Intelligent complete denture design method based on multi-task collaborative deep learning

Abstract

The invention discloses a complete denture intelligent design method based on multi-task collaborative deep learning, and is applied to the technical field of oral rehabilitation. Comprising the following steps: generating an edentulous jaw model after tissue surface buffering through an edentulous jaw buffering area mask self-encoding network; generating a model after undercut filling by adopting an in-place lane direction undercut filling algorithm; generating a complete denture arrangement position based on the multi-rigid body posture prediction network; segmenting an abutment boundary from the edentulous jaw model by using a deep learning model; taking the base boundary and the tooth arrangement position of the complete denture as constraint conditions, constructing a composite loss function in combination with the curvature of the generated surface, and generating a base model by using a conditional generative adversarial network. According to the automatic core tooth arrangement step, errors caused by human factors are reduced, the consistency and reliability of denture design quality are remarkably improved, the wearing comfort and long-term stability of final dentures are enhanced, and popularization and application of a high-quality complete denture repair technology in primary medical institutions are facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oral prosthetics, and more particularly to a full denture intelligent design method based on multi-task collaborative deep learning. BACKGROUND

[0002] The design of full denture for edentulous jaw includes two parts of artificial tooth array design and base design. The design process of artificial tooth array is complex and highly dependent on the experience of technicians. Specifically, the following core challenges are mainly faced. 1. Strong experience dependence and unstable quality: Tooth arrangement needs to consider key anatomical and functional elements such as jaw plane, midline, jaw arch shape, and extensive uniform occlusal contact to disperse occlusal force and prevent excessive local pressure. This process heavily relies on the experience of technicians to make repeated adjustments of the position and posture (six degrees of freedom, 6-DOF) of artificial teeth. Those who lack experience are prone to make errors in key steps such as jaw position determination and occlusal height setting, resulting in problems such as mucosa pain, occlusal discomfort, and low masticatory efficiency after wearing dentures, affecting patient comfort and prosthetic effect; 2. Tedious and inefficient operation: Traditional methods (including some digital software) require technicians to manually fine-tune the 6-DOF pose of each artificial tooth and repeatedly verify its relationship with the opposite teeth, adjacent teeth, and base. The lack of intelligent evaluation and guidance mechanism makes it difficult to obtain immediate feedback from minor adjustments, requiring multiple interactions, consuming time and effort, and making it difficult to standardize the replication of high-quality results, limiting the promotion at the grassroots level. The design of the base requires cushioning of the edentulous jaw tissue surface, and the subjective experience dependence of technicians is also high. Specifically, the following core challenges are mainly faced. 1. Difficulty in determining the cushioning area of the tissue surface: The mucosa covering the surface of the bony protuberance of the edentulous jaw is thin and will cause pain or form a fulcrum affecting the stability of the denture when under pressure. Therefore, it is necessary to cushion these areas to disperse the pressure and improve the wearing comfort of the patient. In addition, the alveolar ridge of the edentulous jaw undergoes bone resorption, resulting in the loosening of the full denture base supported by the alveolar ridge on the labial and buccal sides. Therefore, it is necessary to fill the undercut caused by bone resorption on the labial and buccal sides of the alveolar bone to provide support for the base. Since the cushioning and undercut filling areas of the tissue surface are not obvious, the process of determining these areas is time-consuming and laborious; 2. Strong experience dependence and unstable quality of base boundary: Traditional base design methods require technicians to manually determine the base boundary and repeatedly verify the boundary position to avoid insufficient base fit causing pain or poor retention during swallowing movement, resulting in denture shedding. Therefore, it is necessary to accurately extract the oral anatomical features of the edentulous jaw patient and precisely segment the base boundary to ensure that the tissue surface support area can bear uniform stress and ensure the retention of the denture. Therefore, how to provide a full denture intelligent design method based on multi-task collaborative deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0003] Therefore, the application provides a full-mouth denture intelligent design method based on multi-task cooperative deep learning.

[0004] To achieve the above object, the application provides the following technical scheme.

[0005] A full-mouth denture intelligent design method based on multi-task cooperative deep learning comprises the following steps.

[0006] A no-tooth jaw model after tissue surface buffering is generated through a no-tooth jaw buffering area mask auto-encoding network.

[0007] A support area of the no-tooth jaw model is segmented using a point cloud, normal vectors of all points in the support area are calculated, and a global support area average normal vector is obtained as a seating path direction after normalization and averaging of the normal vectors.

[0008] A deep learning model is used to segment a undercut area of the no-tooth jaw model, a model is formed by projecting the seating path direction and surrounding the undercut area, and a no-tooth jaw model after undercut filling is obtained by splicing the model and the no-tooth jaw model.

[0009] A multi-rigid-body pose prediction network is used to adjust the spatial pose of each tooth template, and the teeth are arranged adaptively according to the anatomical features of the no-tooth jaw patient's oral cavity to obtain a full-mouth denture tooth arrangement position.

[0010] A deep learning model is used to segment a denture border from the no-tooth jaw model after tissue surface buffering.

[0011] The denture border and the full-mouth denture tooth arrangement position are used as constraint conditions, and a composite loss function is constructed in combination with the curvature of the generated surface, and a conditional adversarial generation network is used to generate a denture model.

[0012] Optionally, the generation of the no-tooth jaw model after tissue surface buffering through the no-tooth jaw buffering area mask auto-encoding network specifically comprises the following steps: paired no-tooth jaw model and no-tooth jaw model data after tissue surface buffering are collected to form a no-tooth jaw full-mouth denture buffering data pair; a no-tooth jaw buffering area mask auto-encoding network is trained, the no-tooth jaw buffering area mask auto-encoding network generates a buffering model through learning the geometric feature information of the no-tooth jaw model after tissue surface buffering, generates a buffering model through a bone protuberance area mask, and obtains the no-tooth jaw model after tissue surface buffering.

[0013] Optionally, the adjustment of the spatial pose of each tooth template based on the multi-rigid-body pose prediction network and the adaptive arrangement of the teeth according to the anatomical features of the no-tooth jaw patient's oral cavity to obtain the full-mouth denture tooth arrangement position specifically comprises the following steps:

[0014] The anatomical feature key points of the edentulous jaw model are extracted, the upper and lower incisor papilla-alveolar ridge midpoint and the upper tuberosity-molar posterior pad connecting line are used, the jaw plane position is confirmed using the connecting line midpoint, and the spline interpolation method is used to connect the feature key points of the upper and lower jaws to obtain the upper and lower arch curves;

[0015] The preliminary arrangement centroid and long axis direction of the denture are calculated according to the jaw plane and the upper and lower arch curves, as the 6-DOF pose information of the teeth;

[0016] The adjacent teeth of each tooth are selected as a local pose adjustment set;

[0017] The occlusal distance gradient field of the current tooth and the adjacent teeth is calculated based on the local pose adjustment set, to assist the 6-DOF regression network in adjusting the 6-DOF pose of the current tooth;

[0018] A composite loss function is constructed according to the occlusal contact condition and the posture adjustment constraint;

[0019] The denture pose satisfying the optimal loss function of each local pose adjustment set is calculated to obtain the full denture arrangement position.

[0020] Optionally, the extraction of the anatomical feature key points of the edentulous jaw model specifically comprises: extracting the heat map features of the edentulous jaw model through a feature encoder, inputting the heat map features into a multi-layer perception to obtain category features, inputting the heat map features and the category features into a Transformer key point recognition module to obtain multi-scale heat map features and category features, and inferring the multi-scale heat map features and the category features to obtain the key point positions and key point types of the edentulous jaw model.

[0021] Optionally, inputting the heat map features and the category features into the Transformer key point recognition module to obtain multi-scale heat map features and category features specifically comprises: the input of the Transformer key point recognition module is the category features and the heat map features, adding the position coded heat map features and the category features, combining a self-attention module to enhance the key point heat map position and category of the category features, and a cross-attention module combining heat map information to further enhance the attention heat map of the key position, so that the key point position is more obvious.

[0022] Optionally, constructing the composite loss function according to the occlusal contact condition and the posture adjustment constraint specifically comprises:

[0023] L Total =L reg +w p L pen +w c L gap +w t λ t L trans +wr λ r L rot ;

[0024]

[0025] where L Total is the composite loss, L reg is the network loss, L pen is the penetration depth penalty, L gap is the no-contact penalty, L trans is the 6-DOF position change penalty, L rot is the 6-DOF pose change penalty, w p , w c , w t , w r are the weights of L pen , L gap , L trans , L rot , respectively, λ t , λ r are the pose weights of L trans , L rot , respectively, V and v are the vertices in the set of vertices of the tooth model and the set of vertices, respectively, SDF() is the signed distance function of the vertices, R and t are the rotation matrix and the position vector in the 6-DOF pose parameters of the current tooth, respectively, R0 and t0 are the rotation matrix and the position vector in the initial model pose parameters, respectively, ε c is the contact threshold for controlling the gap depth.

[0026] Optionally, the denture border and the denture arrangement position are taken as constraint conditions, and a composite loss function is constructed in combination with the curvature of the generated surface, and the conditional generative adversarial network is used to generate the denture model, specifically: the conditional generative adversarial network is used to take the artificial dentition model and the denture border as constraint conditions, and construct a composite loss function L 曲率 and L 义齿包绕 , the generator is used to preliminarily generate the complete denture denture model, the loss is calculated based on the surface geometric features of the generated complete denture denture model and the real denture denture model as the discriminator loss, the generator parameters are updated in the next round of training through the back propagation method and the gradient descent algorithm, so that the generation result is closer to the real denture denture model, until the loss function converges, and the complete denture denture model generation is realized.

[0027] Optionally, a composite loss function L 曲率 and L 义齿包绕 is constructed, specifically:

[0028]

[0029] In the formula, L 曲率 With L 义齿包绕 These are the curvature loss function and the denture wrapping rate loss function, respectively. base With v base Let A(v) be the set of vertices in the Kito model and the set of vertices in the set. base ) is the vertex v base The sum of the areas of the neighboring triangles, V N u and v are the vertices of v base The set of adjacent vertices and the vertices in the set, where cot represents the cotangent function. and For the edge (v) on the Kito model base The opposite angles on both sides of u, H0 and R target These are the curvature threshold and the wrapping rate threshold, respectively, P root p and ε are the vertex set and points in the denture model, respectively. δ(p) is the Boolean value for determining the wrapping condition. env This is the wrap-around distance threshold.

[0030] As can be seen from the above technical solution, compared with the prior art, the present invention provides a method for intelligent design of complete dentures based on multi-task collaborative deep learning, which has the following beneficial effects:

[0031] 1. Reduce reliance on experience and improve quality stability: This invention learns from expert experience and combines it with anatomical features to automate the core tooth arrangement steps, reducing errors caused by human factors and significantly improving the consistency and reliability of denture design quality;

[0032] 2. Significantly improve tooth arrangement efficiency: The automated posture adjustment process replaces time-consuming manual fine-tuning, greatly shortening the tooth arrangement cycle and reducing the workload of technicians;

[0033] 3. Optimize occlusal function: Optimization based on occlusal distance gradient field ensures that the denture obtains extensive and uniform occlusal contact, improves denture stability, comfort and chewing efficiency, and reduces clinical modifications;

[0034] 4. Improve the rationality and comfort of the overall denture design: Through intelligent tooth arrangement with base-to-teeth coordination, the generated artificial dentition is stably supported on the base without base penetration. At the same time, the relationship between the long axis of the teeth and the alveolar ridge is considered, which significantly improves the physical rationality and biomechanical adaptability of the denture design. This directly reduces the risk of denture tilting, pressure pain, and fracture caused by improper base-to-teeth relationship, and enhances the wearing comfort and long-term stability of the final denture.

[0035] 5. Promote standardization and popularization: Intelligent methods are easier to standardize, which helps to popularize and apply high-quality complete denture restoration technology in primary healthcare institutions;

[0036] 6. Improve patient experience: shorten the treatment cycle, improve the wearing comfort and functional effect of the final denture. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0038] Figure 1 Flow chart of the full denture intelligent design method of the present application;

[0039] Figure 2 Flow chart of the edentulous jaw buffer area mask auto-encoding network of the present application;

[0040] Figure 3 Flow chart of the multi-rigid-body pose prediction network of the present application;

[0041] Figure 4 Flow chart of the edentulous jaw model dissection key point extraction of the present application;

[0042] Figure 5 Flow chart of the multi-condition constraint generative adversarial network of the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0044] The embodiments of the present application disclose a full denture intelligent design method based on multi-task cooperative deep learning, as shown in Figure 1 The method comprises the following steps:

[0045] An edentulous jaw model after buffer of tissue surface is generated by an edentulous jaw buffer area mask auto-encoding network;

[0046] The bearing area of the edentulous jaw model is segmented using point cloud, the normal vector of all points in the bearing area is calculated, the global bearing area average normal vector is obtained by normalizing and averaging the normal vector, and the global bearing area average normal vector is used as the seating path direction;

[0047] The deep learning model is used to segment the undercut area of the edentulous jaw model, the seating path direction is projected to form a model with the undercut area, and the model is spliced with the edentulous jaw model to obtain an undercut filling model.

[0048] Adjust the spatial pose of each tooth template based on the multi-rigid body pose prediction network, and arrange the teeth according to the anatomical features of the edentulous patient's oral cavity to obtain the full denture arrangement position;

[0049] Using a deep learning model to segment the denture border from the tissue surface buffered edentulous model;

[0050] Taking the denture border and the full denture arrangement position as constraint conditions, and combining the curvature of the generated surface to construct a composite loss function, a conditional adversarial generative network is used to generate a denture model.

[0051] Further, as shown in Figure 2 The tissue surface buffered edentulous model is generated by the edentulous buffer area mask auto-encoding network, which is specifically: collecting pairs of edentulous model and tissue surface buffered edentulous model data to form edentulous full denture buffer data pairs; training the edentulous buffer area mask auto-encoding network, which generates a buffered model for the bone protuberance area mask by learning the geometric feature information of the edentulous tissue surface buffered model, to obtain the tissue surface buffered edentulous model.

[0052] In the embodiment of the present application, the collection of pairs of edentulous model and tissue surface buffered edentulous model data is specifically: obtaining the intraoral scanning three-dimensional model of the edentulous patient, the tissue surface buffered three-dimensional model and the full denture data designed by the expert, and performing orthotopic and down-sampling on the three-dimensional model to obtain the edentulous model and the tissue surface buffered edentulous model data.

[0053] In the embodiment of the present application, the support area of the edentulous model is segmented using PointNet, PointMeta and other related point cloud segmentation algorithms.

[0054] Further, as shown in Figure 3 Adjusting the spatial pose of each tooth template based on the multi-rigid body pose prediction network, and arranging the teeth according to the anatomical features of the edentulous patient's oral cavity to obtain the full denture arrangement position is specifically:

[0055] Extracting the anatomical feature key points of the edentulous model, connecting the upper and lower incisor papilla-alveolar ridge midpoint and the maxillary tubercle-molar pad back pad, using the midpoint of the connecting line to confirm the jaw plane position, and using the spline interpolation method to connect the feature key points of the upper and lower jaws to obtain the upper and lower arch curves;

[0056] Calculating the initial arrangement centroid and long axis direction of the denture according to the jaw plane and the upper and lower arch curves as the 6-DOF pose information of the teeth;

[0057] Selecting the adjacent teeth of each tooth as a local pose adjustment set;

[0058] calculate the occlusion distance gradient field of the current tooth and the adjacent tooth based on the local pose adjustment set, and assist the 6-DOF regression network to adjust the 6-DOF pose of the current tooth;

[0059] construct a composite loss function according to the occlusion contact condition and the pose adjustment constraint;

[0060] calculate the denture pose of each local pose adjustment set satisfying the optimal loss function, and obtain the full denture arrangement position.

[0061] In the embodiment of the present application, a pair of edentulous jaw data and corresponding expert designed arrangement data are collected to form the edentulous jaw full denture arrangement data pair. Then, the expert marks the anatomical feature key points on the edentulous jaw model to assist the determination of the pose of the full denture and construct the edentulous jaw denture pose adjustment data set. The multi-rigid-body pose prediction network is trained based on the above edentulous jaw denture pose adjustment data set. In the deep learning task, the pose of a three-dimensional model in space is often described by 6-DOF, including the position of the center of mass and the direction of the model, and the spatial pose of the object can be directly changed by adjusting the 6-DOF vector of the tooth model, thereby realizing the full denture arrangement function.

[0062] Further, as shown in Figure 4 the extraction of the anatomical feature key points of the edentulous jaw model specifically includes: extracting the heat map features of the edentulous jaw model through a feature encoder, inputting the heat map features into a multi-layer perception to obtain category features, inputting the heat map features and the category features into a Transformer key point recognition module to obtain multi-scale heat map features and category features, and inferring the multi-scale heat map features and the category features to obtain the key point positions and key point types of the edentulous jaw model.

[0063] In the embodiment of the present application, the feature encoder adopts a mesh model such as meshCNN and GCN.

[0064] Further, inputting the heat map features and the category features into the Transformer key point recognition module to obtain multi-scale heat map features and category features specifically includes: the input of the Transformer key point recognition module is the category features and the heat map features, adding the heat map features with position encoding to the category features, combining a self-attention module to enhance the key point heat map position and category of the category features, and combining a cross-attention module to further enhance the attention heat map of the key position, so that the key point position is more obvious.

[0065] The calculation formula of the Transformer key point recognition module is:

[0066] Att E (F1, F2) = BN(Att Self (E posi(F1)+F2, F2)+F2);

[0067] Att H (F1, F2) = W x BN(Att Cross (F1 T , F2) T +F1);

[0068] Att C (F1, F2) = W x (BN(Att Cross (F1 T , F2)+F2));

[0069] F 热图特征 = At H (F 热图特征 , Att E (F 热图特征 , F 类别特征 ));

[0070] F′ 类别特征 = Att C (F 热图特征 , Att E (F 热图特征 , F 类别特征 ));

[0071] In the formula, Att E is a position encoding attention module, Att H is a heat map feature attention module, Att C is a category feature attention module, F1 and F2 are two input features of a Transformer key point recognition module; BN is a batch normalization operation, W is a multi-layer perceptron parameter weight, Att Cross is a cross attention module, F 热图特征 and F 类别特征 are low-dimensional heat map features and category features respectively, F′ 热图特征 and F′ 类别特征 are high-dimensional heat map features and category features respectively.

[0072] Further, according to the occlusion contact condition and the posture adjustment constraint, a composite loss function is constructed, which is specifically:

[0073] L Total = L reg +w p L pen +w c L gap +w t λ t L trans +w r λ r Lrot ;

[0074]

[0075] wherein, L Total is the composite loss, L reg is the network loss, L pen is the penetration depth penalty, L gap is the no-contact penalty, L trans is the 6-DOF position change penalty, L rot is the 6-DOF pose change penalty, w p , w c , w t , w r are the weights of L pen , L gap , L trans , L rot , respectively, λ t , λ r are the pose weights of L trans , L rot , respectively, V and v are the vertices in the set of vertices of the tooth model and the set, respectively, SDF() is the signed distance function for calculating the vertex, R and t are the rotation matrix and the position vector in the 6-DOF pose parameters of the current tooth, respectively, R0 and t0 are the rotation matrix and the position vector of the initial model pose parameters, respectively, ε c is the contact threshold for controlling the gap depth.

[0076] In the embodiments of the present application, the selection of the local pose adjustment set is specifically: according to the tooth arrangement sequence, the order of adjusting the denture pose from mesial to distal, from maxilla to mandible is determined. In this process, the current tooth and the three corresponding teeth of the adjacent teeth on the mesial and distal sides, and the opposite jaw form a local tooth arrangement set T together, and each set contains 6 teeth (if No. 7 is selected, there is no distal tooth, and the set contains 4 teeth).

[0077] Further, as shown in Figure 5 , the denture border and the denture arrangement position are used as constraint conditions, and the curvature of the generated surface is used to construct a composite loss function, and the conditionally generated network is used to generate the denture model, specifically: using the conditional generative adversarial network to use the artificial tooth arrangement model and the denture border as constraint conditions, and constructing a composite loss function L 曲率 and L 义齿包绕The generator is used to initially generate a complete denture base model. The loss is calculated based on the surface geometric features of the generated complete denture base model and the real denture base model as the discriminator loss. The generator parameters are updated in the next round of training through backpropagation and gradient reduction algorithm to make the generated result closer to the real denture base model, until the loss function converges, thus realizing the generation of the complete denture base model.

[0078] Furthermore, construct a composite loss function L 曲率 With L 义齿包绕 Specifically:

[0079]

[0080] In the formula, L 曲率 With L 义齿包绕 These are the curvature loss function and the denture wrapping rate loss function, respectively. base With v base Let A(v) be the set of vertices in the Kito model and the set of vertices in the set. base ) is the vertex v base The sum of the areas of the neighboring triangles, V N u and v are the vertices of v base The set of adjacent vertices and the vertices in the set, where cot represents the cotangent function. and For the edge (v) on the Kito model base The opposite angles on both sides of u, H0 and R target These are the curvature threshold and the wrapping rate threshold, respectively, P root p and ε are the vertex set and points in the denture model, respectively. δ(p) is the Boolean value for determining the wrapping condition. env This is the wrap-around distance threshold.

[0081] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0082] Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent design of complete dentures based on multi-task collaborative deep learning, characterized in that, Includes the following steps: A tissue-buffered edentulous model is generated using an autoencoder network with a masked edentulous region. The support area of ​​the edentulous jaw model is segmented using point cloud. The normal vectors of all points in the support area are calculated. After normalization, the average of the normal vectors is obtained to get the global average normal vector of the support area as the direction of the path of insertion. The undercut region of the edentulous model is segmented using a deep learning model. The model is formed by projecting the path of insertion direction onto the undercut region and then stitching it together with the edentulous model to obtain the undercut-filled model. Based on a multi-rigid-body pose prediction network, the spatial pose of each tooth template is adjusted, and the teeth are arranged adaptively according to the anatomical features of the oral cavity of edentulous patients to obtain the tooth arrangement position of complete dentures. The denture base boundary was segmented from the tissue-buffered edentulous jaw model using a deep learning model. Using the basement boundary and the position of the teeth in the complete denture as constraints, a composite loss function is constructed by combining the curvature of the generated surface, and a conditional adversarial generative network is used to generate the basement model.

2. The intelligent design method for complete dentures based on multi-task collaborative deep learning according to claim 1, characterized in that, The specific steps for generating a tissue-buffered edentulous model using an edentulous buffer region masking autoencoder network are as follows: Data on paired edentulous models and tissue-buffered edentulous models are collected to form edentulous complete denture buffer data pairs; the edentulous buffer region masking autoencoder network is trained, and by learning the geometric features of the tissue-buffered edentulous model, the network generates a buffered model by masking bony prominence regions, thus obtaining the tissue-buffered edentulous model.

3. The intelligent design method for complete dentures based on multi-task collaborative deep learning according to claim 1, characterized in that, Based on a multi-rigid-body pose prediction network, the spatial pose of each tooth template is adjusted, and the teeth are adaptively arranged according to the anatomical features of the oral cavity of edentulous patients, resulting in the specific tooth arrangement positions for complete dentures as follows: Extract key anatomical features from the edentulous jaw model, connect the midpoint of the epiphyseal process of the maxillary and mandibular incisors to the alveolar ridge, and the midpoint of the maxillary tubercle to the molar pad. Use the midpoint of the connecting line to confirm the position of the jaw plane. Use spline interpolation to connect the key features of the maxilla and mandible respectively to obtain the dental arch curves of the maxilla and mandible. The centroid and long axis directions of the denture are initially calculated based on the jaw plane and the curves of the maxillary and mandibular dental arches, serving as the 6-DOF positional information of the teeth; Select the adjacent teeth of each tooth as the local pose adjustment set; The gradient field of the occlusal distance between the current tooth and its neighboring teeth is calculated based on the local pose adjustment set, and the 6-DOF regression network is used to adjust the 6-DOF pose of the current tooth. A composite loss function is constructed based on the biting contact condition and posture adjustment constraints; Calculate the denture posture that satisfies the optimal loss function for each local pose adjustment set, and obtain the denture tooth arrangement position.

4. The intelligent design method for complete dentures based on multi-task collaborative deep learning according to claim 3, characterized in that, The extraction of key anatomical features of the edentulous jaw model is specifically as follows: heat map features of the edentulous jaw model are extracted by a feature encoder, the heat map features are input into a multilayer perceptron to obtain category features, the heat map features and category features are input into a Transformer key point recognition module to obtain multi-scale heat map features and category features, and reasoning is performed on the multi-scale heat map features and category features to obtain the key point location and key point type of the edentulous jaw model.

5. The intelligent design method for complete dentures based on multi-task collaborative deep learning according to claim 4, characterized in that, The heatmap features and category features are input into the Transformer keypoint recognition module to obtain multi-scale heatmap features and category features. Specifically, the input of the Transformer keypoint recognition module is the category features and heatmap features. The location-encoded heatmap features are added to the category features. The self-attention module enhances the keypoint heatmap position and category of the category features. The cross-attention module further enhances the attention heatmap of key positions by combining heatmap information, making the keypoint positions more obvious.

6. The intelligent design method for complete dentures based on multi-task collaborative deep learning according to claim 3, characterized in that, The composite loss function is constructed based on the bite contact condition and posture adjustment constraints as follows: L Total =L reg +w p L pen +w c L gap +w t λ t L trans +w r λ r L rot ; In the formula, L Total For composite loss, L reg For network loss, L pen As a penetration depth penalty, L gap To avoid punishment, L trans For 6-DOF position change penalty, L rot For 6-DOF pose change penalty, w p w c w t w r L respectively pen L gap L trans L rot The weight, λ t , λ r L respectively trans L rot The pose weights are: V and v, where V and v are the set of vertices of the tooth model and the vertices in the set, respectively; SDF() is the signed distance function for calculating the vertices; R and t are the rotation matrix and position vector in the 6-DOF pose parameters of the current tooth, respectively; R0 and t0 are the rotation matrix and position vector in the initial model pose parameters, respectively; ε c The contact threshold is used to control the gap depth.

7. The intelligent design method for complete dentures based on multi-task collaborative deep learning according to claim 1, characterized in that, Using the denture base boundary and the denture tooth arrangement position as constraints, a composite loss function is constructed based on the curvature of the generated surface. A conditional generative adversarial network is then used to generate the denture base model. Specifically, the artificial tooth arrangement model and the denture base boundary are used as constraints, and a composite loss function L is constructed based on the denture base curvature and the denture surface wrapping rate. 曲率 With L 义齿包绕 The generator is used to initially generate a complete denture base model. The loss is calculated based on the surface geometric features of the generated complete denture base model and the real denture base model as the discriminator loss. The generator parameters are updated in the next round of training through backpropagation and gradient reduction algorithm to make the generated result closer to the real denture base model, until the loss function converges, thus realizing the generation of the complete denture base model.

8. The intelligent design method for complete dentures based on multi-task collaborative deep learning according to claim 7, characterized in that, Constructing a composite loss function L 曲率 With L 义齿包绕 Specifically: In the formula, L 曲率 With L 义齿包绕 These are the curvature loss function and the denture wrapping rate loss function, respectively. base With v base Let A(v) be the set of vertices in the Kito model and the set of vertices in the set. base ) is the vertex v base The sum of the areas of the neighboring triangles, V N u and v are the vertices of v base The set of adjacent vertices and the vertices in the set, where cot represents the cotangent function. and For the edge (v) on the Kito model base The opposite angles on both sides of u, H0 and R target These are the curvature threshold and the wrapping rate threshold, respectively, P root p and ε are the vertex set and points in the denture model, respectively. δ(p) is the Boolean value for determining the wrapping condition. env This is the wrap-around distance threshold.

Images