Deep learning and three-dimensional constraint-based design method for paranasal implant
Patent Information
- Application Number
- CN202611529424.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-03
AI Technical Summary
Existing designs for nasal implants suffer from low design efficiency, poor standardization, and insufficient safety and adaptability, making it difficult to achieve fully automated, standardized, and high-precision personalized 3D-printed implant design.
A deep learning-based and three-dimensional constraint-based implant design method for the paranasal region was adopted. The jawbone was segmented and anatomical landmarks were located through a multi-task deep learning network. A reference coordinate system was constructed and a three-dimensional rectangular constraint bounding box was generated. The three-dimensional solid model of the prosthesis, including the pin path design, was generated by combining the jawbone mesh model and anatomical landmarks.
It achieves fully automated, standardized, and high-precision design of nasal implants, improving design efficiency and result consistency, avoiding error accumulation, and ensuring safe adaptation of the implant to the surrounding anatomical structures.
Smart Images
Figure CN123033754A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image data processing technology, and in particular to a method for designing paranasal implants based on deep learning and three-dimensional constraints. Background Technology
[0002] Bony depression in the paranasal region is a common developmental malformation and degenerative aging manifestation in the maxillofacial region. It not only affects the facial aesthetic contour, but in severe cases, it can also be accompanied by nasal base collapse and abnormal ventilation.
[0003] Current paranasal implant design relies heavily on manual interaction, requiring manual completion of jawbone and tooth segmentation, anatomical landmark marking, implant contour delineation, and thickness parameter setting. Therefore, current paranasal implant design faces three major bottlenecks: (1) The design efficiency is low, and a single solution takes several hours, which is difficult to meet the needs of rapid clinical diagnosis and treatment and batch customization; (2) Poor standardization, the shape of the prosthesis depends entirely on the designer's experience, the schemes vary greatly, the quality is uncontrollable, and the subjective bias is significant; (3) Insufficient safety and adaptability, manual contouring is prone to curvature changes, and the soft tissue is prone to step-like and foreign body sensations after surgery. It is also difficult to stably avoid key structures such as tooth roots and infraorbital neurovascular bundles, and there is a risk of intraoperative damage.
[0004] Although some studies in recent years have applied deep learning to jaw segmentation and landmark localization, most of them adopt a step-by-step process of "segmentation first, landmark localization later", which is prone to error accumulation and can only realize a single image processing link, without building a fully automated design link from image input to prosthesis output.
[0005] Therefore, developing a fully automated, standardized, and high-precision method for designing personalized 3D-printed implants for the paranasal region is a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0006] The technical problem to be solved by this invention is to provide a method for designing nasal implants based on deep learning and three-dimensional constraints, which can achieve full-process automation, standardization and high precision.
[0007] To address the aforementioned technical problems, this invention provides a method for designing paranasal implants based on deep learning and three-dimensional constraints, comprising: acquiring preoperative maxillofacial images; segmenting the maxilla, mandible, and teeth in the preoperative maxillofacial images to generate a three-dimensional mesh model of the maxilla, mandible, and teeth; and identifying anatomical landmarks in the preoperative maxillofacial images to generate three-dimensional coordinate information of the anatomical landmarks, wherein the anatomical landmarks include the lowest point of the left infraorbital margin, the lowest point of the right infraorbital margin, the highest point of the left superior margin of the external auditory canal, the highest point of the right superior margin of the external auditory canal, the anterior nasal spine, the left infraorbital foramen, the right infraorbital foramen, and the piriform foramen. The anatomical landmarks include the low point, canine root apex, nasal alar point, and nasal paranasal point. A reference coordinate system is constructed based on these anatomical landmarks. The three-dimensional coordinate information of the anatomical landmarks is transformed into the reference coordinate system. Three-dimensional rectangular constraint bounding boxes for both sides of the prosthesis are constructed based on the anatomical landmarks and the reference coordinate system. The bone surface fitting contours of both sides of the prosthesis are constructed based on the three-dimensional mesh model, anatomical landmarks, and three-dimensional rectangular constraint bounding boxes. A three-dimensional solid model of the prosthesis is generated based on the three-dimensional mesh model of the jawbone, anatomical landmarks, and bone surface fitting contours. The pin path is generated based on the three-dimensional mesh model, three-dimensional rectangular constraint bounding boxes, bone surface fitting contours, and three-dimensional solid model of the prosthesis.
[0008] As an improvement to the above scheme, the steps of segmenting the maxilla, mandible, and teeth in the preoperative maxillofacial image to generate a three-dimensional mesh model of the maxilla, mandible, and teeth, and identifying anatomical landmarks in the preoperative maxillofacial image to generate the three-dimensional coordinate information of the anatomical landmarks include: performing integrated processing on the preoperative maxillofacial image to generate three-dimensional volume data; inputting the three-dimensional volume data into a pre-trained multi-task deep learning network to generate three-dimensional mesh models of the maxilla, mandible, and teeth, and the three-dimensional coordinate information of the anatomical landmarks respectively; the multi-task deep learning network includes a feature extraction module and a parallel segmentation module and a recognition module; the segmentation module outputs a three-dimensional segmentation mask of the maxilla, mandible, and teeth using a voxel-level point-by-point classification method, and performs surface rendering and three-dimensional reconstruction of the three-dimensional segmentation mask using a moving cube algorithm to generate a three-dimensional mesh model of the maxilla, mandible, and teeth; the recognition module outputs a probability heatmap of the anatomical landmarks using a three-dimensional heatmap regression method, and calculates the three-dimensional coordinate information of the anatomical landmarks.
[0009] As an improvement to the above scheme, the step of constructing a reference coordinate system based on the anatomical landmarks includes: using the Frankfurt orbitoauricular plane as the XY plane; constructing the Z-axis using the anterior nasal spine as the necessary point and the unit normal vector of the XY plane as the direction vector; using the intersection of the Z-axis and the XY plane as the origin; orthogonally projecting the line vector connecting the lowest point of the left infraorbital margin to the lowest point of the right infraorbital margin onto the XY plane to form a projection vector, and making the projection vector pass through the origin to construct the X-axis; constructing the unit direction vector of the Y-axis through vector cross product operation according to the right-hand coordinate system principle, and making the Y-axis pass through the origin.
[0010] As an improvement to the above scheme, the step of constructing a three-dimensional rectangular constraint boundary box of the target-side prosthesis based on the anatomical landmarks and reference coordinate system includes: using a plane passing through the target-side infraorbital foramen and parallel to the XY plane as the upper boundary plane; using a plane parallel to the XY plane as the lower boundary plane, and using the larger of the initial Z-axis coordinate value of the lower boundary plane and the Z-axis coordinate value of the anterior nasal spine as the final Z-axis coordinate value of the lower boundary plane, wherein the initial Z-axis coordinate value of the lower boundary plane is the sum of the maximum Z-axis coordinate values of all canine root apexes on the target side and a preset safety distance; using a sagittal plane passing through the target-side infraorbital foramen and parallel to the YZ plane as the outer boundary plane; using a sagittal plane passing through the anterior nasal spine and parallel to the YZ plane as the inner boundary plane; and enclosing the upper boundary plane, lower boundary plane, inner boundary plane, and outer boundary plane to form the three-dimensional rectangular constraint boundary box.
[0011] As an improvement to the above scheme, the step of constructing the bone surface fitting contour of the target-side prosthesis based on the three-dimensional mesh model, anatomical landmarks, and three-dimensional rectangular constraint boundary frame includes: constructing an inner contour line running along the outer edge of the piriform foramen on the target side based on the three-dimensional mesh model; the upper end of the inner contour line intersects with the upper boundary plane of the three-dimensional rectangular constraint boundary frame on the target side to form an upper intersection point; the lower end of the inner contour line intersects with the lower boundary plane of the three-dimensional rectangular constraint boundary frame on the target side to form a lower intersection point; constructing the outer convex point of the outer contour line within the three-dimensional rectangular constraint boundary frame on the target side based on the anatomical landmarks; constructing the thickest point of the prosthesis based on the inner contour line and the outer convex point; constructing an upper transition contour segment based on the upper intersection point and the outer convex point; constructing a lower transition contour segment based on the outer convex point and the lower intersection point; and splicing the upper transition contour segment, the lower transition contour segment, and the inner fitting contour line end to end to form the bone surface fitting contour of the prosthesis.
[0012] As an improvement to the above scheme, the horizontal distance between the outer convex point and the outer boundary plane of the target-side three-dimensional rectangular constraint boundary box is 1 / 2 of the horizontal distance between the outer boundary plane of the target-side three-dimensional rectangular constraint boundary box and the canine root apex with the largest Y-axis coordinate value; the vertical distance between the outer convex point and the lower boundary plane of the target-side three-dimensional rectangular constraint boundary box is 1 / 3 of the vertical distance between the upper and lower boundary planes of the target-side three-dimensional rectangular constraint boundary box; the thickest point of the prosthesis is located on the line connecting the point of maximum curvature of the inner contour line and the outer convex point, and the distance between the point of maximum curvature and the thickest point of the prosthesis is 1 / 3 of the distance between the point of maximum curvature and the outer convex point.
[0013] As an improvement to the above scheme, the step of constructing an upper transition contour segment based on the upper intersection point and the outer convex point, and constructing a lower transition contour segment based on the outer convex point and the lower intersection point includes: arranging the upper intersection point, the outer convex point, and the lower intersection point sequentially to form a feature point sequence; adding a first-end auxiliary point at the beginning of the feature point sequence, wherein the first-end auxiliary point extends along the direction from the outer convex point to the upper intersection point, and the upper intersection point is located at the midpoint between the first-end auxiliary point and the outer convex point; adding a last-end auxiliary point at the end of the feature point sequence, wherein the last-end auxiliary point extends along the direction from the outer convex point to the lower intersection point, and the lower intersection point is located at the midpoint between the outer convex point and the upper convex point. The midpoint position of the first auxiliary point and the tail auxiliary point; constructing curve segments for every two adjacent feature points in the feature point sequence; performing interpolation calculations with the start point, end point, the feature point before the start point, and the feature point after the end point of the curve segment as tangent constraints to adjust the smoothness of the curve segment; adjusting the roundness at the feature points of the curve segment by using a preset tension coefficient; taking the curve segment between the first auxiliary point and the outer convex point within the three-dimensional rectangular constraint boundary as the upper transition contour segment, and taking the curve segment between the outer convex point and the tail auxiliary point within the three-dimensional rectangular constraint boundary as the lower transition contour segment.
[0014] As an improvement to the above scheme, the step of generating a three-dimensional solid model of the prosthesis based on the three-dimensional mesh model of the jawbone, anatomical landmarks, and bone surface fitting contour includes: calculating the reference prosthesis thickness based on the preoperative distance and ideal distance between the thickest point of the prosthesis, the alar point, and the paranasal point; forming a prosthesis bone surface fitting area based on the bone surface fitting contour; constructing a non-uniform prosthesis thickness scalar field based on the reference prosthesis thickness, with the thickest point of the prosthesis as the peak center, the thickness of the prosthesis thickness scalar field decreasing gradually from the peak center to the surrounding contour boundary; biasing and optimizing the mesh vertices of the three-dimensional mesh model of the jawbone based on the prosthesis thickness scalar field to generate a three-dimensional outer surface mesh of the prosthesis; performing triangular mesh edge sealing on the prosthesis bone surface fitting surface and the three-dimensional outer surface mesh of the prosthesis at the closed contour boundary to generate a three-dimensional solid mesh model of the prosthesis, and optimizing the three-dimensional solid mesh model of the prosthesis.
[0015] As an improvement to the above solution, the step of calculating the reference prosthesis thickness based on the preoperative distance and ideal distance between the thickest point of the prosthesis, the alar point, and the paranasal point includes: calculating the preoperative distance between the alar point and the paranasal point in the Y-axis direction; calculating the expected soft tissue increment based on the preoperative distance and the preset ideal distance; calculating the expected bone increment based on the expected soft tissue increment; and calculating the reference prosthesis thickness based on the expected bone increment and the angle between the normal vector of the thickest point of the prosthesis and the sagittal plane of the human body.
[0016] As an improvement to the above solution, the step of generating the pin track based on the three-dimensional mesh model, the three-dimensional rectangular constraint bounding box, the bone surface fitting contour, and the three-dimensional solid model of the prosthesis includes: constructing fixation points of the pin track based on the three-dimensional rectangular constraint bounding box and the bone surface fitting contour; calculating the central axis of the fixation points based on the three-dimensional mesh model of the maxilla; mapping a preset titanium pin model to the reference coordinate system, and constraining the titanium pin model mapped to the reference coordinate system according to the three-dimensional rectangular constraint bounding box to correct the fixation points and the central axis; mapping the titanium pin model to the reference coordinate system according to the corrected fixation points and the central axis, and constructing the pin track entity on the three-dimensional solid model of the prosthesis based on the titanium pin model mapped to the reference coordinate system.
[0017] Implementing this invention has the following beneficial effects: This invention employs an end-to-end multi-task deep learning network, which simultaneously completes the precise segmentation of the maxilla, mandible, and dental tissues and the spatial coordinate localization of anatomical landmarks through parallel segmentation and recognition modules. At the same time, the shared feature extraction module enables feature complementarity between the two types of tasks, avoiding the accumulation of errors caused by step-by-step processing, and improving the efficiency and consistency of the entire process automation.
[0018] This invention employs a two-level constraint strategy, from coarse to fine, to achieve automated and precise delineation of the prosthesis implantation range. The first level is the coarse positioning stage, which generates a three-dimensional rectangular constraint bounding box based on anatomical landmarks and a reference coordinate system to limit the overall implantation boundary of the prosthesis and avoid important anatomical structures. The second level is the precise positioning stage, which generates a smooth and continuous prosthesis bone surface fit contour by combining the local bony anatomy of the piriform foramen. Attached Figure Description
[0019] Figure 1 This is a flowchart of the first embodiment of the paranasal implant design method based on deep learning and three-dimensional constraints of the present invention; Figure 2 This is a schematic diagram of the upper boundary plane, lower boundary plane, and outer boundary plane in this invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It is hereby declared that the directional terms such as up, down, left, right, front, back, inside, and outside used in this text are based solely on the accompanying drawings and are not intended to specifically limit the invention.
[0021] See Figure 1 , Figure 1 The flowchart illustrates an embodiment of the paranasal implant design method based on deep learning and three-dimensional constraints of the present invention, which includes: S101, Obtain preoperative maxillofacial images; Obtain the patient's preoperative maxillofacial images, which are raw CBCT (Cone Beam Computed Tomography) images.
[0022] S102, the maxilla, mandible and teeth in the preoperative maxillofacial images are segmented to generate a three-dimensional mesh model of the maxilla, mandible and teeth, and anatomical landmarks in the preoperative maxillofacial images are identified to generate the three-dimensional coordinate information of the anatomical landmarks; It should be noted that anatomical landmarks include coordinate system construction reference points and prosthesis design constraint points, specifically: The coordinate system is constructed based on the following reference points: the lowest points of the bilateral infraorbital margins (the lowest point of the left infraorbital margin and the lowest point of the right infraorbital margin), the highest points of the bilateral external auditory canals (the highest point of the left external auditory canal and the highest point of the right external auditory canal), and the anterior nasal spine point. Among them, the lowest points of the bilateral infraorbital margins are the lowest extreme points of the bony infraorbital margins, the highest points of the bilateral external auditory canals are the highest extreme points of the bony external auditory canal superior wall, and the anterior nasal spine point is the most anterior extreme point of the anterior end of the anterior nasal spine of the maxilla.
[0023] The design constraints for the prosthesis include: bilateral infraorbital foramina (left and right infraorbital foramina), the lowest point of the piriform foramen, the canine root apex, the alar point, and the paranasal point. Among these, the bilateral infraorbital foramina are the center points of the bony foramina through which the infraorbital neurovascular bundles emerge; the lowest point of the piriform foramen is the lowest bony point in the middle of the lower edge of the piriform foramen; the canine root apex is the position of the apex of the maxillary canine root; the alar point is the soft tissue feature point of the lateral edge of the alar; and the paranasal point is the feature point of the most prominent soft tissue depression in the paranasal region.
[0024] Accordingly, the steps of segmenting the maxilla, mandible, and teeth in the preoperative maxillofacial images to generate a three-dimensional mesh model of the maxilla, mandible, and teeth, and identifying anatomical landmarks in the preoperative maxillofacial images to generate three-dimensional coordinate information of the anatomical landmarks include: (1) Perform integrated processing on preoperative maxillofacial images to generate three-dimensional volumetric data; Specifically, the preoperative maxillofacial images are resampled to a uniform voxel resolution, and the gray values are normalized and mapped to generate standard three-dimensional volume data that conforms to the input specifications of multi-task deep learning networks.
[0025] (2) Input the three-dimensional volume data into a pre-trained multi-task deep learning network to generate three-dimensional mesh models of the maxilla, mandible and teeth and three-dimensional coordinate information of anatomical landmarks respectively; In this invention, the multi-task deep learning network includes a feature extraction module and a segmentation module and a recognition module configured in parallel, wherein: The feature extraction module is used to extract multi-level semantic features from 3D volume data; correspondingly, the multi-level semantic features output by the feature extraction module are output to the segmentation module and the recognition module, respectively. The segmentation module outputs three-dimensional segmentation masks of the maxilla, mandible, and teeth using a voxel-level point-by-point classification method, and performs surface rendering and three-dimensional reconstruction of the three-dimensional segmentation masks using the Marching Cubes algorithm to generate three-dimensional mesh models of the maxilla, mandible, and teeth. The identification module uses a 3D heatmap regression method to output a probability heatmap of anatomical landmarks and calculates the 3D coordinate information of the anatomical landmarks. The 3D coordinate information of the anatomical landmarks is based on the world coordinate system of the original CBCT images. The X-axis of this world coordinate system is coronal, the Y-axis is sagittal, and the Z-axis is vertical, with the spatial measurement unit being millimeters.
[0026] It should be noted that the multi-task deep learning network is pre-trained using a large sample maxillofacial image dataset labeled with the gold standard for maxillary, mandible, and tooth segmentation, as well as the gold standard for anatomical landmark coordinates. Through weighted optimization of the multi-task loss function, the inference accuracy of the two types of tasks is improved simultaneously.
[0027] Therefore, this invention employs an end-to-end multi-task deep learning network, which simultaneously completes the precise segmentation of the maxilla, mandible, and dental tissues and the spatial coordinate localization of anatomical landmarks through parallel segmentation and recognition modules. At the same time, the shared feature extraction module enables feature complementarity between the two types of tasks, avoiding the accumulation of errors caused by step-by-step processing, and improving the efficiency and consistency of the entire process automation.
[0028] S103, Construct a reference coordinate system based on anatomical landmarks; Accordingly, the steps for constructing a reference coordinate system based on anatomical landmarks include: (1) The Frankfurt orbital plane is taken as the XY plane; Using four anatomical landmarks—the lowest point of the left infraorbital margin, the lowest point of the right infraorbital margin, the highest point of the left superior margin of the external auditory canal, and the highest point of the right superior margin of the external auditory canal—as input, the plane equation was obtained by least squares fitting.
[0029] The plane obtained by the above fitting is the Frankfurt orbital plane, which is used as the XY horizontal plane of the reference coordinate system in this invention.
[0030] Calculate the unit normal vector of the XY horizontal plane. :
[0031] In this invention, the unit normal vector is defined to point towards the top of the skull.
[0032] (2) Taking the anterior nasal spine as a necessary point and the unit normal vector of the XY plane as the boundary. Construct a Z-axis (i.e., the vertical axis) for the direction vector; (3) The origin is the intersection of the Z-axis and the XY plane; (4) Take the line vector connecting the lowest point of the left orbital margin to the lowest point of the right orbital margin as the initial direction, and orthogonally project the line vector onto the XY plane to obtain the projection vector located in the XY plane; place the starting point of the projection vector at the origin, and its direction is the positive direction of the X-axis, thereby constructing the X-axis (i.e. the left and right horizontal axes). (5) According to the right-hand coordinate system principle, construct the unit direction vector of the Y-axis through the vector cross product operation, and make the Y-axis pass through the origin.
[0033] The unit direction vector of the Y-axis (i.e., the anterior and posterior sagittal axes) is represented as:
[0034] in, The unit direction vector of the Y-axis. Let Z be the unit direction vector of the Z-axis. is the unit direction vector of the X-axis.
[0035] Therefore, this invention uses the clinically common Frankfurt orbitoauricular plane as the XY plane and the anterior nasal spine as the core positioning feature to construct a vertical axis, forming a three-dimensional orthogonal reference coordinate system conforming to the right-hand rule, with the spatial measurement unit being millimeters. The reference coordinate system constructed in this way is the RAS coordinate system, in which: the positive direction of the X-axis points to the right side of the patient, the positive direction of the Y-axis points to the front (facial side) of the patient, and the positive direction of the Z-axis points to the top of the patient's skull.
[0036] S104 transforms the three-dimensional coordinate information of the anatomical landmarks to the reference coordinate system in S103; S105, construct the three-dimensional rectangular constraint boundary boxes of the two prostheses respectively based on the anatomical landmarks and the reference coordinate system; like Figure 2 As shown, the three-dimensional rectangular constraint bounding box is formed by the upper boundary plane C1, the lower boundary plane C2, the inner boundary plane, and the outer boundary plane C3.
[0037] It should be noted that the prosthesis is divided into a left prosthesis unit and a right prosthesis unit with the midsagittal plane of the human body as the boundary, and the construction method of the three-dimensional rectangular constraint bounding box of the two prosthesis units is the same.
[0038] Accordingly, the steps for constructing the three-dimensional rectangular bounding box of the target-side prosthesis based on anatomical landmarks and a reference coordinate system include: (1) The upper boundary plane is the plane that passes through the infraorbital foramen point on the target side and is parallel to the XY plane; A plane passing through the target infraorbital foramen and parallel to the XY plane is constructed as the upper boundary plane of the three-dimensional rectangular constraint bounding box; this upper boundary plane can effectively limit the upper limit of the vertical implantation of the prosthesis and avoid the prosthesis being too high and affecting the infraorbital neurovascular bundle.
[0039] (2) Take the plane parallel to the XY plane as the lower boundary plane, and take the larger of the initial Z-axis coordinate value of the lower boundary plane and the Z-axis coordinate value of the anterior nasal spine as the final Z-axis coordinate value of the lower boundary plane; The initial Z-axis coordinate of the lower boundary plane is the sum of the maximum Z-axis coordinate of all canine root apex points on the target side and the preset safety distance; Furthermore, the steps for constructing the lower boundary plane include: (2.1) Statistically calculate the maximum coordinate values of the apex points of all canine roots on the target side of the maxilla in the Z-axis direction of the reference coordinate system. That is, the uppermost root apex in the corresponding position; (2.2) Construct a plane parallel to the XY plane as the lower boundary plane. The initial Z-axis coordinate value corresponding to this lower boundary plane is:
[0040] in, This is a preset safe distance; in this embodiment, .
[0041] Therefore, the present invention can reserve a specific vertical safety distance for the apex of the canine root, thereby avoiding damage to the tooth tissue during the implantation of the prosthesis.
[0042] (2.3) Compare the initial Z-axis coordinates of the lower boundary plane with the Z-axis coordinates of the anterior nasal spine, where: (2.3.1) If the initial Z-axis coordinate value of the lower boundary plane is less than the Z-axis coordinate value of the anterior nasal spine (i.e., the position is lower), then the Z-axis coordinate value of the anterior nasal spine is taken as the final Z-axis coordinate value of the lower boundary plane. (2.3.2) Otherwise, retain the initial Z-axis coordinate value of the lower boundary plane as the final Z-axis coordinate value of the lower boundary plane.
[0043] Therefore, the above verification rules can ensure that the lower inner margin of the prosthesis is not lower than the bony structure at the lower margin of the piriform foramen, which meets the anatomical safety constraints.
[0044] (3) The sagittal plane that passes through the infraorbital foramen on the target side and is parallel to the YZ plane is the outer boundary plane; A sagittal plane passing through the target infraorbital foramen and parallel to the YZ plane is constructed as the lateral boundary plane of the three-dimensional rectangular constraint bounding box. This lateral boundary plane can effectively limit the lateral implantation range of the prosthesis to not exceed the anatomical boundary of the paranasal region, while avoiding the course of the infraorbital neurovascular bundle and avoiding damage to important neurovascular structures during surgery.
[0045] (4) The sagittal plane that passes through the anterior nasal spine and is parallel to the YZ plane is the inner boundary plane; A sagittal plane passing through the anterior nasal spine and parallel to the YZ plane is constructed as the inner boundary plane of the three-dimensional rectangular constraint bounding box; this inner boundary plane can effectively limit the inner implantation boundary of the prosthesis from crossing the midline sagittal plane.
[0046] (5) The anterior-posterior direction is naturally defined by the contour of the bone surface; (6) Enclose the upper boundary plane, lower boundary plane, inner boundary plane and outer boundary plane to form a three-dimensional rectangular constraint boundary box.
[0047] Therefore, the present invention automatically generates the boundary planes of a three-dimensional rectangular constraint bounding box based on the anatomical landmarks identified in step S102 and the reference coordinate system established in step S103, so as to limit the overall implantation boundary of the prosthesis, avoid important anatomical structures, and achieve coarse positioning.
[0048] S106, construct the bone surface fitting contours of the two prostheses respectively based on the three-dimensional mesh model, anatomical landmarks and three-dimensional rectangular constraint bounding box; It should be noted that the bone surface fitting contour is formed by the inner contour line, the lower transition contour segment, and the upper transition contour segment closing at their ends.
[0049] Accordingly, the steps for constructing the bone surface fitting contour of the target-side prosthesis based on the three-dimensional mesh model, anatomical landmarks, and three-dimensional rectangular constraint bounding box include: (1) Construct an inner contour line that runs along the outer edge of the pear-shaped hole on the target side according to the three-dimensional mesh model. The upper end of the inner contour line intersects with the upper boundary plane of the three-dimensional rectangular constraint boundary box on the target side to form an upper intersection point, and the lower end of the inner contour line intersects with the lower boundary plane of the three-dimensional rectangular constraint boundary box on the target side to form a lower intersection point. (1.1) Using the bony edge of the piriform foramen as a reference, extract the initial contour curve running along the outer edge of the piriform foramen on the maxillary bone surface; (1.2) Calculate the lateral normal vector at each sampling point on the initial contour curve. The lateral normal vector is perpendicular to the tangent vector of the curve at that point, tangent to the bone surface at that point, and points to the outer side of the nasal cavity (cheek side). Average and normalize the lateral normal vectors of all sampling points to obtain the average normal vector of the inner contour line. (1.3) The initial contour curve is translated along the direction of the average normal vector by a preset offset to obtain the offset contour curve; in this embodiment, the offset is 1 mm, which can ensure that the edge of the prosthesis and the bony edge of the piriform aperture are kept at a safe distance to avoid the prosthesis from protruding into the nasal cavity. (1.4) Project the offset contour curve along the normal of the maxillary bone surface onto the maxillary bone surface so that the offset curve falls on the bone surface to obtain the inner contour line; the inner contour line runs along the outer edge of the piriform aperture, and its upper end intersects with the upper boundary plane to form the upper intersection point, and its lower end intersects with the lower boundary plane to form the lower intersection point.
[0050] (2) Construct the outer convex point of the outer contour line within the three-dimensional rectangular bounding box on the target side based on the anatomical landmarks; The spatial location of the convex point of the outer contour line of the prosthesis is determined by the following rules: X-axis position: The horizontal distance between the outer convex point and the outer boundary plane of the target side three-dimensional rectangular constraint boundary box is 1 / 2 of the horizontal distance between the outer boundary plane of the target side three-dimensional rectangular constraint boundary box and the canine root apex with the largest Y-axis coordinate value; Z-axis position: The vertical distance between the outer convex point and the lower boundary plane of the target-side three-dimensional rectangular constraint boundary box is 1 / 3 of the vertical distance between the upper and lower boundary planes of the target-side three-dimensional rectangular constraint boundary box; Y-axis position: The position of the maximum value of the Y-axis at X and Z points in the 3D mesh model; (3) Construct the thickest point of the prosthesis based on the inner contour line and the outer protrusion; The thickest point of the implant is located on the line connecting the point of maximum curvature of the inner contour and the point of external convexity, and the distance between the point of maximum curvature and the thickest point of the implant is 1 / 3 of the distance between the point of maximum curvature and the point of external convexity.
[0051] In other words, the thickest point of the implant is located at the point where the curvature of the inner contour line is greatest and the outer convex point is about 1 / 3 of the way in from the inner side.
[0052] Therefore, this invention allows the thickest point of the prosthesis to correspond to the most significant anatomical region of the mid-surface depression, while ensuring that the outer side does not exceed the safety boundary; at the same time, the thickest point of the prosthesis corresponds to the core area of the bony depression of the maxillary canine fossa, which can achieve the best soft tissue filling and support effect.
[0053] (4) Construct the upper transition contour segment based on the upper intersection point and the outer convex point, and construct the lower transition contour segment based on the outer convex point and the lower intersection point; Furthermore, the steps of constructing the upper transition contour segment based on the upper intersection point and the outer convex point, and constructing the lower transition contour segment based on the outer convex point and the lower intersection point include: (4.1) Arrange the upper intersection point, the outer convex point and the lower intersection point in sequence to form a feature point sequence; The upper intersection point, the outer convex point, and the lower intersection point are arranged sequentially along the circumference of the prosthesis outline to form an open sequence of feature points.
[0054] (4.2) Add auxiliary points at the beginning of the feature point sequence; To ensure a natural transition of the curve tangent at the upper intersection point, a virtual auxiliary point is added at the beginning of the feature point sequence: the auxiliary point extends along the direction from the outer convex point to the upper intersection point, and the upper intersection point is located at the midpoint between the auxiliary point and the outer convex point.
[0055] (4.3) Add auxiliary points at the end of the feature point sequence; To ensure a natural transition of the curve tangent at the lower intersection point, a virtual tail auxiliary point is added to the end of the feature point sequence: the tail auxiliary point extends along the direction from the outer convex point to the lower intersection point, and the lower intersection point is located at the midpoint between the outer convex point and the tail auxiliary point.
[0056] (4.4) Construct curve segments for every two adjacent feature points in the feature point sequence; (4.5) Interpolation calculations are performed with the starting point, ending point, the preceding feature point of the starting point, and the following feature point of the ending point of the curve segment as tangent constraints to adjust the smoothness of the curve segment. Using the start point, end point, and one adjacent feature point (including the first and last auxiliary points) of the curve segment as tangent constraints, interpolation calculations are performed using cubic Hermite basis functions to generate smooth curve segments.
[0057] (4.6) Adjust the roundness of the feature points of the curve segment by setting the preset tension coefficient; This invention introduces a tension coefficient To control the smoothness of the curve at feature points, The value range is 0.2-0.6; The larger the value, the smoother the curve transition and the more rounded the outline at the feature point; The smaller the value, the closer the curve is to the original broken line shape; in this embodiment, 0.4.
[0058] Accordingly, the parametric expression for a single curve segment is:
[0059] in: The normalized parameter for the curve segment; This is the starting point of the curve segment. This is the endpoint of the curve segment; Let be the tangent vector at the starting point. The tangent vector at the endpoint. , From the tension coefficient The coordinates of the adjacent feature points are calculated. The basis functions for cubic Hermite interpolation are expressed as follows:
[0060]
[0061]
[0062]
[0063] (4.7) The curve segment between the first auxiliary point and the outer convex point and within the three-dimensional rectangular constraint bounding box is taken as the upper transition contour segment, and the curve segment between the outer convex point and the tail auxiliary point and within the three-dimensional rectangular constraint bounding box is taken as the lower transition contour segment.
[0064] Therefore, steps (4.1) to (4.7) are based on the three feature points of the upper intersection point, the outer convex point and the lower intersection point, and adopt the cubic spline interpolation algorithm with tension constraints to generate the upper transition contour segment and the lower transition contour segment respectively.
[0065] (5) The upper transition contour segment, the lower transition contour segment and the inner fitting contour line are spliced together to form the prosthesis bone surface fitting contour.
[0066] The generated upper and lower transition contour segments are projected onto the outer surface of the bone tissue and then spliced together with the inner fitting contour line from beginning to end. After removing the overlapping endpoints, a complete closed prosthesis bone surface fitting contour can be obtained.
[0067] This invention combines the local bony anatomy of the piriform aperture to generate a smooth and continuous prosthesis bone surface fit contour, thereby achieving precise positioning; the contour achieves tangential continuity at the three characteristic points of the upper intersection, the outer convex point and the lower intersection, without sharp angles or curvature abrupt changes, ensuring a natural transition between the prosthesis edge and the surrounding bony structure and soft tissue, resulting in the best postoperative aesthetic effect.
[0068] Therefore, the present invention adopts a two-level constraint strategy from coarse to fine in steps S105 and S106 to achieve automated and precise delineation of the prosthesis implantation range.
[0069] S107, a three-dimensional solid model of the prosthesis is generated based on the three-dimensional mesh model of the jawbone, anatomical landmarks and bone surface fitting contour; Accordingly, the steps for generating a three-dimensional solid model of the prosthesis based on the three-dimensional mesh model of the jawbone, anatomical landmarks, and bone surface fitting contours include: (1) Calculate the baseline implant thickness based on the preoperative distance and ideal distance between the thickest point of the implant, the alar point and the paranasal point; Furthermore, the steps for calculating the baseline implant thickness based on the preoperative distance and ideal distance between the thickest point of the implant, the alar point, and the paranasal point include: (1.1) Calculate the preoperative distance between the patient's actual alar point and paranasal point in the Y-axis direction:
[0070] in, Preoperative distance, The Y-coordinate of the nasal ala point. The value is the Y-axis coordinate of the point beside the nose.
[0071] (1.2) Calculate the expected soft tissue increment based on the preoperative distance and the preset ideal distance:
[0072] in, For the expected increase in soft tissue; For the ideal distance, in this embodiment, .
[0073] (1.3) Calculate the expected bone increment based on the expected soft tissue increment:
[0074] in, This is for the expected increase in bone mass.
[0075] (1.4) Calculate the reference prosthesis thickness based on the expected bone increment and the angle between the normal vector of the thickest point of the prosthesis and the sagittal plane of the human body; Calculate the reference prosthesis thickness based on three-dimensional geometric projection relationships:
[0076] in, As a reference prosthesis thickness, The angle between the normal vector of the thickest point of the prosthesis and the sagittal plane.
[0077] (2) The prosthesis bone surface fitting area is formed according to the bone surface fitting contour; (3) Within the area where the prosthesis fits the bone surface, a non-uniform prosthesis thickness scalar field is constructed with the thickest point of the prosthesis as the peak center based on the reference prosthesis thickness. It should be noted that the thickness of the prosthesis thickness scalar field decreases gradually from the peak center to the surrounding contour boundary, and the thickness drops to 0 at the contour boundary, thereby achieving the shape effect of filling and supporting the core area of the prosthesis and naturally migrating the edge area.
[0078] The specific steps for constructing the scalar field of prosthesis thickness include: (3.1) Geodetic distance normalization Any vertex within the bone surface fitting area of the prosthesis Calculate the geodesic distance from the peak point along the maxillary bone surface. Simultaneously, calculate the shortest geodesic distance from this point along the bone surface to the boundary of the closed contour. Define the normalized distance parameter. :
[0079] in, Corresponding peak point (i.e., the thickest point of the implant). Corresponding to the outline boundary of the prosthesis.
[0080] (3.2) Power function thickness attenuation A power-law decay principle is used to achieve a thickness transition that is slow at first and then fast, at any point. Corresponding prosthesis thickness for:
[0081] in, The decay pattern control index, The preferred value is 2 to 3.
[0082] when When it is smaller (near the center of the peak). The growth is slow and the thickness decreases gradually, ensuring that the core area of the implant has a uniform support volume and maximizing the filling effect of the depression in the middle of the face. when When it approaches 1 (close to the edge of the contour). The growth rate accelerates, and the thickness rapidly decreases to 0, achieving seamless migration between the prosthesis edge and the surrounding bone surface, avoiding postoperative soft tissue issues such as step-like appearance and visible edges.
[0083] (4) In the area where the prosthesis fits the bone surface, the mesh vertices of the three-dimensional mesh model of the jawbone are biased and optimized according to the scalar field of the prosthesis thickness to generate the three-dimensional outer surface mesh of the prosthesis. Based on the three-dimensional mesh model of the maxilla generated in step S102, within the prosthesis bone surface fitting area, the thickness value of each mesh vertex is offset outward along its corresponding bone surface unit normal vector. The vertex coordinate set of the outer surface of the prosthesis is obtained; then the offset vertex set is subjected to mesh retopology and smoothing, and the Laplacian smoothing algorithm is used to optimize the mesh quality of the outer surface, eliminate local curvature abrupt changes, and generate a continuous and smooth three-dimensional outer surface mesh of the prosthesis.
[0084] (5) Triangular mesh edge sealing is performed on the bone surface of the prosthesis and the three-dimensional outer surface mesh of the prosthesis at the closed contour boundary to generate a three-dimensional solid mesh model of the prosthesis, and the three-dimensional solid mesh model of the prosthesis is optimized.
[0085] The specific steps include: (5.1) Generation of closed entities The inner surface (the bone-fitting surface of the prosthesis) and the outer surface (the three-dimensional outer surface mesh of the prosthesis) of the prosthesis are sealed with triangular meshes at the closed contour boundary to generate a complete closed three-dimensional solid mesh model of the polyetheretherketone (personalized 3D printing) prosthesis.
[0086] (5.2) Edge rounding treatment The lateral edges of the implant are smoothed with rounded corners of equal radius, preferably 0.5~1mm, to ensure that the overall edge of the implant is free of sharp corners, reduce postoperative irritation to surrounding soft tissues, and improve the safety of the implantation procedure.
[0087] (5.3) Overall smoothness verification A global curvature check is performed on the three-dimensional solid mesh model of the prosthesis to ensure a continuous transition of surface curvature without local protrusions or depressions, and to ensure that the shape of the prosthesis meets the requirements of facial aesthetics and biomechanics.
[0088] Therefore, the prosthesis can be solidified and its edges optimized through step (5).
[0089] S108 generates the nail path based on the three-dimensional mesh model, the three-dimensional rectangular constraint bounding box, the bone surface fitting contour, and the three-dimensional solid model of the prosthesis.
[0090] Accordingly, the steps for generating the pin track based on the 3D mesh model, the 3D rectangular constraint bounding box, the bone surface fitting contour, and the 3D solid model of the prosthesis include: (1) Construct fixation points for the pin track based on the three-dimensional rectangular bounding box and the bone surface fitting contour; This invention requires the placement of two fixation titanium screws in a unilateral nasal implant. The fixation points of these screws are vertically distributed above and below the thickest point of the implant to ensure uniform distribution of fixation force and prevent implant displacement under stress. All fixation points fall within the internal region of the implant's bone surface contour and meet anatomical safety constraints. The specific positioning rules are as follows: Upper fixation point: Take a horizontal section at 1 / 3 of the distance from the thickest point of the prosthesis to the Z-axis of the upper boundary plane. Within this section, take 1 / 3 of the distance between the two intersection points of the medial contour line and the lateral contour line and the section, close to the medial intersection point. Project this point along the bone surface normal onto the surface of the maxilla to obtain the initial upper fixation point. This position is close to the area with sufficient bone volume on the piriform aperture side, while being far away from the lateral infraorbital neurovascular bundle course area.
[0091] Lower fixation point: Take a horizontal section at 1 / 3 of the distance from the thickest point of the prosthesis to the Z-axis of the lower boundary plane. Within this section, take the intersection of the inner and outer contour lines at the middle 1 / 3 along the inner and outer directions. Project this intersection onto the surface of the maxilla to obtain the initial lower fixation point. This position avoids the vertical projection area of the canine root apex, reducing the risk of tooth damage.
[0092] Therefore, the automatic planning of fixed sites can be achieved through step (1).
[0093] (2) Calculate the central axis of the fixation point based on the three-dimensional mesh model of the maxilla; Using the local bone surface normal vector at the fixation point as the central axis of the screw path, the titanium screw implantation direction is ensured to be perpendicular to the bone surface, maximizing the fixation contact area and initial stability. Specific steps include: (2.1) Taking the upper fixation point and the lower fixation point as the center, the vertices of the bone surface mesh of the maxilla within a radius of 3 mm are respectively taken as the neighborhood point set; (2.2) Calculate the average value of the unit normal vectors of all vertices in the neighborhood point set, and then perform normalization to obtain the average unit normal vector of the local bone surface, with the direction pointing from the prosthesis to the inside of the bone tissue; (2.3) The straight line extending along the average unit normal vector direction from the fixation point is the central axis of the nail channel, ensuring that the nail channel is perpendicular to the local bone surface and avoiding stress concentration after titanium nail implantation.
[0094] Therefore, the automatic calculation of the center axis of the nail track can be achieved through step (2).
[0095] (3) Map the preset titanium nail model to the reference coordinate system, and constrain the titanium nail model mapped to the reference coordinate system according to the three-dimensional rectangular constraint boundary box to correct the fixation point and the central axis. This invention uses clinical standard micro titanium nail specifications (e.g., 2.0 mm in diameter), calls a preset titanium nail 3D model library, and sets multiple safety constraint checks; if the constraints are not met, an automatic fine-tuning mechanism is triggered until all conditions meet the requirements.
[0096] Specifically, after mapping the preset titanium nail model to the reference coordinate system, the following spatial position and geometric constraint verification is performed: Boundary inclusion verification: Check whether the nail body of the pre-mapped titanium nail model is completely contained within the 3D solid model of the prosthesis (except for the part that protrudes from the bottom surface of the prosthesis); at the same time, verify whether the overall spatial envelope of the titanium nail model is strictly limited to the 3D rectangular constraint bounding box to ensure that there is no geometric overrun.
[0097] Safe distance between screws: Calculate the minimum geodesic distance between the intersection of the central axis of two preset titanium screw models on the same side of the prosthesis and the outer side of the bone model. If the distance is less than the minimum safe threshold set clinically (preferably 2.0 mm), it is determined that the constraint is not met, so as to avoid local stress concentration in the prosthesis after implantation, which may lead to fracture.
[0098] Correspondingly, if the initial fixed point does not meet any safety constraint, the fixed point is automatically fine-tuned in 0.5mm increments along the horizontal cross section inwards and in the vertical direction, and the pose registration and interference verification are re-executed until all constraints are met; if no compliant point is found after traversing the preset fine-tuning range, an abnormal signal is output and manual intervention is requested.
[0099] Therefore, step (3) can achieve pin track parameter constraints and automated safety verification.
[0100] (4) Map the titanium nail model to the reference coordinate system according to the corrected fixation point and central axis, and construct the nail track entity on the three-dimensional solid model of the prosthesis according to the titanium nail model mapped to the reference coordinate system.
[0101] Based on the determined anchor points, axial direction, and preset titanium nail model, the nail track entity is automatically generated through high-precision Boolean operations. The specific steps include: (4.1) Model instantiation and pose registration Call the standard titanium nail model from the preset titanium nail 3D model library as the "tool body" and the 3D solid model of the prosthesis as the "target body"; according to the fixation point and central axis determined in step (3), map the titanium nail model to the reference coordinate system; adjust the pose of the titanium nail model so that the outer end face of the nail head of the titanium nail model and the outer surface of the prosthesis design at the fixation point maintain the preset relative position relationship (such as completely flush or maintaining a countersunk hole structure at a specific depth).
[0102] (4.2) Boolean difference operation Perform a Boolean difference operation on the target body and the tool body, that is: subtract the titanium nail model from the 3D solid model of the prosthesis.
[0103] (4.3) Complex nail track forming A solid space with geometric features completely consistent with the titanium nail model is cut off on the three-dimensional solid model of the prosthesis, thereby generating a high-precision matching nail channel containing nail body channels, nail head countersunk holes and thread shape, ensuring seamless fit between the titanium nail and the prosthesis after implantation.
[0104] Furthermore, after the pin track entity is generated, this invention automatically performs geometric and topological quality checks to ensure that the model conforms to 3D printing and clinical assembly specifications. Specific checks include: Mesh integrity check: Check the mesh topology of the inner wall of the nail path after Boolean operation to confirm that there are no broken surfaces, no self-intersections, and no illegal manifold edges, ensuring that the inner wall is smooth and continuous, meeting the requirements of 3D printing.
[0105] Connection surface verification: Verify the connection boundary between the outer end face of the nail channel (nail head) and the outer surface of the prosthesis to confirm that there are no steps, misalignments, or broken surfaces, ensuring a natural transition between the nail head and the prosthesis surface after titanium nail implantation, with no risk of soft tissue irritation.
[0106] Geometric matching verification: Extract key geometric features (such as diameter, countersunk hole depth, and thread profile) from the inner wall of the nail path and compare them with the corresponding parameters of the titanium nail model. Output a nail path design parameter matching report to ensure that the nail path dimensional accuracy meets the prosthesis structure and clinical assembly specifications.
[0107] Therefore, step S108 can automatically complete the titanium screw implantation site planning, screw path direction calculation, safety constraint verification and screw path entity generation based on the generated three-dimensional solid model of the prosthesis and the anatomical structure of the maxilla. No manual interaction is required throughout the process, ensuring that the screw path design complies with the clinical operation specifications of maxillofacial surgery, while avoiding important anatomical structures and achieving stable fixation of the prosthesis.
[0108] In summary, the paranasal implant design method based on deep learning and 3D constraints of this invention can achieve end-to-end operation without human intervention throughout the entire process. Specifically, this invention employs an end-to-end multi-task deep learning network, which simultaneously completes the precise segmentation of the maxilla, mandible, and dental tissues and the spatial coordinate localization of anatomical landmarks through parallel segmentation and recognition modules. Simultaneously, a shared feature extraction module enables feature complementarity between the two types of tasks, avoiding error accumulation caused by step-by-step processing and improving the efficiency and consistency of the entire automated process. Furthermore, this invention employs a two-level constraint strategy from coarse to fine to achieve automated and precise delineation of the prosthesis implantation range. The first level is the coarse localization stage, which generates a 3D rectangular constraint bounding box based on anatomical landmarks and a reference coordinate system to limit the overall implantation boundary of the prosthesis and avoid important anatomical structures. The second level is the precise localization stage, which combines the local bony anatomy of the piriform aperture to generate a smooth and continuous prosthesis bone surface fit contour.
[0109] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A method for designing paranasal implants based on deep learning and three-dimensional constraints, characterized in that, include: Obtain preoperative images of the maxillofacial region; The maxilla, mandible, and teeth in the preoperative maxillofacial images are segmented to generate a three-dimensional mesh model of the maxilla, mandible, and teeth. Anatomical landmarks in the preoperative maxillofacial images are identified to generate three-dimensional coordinate information of the anatomical landmarks. The anatomical landmarks include the lowest point of the left infraorbital margin, the lowest point of the right infraorbital margin, the highest point of the left external auditory canal, the highest point of the right external auditory canal, the anterior nasal spine, the left infraorbital foramen, the right infraorbital foramen, the lowest point of the piriform foramen, the apex of the canine root, the alar point, and the paranasal point. Construct a reference coordinate system based on the anatomical landmarks; Transform the three-dimensional coordinate information of the anatomical landmarks into the reference coordinate system; Based on the anatomical landmarks and the reference coordinate system, construct three-dimensional rectangular constraint bounding boxes for the two prostheses respectively; Based on the three-dimensional mesh model, anatomical landmarks, and three-dimensional rectangular constraint bounding boxes, the bone surface fitting contours of the two prostheses are constructed respectively. A three-dimensional solid model of the prosthesis is generated based on the three-dimensional mesh model of the jawbone, anatomical landmarks, and bone surface fitting contour. The pin path is generated based on the three-dimensional mesh model, the three-dimensional rectangular constraint bounding box, the bone surface fitting contour, and the three-dimensional solid model of the prosthesis.
2. The method for designing paranasal implants based on deep learning and three-dimensional constraints as described in claim 1, characterized in that, The steps of segmenting the maxilla, mandible, and teeth in the preoperative maxillofacial image to generate a three-dimensional mesh model of the maxilla, mandible, and teeth, and identifying anatomical landmarks in the preoperative maxillofacial image to generate the three-dimensional coordinate information of the anatomical landmarks include: The preoperative maxillofacial images are processed in an integrated manner to generate three-dimensional volume data; The three-dimensional volume data is input into a pre-trained multi-task deep learning network to generate three-dimensional mesh models of the maxilla, mandible, and teeth, as well as the three-dimensional coordinate information of the anatomical landmarks. The multi-task deep learning network includes a feature extraction module and a segmentation module and a recognition module set in parallel. The segmentation module outputs the three-dimensional segmentation mask of the maxilla, mandible and teeth using a voxel-level point-by-point classification method, and uses the moving cube algorithm to perform surface rendering and three-dimensional reconstruction of the three-dimensional segmentation mask to generate a three-dimensional mesh model of the maxilla, mandible and teeth. The recognition module outputs a probability heatmap of the anatomical landmark using a three-dimensional heatmap regression method, and calculates the three-dimensional coordinate information of the anatomical landmark.
3. The method for designing paranasal implants based on deep learning and three-dimensional constraints as described in claim 1, characterized in that, The step of constructing a reference coordinate system based on the anatomical landmarks includes: The Frankfurt orbital plane is taken as the XY plane; The Z-axis is constructed using the anterior nasal spine point as the necessary point and the unit normal vector of the XY plane as the direction vector; The origin is the intersection of the Z-axis and the XY plane. Using the line vector connecting the lowest point of the left infraorbital margin to the lowest point of the right infraorbital margin as the initial direction, the line vector is orthogonally projected onto the XY plane to obtain a projection vector located in the XY plane. The starting point of the projection vector is placed at the origin, and the direction of the projection vector is the positive direction of the X-axis to construct the X-axis. According to the right-hand coordinate system principle, the unit direction vector of the Y-axis is constructed by the vector cross product operation, and the Y-axis passes through the origin.
4. The method for designing paranasal implants based on deep learning and three-dimensional constraints as described in claim 3, characterized in that, The steps for constructing a three-dimensional rectangular bounding box of the target-side prosthesis based on the anatomical landmarks and reference coordinate system include: The upper boundary plane is the plane that passes through the infraorbital aperture of the target side and is parallel to the XY plane; The plane parallel to the XY plane is taken as the lower boundary plane, and the larger of the initial Z-axis coordinate value of the lower boundary plane and the Z-axis coordinate value of the anterior nasal spine is taken as the final Z-axis coordinate value of the lower boundary plane. The initial Z-axis coordinate value of the lower boundary plane is the sum of the maximum value of the Z-axis coordinate values of all canine root apexes on the target side and the preset safety distance. The sagittal plane, which passes through the infraorbital foramen of the target side and is parallel to the YZ plane, is the outer boundary plane; The sagittal plane that passes through the anterior nasal spine and is parallel to the YZ plane is the inner boundary plane; The upper boundary plane, lower boundary plane, inner boundary plane, and outer boundary plane are enclosed to form the three-dimensional rectangular constraint boundary box.
5. The method for designing paranasal implants based on deep learning and three-dimensional constraints as described in claim 4, characterized in that, The steps for constructing the bone surface fitting contour of the target side prosthesis based on the three-dimensional mesh model, anatomical landmarks, and three-dimensional rectangular constraint bounding box include: Based on the three-dimensional mesh model, an inner contour line is constructed along the outer edge of the pear-shaped aperture on the target side. The upper end of the inner contour line intersects with the upper boundary plane of the three-dimensional rectangular constraint boundary box on the target side to form an upper intersection point, and the lower end of the inner contour line intersects with the lower boundary plane of the three-dimensional rectangular constraint boundary box on the target side to form a lower intersection point. Based on the anatomical landmarks, construct the outer convex points of the outer contour line within the three-dimensional rectangular bounding box on the target side; Construct the thickest point of the prosthesis based on the inner contour line and the outer protrusion; The upper transition contour segment is constructed based on the upper intersection point and the outer convex point, and the lower transition contour segment is constructed based on the outer convex point and the lower intersection point. The upper transition contour segment, the lower transition contour segment, and the inner fitting contour line are spliced together end to end to form the prosthesis bone surface fitting contour.
6. The method for designing paranasal implants based on deep learning and three-dimensional constraints as described in claim 5, characterized in that, The horizontal distance between the external convex point and the outer boundary plane of the target side three-dimensional rectangular constraint boundary box is 1 / 2 of the horizontal distance between the outer boundary plane of the target side three-dimensional rectangular constraint boundary box and the canine root apex with the largest Y-axis coordinate value; The vertical distance between the outer convex point and the lower boundary plane of the target-side three-dimensional rectangular constraint boundary box is 1 / 3 of the vertical distance between the upper and lower boundary planes of the target-side three-dimensional rectangular constraint boundary box. The thickest point of the prosthesis is located on the line connecting the point of maximum curvature of the inner contour and the point of external convexity, and the distance between the point of maximum curvature and the thickest point of the prosthesis is 1 / 3 of the distance between the point of maximum curvature and the point of external convexity.
7. The method for designing paranasal implants based on deep learning and three-dimensional constraints as described in claim 5, characterized in that, The steps of constructing the upper transition contour segment based on the upper intersection point and the outer convex point, and constructing the lower transition contour segment based on the outer convex point and the lower intersection point include: The upper intersection point, the outer convex point, and the lower intersection point are arranged sequentially to form a feature point sequence; A first-end auxiliary point is added at the beginning of the feature point sequence. The first-end auxiliary point extends along the direction from the outer convex point to the upper intersection point, and the upper intersection point is located at the midpoint between the first-end auxiliary point and the outer convex point. A tail auxiliary point is added at the end of the feature point sequence. The tail auxiliary point extends in the direction from the outer convex point to the lower intersection point, and the lower intersection point is located at the midpoint between the outer convex point and the tail auxiliary point. Construct curve segments for every two adjacent feature points in the feature point sequence; Interpolation calculations are performed using the start point, end point, the preceding feature point of the start point, and the following feature point of the end point of the curve segment as tangent constraints, respectively, to adjust the smoothness of the curve segment. The roundness of the feature points of the curve segment is adjusted by a preset tension coefficient; The curve segment between the first auxiliary point and the outer convex point, and within the three-dimensional rectangular constraint boundary box, is taken as the upper transition contour segment, and the curve segment between the outer convex point and the tail auxiliary point, and within the three-dimensional rectangular constraint boundary box, is taken as the lower transition contour segment.
8. The method for designing paranasal implants based on deep learning and three-dimensional constraints as described in claim 5, characterized in that, The step of generating a three-dimensional solid model of the prosthesis based on the three-dimensional mesh model of the jawbone, anatomical landmarks, and bone surface fitting contour includes: The baseline implant thickness is calculated based on the preoperative distance and ideal distance between the thickest point of the implant, the alar point, and the paranasal point. The prosthesis bone surface fitting area is formed according to the bone surface fitting contour; Within the area where the prosthesis fits the bone surface, a non-uniform prosthesis thickness scalar field is constructed with the thickest point of the prosthesis as the peak center and based on the reference prosthesis thickness. The thickness of the prosthesis thickness scalar field decreases gradually from the peak center to the surrounding contour boundary, first slowly and then quickly. Within the area where the prosthesis bone surface fits, the mesh vertices of the three-dimensional mesh model of the jawbone are biased and optimized according to the prosthesis thickness scalar field to generate the three-dimensional outer surface mesh of the prosthesis. The prosthesis bone surface and the three-dimensional outer surface mesh of the prosthesis are subjected to triangular mesh edge sealing at the closed contour boundary to generate a three-dimensional solid mesh model of the prosthesis, and the three-dimensional solid mesh model of the prosthesis is optimized.
9. The method for designing paranasal implants based on deep learning and three-dimensional constraints as described in claim 8, characterized in that, The step of calculating the baseline implant thickness based on the preoperative distance and ideal distance between the thickest point of the implant, the alar point, and the paranasal point includes: Calculate the preoperative distance between the alar point and the paranasal point in the Y-axis direction; The expected soft tissue increment is calculated based on the preoperative distance and the preset ideal distance; Calculate the expected bone increment based on the expected soft tissue increment; The baseline prosthesis thickness is calculated based on the expected bone increment and the angle between the normal vector of the thickest point of the prosthesis and the sagittal plane of the human body.
10. The method for designing paranasal implants based on deep learning and three-dimensional constraints as described in claim 8, characterized in that, The step of generating the pin track based on the three-dimensional mesh model, the three-dimensional rectangular constraint bounding box, the bone surface fitting contour, and the three-dimensional solid model of the prosthesis includes: The fixation points of the pin track are constructed based on the three-dimensional rectangular bounding box and the bone surface fitting contour; The central axis of the fixation point is calculated based on the three-dimensional mesh model of the maxilla; The preset titanium nail model is mapped to the reference coordinate system, and the titanium nail model mapped to the reference coordinate system is constrained according to the three-dimensional rectangular constraint boundary box to correct the fixation point and the central axis. The titanium nail model is mapped to the reference coordinate system based on the corrected fixation point and central axis, and the nail track entity is constructed on the three-dimensional solid model of the prosthesis based on the titanium nail model mapped to the reference coordinate system.