CN-121812179-B - Deep learning method for analyzing occlusion contact point of oral scanning data

CN121812179BCN 121812179 BCN121812179 BCN 121812179BCN-121812179-B

Abstract

The invention provides an oral scanning data occlusion contact point analysis method adopting deep learning, which realizes point cloud data preprocessing through self-adaptive outlier filtering and rigid registration, models a deformation sensitive area by combining tooth surface curvature and normal vector change, extracts geometric and deformation semantic features by a light-weight deep learning structure, enhances feature resolution by adopting a feature decoupling and attention gating mechanism, reconstructs a priori motion topological graph and a graph roll propagation to improve the time sequence and physiological consistency of candidate contact points, outputs three-dimensional coordinates, intensity probability and tooth position labels of the contact points by a multi-scale regression model, and verifies the auxiliary result by a multi-threshold fusion mechanism.

Inventors

PENG BO
WANG SHAOHONG
XIANG QIAN
MENG SI
ZHANG YAN
LIU JIAQIAN
ZENG SUJUAN

Assignees

广州医科大学附属口腔医院（广州医科大学羊城医院）

Dates

Publication Date: 20260508
Application Date: 20260304

Claims (9)

1. The method for analyzing the occlusion contact point of the oral scanning data by deep learning is characterized by comprising the following steps of: S1, preprocessing original upper and lower jaw oral cavity scanning point cloud data by adopting a data preprocessing module to generate a preprocessed oral cavity scanning point cloud sequence; S2, performing local curvature gradient field modeling processing on the preprocessed oral cavity scanning point cloud sequence, and generating a deformation sensitivity heat map based on a region response characteristic extraction method vector disturbance response mode of the tooth surface under the occlusion pressure; S3, based on the deformation-sensitive heat map and the preprocessing oral cavity scanning point cloud sequence, performing double-flow embedding coding processing, wherein a main path adopts PointMLP structures to extract geometric invariant features, and a bypass path takes the deformation-sensitive heat map as guidance to estimate elastic deformation potential energy distribution to generate geometric feature vectors and deformation semantic embedding vectors; s4, based on the geometric feature vector and the deformation semantic embedded vector, performing deformation-motion decoupling attention gating processing to generate a decoupling characterization vector; S5, based on the decoupling characterization vector, introducing time sequence consistency constraint processing, constructing a priori motion topological graph by utilizing a clinical occlusion physiological rule, mapping candidate contact points to the priori motion topological graph nodes according to anatomical positions, and generating a time sequence consistency candidate contact point set by graph convolution propagation deformation semantic confidence, wherein the method specifically comprises the following steps of: Acquiring an anterior tooth guide angle and a cuspid protection coefficient based on clinical occlusion physiological rule parameters, calculating motion constraint weights among anatomical position nodes, and generating a priori motion topological graph; Performing anatomical position mapping processing on the prior motion topological graph and the candidate contact point set, and associating each candidate contact point to a corresponding topological graph node according to a tooth position label of the candidate contact point to generate a node-contact point mapping relation; Initializing the confidence coefficient distribution of each topological graph node based on the node-contact point mapping relation and the deformation semantic confidence coefficient value, and generating initial confidence coefficient distribution; Performing graph convolution propagation processing on the initial confidence coefficient distribution and the prior motion topological graph, and applying a differentiable graph volume integrating operator to propagate deformation semantic confidence coefficient along a topological edge to generate a propagated confidence coefficient distribution; Performing confidence threshold screening processing based on the transmitted confidence distribution, reserving candidate contact points with confidence higher than a preset threshold, and generating a time sequence consistent candidate contact point set; And S6, performing multi-scale regression prediction processing on the time sequence consistent candidate contact point set, and outputting the three-dimensional coordinates of the contact point, the contact strength probability value and the belonging tooth position label.
2. The method for bite contact point analysis according to claim 1, wherein said step S6 further comprises: and S7, judging whether the time sequence consistency candidate contact point set meets a preset condition or not based on a curvature consistency threshold value, a normal consistency threshold value and a decoupling orthogonality threshold value, if so, generating a final occlusion contact point identification result, otherwise, returning to the time sequence consistency candidate contact point set for dynamic adjustment.
3. The method for bite contact point analysis of oral scan data using deep learning of claim 1, wherein the data preprocessing module integrates an adaptive outlier filtering unit, a dental bow rate guided rigid registration engine, and a coordinate normalizer for bite plane references to perform noise suppression, spatial alignment, and scale normalization processing on raw maxillary and mandibular oral scan point cloud data to generate a preprocessed oral scan point cloud sequence.
4. The method for analyzing occlusion contact points of oral scan data using deep learning according to claim 1, wherein said step S3 specifically comprises: constructing a differentiable surface fitting module based on the deformation sensitivity heat map generated in the step S2, wherein the module models the response characteristics of the micro-concave/micro-convex areas of the tooth surface under the occlusion pressure by utilizing a B spline surface fitting algorithm, estimates the elastic deformation potential energy distribution of each sampling point neighborhood, and generates a differentiable surface fitting module example; Performing point cloud downsampling processing on the preprocessed oral scanning point cloud sequence, extracting local geometric features by applying a PointMLP-structure convolutional neural network layer, integrating multi-scale geometric information by a feature aggregation layer, generating geometric invariant features by using a full-connection layer, and outputting geometric feature vectors; And executing surface fitting processing on the preprocessed oral cavity scanning point cloud sequence based on the differentiable surface fitting module and the deformation sensitivity heat map, and generating a deformation semantic embedding vector.
5. The method for analyzing occlusion contact points of oral scan data by deep learning according to claim 4, wherein the surface fitting process specifically comprises identifying micro-deformation sensitive areas on tooth surfaces according to a deformation sensitive thermal map, calculating elastic deformation potential energy distribution of each sampling point neighborhood by using a differentiable surface fitting module, and converting the elastic deformation potential energy distribution into low-dimensional semantic embedding by linear mapping to generate deformation semantic embedding vectors.
6. The method for analyzing occlusion contact points of oral scan data using deep learning according to claim 1, wherein said step S4 specifically comprises: Based on the dimensional characteristics and distribution statistical parameters of the geometric feature vectors and the deformation semantic embedded vectors, a deformation-motion decoupling attention gating module is constructed, and the deformation-motion decoupling attention gating module integrates a cross entropy constraint contrast learning unit to initialize a feature decoupling processing environment and generate a deformation-motion decoupling attention gating module instance; Based on the deformation-motion decoupling attention gating module instance, performing cross entropy constraint contrast learning processing on the geometric feature vector and the deformation semantic embedding vector, calculating a feature similarity matrix in a potential space, and generating an orthogonalization indicating signal; Based on the orthogonalization indication signal, calculating an inner product loss value of the geometric feature vector and the deformation semantic embedding vector, and generating a gradient updating direction; optimizing parameter configuration of the geometric feature vector and the deformation semantic embedding vector according to the gradient updating direction to generate orthogonalization feature pairs; Based on the orthogonalization feature pairs, fusing the geometric feature vectors and the deformation semantic embedding vectors to generate decoupling characterization vectors.
7. The method of claim 6, wherein the decoupling characterization vector comprises a geometric component that characterizes rigid pose changes and a deformation semantic component that characterizes nonlinear local responses.
8. The method according to claim 1, wherein the multi-scale regression prediction process includes synchronously predicting three-dimensional coordinates of contact points, contact intensity probability values, and belonging dental position labels based on a time-series coincidence candidate contact point set by constructing a multi-branch neural network model including three-dimensional coordinate prediction branches, contact intensity probability prediction branches, and dental position label prediction branches, the three-dimensional coordinate prediction branches processing geometric feature information using a fully-connected neural network structure, the contact intensity probability prediction branches outputting probability distributions using a sigmoid activation function, and the dental position label prediction branches identifying anatomical position categories using a softmax classifier.
9. The method for bite contact point analysis according to claim 2, wherein the curvature consistency threshold is defined as a threshold of a standard deviation of gaussian curvature of a neighborhood of a predicted contact point, the normal consistency threshold is determined according to a contact direction included angle tolerance set in a training phase normal consistency loss function, and the decoupling orthogonality threshold is determined according to an inner product convergence threshold set in a training phase decoupling orthogonality loss function.

Description

Deep learning method for analyzing occlusion contact point of oral scanning data Technical Field The invention relates to the technical field of oral digital medical treatment and three-dimensional point cloud intelligent analysis, in particular to an oral scanning data occlusion contact point analysis method adopting deep learning. Background Currently, in the field of oral digital medical treatment, intelligent analysis technology based on three-dimensional point cloud data is widely focused on digital occlusion relation reconstruction, tooth morphology reconstruction, automatic identification of occlusion contact points and the like. With the popularization of three-dimensional oral scanners and the development of point cloud deep learning technology, the analysis of the occlusion state of teeth by using point cloud data gradually becomes an important information foundation for intelligent oral diagnosis and treatment, correction and repair. The prior art mainly adopts a method based on point cloud geometric feature extraction, semantic segmentation, rigid body registration or mechanical simulation combination to detect the occlusion contact point. In addition, part of technical routes attempt to introduce multi-modal force feedback, pressure sensors, RGB-D multi-view fusion and wearable equipment to acquire synchronous occlusion sequences, so that finer reconstruction of dynamic occlusion behaviors is realized. However, the above-described techniques still face a number of typical challenges; Firstly, the current technology focuses on morphological segmentation of static point cloud, and has limited modeling capability on dynamic process and micro deformation of occlusion itself. Because of the elastic deformation of the tooth surface, such as micro-pits, micro-ridges, etc., during occlusion of the mouth, it is often difficult for a single geometric feature to distinguish between a true occlusion contact point and a false contact region. In addition, the existing semantic segmentation scheme based on the end-to-end neural network generally depends on a large amount of tagged data, the generalization capability is easily influenced by personal anatomical structure differences and acquisition posture disturbance, and deformation response of teeth under dynamic occlusion pressure is difficult to be described; Secondly, a part of the currently disclosed technical paths are used for improving the dynamic occlusion analysis precision by combining with equipment such as a pressure sensor array, a multi-mode motion capture equipment for synchronous acquisition, an external vision mechanics system and the like. The method not only obviously increases the hardware cost, leads to complicated operation flow and is not suitable for popularizing scenes, but also is limited by the consistency of equipment resolution, synchronization precision and data calibration, and is easy to cause semantic information loss or false contact point false detection. Meanwhile, the scheme has poor compatibility with the traditional clinical digital oral workflow, and is difficult to be widely adapted to different types of acquisition instruments or historical oral scanning data; And finally, the existing part of methods adopt modes such as rigid transformation parameter regression, LSTM time sequence network, nonlinear motion simulation and the like to directly return to an occlusion state, but the fitting is very easy to occur due to the fact that the characteristic decoupling of space geometric change and local nonlinear deformation cannot be effectively realized, and the adaptability to complex occlusion paths and physiological variation is limited. The robustness and clinical interpretability of such models in contact point identification are universally insufficient; in addition, although the traditional semantic segmentation or surface matching scheme shows a certain point cloud analysis capability, it is difficult to fully combine occlusion physiological rules, such as time sequence and structure topology prior of front tooth guidance, cusp protection and the like, so that the dynamic discontinuous occlusion scene and non-standard oral morphology generalization capability is poor. Disclosure of Invention The invention aims to solve the technical problems and provides an oral scanning data occlusion contact point analysis method adopting deep learning. The technical scheme of the invention is realized by adopting a deep learning method for analyzing the occlusion contact point of the oral scanning data, which comprises the following steps: S1, receiving a preprocessed upper and lower jaw oral cavity scanning point cloud sequence as input data, wherein the point cloud sequence comprises discrete point cloud frames in different occlusion stages and is used for dynamically identifying subsequent occlusion contact points; s2, performing local curvature gradient field modeling processing on the preprocessed oral cavity scanning point