transformation relationship using the Singular Value Decomposition (SVD) algorithm, thereby completing the coarse registration.

[0028] To meet sub-millimeter-level navigation accuracy requirements, the system further utilizes a feature control point determination unit within the feature point optimization module to automatically select control constraint rules based on prior anatomical knowledge. This determines a set of spatially circumferentially covered and patient-repeatable feature control points on the coarsely aligned preoperative 3D skeletal model surface. These control points are specifically distributed in the posterior margin region of the target vertebral body, the narrowest apex region of the bilateral pedicles, the spinous process tip region, the outermost apex region of the bilateral transverse processes, and the central region of the bilateral articular facets. After the control point set is determined, the iterative optimization unit then executes an improved iterative nearest point (ICP) algorithm based on feature distance weighting for fine registration. Specifically, in each iteration, the algorithm first searches for the candidate corresponding point with the closest spatial Euclidean distance in the intraoperative point cloud data for each feature control point. At the same time, it calculates the cosine similarity of the point pair to be matched on its multi-scale feature vectors. The obtained similarity value range is normalized to 0 to 1, and the system only retains point pairs with a similarity greater than a preset threshold of 0.7 to enter the transformation parameter solution of the current iteration. Using the retained set of weighted corresponding point pairs, the incremental rigid body transformation parameters are solved by the weighted singular value decomposition algorithm, and the cumulative spatial transformation matrix is ​​updated, finally obtaining the precise registration spatial transformation relationship.

[0029] As the surgery progresses, the patient's respiratory movements and the mechanical effects of the surgical instruments continuously alter the actual spatial position and local morphology of the target vertebral body. This invention constructs a dynamic error compensation model, breaking away from the conventional approach of separating respiratory motion compensation and mechanical deformation correction in modeling. It posits an inherent time-varying coupling relationship between deformation response and respiratory phase. To fundamentally represent this coupling effect uniformly, the dynamic error compensation module of this invention constructs a hybrid kernel regression dynamic compensation model based on heterogeneous parameter fusion. It combines real-time acquired respiratory phase parameters with surgical operation parameters as a high-dimensional input vector, directly predicting the required three-dimensional displacement compensation amount at the current moment through a single unified mapping function. The mathematical expression of the model is: And its hybrid kernel function The specific expression and constraints are as follows: In equations (1) and (2) above, This is the three-dimensional displacement compensation vector applied to the fine registration transformation relationship at the current moment; This is the joint input vector of heterogeneous parameters. Its components are specifically divided into components derived from the instantaneous respiratory phase. (Value range 0 to) The spontaneous respiratory motion subvectors constituted by ) and the angle of needle insertion. puncture depth outer diameter of working sleeve Casing advance rate and the cumulative parameter representing the duration of the operation. Jointly constructed mechanical operation subvectors Furthermore, in the formula... The total number of central nodes (core centers), in a typical implementation Take 15 to 25; For the first Each center node corresponds to a three-dimensional weight vector in three orthogonal directions. Preferably, the center point of the support vector kernel... This involves establishing an offline high-fidelity finite element deformation simulation database in advance, containing patients with different BMIs, osteoporosis scores, and variation characteristics. By pre-simulating tens of thousands of sets of bone displacement responses under the combined action of respiratory motion boundary conditions and surgical mechanical stress loads, this database is used for offline joint training and updating to select the most physically and mechanically representative state center points. These center nodes then automatically learn and encode the interaction patterns between different respiratory phases and different mechanical states in the regression space. Furthermore, in equation (2)... To balance the weighting adjustment factor, and This refers to the kernel width parameter, which controls the smoothness of the regressive manifold. During intraoperative online deployment, the system acquires respiratory waveforms from the respiratory phase sensor at a high frequency of 100 times per second and extracts the instantaneous phase in real time. Simultaneously, the current operation parameters are read from the navigation system interface, and the required three-dimensional dynamic compensation amount at the current moment is directly calculated by substituting them into equation (1). The compensation amount is multiplied on the left by the transformation matrix obtained in the fine registration stage in the form of a translation transformation matrix to generate the final registration parameters. As an enhanced implementation, when the working sleeve contains a miniature ultrasound probe, the system can also use ultrasound data to provide direct hard constraints on the location of the cortical bone. The cortical bone contour point set of the target vertebral body posterior margin region obtained using the ultrasound probe is registered with the preoperative 3D model using the ICP algorithm to obtain an independent ultrasound incremental correction vector. And it is incorporated into the dynamic compensation framework in a weighted fusion manner, so that the final local displacement compensation amount is corrected as follows: Among them, the fusion weight The preferred value range is 0.2 to 0.3.

[0030] Finally, the spatial fusion and monitoring module performs comprehensive rendering and closed-loop monitoring of the registration and fusion status. Specifically, the spatial transformation unit receives the final transformation matrix after dynamic compensation. The system performs rigid body space transformation and deformation field interpolation on the preoperative 3D skeletal model. The rendering unit then draws virtual safety boundary lines for navigation guidance on the aligned 3D model and overlays the preoperative model with 40% to 60% semi-transparency onto the intraoperative endoscopic field of view. Furthermore, to ensure the continuous reliability of the intraoperative registration results throughout the process, the registration accuracy evaluation module continuously monitors the registration status in a closed loop, executing parallel judgment logic including feature-corresponding residual index (determining if it exceeds 1.0 mm), feature control point spatial distribution uniformity index (determining if the coverage is less than 60%), and dynamic compensation effectiveness index (determining if the residual displacement exceeds 0.5 mm). Preferably, when any two of the above three evaluation indicators simultaneously exceed their respective preset thresholds, the system determines that the registration system has lost reliability, immediately triggers a registration failure warning, forcibly pauses the overlay rendering of the augmented reality image, and clearly prompts the surgeon to re-acquire C-arm fluoroscopic images to refresh the entire set of registration parameters, thus forming a complete perception-evaluation-intervention closed-loop safety control mechanism.

[0031] Example 2 Furthermore, based on the technical solution of Embodiment 1 above, the present invention further solves the problem of degradation of dynamic compensation accuracy caused by non-stable changes in the intraoperative breathing pattern (such as the patient suddenly transitioning from a regular cycle to shallow and rapid breathing, panting, or brief pauses due to adjustment of anesthesia depth or pain stress) through Embodiment 2.

[0032] Specifically, this embodiment adds a breathing pattern recognition unit to the dynamic error compensation module to construct an online closed-loop recognition-switching adaptive mechanism. This unit specifically uses the continuous breathing waveform signal output by the breathing sensor. As input, a sliding window analysis strategy is employed, truncating a length from the current time point backwards at a frequency of once per second. Waveform segments within a time window. For the waveform data within each window, a set of multidimensional feature vectors for quantitatively characterizing the breathing pattern is calculated online. Among them, the characteristic components The coefficient of variation of the period, which reflects the regularity of the respiratory rhythm, is defined as the ratio of the standard deviation to the mean of the duration of each respiratory cycle within the window. This represents the proportion of inspiratory phase time used to distinguish between thoracic and abdominal breathing characteristics; The peak amplitude variation coefficient of the waveform characterizing the stability of respiratory amplitude; The peak frequency and bandwidth of the power spectral density, reflecting the concentration trend of respiratory rate; and This represents the normalized pulse impulse, which reflects the number of abrupt changes in the waveform slope, and is used to detect abrupt noise caused by sudden coughing, wheezing, or sighing events.

[0033] Based on the aforementioned multidimensional feature vectors, the breathing pattern recognition unit online invokes a pre-trained support vector machine (SVM) classifier to determine and classify the current breathing pattern into a preset category. One of the typical breathing patterns (e.g., steady breathing pattern I, shallow and rapid breathing pattern II, irregular / perturbed breathing pattern III). The classifier outputs a posterior probability estimate or fuzzy membership degree for each pattern, which is used as the pattern assignment weight coefficient. Satisfying the normalization constraint conditions Correspondingly, the dynamic compensation model is pre-configured with multiple sets of parameters trained independently offline during deployment. Each parameter set is obtained through simulation input training under the corresponding breathing mode. During intraoperative online calculation of displacement compensation, the system employs an adaptive smooth interpolation strategy based on mode attribution weights, with the mathematical update expression as follows: Preferably, the respiratory subvector kernel width parameter in formula (2) of Example 1 is... It is also simultaneously changed to an adaptive parameter controlled by weight interpolation: In the formula, In order to prepare for the first The baseline kernel width is determined and stored for each breathing pattern optimization. The technical advantage of this interpolation mechanism is that when the breathing pattern begins to change or the classification confidence is insufficient, adjacent patterns can smoothly transition according to weights, effectively avoiding the compensation step that may be caused by hard switching. Thus, it ensures the temporal continuity and generalization robustness of the dynamic compensation results when the breathing submanifold fluctuates drastically.

[0034] Example 3 Furthermore, based on the technical solutions of Embodiment 1 and Embodiment 2 above, the present invention further expands the registration and correction scope from a single target vertebra to a spinal functional unit covering adjacent segments through Embodiment 3, in order to solve the systemic deviation problem caused by changes in the overall curvature of the spine due to passive adjustment of body position or muscle relaxation during long-segment minimally invasive surgery of multiple segments (such as L3-L5), resulting in relative rotation and translation between adjacent vertebrae.

[0035] Preferably, in this embodiment, the target vertebra and adjacent vertebrae are considered as local rigid bodies, and adjacent vertebrae are connected by intervertebral discs and ligaments, abstracting their biomechanical behavior as an equivalent elastic constraint. Specifically, in the final registration transformation matrix of the target surgical vertebra... Given that the target vertebra has been identified and locked, the feature point selection module adds a multi-segment feature extraction step, which extracts a set of auxiliary feature points from the cortical edge regions of adjacent segments of the target vertebra (such as the inferior endplate of the L3 vertebra and the superior endplate of the L5 vertebra). The spatial location information of these auxiliary feature points during surgery is indirectly obtained through edge detection and stereo matching of two-dimensional perspective images. By constructing the following joint multi-constraint optimization objective function, corresponding correction rigid body space transformation matrices are applied to all adjacent segment models except the target vertebral body. This achieves global elastic curvature correction constrained by prior biomechanical knowledge: In equation (5) above, each mathematical symbol and matrix component has a clear geometric and mechanical calculation definition: This represents the fluoroscopic projection function that maps three-dimensional spatial points onto a two-dimensional plane of intraoperative fluoroscopic images. Its parameter matrix is ​​obtained from the complete geometric calibration process of the C-arm X-ray machine in the anteroposterior and lateral views. For the first The set of coordinates of auxiliary feature points obtained from multiple segmental feature extraction steps on adjacent vertebrae in the three-dimensional model space; These are the two-dimensional feature observation point plane coordinates obtained after edge detection and stereo matching processing of intraoperative fluoroscopic images; To describe the set of topological networks that represent the equivalent biomechanical elastic connectivity between adjacent vertebral bodies, taking L3-L5 cross-segment surgery as an example, it can be explicitly expressed as follows: ; and The first and the The corrected rigid body transformation matrix of each vertebra is to be solved, and the target surgical vertebra transformation matrix has been precisely locked in Implementation Example 1. In this joint optimization, it is treated as a fixed constant and does not participate in the iterative search; and The first and the The three-dimensional coordinates of the geometric center of a cone are specifically defined as the arithmetic mean of the geometric coordinates of all vertices on the triangular mesh model of the cone. The reference relative position vector between the centers of the two vertebral bodies in the preoperative original model is defined as follows: ; The weighting coefficient of the elastic constraint penalty term has the physical and mechanical significance of equivalently simulating and reflecting the inherent physiological and mechanical stiffness of the intervertebral disc-ligament complex. The larger the value, the stronger the algorithm's penalty for deviations from the original preoperative anatomical relative pose, and the more the optimization results tend to maintain the original intersegmental physiological curvature of the spine. The smaller the value, the more the optimization system tends to trust the independent observation information provided by intraoperative real-time fluoroscopic image matching. The typical value range is preferably determined to be 0.5 to 2.0, and the specific value is determined based on preoperative population statistical regression or finite element mechanical simulation.

[0036] Therefore, the first term (projection matching error) of the objective function (5) is driven by the actual two-dimensional image edge during surgery, causing the spatial transformation of adjacent vertebrae to align with the real fluoroscopic projection; the second term (elastic constraint penalty) uses anatomical structural mechanics as a regularization prior to prevent excessive non-physiological relative displacement or misalignment rotation between adjacent vertebrae. The overall optimization balance between the two terms is determined by the weighting coefficients. Dynamic adjustment is applied. Preferably, the system uses the Levenberg-Marquardt algorithm to iteratively solve equation (5) (where the rotational components are parameterized as Lie algebras to ensure the absence of singularities on the rotational group). The optimal correction transformation of adjacent vertebrae is obtained through the solution process. Then, a corresponding rigid transformation is applied to all adjacent segment models except the target vertebral body, while the target vertebral body model remains unchanged by the transformation matrix. The locked precise pose remains unchanged. The updated complete multi-vertebra 3D model is then sent to the registration and fusion module to generate a more comprehensive multi-level virtual safety boundary covering the target segment and adjacent segments, perfectly achieving the mechanical unity of high-precision registration of local target vertebrae and global spinal curvature consistency correction.

[0037] To ensure the stringent real-time performance requirements of all the above embodiments in clinical minimally invasive surgery applications, the computation acceleration module of this invention migrates all computationally intensive tasks, including Gaussian smoothing (parallelizing three-dimensional separable filtering by separating the filter kernel along the three coordinate axes), parallel computation of multi-scale curvature features (accelerated by neighborhood query based on spatial hash index structure), computation of high-dimensional feature distance matrix (calling high-performance matrix operation libraries such as CUDA and CuBLAS), and point-pair search in the ICP algorithm (parallel spatial search based on KD tree structure) and weighted singular value transformation solution (data-parallel reduction algorithm), to the graphics processing unit (GPU) for parallel execution. Under this parallel computing architecture, the execution time of the complete single registration and fusion process, including dynamic error coupling compensation and global elastic constraint correction, is strictly compressed to less than 200 milliseconds (typical average time is only 179ms), fully meeting the stringent clinical requirements of percutaneous endoscopic lumbar discectomy for real-time dynamic tracking and virtual-real fusion stereotactic guidance.

[0038] Example 4 In step S2 of the above embodiment, without external hardware spatial marker constraints, the system uses a deep convolutional neural network (CNN) to extract the pure skeletal edge contour of the target vertebra from intraoperative fluoroscopic images at mutually angled angles, and reconstructs the three-dimensional point cloud structure through multi-view stereo matching under epipolar geometry constraints. However, in actual clinical fluoroscopy, due to the scattering noise of low-dose X-rays, the local gray-level gradient diffusion caused by osteophyte formation at the vertebral edge, and the nonlinear perturbation of the projection matrix caused by the slight intraoperative tremors of the binocular C-arm, traditional stereo matching algorithms based on line-to-line search using rigid epipolar geometry often produce a large number of mismatched points, which in turn leads to artifacts or voids in the three-dimensional reconstructed point cloud near the vertebral foramen edge and facet joint.

[0039] Preferably, in the point cloud reconstruction process of step S2 in this embodiment, the traditional rigid epipolar constraint is extended to a probabilistic epipolar spatiotemporal neighborhood matching model based on manifold constraints. Specifically, the system no longer assumes that the projection matrix is ​​an absolutely static rigid mapping, but models the binocular projection geometry of the C-arm X-ray machine as a nonlinear manifold with perturbation uncertainties.

[0040] When a deep convolutional neural network outputs an orthogonal perspective image Lateral fluoroscopic images After the edge probability map, for the image Any bone edge feature point Its image The corresponding theoretical polar line is ,in The base matrix is ​​used. To compensate for matching drift caused by device jitter and pixel gradient dispersion, this embodiment constructs an epipolar probability space search band based on Mahalanobis distance and kernel density estimation (KDE). Within this search band, the image... Candidate matching points on The mathematical model expression for the complete association of corresponding image points, satisfying the probability density distribution function of the corresponding image points, is as follows: In equation (6) above, the epipolar distance vector function The specific mathematical expression is defined as follows: In equations (6) and (7), the physical and operational definitions of each mathematical symbol and parameter are as follows: Represents the given fundamental matrix Under these conditions, the image Candidate points on With images Feature points on Normalized posterior probability density for the same anatomical landmark (corresponding image point); The first component of the bidirectional asymmetric epipolar distance metric vector represents the point. To the corresponding polar line The algebraic distance, the second component characterizes the first-order geometric disparity correction in the back projection. and These represent the first two algebraic components of the epipolar vector; The intraoperative dynamic covariance matrix is ​​used to quantitatively characterize the image degradation uncertainty caused by the micro-vibration of the C-arm gantry and the pixel spatial location variance introduced by fluoroscopic grayscale noise reduction. This matrix is ​​estimated and updated online in real time during the operation through the constant bone tissue mesh residual of the sterile reference area. It is a feature space similarity modulation operator, specifically defined as the product of the local cosine similarity of two points in the high-level feature map of a deep neural network and the disparity smoothness constraint, used to provide matching support from the perspective of anatomical semantics.

[0041] Furthermore, after solving the probability matching field of the skeletal edge points between images, the system no longer directly calculates the 3D coordinates using the traditional deterministic triangulation formula. Instead, it employs an adaptive view frustum reprojection error minimization algorithm based on maximum a posteriori estimation (MAP). Preferably, the system transforms the reconstruction problem into finding the 3D spatial point in the probability space that minimizes the cumulative reprojection uncertainty of the two perspective images. Its fully correlated parameterized multidimensional energy functional expression is shown in equation (8): In equation (8), For the first A camera projection matrix with a perspective view. It is a nonlinear perspective transformation operator from three-dimensional to two-dimensional; The posterior matching probability derived from equation (6) The pixel weight diagonal matrix obtained by reverse mapping. The surface curvature smoothing regularization coefficient of the point cloud is denoted as . The domain of the local reconstructed manifold representing the surface of the target vertebral body.

[0042] Specifically, by using the Levenberg-Marquardt (LM) optimization algorithm to solve the energy functional (8) in parallel iteration, the point cloud coordinate spikes caused by noise can be dynamically suppressed at the underlying algorithm level. This ensures that even in areas with bone hyperplasia or sparse point cloud, the intraoperative pure bone three-dimensional point cloud structure with excellent topological manifold continuity can still be reconstructed by inverse projection, thereby greatly improving the convergence speed and noise robustness of the heterogeneous coarse and fine registration algorithm in the subsequent steps of Example 1.

[0043] Example 5 In step S4 of the above embodiment, the system introduces a finite element model of spinal canal soft tissue coupled with bony structures, and uses a hybrid radial basis function (RBF) regression model trained based on data-driven and physical mechanics priors to iteratively calculate displacement deformation in the working channel trajectory in real time, so as to generate a dynamically updated model and synchronize the navigation coordinate system. However, in actual surgical operations, when the surgeon uses cutting instruments or dilators to apply strong physical compression or bone resection to the narrow intervertebral foramen, the ligamentum flavum, dural sac, and posterior soft tissues around the lesion vertebra will undergo extremely complex local nonlinear large strain (i.e., a double superposition of geometric nonlinearity and material nonlinearity). At this time, if the regression network is trained purely on offline static simulation samples, it will often cause mesh patch cross-over, topological structure breakage, or geometric distortion (i.e., false penetration) in the agitated area where the instrument edge contacts the soft tissue, resulting in distortion of the shortest spatial distance calculated by the safety warning module.

[0044] Preferably, in the mechanical deformation iterative calculation and coordinate update process of step S4, this embodiment delves into the strain energy dissipation mechanism at the bottom layer of continuous medium mechanics and proposes a triangular mesh topology update algorithm based on Green-Lagrange strain tensor and local volume preservation constraints.

[0045] Specifically, in the preoperative model reconstructed in step S1, the system models the initial triangular mesh surface and its internal elemental units as a continuous medium with viscoelastic mechanical response. When the high-dimensional input vector in step S4... After introducing measured operating loads (such as casing thrust torque and extrusion force), any point inside the unit will be at its initial coordinates. Intraoperative coordinates adjusted to the current moment Its deformation gradient tensor is defined as To accurately describe the geometric invariants of soft tissue under large strain, this embodiment derives the fully correlated Green-Lagrange deformation energy and hyperelastic material kinetic energy balance equation, and its nonlinear finite element dynamic partial differential matrix equation mathematical model is shown in equation (9): In the matrix equation (9) above, the Green-Lagrange strain tensor involved is... With the Second Piola-Kirchhoff Stress Tensor The specific definition of the constitutive relation fully correlated mathematical model expression is as follows: In equations (9), (10), and (11), the physical and engineering calculation definitions of each mathematical term and tensor component are as follows: and These represent the global mass matrix and viscous damping dissipation matrix of the triangular mesh system, respectively, and are used to simulate the dynamic time delay effect and energy dissipation damping of soft tissue under pressure. This represents the cumulative 3D deformation displacement vector of the current triangular mesh node. and These are the first and second time derivatives corresponding to the displacement, respectively; This is the deformation-related nonlinear strain-displacement transformation matrix; The initial geometric volume domain of the element; For unit tensors, For the right Gaugen-Kirchhoff deformation tensor, The rate of change of volume is an invariant; Represents the strain energy density function of Neo-Hookean hyperelastic materials. and These are the soft tissue shear modulus and bulk modulus obtained by converting and mapping preoperative images, respectively. This is a specially designed operator for the incompressible and impenetrable geometric constraints on local cell volume. This is the corresponding penalty factor coefficient, used to forcibly prevent the self-overlap of mesh patches under extreme thrust; The input vector is the result of the high-dimensional operation in step S4. The intraoperative external concentrated stress load vector derived from mapping.

[0046] Furthermore, when performing intraoperative online deformation iteration, the system uses a combination of implicit Newmark time integration step algorithm and Newton-Raphson space iteration method to solve the partial differential matrix equation system (9) at high speed in real time.

[0047] Preferably, within each time step, the algorithm not only calculates the rigid and elastic hybrid displacement of the mesh nodes using Equation (9), but also calculates the shear strain energy dissipation rate of the current local triangular mesh in real time. When the shear distortion energy of a local triangular mesh patch exceeds the material tearing threshold, the topology-preserving element will adaptively trigger local edge flipping and patch remeshing under the constraints of maintaining volume conservation and physical boundary closure, dynamically updating the origin and axis of the navigation coordinate system.

[0048] This physical and mechanical algorithm, based on the balance of kinetic energy and stress energy in a continuous medium under large strain, ensures that even under extremely complex dynamic conditions such as the large-scale expansion and cutting of surgical instruments, the three-dimensional triangular mesh model instance of the navigation is precisely synchronized in spatial morphology with the actual deformation of the dura mater and nerve roots within the spinal canal. This completely eliminates the false dangers or safety blind spots caused by the assumption that the spine is a static rigid body or simple decoupling and superposition in traditional navigation. This lays a solid physical algorithm foundation for the highly specific graded and precise early warning of nerve roots and high-risk soft tissues in the subsequent safety warning module in Embodiment 1.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0050] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An image registration algorithm method for use in percutaneous endoscopic lumbar discectomy, the method comprising: Includes the following steps: Step S1: Acquire preoperative three-dimensional image data and reconstruct a three-dimensional skeletal model of the target vertebra; Step S2: Acquire intraoperative fluoroscopic images and reconstruct the intraoperative three-dimensional point cloud structure; Step S3: Perform initial registration on the 3D skeleton model and the 3D point cloud structure to obtain the initial spatial transformation relationship, and extract the registration feature point set based on the initial spatial transformation relationship to perform fine registration and obtain the fine registration spatial transformation relationship; Step S4: Construct a dynamic error compensation model. Combine the real-time respiratory phase parameters with the surgical operation parameters as the input vector. Predict the three-dimensional displacement compensation amount through a single unified mapping function. Then, use the three-dimensional displacement compensation amount to correct the fine registration spatial transformation relationship and obtain the final registration transformation matrix. Step S5: Use the final registration transformation matrix to spatially align and fuse the 3D skeletal model with the intraoperative images.

2. The algorithm method of image registration in percutaneous endoscopy of intervertebral foramen according to claim 1, characterized in that: In step S2, reconstructing the intraoperative 3D point cloud structure refers to extracting the pure skeletal edge contour of the target vertebra from at least two intraoperative fluoroscopic images at angles to each other using a deep convolutional neural network without external hardware spatial markers, and directly reconstructing the 3D point cloud structure representing the true spatial pose of the bone during surgery through multi-view stereo matching under epipolar geometric constraints.

3. The image registration algorithm method for percutaneous endoscopic discectomy according to claim 1, characterized in that: In step S3, the initial registration specifically includes: Simultaneously, Gaussian smoothing filtering is applied to the 3D skeletal model and 3D point cloud structure at multiple different spatial scales to simultaneously capture the local geometric undulations and overall anatomical orientation of the vertebral surface. Calculate the geometric curvature characteristics of grid nodes at each spatial scale. The geometric curvature characteristics include Gaussian curvature, average curvature, and a high-dimensional local shape histogram descriptor based on the statistical distribution of the angle between the local neighborhood normal vectors. Calculate the similarity matrix between the 3D skeleton model and the 3D point cloud structure in the feature space, match feature descriptors, and use a bidirectional matching strategy and random sampling consistency algorithm to remove abnormal corresponding point pairs; Based on the preserved set of corresponding interior points, the initial rigid space transformation relationship that minimizes the sum of squared Euclidean distances between corresponding point pairs is solved using the singular value decomposition algorithm.

4. The image registration algorithm method for percutaneous endoscopic discectomy according to claim 1, characterized in that: In step S3, the registration feature point set is extracted to perform fine registration, specifically including: Based on prior anatomical knowledge of the spinal structure, control constraint rules are selected, and a set of feature control points that cover the entire spatial circumference and are repeatable across patients is automatically determined on the surface of the coarsely aligned 3D skeletal model. The feature control point set is distributed in the posterior margin region of the target vertebral body, the bilateral pedicle stenosis apex region, the spinous process tip region, the lateral apex region of the transverse process, and the central region of the articular facet. An iterative nearest-point algorithm based on geometric feature distance weighting is executed, using the cosine similarity of point pairs on multi-scale feature vectors as the matching weight. A dual screening mechanism is used to eliminate erroneous correspondences caused by local point cloud missingness or soft tissue occlusion. The cumulative spatial transformation matrix is ​​iteratively updated using the calculated incremental rigid body transformation parameters to obtain the finely registered spatial transformation relationship.

5. The algorithm method of image registration in percutaneous endoscopy of intervertebral foramen according to claim 1, characterized in that: In step S4, the input vector includes at least the instantaneous respiratory phase, the needle insertion angle, the puncture depth, the outer diameter of the working cannula, the cannula advancement rate, and the cumulative parameters representing the duration of the operation. The unified mapping function is a variant structure of the radial basis function network based on data-driven and physical-mechanical prior training, which is used to implicitly establish a cross-coupling model between different respiratory phases and different mechanical compression states of surgical instruments.

6. The algorithm method of image registration in percutaneous endoscopy of intervertebral foramen according to claim 5, characterized in that: The regression equation and the expression for the mixed kernel function of the unified mapping function are as follows: In the formula, This refers to the three-dimensional displacement compensation amount; The input vector is composed of real-time respiratory phase sub-vectors. Surgical operation biomechanical parameter subvector Composed of splicing elements; Representing the The center points of the support vector kernels are determined by pre-updating and training based on an offline high-fidelity finite element simulation database. This is the regression weight vector; This is a balancing weight adjustment factor; and The kernel width parameter is used to adjust the smoothness of the regression space; The total number of core central nodes.

7. The image registration algorithm method for percutaneous endoscopic discectomy according to claim 1, characterized in that: In step S4, the respiratory phase parameters are optimized and corrected in real time by a respiratory pattern recognition unit, specifically including: A sliding window analysis strategy is used to extract continuous respiratory waveform signals, and a multidimensional respiratory feature vector including the period variation coefficient, the proportion of inspiratory phase time, the waveform peak amplitude variation coefficient, the peak frequency of power spectral density, and the number of waveform slope abrupt changes is calculated online. The offline-trained support vector machine classifier performs pattern determination on the multidimensional respiratory feature vector and outputs the pattern assignment weight coefficients for different typical respiratory patterns. Adaptive interpolation updates are performed on multiple pre-set kernel regression parameter sets based on pattern attribution weight coefficients to correct instantaneous registration drift of the dynamic error compensation model during non-stationary breathing switching.

8. The image registration method for percutaneous endoscopy of foraminotomy according to claim 1, wherein: In step S4, a non-rigid space penalty term based on the elastic connection relationship between adjacent vertebral bodies is also introduced, and the registration transformation matrix of the locked target surgical vertebral body is obtained Under the constraints, a joint multi-constraint optimization objective function is constructed: In the formula, For the perspective camera's two-dimensional projection function; This is the set of three-dimensional coordinates of anatomical auxiliary feature points on adjacent non-target vertebrae; These are the corresponding intraoperative two-dimensional feature observation point plane coordinates; It is a set of equivalent biomechanical elastic connections between adjacent vertebral bodies; and Let be the rigid body space transformation matrix of the adjacent vertebrae to be corrected; and These are the coordinates of the geometric center of the cone mesh model; This is the original relative position reference vector before surgery; The weighting coefficient of the elastic constraint penalty term is used to equivalently simulate the physiological stiffness of the intervertebral disc and ligament complex.

9. The image registration method for percutaneous endoscopy of the foramen transpulmonale according to claim 1, characterized in that: In step S5, while the fusion display is being performed, a registration reliability closed-loop continuous monitoring mechanism is initiated to calculate the feature correspondence residual index, the spatial distribution uniformity index of feature control points in the dissected area, and the residual displacement index between consecutive frames after dynamic compensation. When at least two of the three indices exceed their respective preset safety and reliability thresholds, the system determines that the registration has failed, immediately triggers a registration failure warning, and forcibly suspends the fusion display of the augmented reality image.

10. A percutaneous endoscopic discectomy (PED) image registration system for implementing the method described in any one of claims 1 to 9, characterized in that: include: The preoperative image acquisition and reconstruction module is used to receive preoperative three-dimensional image data and extract a smooth, pure bony triangular mesh model of the target vertebral body for reconstruction. The intraoperative image acquisition and point cloud construction module is used to acquire multi-angle fluoroscopic image sequences in surgical scenarios without hardware marker constraints, and reconstruct the intraoperative three-dimensional pure bone point cloud structure based on the depth segmentation matching algorithm. The multi-scale coarse and fine registration module is used to match and solve the initial spatial transformation relationship under high-dimensional shape descriptors at multiple spatial scales, and perform fine registration with feature distance weighting based on anatomical prior feature control points to output the fine registration spatial transformation relationship. The high-dimensional coupled dynamic compensation module is used to construct a radial basis hybrid kernel regression model. It combines the real-time respiratory phase vector with multiple surgical instrument operation mechanical parameters to construct a high-dimensional input vector. It implicitly predicts the three-dimensional displacement compensation amount of coupled deformation with a single mapping function, thereby correcting the output final registration transformation matrix. The spatial alignment and closed-loop monitoring module is used to perform rigid body transformation and deformation rendering to achieve augmented reality 3D stereo fusion guidance, and to perform safe closed-loop disconnection control based on the parallel decision logic of feature correspondence residuals, distribution uniformity and remaining displacement to ensure the accuracy and reliability of the algorithm.

Images