Intelligent generation method of personalized adaptive parameters for denture

Abstract

The application relates to the technical field of oral digitization, and discloses a method for intelligently generating personalized adaptive parameters of dentures, which comprises the following steps: acquiring multi-modal oral data of a patient and extracting initial features of each mode, taking a four-level dental-jaw anatomical topological structure as a benchmark, performing cross-modal attention fusion with a topological weight to obtain a global oral feature vector, inputting the global feature into a field fine-tuning large language model to infer and output structured denture design constraint instructions, taking clinical contraindication hard rules as boundaries, combining semantic soft constraints, performing three-level hierarchical deduction to generate initial denture design parameters, and finally outputting final parameters through semantic-numerical bidirectional iteration checking. The application deeply combines an AI algorithm with characteristics of the denture field, can adaptively generate personalized denture parameters considering aesthetics, mechanics and technology, improves design efficiency and standardization, and is suitable for various denture digital design scenes.

Description

Technical Field

[0001] This application relates to the technical field of digital oral cavity, and in particular to a method for intelligent generation of personalized adaptive parameters for dentures. Background Technology

[0002] With the popularization of digital dental technology, denture restoration has gradually achieved digital design and processing. The rationality of denture design parameters directly determines the clinical fit, aesthetic effect, and lifespan of the restoration. Currently, in the denture design process, various parameters such as crown morphology, occlusal space, transgingival contour, and processing allowance rely heavily on manual adjustments by technicians based on their personal experience. This requires simultaneous consideration of the patient's oral condition, restorative needs, and processing requirements, placing extremely high demands on the technician's professional skills and experience. This not only results in low design efficiency but also significant fluctuations in parameter quality due to the influence of personnel skill levels, easily leading to repeated design modifications and high rework rates.

[0003] Some digital design systems use fixed template matching to generate basic parameters, but these systems can only perform simple calculations based on single 3D model geometric data. They cannot integrate multi-dimensional information such as patient medical records, oral imaging, and occlusal physiology for comprehensive judgment. Furthermore, the lack of precise correspondence between multimodal data and insufficient feature alignment accuracy can easily lead to deviations in parameter generation. Meanwhile, existing parameter generation schemes mostly adopt a unidirectional rule mapping model, with rigid parameter adjustment logic. This makes it impossible to simultaneously balance the coupled requirements of aesthetic form, mechanical strength, and fabrication feasibility. Complex cases still require extensive manual correction, and parameter verification results cannot be used to back-optimize the design logic, making it difficult to form an adaptive optimization mechanism. This hinders the standardization and large-scale application of digital denture design. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a method for intelligently generating personalized adaptive parameters for dentures, employing the following technical solution: The intelligent generation method for personalized adaptive parameters of dentures includes the following steps: Acquire multimodal oral data of patients and perform standardized preprocessing to extract initial feature vectors for each modality; Using the four-level anatomical topology of the jaw as the alignment benchmark, cross-modal attention fusion with topological weights is performed on the initial feature vectors of each modality to obtain the global oral cavity feature vector; The global oral feature vector is input into a large language model fine-tuned in the field of prosthodontics, and the structured denture design constraint instructions are output through semantic reasoning. The initial denture design parameters are generated through three-level hierarchical deduction, using hard rules of clinical contraindications as boundary constraints and combining the semantic soft constraints of denture design constraint instructions. The initial denture design parameters are subjected to process and mechanical verification. If the verification fails, the non-compliant numerical values ​​are converted into semantic correction instructions and input back into the large language model. The parameters are regenerated after iterative adjustment until the verification passes and the final denture design parameters are output.

[0005] Optionally, the multimodal oral data includes text modal data, image modal data, three-dimensional model modal data, and physiological parameter modal data; the text modal data includes patient medical records, periodontal condition, allergy history, restoration requests, and previous restoration records, and semantic feature vectors are extracted after preprocessing; the image modal data includes intraoral photographs, X-rays, and CBCT tomographic images, and image feature vectors of bone volume, gingival morphology, and adjacent tooth distribution are extracted using a convolutional neural network; the three-dimensional model modal data is an intraoral scan of a three-dimensional dental model, and geometric feature vectors of shoulder morphology, tooth loss, occlusal gap, and surface features are extracted using a point cloud feature extraction network; the physiological parameter modal data includes occlusal motion trajectory, occlusal force range, and oral opening degree, and physiological feature vectors are extracted using a temporal encoder.

[0006] Optionally, the four-level anatomical topology of the jaw includes the crown layer, shoulder layer, periodontal layer, and bone tissue layer, with each modality's initial feature vector corresponding to a feature region at a different anatomical level; the cross-modal attention fusion calculation formula with topological weights is as follows: ; Where Q is the feature query vector of the current anatomical level, K is the key vector of the initial features of each modality, and V is the value vector of the initial features of each modality. Let be the dimension of the key vector. Here, represents the topological weight matrix corresponding to the anatomical level, ⊙ represents the Hadamard product operation, and Softmax is the normalized exponential function.

[0007] Optionally, the large language model fine-tuned in the field of prosthodontics is constructed in the following way: using an open-source lightweight large language model as the base, a corpus of prosthodontics as the training data, and a low-rank adaptation method for domain fine-tuning; the corpus of prosthodontics includes annotated prosthodontic medical records, tooth morphology standards, occlusal medical rules and dental technician processing specifications, and the fine-tuning tasks include generating denture design constraint instructions and judging the rationality of design schemes.

[0008] Optionally, the denture design constraint instructions are structured semantic tags, including defect type tags, aesthetic requirement level, occlusal tightness, material contraindication range and process suitability type, and do not contain direct numerical parameters.

[0009] Optionally, the clinical contraindication hard rules include material contraindication rules, occlusal safety threshold rules, and processing limit parameter rules, which are the mandatory boundaries for parameter generation; the three-level hierarchical deduction sequentially generates basic geometric parameters, functional mechanical parameters, and process adaptation parameters, with the output of the previous level parameters serving as the constraint conditions for the next level parameters.

[0010] Optionally, the generated initial denture design parameters include morphological parameters, mechanical parameters, process parameters, and material parameters; wherein the morphological parameters include crown curvature, cusp inclination, interproximal space, and transgingival profile; the mechanical parameters include occlusal contact points, occlusal space, and load-bearing thickness; the process parameters include cutting allowance, printing layer thickness, and support placement threshold; and the material parameters include the type of suitable material and hardness parameters.

[0011] Optionally, the process and mechanical verification includes processing feasibility verification and mechanical safety verification; processing feasibility verification compares the initial denture design parameters with the process threshold of the corresponding processing equipment, and mechanical safety verification confirms that the stress distribution meets the material safety standards through simplified finite element calculation; when the verification fails, the numerical boundaries of non-compliant items are automatically extracted, converted into corresponding semantic constraint labels, and fed back to the large language model.

[0012] Optionally, it also includes incremental optimization steps: collecting labeled clinical rework data and repair effect evaluation data, periodically making incremental fine-tuning to the large language model, and updating the adjustment coefficients of hard rule boundaries and hierarchical inference to continuously optimize the adaptability of parameter generation.

[0013] In summary, this application includes at least the following beneficial technical effects: By fusing cross-modal features based on the anatomical topology of the jaw, this application can achieve precise alignment of multi-source heterogeneous data by relying on the anatomical hierarchy unique to the field of prosthetic restoration, effectively improving the matching accuracy of multi-modal features and overcoming the problem of insufficient alignment accuracy of general cross-modal algorithms in oral scenarios, providing comprehensive and accurate data support for parameter generation. A three-level hierarchical deduction mechanism combining hard rules of clinical contraindications and soft semantic constraints is adopted. This mechanism ensures the clinical safety and processing feasibility of parameters through hard rules, while enabling personalized adaptive adjustment of parameters through semantic reasoning of a large language model. Simultaneously, the hierarchical deduction method solves the technical problem of balancing multi-parameter coupling, ensuring that the generated parameters simultaneously meet multiple requirements of aesthetics, function, and process. A bidirectional iterative calibration closed loop is constructed between semantic instructions and numerical parameters, automatically converting numerical verification results into semantic constraints to feed back into the model, achieving automatic correction of design logic without manual intervention to adjust rules, effectively improving the accuracy and efficiency of parameter generation. This application deeply integrates artificial intelligence algorithms with the domain characteristics of denture design, rather than simply applying general technologies. It can effectively reduce reliance on the experience of senior technicians, improve the standardization and efficiency of denture design, and can be widely applied to various digital denture design scenarios. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall process of the intelligent generation method for personalized adaptive parameters of dentures in this application. Detailed Implementation

[0015] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0016] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0017] like Figure 1 As shown, the specific implementation steps of the personalized adaptive parameter intelligent generation method for dentures described in this application are as follows: Step S1: Multimodal Oral Data Acquisition and Preprocessing. Four different modalities of oral data were acquired from the patient. Standardized preprocessing was performed on each modality to extract the initial feature vectors for that modality. Text modal data was collected through an electronic medical record system, including unstructured text content such as basic patient information, periodontal condition, history of material allergies, restoration requests, and previous restoration records. This text was encoded and embedded using a word segmenter corresponding to the base model, yielding a 768-dimensional semantic feature vector. Image modal data included intraoral clinical photographs, panoramic X-rays, and CBCT tomographic images. Intraoral photographs and X-rays underwent feature extraction using a ResNet50 convolutional neural network to obtain two-dimensional image features of gingival morphology, adjacent tooth arrangement, and the extent of tooth defects. CBCT images were processed using a three-dimensional convolutional network to extract bone tissue features such as bone height, bone width, and nerve canal location corresponding to the target tooth position. Finally, these features were stitched together to obtain a unified-dimensional image feature vector. The 3D model modal data consisted of a 3D mesh model of the jaw and teeth acquired by an intraoral scanner. Preprocessing was performed first, including point cloud downsampling, denoising, and a unified coordinate system. Then, the PointNet++ point cloud feature extraction network was used to extract the shoulder edge morphology, tooth defect volume, occlusal gap of the opposing teeth, and surface features of the adjacent teeth's contact surfaces, outputting a 1024-dimensional geometric feature vector. Physiological parameter modal data was acquired using a mandibular motion trajectory analyzer and an occlusal force testing instrument, including temporal data such as the patient's maximum open bite, lateral occlusal motion trajectory, and centric occlusal force distribution range. This data was encoded by a Transformer temporal encoder to obtain a 256-dimensional physiological feature vector.

[0018] Step S2: Cross-modal feature fusion with topological weights. Using the four-level anatomical topology of the jaw as the alignment benchmark, a cross-modal attention fusion module with topological weights is constructed to perform hierarchical alignment and fusion of the initial feature vectors of the four modalities, generating a global oral feature vector. The four-level anatomical topology of the jaw, from shallow to deep, consists of the crown layer, shoulder layer, periodontal layer, and bone tissue layer. The crown layer corresponds to the crown geometry of the 3D model and the appearance features of the intraoral photograph; the shoulder layer corresponds to the shoulder features of the 3D model and the tooth edge features in the image; the periodontal layer corresponds to the gingival features in the image and the periodontal condition information in the text; and the bone tissue layer corresponds to the bone features of the CBCT image and the bone volume information in the text.

[0019] During the fusion process, for each anatomical level, features of the corresponding region in each modality are extracted as key vector K and value vector V. The baseline anatomical features of that level are used as the query vector Q, and an attention value with topological weights is calculated. The specific calculation formula is as follows: ; In the formula, Q is the feature query vector of the current anatomical level, with a dimension of 1× K is the feature key vector corresponding to the current level in a single modality, with dimensions n× , where n is the sequence length of the modality feature; V is the feature value vector corresponding to the current level in a single modality, with the same dimension as K; The dimension of the key vector is used to scale the dot product result to avoid excessively large values; This is the topological weight matrix corresponding to the current anatomical level. Each element in the matrix represents the weight percentage of different feature regions within that level. Initial values ​​can be preset based on clinical experience and optimized through model training. ⊙ represents the Hadamard product operation, used to multiply the topological weights element-wise with the attention score. Softmax is the normalization exponential function, used to normalize the final attention weights to the 0-1 range. After completing the attention calculation for four levels, the fused features from each level are concatenated according to their dimensions to obtain a 2048-dimensional global oral feature vector.

[0020] Step S3: Semantic Reasoning of the Large Language Model Generates Design Constraint Instructions. The global oral feature vector and patient tooth position identification information are input into the large language model fine-tuned for the dental prosthodontics domain. The model outputs structured denture design constraint instructions through internal semantic reasoning. The large language model is based on an open-source lightweight large language model and uses a low-rank adaptation method for domain fine-tuning. During fine-tuning, the backbone parameters of the base model are frozen, and only the low-rank matrix of the attention layer is trained to reduce training costs and deployment barriers. The dental prosthodontics corpus used for training contains clinical prosthodontic cases annotated by licensed physicians. Each case corresponds to the defect type, restoration plan, design points, parameter constraints, and process requirements. It also incorporates professional materials such as tooth anatomy standards, occlusal design medical rules, and denture processing specifications. Fine-tuning sets up two parallel tasks: the first is to output structured design constraint instructions based on the input case features, and the second is to determine the clinical rationality of a given design plan. The model training uses a cross-entropy loss function, with 8 to 12 training rounds and an initial learning rate of 2e-4.

[0021] The final denture design constraint instructions output by the model are in the form of structured semantic labels, specifically including five categories of constraints: defect type classification, aesthetic requirement level, occlusal tightness, material contraindication range, and process adaptation type. The model does not directly output numerical parameters throughout the process, which fully leverages the semantic understanding and clinical logical reasoning capabilities of the large language model while avoiding the shortcomings of insufficient precision in directly outputting numerical values.

[0022] Step S4: The three-level hierarchical parameter deduction under dual constraints uses hard clinical contraindication rules as mandatory boundaries, combined with semantic soft constraints output by the large language model, to generate initial denture design parameters through a three-level hierarchical deduction mechanism. The hard clinical contraindication rules are pre-built into the rule base, including material contraindication rules, occlusal safety threshold rules, and processing limit parameter rules. These are unbreakable mandatory constraints. For example, patients with a history of nickel alloy allergy are directly excluded from nickel-containing materials, the minimum occlusal gap must not be less than 0.02 mm, and the minimum wall thickness for machining must not be less than 0.5 mm, etc. During the parameter generation process, all values ​​must fall within the feasible domain defined by the hard rules.

[0023] The three-level hierarchical deduction adopts the logic of "benchmark parameter anchoring - semantic label matching - hard rule boundary verification - hierarchical constraint transfer" and executes them sequentially. The output of the previous level serves as the constraint condition for the next level. Specifically, the first level generates basic geometric morphological parameters. The inputs are the target tooth position identifier, defect type label, aesthetic requirement level label, and adjacent tooth benchmark anatomical data extracted from the intraoral scan 3D model. A standard tooth parameter benchmark library based on the industry standard of prosthodontics is pre-constructed, which includes the standard values ​​and reasonable ranges of morphological parameters for each tooth position in the entire dentition.

[0024] During the simulation, the corresponding baseline parameters are first retrieved based on the target tooth position; then, the parameter compensation amount is determined according to the defect type label; and the morphological adjustment coefficient is matched according to the aesthetic requirement level. When the aesthetic requirement level is high, the labial curvature values ​​of adjacent healthy natural teeth are extracted from the intraoral 3D model as a reference benchmark. The least squares method is used to fit the crown curvature curve of the target tooth position, controlling the curvature deviation between the restoration and adjacent teeth to ≤5%. Simultaneously, the transgingival contour morphology is optimized to ensure coordination with the gingival margin of adjacent teeth. After all parameters are calculated, the boundaries are verified through hard rules, and morphological parameters such as crown curvature, cusp inclination, interproximal gap, and transgingival contour are output.

[0025] The second level generates functional biomechanical parameters. The inputs are the morphological parameters, occlusal tightness labels, occlusal force range parameters, and occlusal motion trajectory data extracted from the first level. Based on the tooth morphology generated in the first level, a biomechanical safety factor is matched to the occlusal force level; cases with high occlusal force have a 1.2-fold increase in safety factor. The occlusal gap baseline value is adjusted according to the occlusal tightness; when the occlusion is tight, the centric occlusal gap is reduced from the standard 0.08mm to 0.03~0.05mm, while simultaneously increasing the minimum thickness of the load-bearing zone. The occlusal contact point is based on the standard cusp-fossa alignment distribution rules, adapted to the individual patient's occlusal trajectory. After this level of parameters is validated by the hard rule of occlusal safety threshold, the outputs biomechanical parameters such as occlusal contact point, occlusal gap, and minimum thickness of the load-bearing zone.

[0026] The third level generates process adaptation parameters. The inputs are the morphological and mechanical parameters, process adaptation type label, and adaptable material type output from the first two levels. Based on the processing technology type, a corresponding process allowance benchmark is matched: 0.2~0.3mm single-sided cutting allowance for CNC cutting processes, and 0.1~0.15mm printing layer thickness and 30°~45° support placement angle threshold for 3D printing processes. The polishing allowance is adjusted based on material hardness: 0.05mm polishing allowance for ceramic materials and 0.1mm polishing allowance for resin materials. The minimum wall thickness parameters determined in the first two levels serve as constraint boundaries for the process parameters, ensuring that the finished product wall thickness meets mechanical safety requirements. The final output includes cutting allowance, printing layer thickness, support placement threshold, polishing allowance, and other process parameters, while also matching the corresponding repair material type and hardness parameters.

[0027] During the simulation process, if the lower-level simulation finds that the parameters of the upper level cannot meet the process or mechanical requirements, the parameters within this level are first fine-tuned. If the parameters are still not met after fine-tuning, feedback is sent to the next higher level for adjustment. Finally, the initial denture design parameters that take into account aesthetics, function and process requirements are obtained within the hard rule boundaries.

[0028] Step S5: Two-way iterative verification and result output. Perform processing feasibility verification and mechanical safety verification on the generated initial denture design parameters. Processing feasibility verification compares each process parameter with the corresponding processing equipment's process threshold to confirm that the parameters meet the equipment's processing capability requirements; mechanical safety verification uses the built-in simplified finite element rapid calculation module to calculate the stress distribution of the denture structure based on the patient's occlusal force parameters to confirm that the maximum stress value is lower than the fracture strength threshold of the corresponding material.

[0029] If all validations pass, the current parameters are output as the final denture design parameters and can be directly imported into denture design software for subsequent 3D modeling. If any validation fails, the system automatically extracts the numerical boundary requirements of the non-compliant items and converts them into corresponding semantic constraint labels. For example, "insufficient thickness in the occlusal bearing area" is converted into the semantic instruction "the mechanical strength of the occlusal area needs to be improved." This is then input back into the large language model, which adjusts the design constraint strategy and re-executes the parameter generation step, forming an iterative closed loop of "semantic instruction - numerical parameter - validation feedback - semantic correction" until all parameters pass validation.

[0030] In another embodiment of this application, incremental optimization steps for the model and rules are also included: regularly collecting rework case data and restoration effect evaluation data from the clinical end, which are then annotated by professional physicians and added to the oral restoration professional corpus; incremental fine-tuning of the large language model is performed every 3 months; at the same time, based on the parameter adaptation of clinical feedback, the boundary values ​​of clinical contraindication hard rules and the adjustment coefficients of hierarchical inference are updated, so that the parameter generation results continuously meet the actual clinical needs and continuously improve the adaptability and accuracy of the solution.

[0031] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for intelligently generating personalized adaptive parameters for dentures, characterized in that, Includes the following steps: Acquire multimodal oral data of patients and perform standardized preprocessing to extract initial feature vectors for each modality; Using the four-level anatomical topology of the jaw as the alignment benchmark, cross-modal attention fusion with topological weights is performed on the initial feature vectors of each modality to obtain the global oral cavity feature vector; The global oral feature vector is input into a large language model fine-tuned in the field of prosthodontics, and the structured denture design constraint instructions are output through semantic reasoning. The initial denture design parameters are generated through three-level hierarchical deduction, using hard rules of clinical contraindications as boundary constraints and combining the semantic soft constraints of denture design constraint instructions. The initial denture design parameters are subjected to process and mechanical verification. If the verification fails, the non-compliant numerical items are converted into semantic correction instructions and input into the large language model. The parameters are regenerated after iterative adjustment until the verification passes and the final denture design parameters are output. The four-level anatomical topology of the jaw includes the crown layer, shoulder layer, periodontal layer, and bone tissue layer. The initial feature vector of each modality corresponds to the feature region of different anatomical levels. The cross-modal attention fusion calculation formula with topological weights is as follows: ； Where Q is the feature query vector of the current anatomical level, K is the key vector of the initial features of each modality, and V is the value vector of the initial features of each modality. Let be the dimension of the key vector. Here is the topological weight matrix corresponding to the anatomical level, ⊙ represents the Hadamard product operation, and Softmax is the normalized exponential function; The clinical contraindication rules include material contraindication rules, occlusal safety threshold rules, and processing limit parameter rules, which serve as the mandatory boundaries for parameter generation; the three-level hierarchical deduction sequentially generates basic geometric parameters, functional mechanical parameters, and process adaptation parameters, with the output of the parameters at the previous level serving as the constraints for the parameters at the next level. The initial denture design parameters include morphological parameters, mechanical parameters, process parameters, and material parameters. Among them, morphological parameters include crown curvature, cusp inclination, interproximal space, and transgingival contour; mechanical parameters include occlusal contact points, occlusal space, and load-bearing thickness; process parameters include cutting allowance, printing layer thickness, and support placement threshold; and material parameters include the type of suitable material and hardness parameters.

2. The method for intelligent generation of personalized adaptive parameters for dentures according to claim 1, characterized in that, The multimodal oral data includes text modal data, image modal data, 3D model modal data, and physiological parameter modal data. The text modal data includes patient medical records, periodontal condition, allergy history, restoration requests, and previous restoration records, and semantic feature vectors are extracted after preprocessing. The image modal data includes intraoral photographs, X-rays, and CBCT tomographic images, and image feature vectors of bone volume, gingival morphology, and adjacent tooth distribution are extracted using a convolutional neural network. The 3D model modal data is an intraoral scan of a 3D dental model, and geometric feature vectors of shoulder morphology, tooth loss, occlusal gap, and surface features are extracted using a point cloud feature extraction network. The physiological parameter modal data includes the bite motion trajectory, bite force range and oral cavity opening degree, and physiological feature vectors are extracted by a time encoder.

3. The method for intelligent generation of personalized adaptive parameters for dentures according to claim 1, characterized in that, The large language model fine-tuned in the field of prosthodontics is constructed in the following way: based on an open-source lightweight large language model, using a corpus of prosthodontics as training data, and employing a low-rank adaptation method for domain fine-tuning; the corpus of prosthodontics includes annotated prosthodontic medical records, tooth morphology standards, occlusal medical rules and dental technician processing specifications, and the fine-tuning tasks include generating denture design constraint instructions and judging the rationality of design schemes.

4. The method for intelligent generation of personalized adaptive parameters for dentures according to claim 1, characterized in that, The denture design constraints are structured semantic tags, including defect type tags, aesthetic requirement level, occlusal tightness, material contraindications, and process compatibility type, but do not contain direct numerical parameters.

5. The method for intelligent generation of personalized adaptive parameters for dentures according to claim 1, characterized in that, The process and mechanical verification includes processing feasibility verification and mechanical safety verification. Processing feasibility verification compares the initial denture design parameters with the process threshold of the corresponding processing equipment. Mechanical safety verification confirms that the stress distribution meets the material safety standards through simplified finite element calculation. When the verification fails, the numerical boundaries of non-compliant items are automatically extracted, converted into corresponding semantic constraint labels, and fed back to the large language model.

6. The method for intelligent generation of personalized adaptive parameters for dentures according to claim 1, characterized in that, It also includes incremental optimization steps: collecting labeled clinical rework data and repair effect evaluation data, regularly making incremental fine adjustments to the large language model, and updating the adjustment coefficients of hard rule boundaries and hierarchical inference to continuously optimize the adaptability of parameter generation.

Images