A method and system for processing an intraoral scan image

Abstract

This invention discloses a method and system for processing oral cavity scanning images. The method includes: acquiring three-dimensional scanning data of the oral cavity region of a patient to generate an original grayscale image sequence; extracting an initial coordinate set of multiple anatomical landmarks, including the alveolar ridge crest and nerve canal, from the original grayscale image sequence based on a convolutional neural network; dividing the anatomical landmarks in the initial coordinate set into voxel cubes and filtering low-confidence points; calculating the change in a joint function using the gradient magnitude of the low-confidence points to correct the coordinates of the anatomical landmarks; and obtaining the appropriate implant model based on the anatomical landmarks of the alveolar ridge crest with corrected coordinates. This solution, by combining landmark recognition with gradient optimization correction, not only improves the accuracy and stability of anatomical landmark positioning but also provides a reliable three-dimensional reference for oral implant surgery, helping to reduce implant risks, ensuring that the implant can be selected with a suitable model, and ensuring the long-term stability of the implant.

Description

Technical Field

[0001] This invention relates to the field of oral cavity, and more particularly to a method and system for processing oral cavity scanning images. Background Technology

[0002] Dental implant surgery is a common method for restoring missing teeth. The safety of the surgery and the long-term stability of the implants highly depend on accurate preoperative anatomical information. Traditional oral imaging typically uses two-dimensional X-rays or three-dimensional oral CT scans to obtain images of the alveolar bone and adjacent structures. Two-dimensional images have limitations in planar projection, making it difficult to accurately reflect the three-dimensional spatial positions of key anatomical structures such as the alveolar bone, nerve canal, and maxillary sinus floor, easily leading to deviations between preoperative planning and actual operation.

[0003] Current 3D CT data processing methods primarily rely on manual annotation or simple threshold segmentation to identify structures such as alveolar bone, nerve canals, and the floor of the maxillary sinus. Manual annotation is not only time-consuming but also heavily influenced by the operator's experience, making it difficult to guarantee annotation accuracy and consistency. Simple threshold segmentation methods are easily affected by image noise, bone complexity, and artifacts, leading to inaccurate localization of key anatomical locations. Especially at critical anatomical points such as the alveolar ridge crest, the course of the inferior alveolar nerve canal, the mental foramen, and the floor of the maxillary sinus, even minor deviations can directly affect the angle, direction, and depth of implant placement, increasing the risk of intraoperative nerve damage and implant failure.

[0004] Although Convolutional Neural Networks (CNNs) can automatically identify key anatomical structures and improve the efficiency of landmark localization using large-scale training data in oral medical image analysis, they still have shortcomings in practical applications. For some regions with unclear grayscale information or complex structures, the initial anatomical landmarks generated by CNNs may exhibit positional deviations or instability, especially in low-confidence regions, potentially deviating from the true anatomical location. Therefore, further optimization and correction of the initial coordinates are needed to ensure the spatial accuracy and reliability of the anatomical landmarks. This also necessitates determining the appropriate implant type based on the anatomical landmarks and oral CT scan images. In summary, overcoming these technical problems is a pressing issue that those skilled in the art must address. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for processing oral cavity scan images, which solves the above-mentioned technical problems pointed out in the prior art.

[0006] This invention provides a method for processing oral cavity scan images, the specific operation steps of which include:

[0007] Three-dimensional scanning data of the patient's oral cavity region is collected to generate an original grayscale image sequence; an initial coordinate set of multiple anatomical landmarks, including the alveolar ridge crest and nerve canal, is extracted from the original grayscale image sequence based on a convolutional neural network;

[0008] The anatomical landmarks in the initial coordinate set are used to divide the voxel cubes, and low-confidence points are selected. The gradient magnitude of the low-confidence points is used to calculate the change in the joint function and correct the coordinates of the anatomical landmarks. The appropriate implant model is obtained based on the anatomical landmarks of the alveolar ridge crest with the corrected coordinates.

[0009] Preferably, the anatomical landmarks in the initial coordinate set are used to divide the data into voxel cubes, and low-confidence points are selected. The gradient magnitudes of the low-confidence points are used to calculate the change in the joint function, and the coordinates of the anatomical landmarks are corrected. The specific steps are as follows:

[0010] Select any one of the anatomical landmarks in the initial coordinate set as the central anatomical landmark, and divide the voxel cube with the central anatomical landmark; traverse each voxel in the neighborhood of the voxel cube, determine the gray value of the voxel in each neighborhood, and construct a local gray-level distribution matrix.

[0011] The local gray-level distribution matrix is ​​decomposed using principal component analysis to reduce its dimensionality, yielding the maximum and minimum gray-level eigenvalues; the gray-level eigenvalue ratio between the maximum and minimum gray-level eigenvalues ​​is then calculated.

[0012] A low confidence threshold is preset; it is then determined whether the gray-scale feature ratio is less than the low confidence threshold.

[0013] If so, then the central anatomical landmark is classified as a low-confidence point;

[0014] For each low-confidence point in the neighborhood, the gradient magnitude in each direction is calculated, and the maximum gradient magnitude is selected as the image feature term; the prior position of the anatomical landmark point in the neighborhood is obtained; and the prior position is converted into an anatomical constraint term.

[0015] The image feature terms and anatomical constraint terms are weighted and combined to obtain the first joint function;

[0016] Preset convergence threshold and search step size; use the quasi-Newton method to perform a one-dimensional line search on the gradient principal direction of the current low confidence point according to the search step size to determine the latest position of the current low confidence point and obtain candidate new points; return the candidate new points to the above steps, use them as new central dissection landmarks to divide the neighborhood, and calculate the second joint function;

[0017] The change in the joint function is obtained by calculating the first joint function and the second joint function;

[0018] Determine whether the change in the joint function is less than the convergence threshold;

[0019] If yes, stop the iteration and output the coordinates of the candidate new point as the new coordinates of the central anatomical landmark; if no, return the candidate new point to the above steps and continue iterating until the change of the joint function converges to obtain the new coordinates of the central anatomical landmark.

[0020] Preferably, the appropriate implant model is obtained based on the anatomical landmarks of the alveolar ridge crest of the calibration coordinates. The specific operation steps are as follows:

[0021] A historical oral anatomy standard size library is pre-built, and gray-level histograms of anatomical landmarks are obtained. A sliding window is applied to the gray-level histograms to construct kernel density curves and further calculate local minima. Based on the local minima, the bone tissue boundary contours corresponding to the voxels are extracted. Based on the type of the anatomical landmarks, voxels of the inferior wall of the maxillary sinus are identified within the bone tissue boundary contours, and the normal vector of the sinus wall center point is calculated. Based on the normal vector of the sinus wall center point, the voxel direction of the alveolar bone is searched, and the vertical thickness and gray-level change rate are calculated. The vertical thickness and gray-level change rate are normalized to obtain a multidimensional distance feature vector. Based on the multidimensional distance feature vector and the anatomical landmarks, the axial direction vector of the anatomical landmarks is determined. Based on the anatomical landmarks of the alveolar ridge crest along the direction of the axial direction vector, implant simulation is performed to obtain the appropriate implant model.

[0022] Preferably, a historical oral anatomy standard size library is pre-constructed, gray-level histograms of anatomical landmarks are obtained, a sliding window is applied to the gray-level histograms, a kernel density curve is constructed, and local minima are further calculated; the bone tissue boundary contours corresponding to the voxels are extracted based on the local minima. The specific operation steps are as follows:

[0023] A historical oral anatomy standard size library is pre-constructed; the point type of the anatomical landmark is determined according to the historical oral anatomy standard size library, and the preset radius and shape of the corresponding point type are extracted to form a spherical voxel set; the gray value occurrence frequency of the voxels in the spherical voxel set is counted to construct a gray-level histogram; a sliding window is applied to the gray-level histogram, and the kernel density is calculated by weighted averaging of the gray values ​​of the voxels in each sliding window; the kernel density curve is obtained by fitting the voxel kernel densities of all sliding windows.

[0024] The minimum gray value of adjacent voxels in the kernel density curve is selected as the local minimum; the minimum average value of all local minimum values ​​is calculated; if the voxel gray value is greater than or equal to the minimum average value, the voxel is determined to be bone tissue; if the voxel gray value is less than the minimum average value, the voxel is determined to be soft tissue; the voxels of bone tissue and soft tissue are combined to form a voxel binary image; morphological closing operation is performed on the voxel binary image to extract the boundary contour of bone tissue.

[0025] Preferably, based on the type of the anatomical landmark, the voxels of the inferior wall of the maxillary sinus in the bone tissue boundary contour are identified, and the normal vector of the sinus wall center point is calculated; the voxel direction of the alveolar bone is searched based on the normal vector of the sinus wall center point, and the vertical thickness and grayscale change rate are calculated; the vertical thickness and grayscale change rate are normalized to obtain a multidimensional distance feature vector; the axial direction vector of the anatomical landmark is determined based on the multidimensional distance feature vector and the anatomical landmark; implant simulation is performed along the direction of the axial direction vector of the anatomical landmark of the alveolar ridge crest to obtain the appropriate implant model. The specific operation steps are as follows:

[0026] Based on the type of anatomical landmarks, voxels of the inferior wall of the maxillary sinus in the bone tissue boundary contour are identified; the coordinates of the center point of the sinus wall are obtained for all voxels of the inferior wall of the maxillary sinus, forming a set of three-dimensional coordinate points of the sinus wall;

[0027] In the bone tissue boundary contour, the voxels adjacent to the three-dimensional coordinate point set of the inferior wall of the maxillary sinus are extracted to obtain the voxels of the alveolar bone, and the coordinates of the alveolar bone center point of the alveolar bone voxel are obtained to form a three-dimensional coordinate point set of the alveolar bone.

[0028] Fit the vertical direction of the sinus wall center point in the set of three-dimensional coordinate points of the sinus wall, and calculate the normal vector of the sinus wall center point;

[0029] Starting from the center point of the sinus wall, construct a vector ray along the normal vector. In the set of three-dimensional coordinate points of the alveolar bone, calculate the coordinate distance between the coordinates of each alveolar bone center point and the coordinates of the vector ray, and search for the alveolar bone center point with the closest coordinate distance as the first node of the alveolar bone. Calculate the straight-line distance between the coordinates of the first node of the alveolar bone and the coordinates of the center point of the sinus wall as the vertical thickness.

[0030] Connect the first node of the alveolar bone to the center point of the sinus wall to form a three-dimensional line segment; uniformly sample the three-dimensional line segment according to a pre-set fixed step size to obtain a sampling point sequence;

[0031] The grayscale value of each sampling point is calculated, and the standard deviation of the grayscale value of the three-dimensional line segment is calculated as the grayscale change rate; the vertical thickness of the straight-line distance to the center point of the sinus wall in the three-dimensional coordinate point set of the sinus wall and the grayscale change rate of the three-dimensional line segment are obtained; the mean of all vertical thicknesses and the mean of all grayscale change rates are calculated, and the normalized weighted sum is obtained to obtain a two-dimensional distance vector; based on the two-dimensional distance vector and the three-dimensional coordinates of the center point of the sinus wall;

[0032] A three-dimensional mesh implantation model is constructed using the anatomical landmarks and multidimensional distance feature vectors. An initial implantation tendency vector for the anatomical landmarks at the alveolar ridge crest is obtained based on the three-dimensional mesh implantation model, and buccal-lingual and mesiodistal angles are preset. Preset candidate angle combinations are performed based on the buccal-lingual and mesiodistal angles, and a sparsity index is calculated. The optimal Pareto candidate angle combination is selected based on the sparsity index, and an axial direction vector is calculated. Using the anatomical landmarks at the alveolar ridge crest as the intraosseous entry point coordinates, implant fitting simulation is performed along the axial direction vector to obtain the appropriate implant model.

[0033] Preferably, a three-dimensional mesh implantation model is constructed using the anatomical landmarks and multidimensional distance feature vectors; the initial implantation tendency vector of the anatomical landmarks on the alveolar ridge crest is obtained based on the three-dimensional mesh implantation model, and the buccal-lingual angle and mesiodistal angle are preset; a preset candidate angle combination is performed based on the buccal-lingual angle and mesiodistal angle, and the sparsity index is calculated. The specific operation steps are as follows:

[0034] The anatomical landmarks are used as anchor points in three-dimensional space; a three-dimensional mesh implantation model is constructed based on the multi-dimensional distance feature vector and the anchor points; the anatomical landmark of the alveolar ridge is found in the three-dimensional mesh implantation model as the simulation center point, and a quadratic surface is fitted to all mesh points in the three-dimensional mesh implantation model within a preset neighborhood radius;

[0035] For each grid point in the quadratic surface, obtain the corresponding voxel gray value, and calculate the principal curvature based on the voxel gray value; select the minimum principal curvature as the bone filling direction;

[0036] A vector ray is drawn along the bone filling direction to serve as the initial implantation tendency vector;

[0037] The major axis direction of the coordinates of the teeth in the vicinity of the anatomical landmark of the alveolar ridge crest is obtained from the three-dimensional mesh implantation model; the direction of travel of the anatomical landmark of the nerve canal is obtained.

[0038] Preset the buccal-lingual angle and mesio-distal angle based on the initial implantation tendency vector;

[0039] A preset number of candidate angle combinations is determined based on the buccal-tongue angle and the mesio-distal angle.

[0040] The implantation direction vector for each candidate angle combination is calculated based on the initial implantation tendency vector and the buccal-lingual angle and mesiodistal angle. The vector deviation for each candidate angle combination is calculated using the implantation direction vector and the initial implantation tendency vector. The directional angle between the vector deviation and the major axis direction is calculated for each candidate angle combination.

[0041] If the vector deviation and direction angle of the candidate angle combination are greater than those of another candidate angle combination, then the candidate angle combination is taken as the first frontier; the remaining candidate angle combinations are taken as the second frontier; and the sparsity index is calculated for the candidate angle combinations in the same frontier.

[0042] Preferably, the axis direction vector is calculated by selecting the optimal Pareto candidate angle combination based on the sparsity index. The specific operation steps are as follows:

[0043] In the first frontier, candidate angle combinations with the maximum sparsity index are selected as guiding directions; new candidate angle combinations are generated using the candidate angle combinations according to the guiding direction and a preset search step size; a neural tube avoidance threshold is preset according to the traveling direction; and the corresponding implantation direction vector of the new candidate angle combination is calculated.

[0044] Using the new candidate angle combination, the implant is simulated along the corresponding implantation direction vector, and the shortest distance between the implant and the anatomical landmark of the neural tube is calculated; if the shortest distance is greater than or equal to the neural tube avoidance threshold, the optimal Pareto is obtained.

[0045] If the latest position of the candidate angle combination is less than the neural tube avoidance threshold, then the search step size is reset for each candidate angle combination and the iteration continues until the optimal Pareto is obtained.

[0046] When the number of iterations equals the preset iteration threshold, the iteration stops; determine the ideal candidate angle combination, calculate the Euclidean distance between all the optimal Pareto values ​​and the ideal candidate angle combination, and select the candidate angle combination corresponding to the minimum Euclidean distance; calculate the axis direction vector based on the candidate angle combination corresponding to the minimum Euclidean distance and the initial implantation tendency vector.

[0047] Preferably, using the anatomical landmark of the alveolar ridge crest as the coordinate of the intraosseous entry point, implant fitting simulation is performed along the vector direction of the axis to obtain the fitting implant model. The specific operation steps are as follows:

[0048] The anatomical landmarks of the alveolar ridge crest are used as the coordinates of the intraosseous entry point.

[0049] Construct an implant model adaptation library; starting from the coordinates of the intraosseous inlet point, select implants from the implant model adaptation library and simulate the extension of the axial direction vector along the preset implant standard length to obtain the implant central axis; according to the preset path sampling step size, uniformly sample the implant central axis to obtain multiple axis sampling nodes; construct a cross-section perpendicular to the implant central axis for each axis sampling node to form a bone cross-section sampling plane;

[0050] According to the preset circular radius, a circular region is defined on the bone cross-section sampling plane; the voxel gray value of each circular region is counted, and the average value is calculated as the bone density value;

[0051] Calculate the shortest distance between the circular region and the anatomical landmark;

[0052] Preset bone quality threshold and safety distance threshold; if the bone density value is less than the bone quality threshold, the simulated implantation is determined to have a risk type; return to the above steps, reselect the implant from the implant model adaptation library until the simulated implantation bone density value is less than the bone quality threshold;

[0053] If the shortest distance is less than the safe distance threshold, the implantation is determined to be risky; return to the above steps, reselect the direction of the principal curvature, adjust the initial implantation tendency vector until the implantation is completed, and record the adaptation parameters.

[0054] Preferably, the adaptation parameters include the coordinates of the intraosseous entry point, candidate angle combinations and axial direction vectors, and the type of implant to be adapted.

[0055] Accordingly, the present invention also proposes an oral cavity scanning image processing system, comprising: an acquisition module; and an analysis module;

[0056] The acquisition module is used to acquire three-dimensional scanning data of the patient's oral cavity area and generate an original grayscale image sequence; and to extract an initial coordinate set of multiple anatomical landmarks, including the alveolar ridge crest and nerve canal, from the original grayscale image sequence based on a convolutional neural network.

[0057] The analysis module is used to filter anatomical landmarks in the initial coordinate set to divide voxel cubes and filter low-confidence points; to calculate the change in the joint function using the gradient magnitude of the low-confidence points and to correct the coordinates of the anatomical landmarks; and to obtain the appropriate implant model based on the anatomical landmarks of the alveolar ridge crest with the corrected coordinates.

[0058] Compared with the prior art, the embodiments of the present invention have at least the following technical advantages:

[0059] Analysis of the oral cavity scanning image processing method and system provided by the present invention shows that, in specific applications, a three-dimensional scan of the patient's oral cavity is performed to generate an original grayscale image sequence. Each voxel represents a tiny cube and its density grayscale value. Through layer-by-layer slicing, a continuous two-dimensional image sequence is formed, constituting a three-dimensional voxel block. The complete three-dimensional grayscale data is input into a pre-trained CNN model to automatically extract the initial three-dimensional coordinates of key anatomical landmarks such as the alveolar ridge crest, mandibular nerve canal, maxillary sinus floor, and mental foramen. The alveolar ridge crest serves as the entry reference point for implant placement, the nerve canal path is used to avoid nerve damage during surgery, the maxillary sinus floor is used to assess bone height, and the mental foramen indicates the location of the nerve exit.

[0060] Furthermore, this scheme maps the initial marker points to a three-dimensional voxel space and filters out low-confidence points; it uses the voxel gray-level gradient magnitude to calculate the change of the joint function and performs local optimization correction on the low-confidence points, achieving precise fine-tuning of the marker point coordinates; this step can effectively correct the initial coordinate deviation caused by scanning noise, complex bone, or image artifacts, making the positions of key anatomical points such as alveolar ridge crest and nerve canal more closely match the real anatomical structure;

[0061] Furthermore, this scheme generates implant parameters, including implantation angle, direction, and depth, using the corrected alveolar ridge coordinates to ensure that the implant path planning conforms to the actual anatomy, significantly improving implantation accuracy and surgical safety. This scheme combines the identification of landmarks with gradient optimization correction, which not only improves the accuracy and stability of anatomical landmark positioning but also provides a reliable three-dimensional reference for oral implant surgery, helping to reduce implantation risks, ensuring that the implant can be selected with a suitable model, and ensuring the long-term stability of the implant. Attached Figure Description

[0062] Figure 1 This is a flowchart of the main process of an oral cavity scan image processing method according to Embodiment 1;

[0063] Figure 2 This is a flowchart of the joint function variation of an oral cavity scan image processing method according to Embodiment 1;

[0064] Figure 3 This is a flowchart of a multidimensional distance vector in an oral cavity scan image processing method according to Embodiment 1;

[0065] Figure 4 This is a schematic diagram of the valley points in an oral cavity scan image processing method according to Embodiment 1;

[0066] Figure 5 This is a schematic diagram of the vector rays in an oral cavity scan image processing method according to Embodiment 1;

[0067] Figure 6This is a flowchart illustrating the shortest distance in an oral cavity scan image processing method according to Embodiment 1.

[0068] Figure 7 This is a schematic diagram of the implantation angle for an oral cavity scanning image processing method according to Embodiment 1.

[0069] Figure 8 This is a flowchart of an oral cavity scanning image processing system according to Embodiment 2;

[0070] Labels: Acquisition Module 10; Analysis Module 20. Detailed Implementation

[0071] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0072] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0073] Example 1

[0074] like Figure 1 As shown, the present invention provides a method for processing oral cavity scan images, the specific operation steps of which are as follows:

[0075] S10: Collect three-dimensional scanning data of the patient's oral cavity area to generate an original grayscale image sequence; extract the initial coordinate set of multiple anatomical landmarks, including the alveolar ridge crest and the nerve canal, from the original grayscale image sequence based on a convolutional neural network (i.e., the initial three-dimensional coordinates of the anatomical landmarks automatically identified by the convolutional neural network model); In this scheme, although the nerve itself is soft tissue, the nerve canal, as an anatomical structure, has very clear features in the images, divided into the inferior alveolar nerve (i.e., the actual nerve fiber bundle, which is indeed soft tissue and is not directly visualized on CT) and the mandibular nerve canal (i.e., a bony canal within the mandible with a clear cortical bone wall), forming very obvious ring-shaped or tubular features on CT images;

[0076] It should be noted that in the collected 3D data, each small cube is called a voxel; each voxel has a value representing the density of that point, which is the grayscale value; the system slices each small cube of the 3D data layer by layer to form multiple consecutive superimposed 2D images, and each 2D image is a bone section sampling plane (that is to say, the entire cube is 3D data, which is formed by the continuous superposition of a bunch of 2D images).

[0077] Prepare a convolutional neural network model that has been trained on tens of thousands of oral CT images to automatically pinpoint key anatomical locations on the patient's new data (e.g., the highest point of the alveolar bone, the center point of the neural canal; the extracted anatomical landmarks are in three-dimensional coordinates because each original grayscale image sequence is a three-dimensional voxel block, forming a three-dimensional whole, while individual original grayscale images are two-dimensional images; therefore, the entire original grayscale image sequence (i.e., a three-dimensional voxel block) is input into the model, and the anatomical landmarks in three-dimensional coordinates are output). The anatomical landmarks include the alveolar ridge crest, the highest point of the alveolar bone, which serves as the entry point reference for implantation; the neural canal, specifically the inferior alveolar nerve canal, the neurovascular bundle channel within the mandible, with key points along its path (such as the mental foramen and mandibular foramen) to avoid nerve damage during surgery; the floor of the maxillary sinus, the inferior wall of the maxillary sinus, used to assess available bone height; and the mental foramen, the nerve exit point on the anterolateral side of the mandible.

[0078] S20: Filter the anatomical landmarks in the initial coordinate set to divide the voxel cube, and filter the low confidence points; use the gradient magnitude of the low confidence points to calculate the change of the joint function and correct the coordinates of the anatomical landmarks; obtain the appropriate implant model based on the anatomical landmarks of the alveolar ridge crest with corrected coordinates.

[0079] It should be noted that the initial coordinate set is mapped onto a voxel cube (i.e., a voxel is the smallest cubic unit in 3D data, with a corresponding grayscale value), and low-confidence points are filtered out. For low-confidence points, the change in the joint function is calculated through the voxel grayscale gradient magnitude (which can be understood as a local optimization algorithm) to correct the 3D coordinates of anatomical landmarks. The corrected alveolar ridge crest coordinates are used as the implantation reference point to generate implantation parameters (such as implantation angle, direction, and implant type (length, diameter)). This corrects inaccurate or unstable points in the initial CNN output, improves the accuracy of landmarks, ensures the reliable position of the alveolar ridge crest, and avoids implantation angle or position errors caused by initial coordinate deviations. The resulting adaptation parameters are more consistent with the actual anatomy, providing a solid foundation for implant path planning. At the same time, the precise correction of this scheme can ensure that the implantation position is completely matched with the patient's anatomical structure, thereby significantly improving the stability of the implant and avoiding nerve damage or implantation failure due to improper operation.

[0080] In particular, low-confidence points may deviate from their actual anatomical positions due to scanning noise, complex bone structure, and image artifacts. The gradient information of voxel grayscale is used to fine-tune the position of the landmark points to make them more consistent with the actual bone surface or the direction of the nerve canal. Even small deviations in the position of alveolar ridge crest, nerve canal, etc., may affect implant stability or cause nerve damage. The corrected landmark points provide a safe and reliable implantation reference.

[0081] like Figure 2As shown, in step S20, anatomical landmarks in the initial coordinate set are screened to divide voxel cubes, and low-confidence points are screened; the gradient magnitude of the low-confidence points is used to calculate the change in the joint function to correct the coordinates of the anatomical landmarks; the appropriate implant model is obtained based on the anatomical landmarks of the alveolar ridge crest with corrected coordinates. The specific operation steps are as follows:

[0082] S21: Select any one of the anatomical landmarks in the initial coordinate set as the central anatomical landmark, and divide the voxel cube (e.g., 10×10×10, the size of this area is usually preset and slightly larger than the size of the target anatomical structure) using the central anatomical landmark.

[0083] Traverse each voxel in the neighborhood of the voxel cube, determine the gray value of each voxel in the neighborhood, and construct a local gray distribution matrix (that is, include the coordinate difference of each voxel, form an N×3 matrix, the number of rows is equal to the number of voxels in the neighborhood N; the matrix records the coordinates and gray value of each voxel in the voxel cube).

[0084] Principal component analysis (PCA) is used to perform dimensionality reduction decomposition on the local gray-level distribution matrix to obtain the maximum and minimum gray-level eigenvalues ​​(i.e., the covariance matrix of the local gray-level distribution matrix is ​​calculated, and then dimensionality reduction decomposition is performed to obtain three gray-level eigenvalues ​​λ1≥λ2≥λ3; λ1 (the maximum eigenvalue, representing the maximum eigenvalue obtained from PCA of the local gray-level distribution matrix, representing the direction and intensity of the most drastic gray-level change) and λ3 (the minimum eigenvalue, representing the minimum eigenvalue obtained from PCA, representing the direction and intensity of the most gradual gray-level change) are selected).

[0085] Calculate the gray-scale feature ratio between the maximum and minimum gray-scale feature values ​​(i.e., the ratio of the maximum to the minimum gray-scale feature values ​​(possibly λ3 / λ1), used to determine the confidence level of the marker point);

[0086] A low confidence threshold is preset (i.e., the threshold needs to be empirically calibrated based on image quality and anatomical location, and is used to compare with the gray-scale feature ratio to determine whether the point is a low confidence point); it is then determined whether the gray-scale feature ratio is less than the low confidence threshold.

[0087] If yes, then the central anatomical landmark is classified as a low-confidence point (i.e., if no; the central anatomical landmark is considered to be located in a well-defined area, and the original coordinates are accepted; that is, if the ratio is very small (e.g., close to 0), it means that the difference between the direction of the greatest change and the direction of the least change of the central anatomical landmark is huge, which means that the central anatomical landmark is likely located on a very clear surface structure (such as the surface of the bone cortex), and is represented as a high-confidence landmark; if the ratio is close to 1 (i.e., λ1≈λ2≈λ3), it means that the gray-level changes in each direction are similar, which means that the central anatomical landmark is located in a uniform area (such as inside soft tissue), and has low confidence; that is, low-confidence points are anatomical landmarks that may be inaccurately located, determined based on the gray-level feature ratio).

[0088] It should be noted that, in the above steps, by screening initial marker points and calculating the feature values ​​of grayscale changes, high-confidence regions and low-confidence regions in the image can be identified. The grayscale feature ratio reflects whether certain regions in the image have clear boundaries or uniform tissue structure. For example, edge regions usually have large grayscale changes and are therefore judged as high-confidence points, while uniform soft tissue is judged as low-confidence regions.

[0089] Anatomical structures in the original grayscale image sequence have different grayscale features at edges (such as bone surfaces) and in uniform regions (such as soft tissues); the step calculates the maximum and minimum grayscale feature values ​​to effectively determine whether the landmark is located at a clear boundary or in a uniform region, thereby evaluating the confidence of the landmark.

[0090] S22: For each low-confidence point in a neighborhood, calculate the gradient magnitude in each direction (i.e., a scalar value reflecting the intensity of gray-level changes in the image; the gradient magnitude is larger at the edges), and select the maximum gradient magnitude as the image feature term (i.e., the direction with the largest gradient magnitude in all directions within a certain neighborhood, representing the most likely boundary direction, which represents the most likely boundary direction in the image; the goal is to make the optimized low-confidence point fall on the ridge with the largest gradient, which also uses local gray-level gradient information to guide the point to move towards the image's significant feature area (such as the edge); that is, this term guides the low-confidence point to move towards the area with the most drastic gray-level changes in the image (i.e., the bone edge); this is achieved by calculating the gradient magnitude in the neighborhood of the low-confidence point, in order to make the low-confidence point tend towards the position with the larger gradient during the minimization process).

[0091] The process involves obtaining the prior positions of anatomical landmarks in the neighborhood (i.e., predicting the reasonable location a point should be in based on anatomical knowledge or relationships with neighboring points; for example, if correcting an anatomical landmark on the mandibular canal, its position should not deviate too far from the trend of the line connecting several adjacent canal points, nor should it extend outside the bone); converting this prior position into anatomical constraints (i.e., distance constraints, angle constraints, etc., between the current point and neighboring points, as long as they can be converted into a mathematical form; simultaneously, based on the prior positional relationships of neighboring anatomical landmarks, constraining the movement range and anatomical rationality of the point; calculating the specific value of the energy function at the current point and its gradient direction; this term ensures that the optimized point does not deviate too far from its proper anatomical structure, maintaining overall anatomical rationality; the specific form of this term depends on the type of landmark and its relationship with neighboring points).

[0092] The image feature terms and anatomical constraint terms are weighted and combined to obtain the first joint function (that is, for each anatomical landmark determined to have low confidence, an energy function needs to be constructed, and its more precise location is found by minimizing the function; this function is called the joint function; the function formed by the weighted combination of image feature terms and mathematical constraint terms is the objective function of the process of optimizing the coordinates of anatomical landmarks).

[0093] It should be noted that the above steps, by calculating the gradient magnitude of low-confidence points and guiding these points toward the most prominent boundary regions (such as bone edges) in the image, help to accurately calibrate the positions of these landmark points, making them more consistent with anatomical structures. The gradient magnitude in the steps reflects the intensity of grayscale changes in the image, which are usually most dramatic in the edge regions of anatomical structures. Using gradient information to guide the optimized points toward these obvious boundary regions improves the accuracy of landmark point localization.

[0094] S23: Preset convergence threshold (i.e., a preset small value; when the change of the joint function is less than this value, the iteration stops, and it is considered that the point has converged to the optimal position) and search step size (i.e., the step size moved along the search direction each time during the optimization search process).

[0095] The quasi-Newton method is used to perform a one-dimensional line search on the gradient principal direction of the current low confidence point according to the search step size, to determine the latest position of the current low confidence point and obtain candidate new points (i.e., new position points obtained by moving the current point during the optimization iteration process).

[0096] Return the candidate new point to step S21 above, divide the neighborhood as a new central anatomical landmark, and calculate the second joint function;

[0097] The first joint function and the second joint function are calculated to obtain the change in the joint function (i.e., the absolute value of the difference between the joint function values ​​before and after the iteration, which is used to determine whether the optimization has converged).

[0098] Determine whether the change in the joint function is less than the convergence threshold;

[0099] If so, stop the iteration and output the coordinates of the candidate new point as the new coordinates of the central anatomical landmark (that is, the coordinates of all anatomical landmarks have been corrected to new coordinates).

[0100] If not, return the candidate new point to step S22 above and continue iterating until the change of the joint function converges to obtain the new coordinates of the central anatomical landmark point;

[0101] It should be noted that in the above steps, the positions of low-confidence points are iteratively updated using optimization algorithms such as the quasi-Newton method, gradually optimizing their coordinates. The quasi-Newton method ensures that each iteration approaches the correct direction by calculating the joint function change between the current point and the candidate new point, until convergence. The quasi-Newton method is an efficient numerical optimization method that can accurately update the positions of anatomical landmarks through gradients and second-order information. It is used to automatically find the optimal landmark positions in cases of complex anatomical structures, thereby improving the accuracy of image analysis.

[0102] S24: A historical oral anatomy standard size library is pre-constructed; grayscale histograms of anatomical landmark types are obtained; a sliding window is applied to the grayscale histograms to construct kernel density curves and further calculate local minima; the bone tissue boundary contours corresponding to the voxels are extracted based on the local minima; voxels of the inferior wall of the maxillary sinus are identified in the bone tissue boundary contours based on the types of the anatomical landmarks, and the normal vector of the sinus wall center point is calculated; the voxel direction of the alveolar bone is searched based on the normal vector of the sinus wall center point, and the vertical thickness and grayscale change rate are calculated; the vertical thickness and grayscale change rate are normalized to obtain a multidimensional distance feature vector; the axial direction vector of the anatomical landmark is determined based on the multidimensional distance feature vector and the anatomical landmark; implant simulation is performed along the axial direction vector of the anatomical landmark of the alveolar ridge crest to obtain the appropriate implant model;

[0103] It should be noted that the above steps, through the analysis of a pre-constructed standard size library and local grayscale histograms, can identify the boundary contours of bone tissue and further determine the position and orientation of anatomical landmarks. This is crucial for the subsequent implantation process, enabling precise positioning and labeling of the target anatomical structure. The historical oral anatomy standard size library provides standard anatomical structure information. Through the analysis of grayscale histograms and local minima, it is used to effectively extract the contours of bone tissue. The standardization process, combined with actual image information, can ensure the accuracy and stability of the landmarks.

[0104] The coordinates of the landmarks mentioned above are not enough to plan a safe and effective implantation path. This step aims to extract the axial direction vector for guiding implant placement based on the corrected anatomical landmarks, especially the inferior wall of the maxillary sinus and the alveolar ridge crest, by analyzing the boundary features, thickness and grayscale changes of local bone tissue. This provides key directional basis for subsequent implantation planning.

[0105] like Figure 3 As shown, in step S24, a historical oral anatomy standard size library is pre-constructed, and grayscale histograms of anatomical landmarks are obtained. A sliding window is applied to the grayscale histograms to construct kernel density curves and further calculate local minima. Based on the local minima, the bone tissue boundary contours corresponding to the voxels are extracted. Based on the type of the anatomical landmarks, voxels of the inferior wall of the maxillary sinus are identified within the bone tissue boundary contours, and the sinus wall center point normal vector is calculated. Based on the sinus wall center point normal vector, the voxel direction of the alveolar bone is searched, and the vertical thickness and grayscale change rate are calculated. The vertical thickness and grayscale change rate are normalized to obtain a multidimensional distance feature vector. Based on the multidimensional distance feature vector and the anatomical landmarks, the axial direction vector of the anatomical landmarks is determined. Based on the anatomical landmarks of the alveolar ridge crest along the direction of the axial direction vector, implant simulation is performed to obtain the appropriate implant model. The specific operation steps are as follows:

[0106] S241: Pre-construct a historical oral anatomy standard size library (i.e., the database contains, for example, a sphere with a radius of 8mm centered on the maxillary sinus floor point; and a cylinder with a radius of 3mm along the nerve direction centered on the mandibular nerve canal point; where the region of interest is represented as a dynamically generated three-dimensional space centered on the corrected anatomical landmarks, that is, the region divided from the original image).

[0107] The point type of the anatomical landmark is determined based on the historical oral anatomy standard size library, and the preset radius and shape of the corresponding point type are extracted (i.e., with the above example as a reference) to form a set of spherical voxels;

[0108] The grayscale values ​​of voxels in the spherical voxel set are counted to determine the frequency of occurrence, and a grayscale histogram is constructed (i.e., there are usually several peaks or troughs, for example, the lowest peak corresponds to air (low grayscale value); the middle peak corresponds to soft tissue, fat or muscle; the higher peak corresponds to cancellous bone; and the highest peak corresponds to compact bone).

[0109] A sliding window is applied to the grayscale histogram, and the kernel density is calculated by weighted averaging of the grayscale values ​​of the voxels in each sliding window. The kernel density of all voxels in the sliding windows is then fitted to obtain a kernel density curve (i.e., a continuous curve obtained after smoothing the grayscale histogram (kernel density estimation), representing the probability distribution of grayscale values; since the original histogram is based on discrete grayscale values, the lines often appear jagged and not smooth; if one directly searches for troughs in these jagged edges, it is easy to find many false troughs; therefore, kernel density estimation can be understood as using a smooth window to slide on the histogram, weighted averaging the values ​​of each point, thereby generating a smooth curve; it more realistically reflects the probability distribution of various grayscale values ​​in the region; high points on the curve represent high probabilities of the grayscale value, and low points represent low probabilities).

[0110] The minimum gray value among adjacent voxels in the nuclear density curve is selected as the local minimum (i.e., the trough). Troughs on the nuclear density curve typically represent the gray-level boundaries between different tissues (such as bone and soft tissue). Figure 4 (as shown)

[0111] It should be noted that the above steps construct a dataset containing historical standard sizes of oral anatomy (such as the inferior wall of the maxillary sinus and the mandibular canal) to improve the accuracy of landmark localization. This is because the standard size library allows for the use of appropriate region extraction methods (such as spheres or cylinders) for different types of anatomical structures. Dynamically generating three-dimensional subspaces makes the processing more flexible and adaptable to different patients or scan data.

[0112] S242: Calculate the minimum average of all local minima (i.e., represent the average of all local minima as an adaptive threshold for binarizing voxels into bone tissue (1) and soft tissue (0));

[0113] If the voxel gray value is greater than or equal to the minimum mean, then the voxel is determined to be bone tissue (that is, the voxel gray value is marked as 1).

[0114] If the voxel gray value is less than the minimum average value, the voxel is determined to be soft tissue (i.e., or background, and the voxel gray value is marked as 0).

[0115] A voxel image (i.e., a three-dimensional binary image) is constructed by combining voxels of bone tissue and soft tissue; morphological closing operations are performed on the voxel image to extract the boundary contour of bone tissue (i.e., the bone tissue surface extracted from the voxel image).

[0116] It should be noted that the above steps construct a gray-level histogram by statistically analyzing the gray values ​​of the spherical voxel set, and use a sliding window to estimate the kernel density of the gray values ​​to obtain a smooth curve. The steps analyze the local minima in this curve to classify voxels into bone tissue and soft tissue. The analysis of the gray-level histogram helps to distinguish the gray-level characteristics of different tissues. The kernel density estimation smooth curve is used to eliminate noise and jagged edges in the gray-level histogram, ensuring more accurate classification of gray values, avoiding misclassification of tissue types, and helping the system accurately distinguish between bone tissue and soft tissue.

[0117] S243: Identify voxels of the inferior wall of the maxillary sinus in the bone tissue boundary contour according to the type of anatomical landmarks (i.e., in the above steps, take the anatomical landmarks of the inferior wall of the maxillary sinus in S241 as an example to correct the coordinates of the anatomical landmarks, and use morphological and grayscale features (the grayscale value of the sinus wall cortex is very high) to identify voxels in the bone tissue belonging to the inferior wall of the maxillary sinus).

[0118] The coordinates of the center point of the sinus wall are obtained from the voxels of all inferior walls of the maxillary sinus, forming a set of three-dimensional coordinate points of the sinus wall (i.e., the coordinates of each center point represent the position on the sinus wall).

[0119] In the bone tissue boundary contour, voxels adjacent to the three-dimensional coordinate point set of the inferior wall of the maxillary sinus are extracted to obtain alveolar bone voxels, and the coordinates of the alveolar bone center point of the alveolar bone voxels are obtained to form a three-dimensional coordinate point set of the alveolar bone (i.e., the coordinates of each center point represent the position on the alveolar bone; at the same time, it has been found that in the anatomical positional relationship in the medical field, the alveolar bone is located below the inferior wall of the maxillary sinus, so the alveolar bone voxels can be directly extracted in the area immediately below the sinus wall).

[0120] Fit the vertical direction of the center point of the sinus wall in the set of three-dimensional coordinate points of the sinus wall, and calculate the normal vector of the center point of the sinus wall (that is, the orientation of the sinus wall at that point, pointing out of the sinus cavity or towards the alveolar bone; the vertical direction is the alveolar bone below the sinus wall).

[0121] Starting from the center point of the sinus wall, construct a vector ray along the normal vector (i.e., search in the direction of the normal vector, because the sinus wall and alveolar bone are adjacent structures, and searching along the normal vector is necessary to accurately measure the vertical distance between them). In the set of three-dimensional coordinate points of the alveolar bone, calculate the coordinate distance between the coordinates of each alveolar bone center point and the coordinates of the vector ray (i.e., calculate the closest distance between the coordinates of the sinus wall center point and the coordinates of the alveolar bone center point). Search for the alveolar bone center point with the closest coordinate distance and use it as the first node of the alveolar bone (i.e., the center point of the alveolar bone surface that is perpendicularly closest to the center point of the sinus wall).

[0122] Calculate the straight-line distance between the coordinates of the first node of the alveolar bone and the coordinates of the center point of the sinus wall, and use it as the vertical thickness (that is, the thickness of the sinus wall to the surface of the alveolar bone).

[0123] It should be noted that the above steps, through morphological and grayscale features, can identify voxels of the inferior wall of the maxillary sinus from the boundary contour of bone tissue, and further extract voxels of the alveolar bone; the above steps, by calculating the normal vector of the center point of the sinus wall, extract voxels of the alveolar bone along this direction, obtaining the coordinate set of the alveolar bone; this step helps to establish a precise anatomical positional relationship. By modeling the spatial relationship between the inferior wall of the maxillary sinus and the alveolar bone, the position of the alveolar bone can be accurately extracted, such as... Figure 5 As shown;

[0124] S244: Connect the first node of the alveolar bone to the center point of the sinus wall to form a three-dimensional line segment (i.e., one end of the path is dense sinus wall cortical bone, and the other end is relatively loose alveolar bone); uniformly sample the three-dimensional line segment according to a pre-set fixed step size to obtain a sampling point sequence;

[0125] The grayscale value of each sampling point is calculated, and the standard deviation of the grayscale value of the three-dimensional line segment is calculated as the grayscale change rate (i.e., the standard deviation of the grayscale value of the sampling point along the line connecting the center point of the sinus wall to the first node of the alveolar bone, reflecting whether the transition from bone to soft tissue is clear or blurred; if the boundary is clear, the grayscale value sequence along the path will have a drastic cliff-like change, and the variance of the sequence will be very large; if the boundary is blurred or there is inflammation or bone thinning, the grayscale value may decrease slowly, and the variance of the sequence will be relatively small; the larger the grayscale change rate, the more drastic the grayscale change and the clearer the boundary; that is to say, the grayscale change rate is mainly used to assess whether the transition from the sinus wall to the alveolar bone is clear or blurred).

[0126] Obtain the vertical thickness of the straight-line distance to the center point of each sinus wall in the three-dimensional coordinate point set of the sinus wall and the gray-scale change rate of the three-dimensional line segment;

[0127] The mean of all vertical thicknesses and the mean of all grayscale change rates are calculated and then normalized and weighted to obtain a two-dimensional distance vector (which comprehensively describes the boundary quality). A multi-dimensional distance vector (which forms a five-dimensional feature vector) is then constructed based on the two-dimensional distance vector and the three-dimensional coordinates of the sinus wall center point (i.e., the coordinates of the voxel (xi, yi, zi)).

[0128] It should be noted that the above process involves forming a three-dimensional line segment by connecting the center point of the sinus wall and the first node of the alveolar bone, and uniformly sampling along this line segment to calculate the grayscale value and grayscale change rate at each sampling point. The grayscale change rate reflects the smoothness of the transition from bone tissue to soft tissue and the clarity of the boundary. The grayscale change rate is a key indicator for assessing boundary clarity and helps determine whether the boundary between the sinus wall and the alveolar bone is clear. If the boundary is clear, the grayscale change rate is large, which is crucial for subsequent implantation path selection. This step calculates the change rate to ensure that the implantation path does not cross ambiguous areas or introduce errors.

[0129] S245: Construct a three-dimensional mesh implantation model using the anatomical landmarks and multidimensional distance feature vectors; obtain the initial implantation tendency vector of the anatomical landmarks of the alveolar ridge crest based on the three-dimensional mesh implantation model, and preset the buccal-lingual angle and mesiodistal angle; perform preset candidate angle combinations based on the buccal-lingual angle and mesiodistal angle, and calculate the sparsity index; select the optimal Pareto candidate angle combination based on the sparsity index and calculate the axial direction vector; use the anatomical landmarks of the alveolar ridge crest as the intraosseous entry point coordinates, and perform implant adaptation simulation along the axial direction vector to obtain the adapted implant model;

[0130] It should be noted that the above-mentioned steps involve constructing a three-dimensional mesh implantation model based on anatomical landmarks and multi-dimensional distance feature vectors (i.e., a digital three-dimensional model reconstructed based on anatomical landmarks and features (such as five-dimensional point clouds) that can be used for implantation planning; that is, an intuitive visualization model used for implantation); obtaining the axial direction vector of the anatomical landmarks at the alveolar ridge crest for implantation in that direction, and obtaining the required adaptation parameters; and using the three-dimensional mesh model and precise anatomical landmarks to provide detailed positioning and orientation information for dental implantation. The axial direction vector ensures that the implantation process follows the most suitable path, reducing intraoperative errors and uncertainties.

[0131] The above employs a surface reconstruction algorithm that supports weighted input (such as weighted Poisson surface reconstruction). In this algorithm, the contribution of each point (xi, yi, zi) to the final surface generation is no longer equal, but determined by its corresponding multidimensional distance feature vector. Thickness weight is used as a weight; regions with greater thickness are assigned higher weights during reconstruction to ensure higher accuracy in these key areas. Boundary clarity weight can also be used as a weight; regions with high grayscale change rates (close to 1, clear boundaries) should be more rigorously fitted during reconstruction to preserve sharp anatomical boundaries. A triangular mesh model is then generated through the weighted surface reconstruction algorithm, which is the three-dimensional mesh implantation model. Compared with ordinary models, it has higher geometric accuracy and feature fidelity in regions with sufficient bone volume (large thickness) and clear boundaries (large grayscale change rate), providing a more reliable geometric basis for subsequent implantation path planning. The specific construction of the three-dimensional mesh implantation model is common knowledge and will not be elaborated further.

[0132] like Figure 6 As shown, in step S245, a three-dimensional mesh implantation model is constructed using the anatomical landmarks and multidimensional distance feature vectors; the initial implantation tendency vector of the anatomical landmarks of the alveolar ridge crest is obtained based on the three-dimensional mesh implantation model, and the buccal-lingual angle and mesiodistal angle are preset; a preset candidate angle combination is performed based on the buccal-lingual angle and mesiodistal angle, and a sparsity index is calculated; the optimal Pareto candidate angle combination is selected based on the sparsity index, and the axial direction vector is calculated; using the anatomical landmarks of the alveolar ridge crest as the coordinates of the intraosseous entry point, implant fitting simulation is performed along the axial direction vector to obtain the fitted implant model. The specific operation steps are as follows:

[0133] S2451: Use the anatomical landmarks as anchor points in three-dimensional space; construct a three-dimensional mesh planting model based on the multi-dimensional distance feature vector and the anchor points;

[0134] In the three-dimensional mesh implantation model, the anatomical landmark of the alveolar ridge crest is found as the simulation center point. A quadratic surface is fitted to all mesh points in the three-dimensional mesh implantation model within a preset neighborhood radius (that is, a complex curved surface is formed based on these distributed mesh points, and the mesh points within the radius form a set of points, representing the bone surface morphology near the alveolar ridge crest).

[0135] For each grid point in the quadratic surface, obtain the corresponding voxel gray value, and calculate the principal curvature based on the voxel gray value; select the minimum principal curvature as the bone filling direction (i.e., curvature is a quantity that describes the degree of bending of the surface of the three-dimensional model at a certain point; the maximum principal curvature corresponds to the direction of the most severe bending of the surface, and the minimum principal curvature corresponds to the direction of the gentlest bending of the surface; on the alveolar ridge crest, the direction of the maximum principal curvature usually points to the ridge direction of the alveolar ridge, while the direction perpendicular to it (the direction of the minimum principal curvature) points to the direction of the bone wall thickness).

[0136] It should be noted that in the above steps, anatomical landmarks are used as anchor points, and a three-dimensional mesh implantation model is constructed using multidimensional distance feature vectors. Then, a quadratic surface fitting is performed on the bone surface near the alveolar ridge crest to obtain a complex curved surface. The steps analyze the principal curvature of these mesh points to identify the direction of the most severe bone surface curvature (maximum principal curvature) and the gentlest direction (minimum principal curvature). Quadratic surface fitting helps to more accurately describe the bone surface morphology of the alveolar ridge crest, providing accurate geometric information for implantation path calculation. The calculation of principal curvature identifies the most suitable implantation direction, helping to determine the ideal implantation angle. The analysis of principal curvature also helps to optimize the fit between the implant and bone tissue.

[0137] S2452: Generate a vector ray (i.e., the normal vector direction line) for the bone filling direction (i.e., the bone wall thickness direction) as the initial implantation tendency vector (i.e., the implantation direction initially determined based on the minimum principal curvature direction; that is, find the direction that best fits the local bone surface of the alveolar ridge crest and points deep into the bone tissue. This direction is usually the direction with the most abundant bone and the most stable implantation).

[0138] The three-dimensional mesh implant model is used to obtain the long axis direction of the coordinates of adjacent teeth of the anatomical landmarks of the alveolar ridge (i.e., the ideal direction of the implant is as parallel as possible to the long axis of the adjacent teeth to obtain the best occlusal force and aesthetic effect); the course direction of the anatomical landmarks of the nerve canal (i.e., mainly the course direction and coordinates of the mandibular nerve canal) is obtained.

[0139] Based on the initial implantation tendency vector, preset buccal-lingual angle θ (i.e., the deflection angle in the buccal-lingual plane (e.g., -10° to +10°)) and mesio-distal angle ϕ (i.e., the deflection angle in the mesio-distal plane (e.g., -10° to +10°) are calculated.

[0140] The number of preset candidate angle combinations is determined based on the buccal-tongue angle θ and the mesio-distal angle ϕ (i.e., the position of each candidate angle combination represents a combination between the buccal-tongue angle θ and the mesio-distal angle ϕ, such as a combination of buccal-tongue angle -10° and mesio-distal angle ϕ-10° or buccal-tongue angle -10° and mesio-distal angle ϕ+10°, etc.).

[0141] Based on the initial implantation tendency vector and the buccal-lingual angle and mesiodistal angle, the implantation direction vector (i.e., the first objective function value f1, as shown in the formula) for each candidate angle combination is calculated. ;in, The rotation matrix is ​​used to rotate the initial implantation tendency vector to obtain the implantation direction vector corresponding to each candidate angle combination. (represented as the initial implantation tendency vector).

[0142] The vector deviation of each candidate angle combination is calculated using the implantation direction vector and the initial implantation tendency vector (i.e., the second objective function value f2, which represents the angle between the implantation direction corresponding to the current candidate angle combination and the initial implantation tendency vector, reflecting the degree of deviation from the initial bone morphology preferred direction).

[0143] The directional angle between the vector deviation and the long axis direction is calculated for each candidate angle combination (that is, this value represents the angle between the implantation direction corresponding to the current candidate angle combination and the long axis direction of the adjacent tooth, reflecting the degree of conformity with the parallelism of the adjacent tooth).

[0144] If the vector deviation and direction angle of a candidate angle combination are greater than those of another candidate angle combination, then the candidate angle combination is designated as the first frontier; the remaining candidate angle combinations are designated as the second frontier (that is, all candidate angle combinations are sorted non-dominated according to the first objective function value f1 and the second objective function value f2. If both objective function values ​​of candidate angle combination A are better than those of candidate angle combination B, then A dominates B; if one objective of candidate angle combination A is better than B and the other is worse than B, then A and B do not dominate each other; identify the Pareto optimal solution set in the current population, that is, those candidate angles that cannot be completely surpassed by other solutions on any objective).

[0145] For candidate angle combinations within the same frontier, a sparsity index is calculated (i.e., the sparser the solution in the region where the candidate angle combination is located, the more it should be retained; the calculation formula is...). ;in, and It is the objective function value of the adjacent candidate angle combinations of candidate angle combination j after sorting by the k-th objective function in the frontier. and These are the maximum and minimum values ​​of the k-th objective function in the frontier; the sparsity index indicates that the larger the value, the sparser the solutions in the region surrounding the candidate angle combination, meaning that the exploration of this region is insufficient and should be prioritized to maintain the diversity of solutions, avoid candidate angles clustering in local regions, and maintain the uniform distribution of the Pareto frontier.

[0146] In the first frontier, candidate angle combinations with the largest sparsity index are selected as guiding directions C (i.e., candidate angle combinations with the largest sparsity index are located in the sparsest region of the solution, and using them as guidance can drive the search to the unexplored region, which helps to discover more diverse Pareto optimal solutions; the larger the sparsity index of the candidate angle combination, the higher the probability of it being selected as the global optimum, which helps to guide the candidate angle combination to search for the sparse region of the solution).

[0147] The candidate angle combination is generated according to the guiding direction C and the preset search step size;

[0148] Based on the predetermined neural tube avoidance threshold (i.e., avoiding neural tubes and preventing contact), calculate the corresponding implantation direction vector for the new candidate angle combination;

[0149] Using the new candidate angle combination, the implant is simulated along the corresponding implantation direction vector, and the shortest distance between the implant and the anatomical landmark of the neural tube is calculated; if the shortest distance is greater than or equal to the neural tube avoidance threshold, the optimal Pareto is obtained.

[0150] If the latest position of the candidate angle combination is less than the neural tube avoidance threshold, then the search step size is reset for each candidate angle combination and the iteration continues until the optimal Pareto is obtained.

[0151] When the number of iterations equals the preset iteration threshold, the iteration stops; determine the ideal candidate angle combination (i.e., the minimum value of each of the two targets in the entire Pareto front), calculate the Euclidean distance between all the optimal Pareto and the ideal candidate angle combinations, and select the candidate angle combination corresponding to the minimum Euclidean distance (i.e., determine the ideal candidate angle combination, i.e., the minimum value of each of the two targets in the entire Pareto front).

[0152] The axial direction vector is calculated by combining the candidate angle combination corresponding to the minimum Euclidean distance with the initial implantation tendency vector. (That is, the optimal solution of the candidate angle combination of the minimum Euclidean distance is calculated with the initial implantation tendency vector to obtain the final implantation direction. The calculation formula is the same as the above calculation of the implantation direction vector.) In other words, the optimal implantation direction is obtained after correcting the initial implantation tendency vector by combining the long axis of the adjacent teeth and the direction of the nerve canal. That is, under the premise of ensuring absolute safety, the final implantation direction (i.e., the axial direction vector) is made as close as possible to the initial implantation tendency vector, while satisfying the parallelism constraint of the adjacent teeth (i.e., the long axis direction) as much as possible. The implantation is simulated in different directions, and the constraint of nerve canal avoidance is applied to simulate that the spatial distance between any part of the implant and the nerve canal must be greater than the preset nerve safety threshold (e.g., 2mm). Under the premise of satisfying the hard constraint of nerve avoidance, the balance point with the smallest deviation from the initial tendency vector and the smallest angle with the parallelism constraint of the adjacent teeth is found to obtain the optimal implantation axial direction vector.

[0153] It should be noted that a vector ray is constructed based on the direction of minimum principal curvature (i.e., the direction of gentle bone surface flatness) as the initial implantation tendency vector. This angle is further corrected to ensure that the implantation direction is parallel to the long axis of the adjacent tooth, while avoiding conflict with critical structures such as the nerve canal. The selection of the implantation angle is crucial; the correct angle not only ensures the stability of the implant but also ensures a uniform distribution of the biomechanical load within the bone. Taking into account the orientation of the long axis of the adjacent tooth and the nerve canal can maximize the implantation success rate and avoid damage. Figure 7 As shown;

[0154] S2453: Use the anatomical landmark of the alveolar ridge crest as the coordinates of the intraosseous entry point (that is, in clinical medicine, the alveolar ridge crest is the highest point of the alveolar bone and the entry point for the implant to enter the bone tissue. The neck of the implant is usually located at this position, and the implant is inserted into the bone from here).

[0155] Construct an implant model adaptation library (i.e., containing a combination of various diameters, lengths, and materials);

[0156] Starting from the coordinates of the intraosseous inlet point, implants are selected from the implant model matching library and simulated along the axial direction vector (i.e., the optimal implantation direction) to the preset standard implant length (e.g., 12mm). Implants are selected from the implant model matching library from combinations of minimum diameter and length (i.e., this allows for the implantation of implants from standard sizes (i.e., the length and width of conventional human implants) to smaller sizes, thereby avoiding damage to the bone quality and nerve canal of the implantee during the simulation). This straight line is the initial implantation path, representing the ideal central axis of the implant. The implant central axis (i.e., a straight line extending from the alveolar ridge crest along the axial direction vector to the predetermined length, representing the ideal centerline of the implant) is obtained.

[0157] According to the preset path sampling step size, the implantation center axis is uniformly sampled to obtain multiple axis sampling nodes; for each axis sampling node, a cross section perpendicular to the implantation center axis is constructed to form a bone cross section sampling plane (that is, a plane that can reflect the safe distance between the implantation path and the bone tissue).

[0158] According to the preset circular radius, a circular area is defined on the bone cross-section sampling plane; the voxel gray value of each circular area (i.e. reflecting bone density) is counted, and the average value is calculated as the bone density value (i.e., the average value of voxel gray values ​​in the circular area on the cross-section along the implantation center axis).

[0159] Calculate the shortest distance between the circular region and the anatomical landmarks (i.e., the anatomical landmarks of the maxillary sinus wall and the neural canal);

[0160] It should be noted that, starting from the alveolar ridge crest, the initial implantation path is formed by extending along the modified axis to the standard length of the implant. Uniform sampling is performed along this path, and the cross-section of each axial sampling node is calculated to assess bone density and safety distance. The planning of the initial implantation path is crucial to the entire implantation process. This step, by planning the path along the optimal axial direction, ensures the stability and effectiveness of the implant. The bone density assessment of the axial sampling nodes ensures that the bone at each sampling point is sufficient to support the implant, while the calculation of the safety distance avoids conflicts with adjacent structures (such as the neural canal, maxillary sinus wall, etc.).

[0161] S2454: Preset bone quality threshold (i.e., requiring an average gray value of not less than 300 HU (representing sufficient bone density)) and safety distance threshold (i.e., requiring any part of the implant to be not less than 2 mm away from the neural canal).

[0162] If the bone density value is less than the bone quality threshold, the simulated implant is determined to be of risk type A (i.e., the bone quality of this layer is too poor, which may cause the implant to be unstable).

[0163] Returning to step S2453 above, reselect an implant from the implant model matching library until the simulated implanted bone density value is less than the bone quality threshold (that is, in the case of estimated fragility, by adjusting the implant model, select an implant with a smaller length than the model in S2453, such as changing 12mm to 10mm, to make the implant stable).

[0164] If the shortest distance is less than the safe distance threshold, the implantation is determined to have risk type B (i.e., the layer is determined to have the risk of damaging important structures such as the neural tube).

[0165] Returning to step S2451 above, the direction of the principal curvature is reselected, the initial implantation tendency vector is adjusted until the implantation is completed, and the adaptation parameters are recorded (that is, from the direction of the minimum principal curvature, the direction corresponding to the second minimum principal curvature is reselected, and the implantation angle is finely adjusted to avoid collision with the neural tube).

[0166] The adaptation parameters include the coordinates of the intraosseous entry point, candidate angle combinations and axial direction vector, and the type of implant to be adapted (diameter and length).

[0167] It should be noted that a risk assessment is performed on each sampling point based on preset bone quality thresholds (e.g., grayscale value not less than 300 HU) and safety distance thresholds (e.g., minimum distance from the neural canal not less than 2 mm). If insufficient bone density or excessive proximity to important structures is found, it will be marked as the corresponding risk type (Type A: poor bone quality; Type B: excessive proximity to structures such as the neural canal). Based on the assessment results, the implantation path is adjusted to avoid risk areas. Risk assessment is a crucial step in ensuring successful implantation. Insufficient bone quality may lead to implant instability, while excessive proximity to important structures (such as the neural canal or sinus wall) may cause complications. By dynamically adjusting the implantation angle and path, this procedure ensures that the final implantation position not only meets biomechanical requirements but also avoids damage to surrounding important structures.

[0168] like Figure 8 As shown, the present invention also provides an oral cavity scanning image processing system, including: an acquisition module 10; and an analysis module 20;

[0169] The acquisition module 10 is used to acquire three-dimensional scanning data of the oral cavity region of the patient and generate an original grayscale image sequence; and to extract an initial coordinate set of multiple anatomical landmarks, including the alveolar ridge crest and nerve canal, from the original grayscale image sequence based on a convolutional neural network.

[0170] The analysis module 20 is used to screen anatomical landmarks in the initial coordinate set to divide voxel cubes and screen low-confidence points; calculate the change of joint function using the gradient magnitude of the low-confidence points to correct the coordinates of the anatomical landmarks; and obtain the appropriate implant model based on the anatomical landmarks of the alveolar ridge crest with corrected coordinates.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for processing oral cavity scan images, characterized in that, The specific operating steps include: Three-dimensional scanning data of the patient's oral cavity region is collected to generate an original grayscale image sequence; an initial coordinate set of multiple anatomical landmarks, including the alveolar ridge crest and nerve canal, is extracted from the original grayscale image sequence based on a convolutional neural network; The anatomical landmarks in the initial coordinate set are used to divide the voxel cubes, and low-confidence points are selected. The gradient magnitude of the low-confidence points is used to calculate the change in the joint function and correct the coordinates of the anatomical landmarks. The appropriate implant model is obtained based on the anatomical landmarks of the alveolar ridge crest with the corrected coordinates.

2. The oral cavity scan image processing method according to claim 1, characterized in that, The anatomical landmarks in the initial coordinate set are used to divide the data into voxel cubes, and low-confidence points are selected. The gradient magnitudes of the low-confidence points are used to calculate the change in the joint function, and the coordinates of the anatomical landmarks are corrected. The specific steps are as follows: Select any one of the anatomical landmarks in the initial coordinate set as the central anatomical landmark, and divide the voxel cube with the central anatomical landmark; traverse each voxel in the neighborhood of the voxel cube, determine the gray value of the voxel in each neighborhood, and construct a local gray-level distribution matrix. Principal component analysis is used to perform dimensionality reduction decomposition on the local gray-level distribution matrix to obtain the maximum and minimum gray-level eigenvalues. Calculate the grayscale feature ratio between the maximum and minimum grayscale feature values; A low confidence threshold is preset; it is then determined whether the gray-scale feature ratio is less than the low confidence threshold. If so, then the central anatomical landmark is classified as a low-confidence point; For each low-confidence point in the neighborhood, the gradient magnitude in each direction is calculated, and the maximum gradient magnitude is selected as the image feature term; the prior position of the anatomical landmark point in the neighborhood is obtained; and the prior position is converted into an anatomical constraint term. The image feature terms and anatomical constraint terms are weighted and combined to obtain the first joint function; Preset convergence threshold and search step size; use the quasi-Newton method to perform a one-dimensional line search on the gradient principal direction of the current low confidence point according to the search step size to determine the latest position of the current low confidence point and obtain candidate new points; return the candidate new points to the above steps, use them as new central dissection landmarks to divide the neighborhood, and calculate the second joint function; The change in the joint function is obtained by calculating the first joint function and the second joint function; Determine whether the change in the joint function is less than the convergence threshold; If yes, stop the iteration and output the coordinates of the candidate new point as the new coordinates of the central anatomical landmark; if no, return the candidate new point to the above steps and continue iterating until the change of the joint function converges to obtain the new coordinates of the central anatomical landmark.

3. The oral cavity scan image processing method according to claim 2, characterized in that, The appropriate implant model is obtained based on the anatomical landmarks of the alveolar ridge crest using the corrected coordinates. The specific steps are as follows: A historical oral anatomy standard size library is pre-built, and gray-level histograms of anatomical landmarks are obtained. A sliding window is applied to the gray-level histograms to construct kernel density curves and further calculate local minima. Based on the local minima, the bone tissue boundary contours corresponding to the voxels are extracted. Based on the type of the anatomical landmarks, voxels of the inferior wall of the maxillary sinus are identified within the bone tissue boundary contours, and the normal vector of the sinus wall center point is calculated. Based on the normal vector of the sinus wall center point, the voxel direction of the alveolar bone is searched, and the vertical thickness and gray-level change rate are calculated. The vertical thickness and gray-level change rate are normalized to obtain a multidimensional distance feature vector. Based on the multidimensional distance feature vector and the anatomical landmarks, the axial direction vector of the anatomical landmarks is determined. Based on the anatomical landmarks of the alveolar ridge crest along the direction of the axial direction vector, implant simulation is performed to obtain the appropriate implant model.

4. The oral cavity scan image processing method according to claim 33, characterized in that, A historical oral anatomy standard size library is pre-built, and gray-level histograms of anatomical landmarks are obtained. A sliding window is applied to the gray-level histograms to construct kernel density curves and further calculate local minima. Based on the local minima, the bone tissue boundary contours corresponding to the voxels are extracted. The specific operation steps are as follows: A historical oral anatomy standard size library is pre-constructed; the point type of the anatomical landmark is determined according to the historical oral anatomy standard size library, and the preset radius and shape of the corresponding point type are extracted to form a spherical voxel set; the gray value occurrence frequency of the voxels in the spherical voxel set is counted to construct a gray-level histogram; a sliding window is applied to the gray-level histogram, and the kernel density is calculated by weighted averaging of the gray values ​​of the voxels in each sliding window; the kernel density of the voxels in all sliding windows is fitted to obtain a kernel density curve; The minimum gray value of adjacent voxels in the kernel density curve is selected as the local minimum; the minimum average value of all local minimum values ​​is calculated; if the voxel gray value is greater than or equal to the minimum average value, the voxel is determined to be bone tissue; if the voxel gray value is less than the minimum average value, the voxel is determined to be soft tissue; the voxels of bone tissue and soft tissue are combined to form a voxel binary image; morphological closing operation is performed on the voxel binary image to extract the boundary contour of bone tissue.

5. The oral cavity scan image processing method according to claim 4, characterized in that, Based on the type of anatomical landmarks, voxels of the inferior wall of the maxillary sinus in the bone tissue boundary contour are identified, and the normal vector of the sinus wall center point is calculated. The voxel direction of the alveolar bone is searched based on the normal vector of the sinus wall center point, and the vertical thickness and grayscale change rate are calculated. The vertical thickness and grayscale change rate are normalized to obtain a multidimensional distance feature vector. The axial direction vector of the anatomical landmarks is determined based on the multidimensional distance feature vector and the anatomical landmarks. Implant simulation is performed along the direction of the axial direction vector of the anatomical landmarks at the alveolar ridge crest to obtain the appropriate implant model. The specific operation steps are as follows: Based on the type of anatomical landmarks, voxels of the inferior wall of the maxillary sinus in the bone tissue boundary contour are identified; the coordinates of the center point of the sinus wall are obtained for all voxels of the inferior wall of the maxillary sinus, forming a set of three-dimensional coordinate points of the sinus wall; In the bone tissue boundary contour, the voxels adjacent to the three-dimensional coordinate point set of the inferior wall of the maxillary sinus are extracted to obtain the voxels of the alveolar bone, and the coordinates of the alveolar bone center point of the alveolar bone voxel are obtained to form a three-dimensional coordinate point set of the alveolar bone. Fit the vertical direction of the sinus wall center point in the set of three-dimensional coordinate points of the sinus wall, and calculate the normal vector of the sinus wall center point; Starting from the center point of the sinus wall, construct a vector ray along the normal vector. In the set of three-dimensional coordinate points of the alveolar bone, calculate the coordinate distance between the coordinates of each alveolar bone center point and the coordinates of the vector ray, and search for the alveolar bone center point with the closest coordinate distance as the first node of the alveolar bone. Calculate the straight-line distance between the coordinates of the first node of the alveolar bone and the coordinates of the center point of the sinus wall as the vertical thickness. Connect the first node of the alveolar bone to the center point of the sinus wall to form a three-dimensional line segment; The three-dimensional line segment is sampled uniformly at a pre-set fixed step size to obtain a sampling point sequence; The grayscale value of each sampling point is calculated, and the standard deviation of the grayscale value of the three-dimensional line segment is calculated as the grayscale change rate. Obtain the vertical thickness of the straight-line distance to the center point of the sinus wall in the three-dimensional coordinate point set of the sinus wall and the gray-scale change rate of the three-dimensional line segment; calculate the mean of all vertical thicknesses and the mean of all gray-scale change rates, and perform a normalized weighted summation to obtain a two-dimensional distance vector; then, based on the two-dimensional distance vector and the three-dimensional coordinates of the center point of the sinus wall... A three-dimensional mesh implantation model is constructed using the anatomical landmarks and multidimensional distance feature vectors. An initial implantation tendency vector for the anatomical landmarks at the alveolar ridge crest is obtained based on the three-dimensional mesh implantation model, and buccal-lingual and mesiodistal angles are preset. Preset candidate angle combinations are generated based on the buccal-lingual and mesiodistal angles, and a sparsity index is calculated. The optimal Pareto candidate angle combination is selected based on the sparsity index, and an axial direction vector is calculated. Using the anatomical landmarks at the alveolar ridge crest as the intraosseous entry point coordinates, implant fitting simulation is performed along the axial direction vector to obtain the appropriate implant model.

6. The oral cavity scan image processing method according to claim 5, characterized in that, A three-dimensional mesh implantation model is constructed using the anatomical landmarks and multidimensional distance feature vectors. Based on the three-dimensional mesh implantation model, the initial implantation tendency vector of the anatomical landmarks at the alveolar ridge crest is obtained, and the buccal-lingual angle and mesiodistal angle are preset. Based on the buccal-lingual angle and mesiodistal angle, preset candidate angle combinations are performed, and the sparsity index is calculated. The specific operation steps are as follows: The anatomical landmarks are used as anchor points in three-dimensional space; a three-dimensional mesh implantation model is constructed based on the multi-dimensional distance feature vector and the anchor points; the anatomical landmark of the alveolar ridge is found in the three-dimensional mesh implantation model as the simulation center point, and a quadratic surface is fitted to all mesh points in the three-dimensional mesh implantation model within a preset neighborhood radius; For each grid point in the quadratic surface, obtain the corresponding voxel gray value, and calculate the principal curvature based on the voxel gray value; select the minimum principal curvature as the bone filling direction; A vector ray is drawn along the bone filling direction to serve as the initial implantation tendency vector; The major axis direction of the coordinates of the teeth in the vicinity of the anatomical landmarks of the alveolar ridge crest is obtained from the three-dimensional mesh implantation model; the direction of travel of the anatomical landmarks of the nerve canal is obtained; Preset the buccal-lingual angle and mesio-distal angle based on the initial implantation tendency vector; A preset number of candidate angle combinations is determined based on the buccal-tongue angle and the mesio-distal angle. The implantation direction vector for each candidate angle combination is calculated based on the initial implantation tendency vector and the buccal-lingual angle and mesiodistal angle. The vector deviation for each candidate angle combination is calculated using the implantation direction vector and the initial implantation tendency vector. The directional angle between the vector deviation and the major axis direction is calculated for each candidate angle combination. If the vector deviation and direction angle of the candidate angle combination are greater than those of another candidate angle combination, then the candidate angle combination is taken as the first frontier; the remaining candidate angle combinations are taken as the second frontier; and the sparsity index is calculated for the candidate angle combinations in the same frontier.

7. The oral cavity scan image processing method according to claim 6, characterized in that, The axis direction vector is calculated by selecting the optimal Pareto candidate angle combination based on the sparsity index. The specific operation steps are as follows: In the first frontier, candidate angle combinations with the maximum sparsity index are selected as guiding directions; new candidate angle combinations are generated using the candidate angle combinations according to the guiding direction and a preset search step size; a neural tube avoidance threshold is preset according to the walking direction. Calculate the corresponding implantation direction vector for the new candidate angle combination; Using the new candidate angle combination, the implant is simulated along the corresponding implantation direction vector, and the shortest distance between the implant and the anatomical landmark of the neural tube is calculated; if the shortest distance is greater than or equal to the neural tube avoidance threshold, the optimal Pareto is obtained. If the latest position of the candidate angle combination is less than the neural tube avoidance threshold, then the search step size is reset for each candidate angle combination and the iteration continues until the optimal Pareto is obtained. When the number of iterations equals the preset iteration threshold, the iteration stops; determine the ideal candidate angle combination, calculate the Euclidean distance between all the optimal Pareto values ​​and the ideal candidate angle combination, and select the candidate angle combination corresponding to the minimum Euclidean distance; calculate the axis direction vector based on the candidate angle combination corresponding to the minimum Euclidean distance and the initial implantation tendency vector.

8. The oral cavity scan image processing method according to claim 7, characterized in that, Using the anatomical landmark at the alveolar ridge crest as the coordinates of the intraosseous entry point, implant fitting simulation is performed along the vector direction of the axis to obtain the suitable implant model. The specific operation steps are as follows: The anatomical landmarks of the alveolar ridge crest are used as the coordinates of the intraosseous entry point. Construct an implant model adaptation library; starting from the coordinates of the intraosseous inlet point, select implants from the implant model adaptation library and simulate the extension of the axial direction vector along the preset implant standard length to obtain the implant center axis; According to the preset path sampling step size, the implantation center axis is uniformly sampled to obtain multiple axis sampling nodes; for each axis sampling node, a cross section perpendicular to the implantation center axis is constructed to form a bone cross section sampling plane; According to the preset circular radius, a circular region is defined on the bone cross-section sampling plane; the voxel gray value of each circular region is counted, and the average value is calculated as the bone density value; Calculate the shortest distance between the circular region and the anatomical landmark; A preset bone quality threshold and a safety distance threshold are set; if the bone density value is less than the bone quality threshold, the simulated implant is determined to have a risk type. Returning to the steps above, reselect an implant from the implant model matching library until the simulated implanted bone density value is less than the bone quality threshold. If the shortest distance is less than the safe distance threshold, then the planting is determined to be risky. Returning to the above steps, reselect the direction of the principal curvature, adjust the initial implantation tendency vector until the implantation is completed, and record the adaptation parameters.

9. The oral cavity scan image processing method according to claim 8, characterized in that, The adaptation parameters include the coordinates of the intraosseous entry point, candidate angle combinations and axial direction vectors, and the type of implant to be adapted.

10. A dental scanning image processing system, characterized in that, include: Data acquisition module; Analysis module; The acquisition module is used to acquire three-dimensional scanning data of the patient's oral cavity area and generate an original grayscale image sequence; and to extract an initial coordinate set of multiple anatomical landmarks, including the alveolar ridge crest and nerve canal, from the original grayscale image sequence based on a convolutional neural network. The analysis module is used to filter anatomical landmarks in the initial coordinate set to divide voxel cubes and filter low-confidence points; to calculate the change in the joint function using the gradient magnitude of the low-confidence points and to correct the coordinates of the anatomical landmarks; and to obtain the appropriate implant model based on the anatomical landmarks of the alveolar ridge crest with the corrected coordinates.

Images