CN-122023708-A - Tooth mesh automatic generation method based on depth diffusion model

CN122023708ACN 122023708 ACN122023708 ACN 122023708ACN-122023708-A

Abstract

The present invention proposes an end-to-end dental auto-generation framework TLDF (Tooth Latent Diffusion Framework) designed specifically for dental and crown generation in intraoral scanner (IOS) mode. The framework is based on a newly proposed potential space depth diffusion model (TLDM), potential space features are mined by utilizing a large number of tooth Mesh data, and the proximity relation of the data in the Mesh potential space is analyzed through a multi-head attention mechanism, so that high-precision tooth and crown generation and reconstruction are realized. The technical process comprises three main steps of firstly inputting Mesh scanning data of upper jaw and lower jaw into TLDF frame, extracting vertex coordinates, normal vector and position relation of grid data, and obtaining adjacent dental crown information containing target dental crown through adjacent relation as input data. The 3D geometric features are then transformed into multi-level latent spatial features by position embedding (position embedding) using a feature transformation module (Encoder), which are combined to form latent representation vectors by a multi-headed attention mechanism. Finally, the cascade-enhanced potential representation vectors are input TLDM, enabling 3D generation of teeth and crowns by means of a three-dimensional multi-layer perceptron (MLP). TLDM progressively refine the input data and focus on the structural relationships within the mouth, thereby completing accurate tooth and crown reconstruction.

Inventors

Request for anonymity
Request for anonymity
Request for anonymity
Request for anonymity
Request for anonymity

Assignees

北京新牙科技有限公司

Dates

Publication Date: 20260512
Application Date: 20241112

Claims (10)

1. A tooth mesh automatic generation method based on a depth diffusion model is characterized by comprising the following steps: inputting Mesh scanning data in an IOS mode of the upper jaw and the lower jaw into a TLDF framework, and obtaining vertex coordinates, normal vectors and relative position relations of grid data; Step two, extracting adjacent relation, namely identifying adjacent dental crowns of the target dental crowns through the adjacent relation, and acquiring three groups of adjacent dental crown information as a part of input data; inputting the 3D geometric features adjacent to the dental crowns into a feature conversion module (Encoder) and generating multi-level potential spatial features through position embedding (position embedding); Step four, generating potential representation vectors, namely analyzing the proximity relation of data in a potential space through a multi-head attention mechanism module, and integrating the multi-stage potential space features to generate the potential representation vectors; Inputting the enhanced potential representation vector into TLDM model, and generating 3D structure by using three-dimensional multi-layer perceptron (MLP); step six, gradually refining and generating: TLDM progressively refines the input data, and based on the oral structure relationship, accurate tooth and crown generation and reconstruction are completed.
2. The method for automatically generating a tooth mesh based on a depth diffusion model according to claim 1, wherein in TLDF (Tooth Lantent Diffussion Framework) frames, input data are IOS scanning data of an upper jaw or a lower jaw of an oral cavity, the IOS scanning data comprise adjacent teeth and bases, output data are generated crown mesh comprising the adjacent teeth, and the generated result is recombined into a mesh result of a single crown.
3. The automatic tooth mesh generating method based on the depth diffusion model according to claim 2, wherein the generating result is a crown and a face tooth and a side tooth on both sides of the crown, and the crown can be extracted independently.
4. The method for automatically generating the tooth mesh based on the depth diffusion model of claim 1, wherein in the step one, the experimental training process of the IOS scan data is as follows: The initial learning rate is set to be 4e-4, the L2 decoupling regularized attenuation is carried out by using AdamW optimization algorithm with the data batch size of 16, the initial attenuation rate is 0.001, and the final attenuation rate is-1. The Mesh data is downsampled by using 10000 sampling space, and the sample expansion is performed by using a whole random space transformation mode.
5. The automatic tooth mesh generation method based on the depth diffusion model of claim 4, wherein losses occur in the experimental training process of IOS scanning data, and the loss function is expressed by SSIM, and the specific expression is: μx, μy, represents the average position of the Mesh midpoint, σ represents the distribution and variation of the Mesh midpoint around the average. C1, C2 are small constants to avoid instability when the denominator is close to zero.
6. The automatic tooth mesh generation method based on the depth diffusion model according to claim 5, wherein PSNR is used as an evaluation index, and the specific expression is as follows: Where L represents the dynamic range of Mesh midpoints and N represents the number of Mesh midpoints.
7. The automatic tooth mesh generation method based on the depth diffusion model of claim 1, wherein in the third step, the conversion flow of the feature conversion module is as follows: 3D geometric features of the adjacent dental crowns, including vertex coordinates, normal vectors, and relative positional relationships, are input to a feature transformation module (Encoder). In this module, features are combined with geometric features by positional embedding (position embedding) to form a multi-level, multi-level representation of potential spatial features.
8. The method for automatically generating a tooth mesh based on a depth diffusion model according to claim 7, wherein in the third step, potential spatial features are refined step by step, and the spatial position of the target crown and the structural information of the adjacent crowns are integrated, so that high-quality basic features are prepared for the subsequent potential representation vector generation stage.
9. The method of automatic tooth mesh generation based on depth Diffusion model of claim 1, wherein in step four, a solution Diffusion is adopted as a baseline method, and a Multi-head Attention module (Multi-head Attention) is added in the baseline method to be contained in TLDM, so as to adjust a specific region of the global neural network for selectively focusing IOS data of an oral cavity, the method is as follows: the input embedding converts the data into vector representation, and after layer normalization and linear transformation, the data enters a multi-head self-attention module to capture the interrelationship among the features and generate rich context information.
10. The automatic tooth mesh generation method based on the depth diffusion model according to claim 1 is characterized in that in the fifth step, layer normalization is performed to balance the scale of data and ensure stable model processing. The normalized features are passed into a multi-layer perceptron (MLP) to further extract complex nonlinear features through a series of fully connected layers. And finally, superposing the output of the multi-layer sensor and the initial input characteristics through residual connection, so that original characteristic information is reserved, the gradient is helped to smoothly flow in a network, the problem of gradient disappearance is prevented, and the deep learning capacity and the characteristic expression effect of the model are improved.

Description

Tooth mesh automatic generation method based on depth diffusion model Technical Field The invention relates to the technical field of medical image processing, in particular to a tooth mesh automatic generation method based on a depth diffusion model. Background In dental clinical practice, obtaining accurate three-dimensional (3D) tooth surface models is critical to effective treatment planning. Traditionally, this has been accomplished by scanning a physical impression (plaster model) of the dental arch. However, this process has several drawbacks. First, it may be uncomfortable for the patient, especially children, and may cause allergic reactions due to the chemical composition of certain impression materials. Second, it is inherently inefficient. The advent of advanced intraoral scanners (IOS) has alleviated these challenges by directly reconstructing digital tooth surface models, thereby improving patient comfort and clinical efficiency. Nevertheless, despite the advantages of IOS technology, generating accurate 3D models of individual teeth and crowns from these scans remains a complex task. This complexity stems from several factors: 1. Inter-patient anatomic variability-tooth morphology shows significant differences between individuals, even within the same dentition. 2. The irregularities caused by the misjaw deformity, such as misalignment and crowding, often lead to abnormal tooth appearance, thus complicating the segmentation process. The limitations of ios technology are that intraoral scanning typically involves irregularly shaped non-dental structures (e.g., gingival tissue) that need to be distinguished. Furthermore, due to limited accessibility of the optical scanner, full capture of the posterior tooth regions (e.g., second and third molars) can be challenging. The present disclosure is primarily directed to accurate generation of teeth and crown mesh in IOS based on depth diffusion models. While the depth diffusion model has shown considerable success in generating tasks such as natural image synthesis, its application in dental modeling has not yet been explored, mainly because of the scarcity of open tooth mesh data sets obtained from IOS. This scarcity is further exacerbated by the significant differences in tooth size, orientation and rate of tooth loss between patients. Therefore, it is important to develop an efficient, stable, automated, end-to-end framework to generate accurate and reliable tooth and crown models. Our proposed approach aims to address this challenge In the fields of general computer vision and artificial intelligence, there are various deep learning generation methods based on 3D structures. Early PointNet learned for translation-independent geometry through shared multi-layer perceptron (MLP) and global maximization pooling of point cloud data, but neglected spatial dependencies. To improve this problem PointNet ++ introduced modeling of the local geometry background, and learned the local geometry feature group by group through hierarchical point groups, capturing finer local dependencies. RSNet then order the points with directional information into a sequence and apply a Recurrent Neural Network (RNN) to enhance the model's ability to understand the spatial relationship between the different parts of the 3D shape. In addition, convolutional neural networks (e.g., pointCNN, pointNet, meshNet) based on voxelized and multi-view rendering, while improving the learning effect of surface information, the voxelized or multi-view rendering inevitably introduces artifacts, resulting in partial dependency loss. Based on the model (such as DCPR-GAN) for generating the countermeasure network, the context relation of the teeth and surrounding structures is difficult to integrate due to the limitation of the convolution network, and the countermeasure training characteristic makes the model have poor effect in processing occlusion and irregularity, so that the model is difficult to be suitable for complex tooth shape generating tasks. Disclosure of Invention The invention aims to provide a tooth mesh automatic generation method based on a depth diffusion model, which aims to solve the problems of local geometric modeling of point cloud data, spatial relationship understanding of 3D shapes and surface information capturing and context dependence in the complex tooth shape generation process in the traditional automatic generation method. Therefore, the invention provides a tooth mesh automatic generation method based on a depth diffusion model, which comprises the following steps: inputting Mesh scanning data in an IOS mode of the upper jaw and the lower jaw into a TLDF framework, and obtaining vertex coordinates, normal vectors and relative position relations of grid data; Step two, extracting adjacent relation, namely identifying adjacent dental crowns of the target dental crowns through the adjacent relation, and acquiring three groups of adjacent dental crown informatio