An artificial intelligence-based orthodontic path planning method

Abstract

The application discloses a kind of orthodontic path planning methods based on artificial intelligence.Method includes: scanning the whole teeth to be orthodontic to construct tooth model data, then obtain the initial pose of tooth;From tooth model data, tooth initial point cloud is extracted and is preprocessed to obtain the whole teeth point cloud;Orthodontic path planning model is constructed, the initial pose of tooth, preset target pose and the whole teeth point cloud are input into model, and after processing, the initial tooth movement of each step is output, and using heuristic method is post-processed, and the final tooth movement path is obtained, to realize orthodontic path planning.The method of the application is simple, easy and fast to realize, solves the problem that the accuracy of other artificial intelligence orthodontic methods is low and the planned path does not comply with the biomechanical movement rule, significantly improves the accuracy of orthodontic, and is convenient for promotion.

Description

Technical Field

[0001] This invention relates to a method for orthodontic pathway planning, specifically an artificial intelligence-based method for orthodontic pathway planning. Background Technology

[0002] Orthodontic pathway planning is fundamental to modern treatment, aiming to improve function and aesthetics by achieving ideal tooth alignment and occlusion through controlled tooth movement. Therefore, developing automated orthodontic treatment planning methods is crucial for improving clinical efficiency and expanding accessibility to orthodontic care.

[0003] Early research on automated tooth alignment focused primarily on predicting the final position of teeth. Specifically, these methods predicted the optimal final position based on the patient's initial tooth position. However, a complete orthodontic process also requires predicting the progressive morphological adaptation of the dentition throughout the treatment trajectory. In clinical practice, orthodontists must invest significant time in designing sequential movement patterns and precise treatment trajectories, implementing minute incremental adjustments (typically spanning 50 to 100 independent phases) to mitigate the risk of periodontal ligament strain and avoid significant patient discomfort associated with rapid orthodontic movement.

[0004] Recent studies have recognized this challenge and attempted to predict all intermediate steps based on initial and final tooth positions. However, these methods overlook a crucial point: the entire tooth alignment process is essentially a classic state transition problem, where tooth states and their transitions exhibit strong correlations; that is, planning tooth movement in the current step depends on all previous steps. Furthermore, the state transition process in tooth alignment is extremely complex, requiring comprehensive consideration of the morphology of each tooth at every step, as well as the overall alignment configuration, to determine subsequent positions. This requirement forces models to analyze complete dentition point cloud information at each step to accurately interpret the current alignment state. Incorporating such analysis into existing frameworks imposes a significant computational burden. Summary of the Invention

[0005] To address the problems existing in the background art, the present invention provides an artificial intelligence-based orthodontic path planning method.

[0006] The technical solution adopted in this invention is:

[0007] The artificial intelligence-based orthodontic path planning method of the present invention includes:

[0008] Step 1: Scan the entire set of teeth to be orthodontized to construct a tooth mesh model data, and obtain the initial pose of the teeth based on the tooth model data.

[0009] Step 2: Extract the initial point cloud of the teeth from the tooth model data and perform preprocessing to obtain the point cloud of the entire set of teeth.

[0010] Step 3: Construct and train an AI model based on bidirectional autoregressive prediction to obtain a dental orthodontic path planning model. Input the initial tooth pose, preset target pose, and whole-mouth tooth point cloud into the dental orthodontic path planning model for processing. After processing, the model outputs the initial tooth movement amount for each step. Heuristic methods are then used to post-process the initial tooth movement amounts to construct the final tooth movement path, thus achieving dental orthodontic path planning. Both the initial tooth pose and the preset target pose include the centroid coordinates in the whole-mouth tooth center coordinate system and the local coordinate system of each tooth from the dental model data.

[0011] In step two, the preprocessing specifically involves first downsampling the initial point cloud of each tooth using the Farthest Point Sampling (FPS) method, then normalizing the initial point cloud of the entire set of teeth, and finally obtaining the point cloud of the entire set of teeth.

[0012] In step three, the artificial intelligence model includes a tooth point cloud perceptron and a tooth transformation predictor connected in sequence. The tooth point cloud perceptron includes a pre-trained point cloud feature extractor and a Mamba-based hierarchical sequence modeling module connected in sequence. The artificial intelligence model is pre-set with a preset number of movement steps L. The initial pose and preset target pose of the teeth input to the artificial intelligence model are processed by a simple fully connected network to extract the initial pose code and target pose code. The entire tooth point cloud is processed by the pre-trained point cloud feature extractor to extract the point cloud feature code sequence. The initial pose code and the point cloud feature code sequence are concatenated to obtain the initial concatenation result, which is then input into the Mamba-based hierarchical sequence modeling module. The target pose code and the point cloud feature code sequence are concatenated to obtain the target concatenation result, which is then input into the Mamba-based hierarchical sequence modeling module. The inter-point information of the initial concatenation result and the target concatenation result are extracted respectively and then input into the tooth transformation predictor for bidirectional autoregressive prediction. The initial pose of the teeth is used as step 0 of the movement path, and the preset target pose is used as step L of the movement path. In the i-th prediction, the model receives the tooth poses from steps 0 to i-1 and from L-i+1 to L as input. The tooth transformation predictor outputs the tooth displacement and rotation amounts from steps i and Li, thus obtaining the tooth poses from steps i and Li. The poses are then returned to the tooth point cloud perceptron and the tooth transformation predictor for processing until the tooth displacement and rotation amounts from all steps are obtained as the initial tooth movement amounts.

[0013] The point cloud feature extractor specifically employs a PointNet deep learning model, a multi-layered PointNet++ deep learning model, or a Dilate-Gated Convolutional Neural Network (DGCNN) model. The point cloud feature extractor has been pre-trained and does not require retraining.

[0014] The Mamba-based hierarchical sequence modeling module comprises two sequentially connected sequence modules. Each sequence module includes, in turn, a position encoding processing module, an interdental scanning processing module, and a point-scale state-space model (SSM). The first sequence module's aggregation layer consists of a sequence feature aggregation layer and a point cloud feature aggregation layer connected sequentially. The second sequence module's aggregation layer only includes a sequence feature aggregation layer. The first sequence module's aggregation layer includes the aggregation of four rearranged point cloud feature sequences (sequence feature aggregation layer) and the aggregation of all point features in the same tooth (point cloud feature aggregation layer); the second sequence module's aggregation layer only includes the aggregation of four rearranged point cloud feature sequences (sequence feature aggregation layer). For each point cloud feature input by the Mamba-based hierarchical sequence modeling module, it is first processed by the position encoding of the first sequence module, and then interdental scanning is performed according to the scanning order of top to bottom, bottom to top, left to right, and right to left to obtain four rearranged point cloud sequences. Each rearranged point cloud sequence is then processed by the point-scale state space model (SSM) to extract inter-point feature sequences, and then the four inter-point features are aggregated by the first sequence feature aggregation layer. During aggregation, each rearranged point cloud sequence is first normalized and then passed through a fully connected layer and an activation function SiLU (Sigmoid Linear Array). Unit) obtains the weight of each point, resulting in four weight sequences. The feature sequences between the four points and the corresponding weight sequences are multiplied element by element to obtain four weighted feature sequences. Finally, the four weighted feature sequences are added element by element to obtain the sequence feature aggregation result. The first sequence module also goes through a point cloud feature aggregation layer, which takes the average of the features of all points of the same tooth to aggregate and obtain a tooth-level feature sequence. The tooth-level feature sequence is input into the second sequence module for processing and outputs a feature sequence that integrates the information between teeth.

[0015] The described tooth transformation predictor is a Transformer-based bidirectional autoregressive predictor, comprising layer normalization, several layers of bidirectional self-attention BiMHSA, and an action distribution decoder connected in sequence. The tooth transformation predictor predicts from the initial tooth pose to a preset target pose, and simultaneously predicts from the preset target pose to the initial pose. Each prediction outputs the tooth displacement and rotation at step i and step Li, including the mean and variance of the displacement and rotation along the X, Y, and Z axes, thereby obtaining the tooth pose at steps i and Li. The tooth poses at steps i and Li are returned to the tooth point cloud perceptron for processing, and the feature sequences at steps i and Li are output. These feature sequences are then concatenated with the previously obtained feature sequences from steps 0 to i-1 and from L-i+1 to L, and then input back into the tooth transformation predictor for processing, outputting the tooth displacement and rotation at steps i+1 and Li-1, until the tooth displacement and rotation of all steps are obtained as the initial tooth movement.

[0016] The aforementioned bidirectional self-attention BiMHSA is specifically a multi-head self-attention mechanism (MHSA) employing a bidirectional masking mechanism. The output and input of the first layer of bidirectional self-attention BiMHSA are residually concatenated before being input into the second layer of bidirectional self-attention BiMHSA for processing. The output of each subsequent layer of bidirectional self-attention BiMHSA is residually concatenated with the output of the previous layer of bidirectional self-attention BiMHSA before being input into the next layer of bidirectional self-attention BiMHSA for processing. Specifically, a four-layer bidirectional self-attention BiMHSA can be used.

[0017] The motion distribution decoder specifically includes a Transformer decoder and an upsampling operation connected in sequence. The input of the motion distribution decoder is processed by the Transformer decoder and outputs a six-degree-of-freedom Gaussian motion distribution of the tooth. Specifically, the six-degree-of-freedom motion of the tooth is first obtained, including the displacement and rotation of the tooth in three axes. Then, the six means and six variances of the six-degree-of-freedom motion are obtained, which are then used as the six-degree-of-freedom Gaussian motion distribution. Then, the next displacement and rotation of the tooth are selected by sampling.

[0018] In step three, during training, the AI ​​model uses a movement loss function to train the tooth point cloud perceptron and the tooth transformation predictor. Specifically, the movement loss function is composed of a weighted sum of translation and rotation losses, both of which employ negative log-likelihood loss functions. The input to the movement loss function is the mean and variance of each of the six degrees of freedom actions obtained from the action distribution decoder. Training is complete when the movement loss function converges. During training, the AI ​​model first establishes a training set, acquiring the initial poses, preset target poses, and tooth point clouds of several orthodontic teeth as input, and the orthodontic movement paths of the teeth as output, before training the AI ​​model.

[0019] In step three, the displacement and rotation of the teeth output in each step are processed using a heuristic method. Specifically, when the displacement of the teeth is less than a preset displacement threshold, the displacement is set to 0, and when the rotation of the teeth is less than their respective preset rotation thresholds, the rotation is set to 0.

[0020] This invention proposes an artificial intelligence-based orthodontic path planning method, specifically employing an autoregressive model capable of progressively predicting all intermediate deformations. To overcome the computational burden associated with processing large amounts of point cloud data while maintaining accuracy, a cascaded architecture is also designed, decoupling each step of tooth alignment into two stages: tooth state perception and tooth movement prediction.

[0021] In the tooth state perception stage: the state-of-the-art, efficient sequence modeling method Mamba is employed to model the entire point cloud sequence, enabling the perception of tooth alignment from the vast and complex dental point cloud. To overcome the inherent unidirectional and one-dimensional modeling characteristics of the Mamba model, a tooth position scanning operation and a hierarchical structure within Mamba are further proposed, making it more suitable for encoding dental point cloud data.

[0022] In the tooth movement prediction stage: state transitions are predicted autoregressively, conditioned on previous steps. To further mitigate the problem of movement error accumulation that typically increases with the number of permutation steps, a novel bidirectional Transformer is designed, equipped with a stochastic action decoder. This model can simultaneously predict permutation schemes from both the initial and target tooth configurations, progressively moving towards intermediate states.

[0023] The beneficial effects of this invention are:

[0024] Compared to traditional models that predict all intermediate steps based on initial and final tooth positions, the method of this invention comprehensively considers the morphology of each tooth and the overall arrangement configuration in each step to determine the subsequent tooth movement path. The method of this invention is simple in steps, convenient and fast to implement, and solves the problems of low accuracy and non-compliance of biomechanical movement rules in other artificial intelligence orthodontic methods. It significantly improves the accuracy of orthodontics, effectively reduces prediction errors, improves computational efficiency, and can quickly generate orthodontic paths that are more in line with clinical practice, making it easy to promote. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention;

[0026] Figure 2 This is the overall network structure diagram of the model of this invention;

[0027] Figure 3 This is a schematic diagram of the coordinate axes and point cloud sequence of the oral cavity model of the present invention.

[0028] Figure 4 This is a single-layer network structure diagram of the hierarchical sequence modeling module of the present invention, wherein, Figure 4 (a) is a single-layer structure diagram of the hierarchical sequence modeling module of the present invention. Figure 4 (b) is a structural diagram of the Mamba architecture SSM in the hierarchical sequence modeling module of the present invention;

[0029] Figure 5 This is a diagram of the self-attention mechanism, in which, Figure 5 (a) is a diagram of a one-way self-attention mechanism. Figure 5 (b) is a diagram of the bidirectional self-attention mechanism used in the tooth transformation predictor of the present invention;

[0030] Figure 6 These are planning result diagrams of several embodiments of the present invention, wherein, Figure 6 (a) is a first set of planning result diagrams of an embodiment of the present invention. Figure 6 (b) is a second set of planning result diagrams of an embodiment of the present invention;

[0031] Figure 7 These are planning result diagrams of several embodiments of the present invention, wherein, Figure 7 (a) is the third set of planning result diagrams of an embodiment of the present invention. Figure 7 (b) is the fourth set of planning result diagrams of the embodiment of the present invention.

[0032] Figure 8 This is a complete planning result diagram of the present invention. Detailed Implementation

[0033] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0034] like Figure 1 As shown, the artificial intelligence-based orthodontic path planning method of the present invention is as follows:

[0035] 1. Obtain the initial and target tooth poses, as well as a tooth mesh model, of the patient to be orthodontized. Specifically, this involves scanning the entire set of teeth to be orthodontized using intraoral scanning or cone-beam computed tomography (CBCT) to construct the tooth mesh model data. The initial tooth pose is then obtained based on this model data. Taking a patient with 28 teeth as an example, the initial centroid coordinates and local coordinates of all 28 teeth, as well as the target centroid coordinates and local coordinates of all 28 teeth, are obtained. Correspondingly, the final step sequence to be obtained is a fixed-length sequence of tooth poses.

[0036] 2. Extract the initial point cloud of teeth from the tooth mesh model data, then downsample the initial point cloud of each tooth using the farthest point sampling (FPS) method, and then normalize the initial point cloud of the entire set of teeth to reduce the number of points in each tooth point cloud model to 512, which facilitates subsequent calculations and finally obtains the point cloud of the entire set of teeth.

[0037] 3. For example Figure 2As shown, an artificial intelligence model based on bidirectional autoregressive prediction is constructed and trained to obtain a dental orthodontic path planning model. The artificial intelligence model includes a tooth point cloud perceptron and a tooth transformation predictor connected in sequence. The tooth point cloud perceptron includes a pre-trained point cloud feature extractor and a Mamba-based hierarchical sequence modeling module connected in sequence. The artificial intelligence model is pre-set with a preset number of movement steps L. The initial pose and preset target pose of the teeth input to the artificial intelligence model are processed through a simple fully connected network to extract the initial pose code and the target pose code. The entire tooth point cloud is processed by the pre-trained point cloud feature extractor to extract the point cloud feature encoding sequence. The initial pose code and point cloud feature code are then processed together. After the encoded sequences are concatenated to obtain the initial concatenation result, they are input into the Mamba-based hierarchical sequence modeling module. The pose encoding is at the tooth level, and the point cloud feature encoding sequence is at the point level. During concatenation, the pose encoding of each tooth is concatenated to the feature encoding of that point based on which tooth it belongs to. After the target pose encoding and point cloud feature encoding sequences are concatenated to obtain the target concatenation result, they are input into the Mamba-based hierarchical sequence modeling module. The inter-point information of the initial concatenation result and the target concatenation result are extracted respectively and then input into the tooth transformation predictor for bidirectional autoregressive prediction. The initial pose of the tooth is used as the 0th step of the movement path, and the preset target pose is used as the Lth step of the movement path. In the i-th prediction, the model receives the tooth poses from steps 0 to i-1 and from L-i+1 to L as input. The tooth transformation predictor outputs the tooth displacement and rotation amounts from steps i and Li, thus obtaining the tooth poses from steps i and Li. The poses are then returned to the tooth point cloud perceptron and the tooth transformation predictor for processing until the tooth displacement and rotation amounts from all steps are obtained as the initial tooth movement amounts.

[0038] The point cloud feature extractor specifically employs a PointNet deep learning model, a multi-layered PointNet++ deep learning model, or a Dilated Gate Convolutional Neural Network (DGCNN) model. The point cloud feature extractor has been pre-trained and does not require retraining.

[0039] The Mamba-based hierarchical sequence modeling module comprises two sequentially connected sequence modules. Each sequence module includes, in turn, a position encoding process, an interdental scan process, a point-scale state space model (SSM), and an aggregation layer. The aggregation layer of the first sequence module consists of a sequentially connected sequence feature aggregation layer and a point cloud feature aggregation layer. The aggregation layer of the second sequence module consists only of a sequence feature aggregation layer. The aggregation layer of the first sequence module includes the aggregation of four rearranged point cloud feature sequences (sequence feature aggregation layer) and the aggregation of all point features in the same tooth (point cloud feature aggregation layer). The aggregation layer of the second sequence module consists only of the aggregation of four rearranged point cloud feature sequences (sequence feature aggregation layer). For each point cloud feature input to the Mamba-based hierarchical sequence modeling module, it is first processed by the position encoding process of the first sequence module, and then processed in a top-to-bottom, bottom-to-top, left-to-right order. Interdental scanning is performed in a right-to-left scanning sequence to obtain four rearranged point cloud sequences. Each rearranged point cloud sequence is then processed by a point-scale state space model (SSM) to extract inter-point feature sequences. These four inter-point features are then aggregated through a first sequence feature aggregation layer. During aggregation, each rearranged point cloud sequence is first normalized and then passes through a fully connected layer and an activation function (SiLU) to obtain the weight of each point, resulting in four weight sequences. The four inter-point feature sequences and their corresponding weight sequences are multiplied element-wise to obtain four weighted feature sequences. Finally, the four weighted feature sequences are summed element-wise to obtain the sequence feature aggregation result. The first sequence module also passes through a point cloud feature aggregation layer, which averages the features of all points in the same tooth to obtain a tooth-level feature sequence. This tooth-level feature sequence is then processed in the second sequence module and output as a feature sequence that fuses inter-dental information.

[0040] The tooth transformation predictor is a Transformer-based bidirectional autoregressive predictor, consisting of sequentially connected layer normalization, several layers of bidirectional self-attention BiMHSA, and an action distribution decoder. The tooth transformation predictor predicts from the initial tooth pose to a preset target pose, and simultaneously predicts from the preset target pose to the initial pose. Each prediction outputs the tooth displacement and rotation at step i and step Li, including the mean and variance of the displacement and rotation along the X, Y, and Z axes, thus obtaining the tooth pose at steps i and Li. The tooth poses at steps i and Li are returned to the tooth point cloud perceptron for processing, and the feature sequences at steps i and Li are output. These feature sequences are then concatenated with the previously obtained feature sequences from steps 0 to i-1 and from L-i+1 to L, and then input back into the tooth transformation predictor for processing, outputting the tooth displacement and rotation at steps i+1 and Li-1, until the tooth displacement and rotation of all steps are obtained as the initial tooth movement.

[0041] For example, the initial pose of the teeth and the preset target pose are input to obtain the feature sequences of step 0 and step L. After processing, the tooth displacement and rotation of step 1 and step L-1 are output, thus obtaining the tooth pose of step 1 and step L-1. Then, the tooth point cloud perceptron is returned to process and output the feature sequences of step 1 and step L-1. The feature sequences of step 0, step 1, step L and step L-1 are then concatenated and input into the tooth transformation predictor for the next prediction. Intermediate steps can be left empty. Each time, the feature sequences are concatenated with all previous feature sequences and then returned to the tooth transformation predictor for processing until the tooth displacement and rotation of all steps are obtained. After L / 2 iterations, the complete orthodontic path can be obtained.

[0042] Bidirectional self-attention BiMHSA is specifically a multi-head attention mechanism MHSA that uses a bidirectional masking mechanism. The output and input of the first layer of bidirectional self-attention BiMHSA are residually concatenated and then input into the second layer of bidirectional self-attention BiMHSA for processing. The output of each subsequent layer of bidirectional self-attention BiMHSA is residually concatenated with the output of the previous layer of bidirectional self-attention BiMHSA before being input into the next layer of bidirectional self-attention BiMHSA for processing. Specifically, a four-layer bidirectional self-attention BiMHSA can be used.

[0043] During model training, for efficiency reasons, data from steps 0 to L are simultaneously input into the model for parallel training. The model needs to make predictions for any step from 0 to L at the same time, but the input for each prediction step is different. For example, the prediction for step 5 requires data from steps 0 to 4, and the prediction for step 7 requires data from steps 0 to 6. Without a mask, the model would see all the data from steps 0 to L when making predictions for any step, causing the inputs for prediction and training to become misaligned. Traditional one-way masks are also incompatible with the model structure of this patent, so a two-way mask mechanism is introduced. When predicting step 5, in addition to the data from steps 0 to 4, data from steps L-4 to L are also provided. Compared to traditional multi-head self-attention computation, BiMHSA uses a unique two-way mask mechanism that not only pays attention to all forward information from the initial position to the current step, but also to all backward information from the target position to the current step.

[0044] The action distribution decoder specifically comprises a Transformer decoder and a sampling operation connected in sequence. The input to the action distribution decoder is processed by the Transformer decoder, outputting a six-degree-of-freedom Gaussian action distribution of the tooth. Specifically, it first obtains the six-degree-of-freedom actions of the tooth, including displacement and rotation in three axes. Then, it obtains the six means and six variances of the six-degree-of-freedom actions, which are then used as the six-degree-of-freedom Gaussian action distribution. Finally, it selects the next displacement and rotation of the tooth through sampling. Different sampling methods can be used to select the next displacement and rotation of the tooth from the distribution output by the action distribution decoder, including random sampling, probability-based sampling, and mean-based sampling.

[0045] Compared to traditional Transformer decoders, this invention decodes not a fixed action, but a distribution of actions. Specifically, the tooth transformation used is a 6-DOF action, including the translation and rotation of the teeth in three axes. The output of the action distribution decoder is 6 mean and 6 variance, representing a Gaussian action distribution of 6 degrees of freedom. The tooth action needs to be sampled from these 6 distributions.

[0046] During training, the AI ​​model employs a movement loss function to train the tooth point cloud perceptron and tooth transformation predictor. Specifically, the movement loss function is composed of a weighted sum of translation and rotation losses, both using negative log-likelihood loss functions. The input to the movement loss function is the mean and variance of each of the six degrees of freedom actions obtained from the action distribution decoder. Training is complete when the movement loss function converges. During training, a training set is first established, using the initial poses, preset target poses, and tooth point clouds of several orthodontic teeth as inputs, and the orthodontic movement paths of the teeth as outputs, before training the AI ​​model.

[0047] The initial tooth pose, preset target pose, and whole-mouth tooth point cloud are input into the orthodontic path planning model for processing. After processing, the model outputs the initial tooth movement amount for each step. Heuristic methods are then used to post-process these initial tooth movement amounts. Specifically, for each step's output tooth displacement and rotation, the heuristic method sets the displacement to 0 if it is less than a preset displacement threshold, and sets the rotation to 0 if it is less than its respective preset rotation threshold. Finally, the tooth movement path is constructed, thus achieving orthodontic path planning. Both the initial and preset target tooth poses include the centroid coordinates in the whole-mouth tooth center coordinate system and the local coordinate system of each tooth from the tooth model data.

[0048] In practice, the simplified point cloud and input tooth pose are first processed by a tooth point cloud perceptron to obtain fine-grained tooth feature encodings. The tooth point cloud perceptron includes a feature extractor and a hierarchical sequence modeling module. The former extracts the feature encoding of each point in the point cloud, while the latter extracts inter-point information to obtain fine-grained tooth feature encodings. The tooth feature encodings for all tooth movement steps obtained by the tooth point cloud perceptron are input into a tooth transformation predictor to predict the next tooth movement (including the step after moving from the initial pose to the target pose, and the step before moving from the target pose to the initial pose). The tooth transformation predictor includes a Transformer network based on a bidirectional self-attention mechanism and an action distribution decoder. The former obtains inter-step information for all input tooth movement steps, while the latter predicts tooth movement.

[0049] Specifically, after simplifying the point cloud, the initial pose of the teeth, and the target pose input model, the following processing and calculations are required:

[0050] 4. Input the obtained simplified point cloud into the feature extractor to extract the feature code of each point in the tooth point cloud, forming the point cloud feature code sequence X. g The loss function used for training the feature extractor is the chamfer distance between the reconstructed point cloud and the original point cloud. Simultaneously, the initial and target poses of the teeth need to be input into a pose encoder to extract the initial pose code x of the teeth. i and target pose encoding x t The pose encoder used here is a simple fully connected network, not a pre-trained network, but one that is optimized together with subsequent network layers.

[0051] The obtained tooth point cloud feature encoding sequence X g Separately with the initial pose encoding x i and target pose encoding x t Before stitching, since the pose coding is at the tooth level, it needs to be expanded to the point-level pose coding X. i / X t Each point belonging to the same tooth has the same pose code, and then X... g and X respectively i X t By splicing, an initial feature sequence F of all points in the tooth is obtained. i and target feature sequence F t F i and F t Both are T×N×d vectors, where T is the number of teeth, N is the number of point clouds, and d is the dimension of the point features, i.e., the tooth point cloud feature encoding sequence X. gPoint-level pose coding X i / X t The sum of dimensions. Taking 28 teeth as an example, T=28, N=512, X... g dimensional d g =128, X i and X t dimensional d i =128, point feature F i / F t The dimension d=256, that is, F i =[X g X i ], F t =[X g X t ].

[0052] 5. Obtain the initial feature sequence F of all points. i and target feature sequence F t Predicting tooth transition movements in the input model requires predicting both a three-dimensional translation and a three-dimensional rotation, representing the tooth's displacement and rotation angle in each axis. Specifically, F i and F t First, a hierarchical sequence modeling module based on a Mamba network will be used to extract the relationships between points, capturing fine-grained tooth geometric features. First, based on F... i and F t For each point in the sequence, its positional code is calculated and fused into the point features. Specifically, for the point feature sequence F... i / F t For the j-th point feature f, calculate its position code p and fuse it into f. The calculation methods include, but are not limited to, Rotary Position Embedding and Attention with Linear Biases.

[0053] After adding position encoding, the point cloud feature sequence F is then processed according to the point cloud scanning order. i and F t Rearrange the scans, in the following order: Figure 3As shown. Specifically, taking 28 teeth as an example, firstly, according to the FDI (Fédération Dentaire Internationale) tooth position representation, the teeth are numbered 11-17, 21-27, 31-37, and 41-47 in order from the central incisors outwards. Then, the tooth point cloud feature sequences are rearranged in four different orders: 47->37->...->17->...->27, 47->17->46->16->...37->27, and their corresponding reverse orders. Next, the four rearranged point cloud feature sequences are input into a point-scale SSM to extract the relationships between points. Then, the feature sequences processed by the four SSMs are aggregated. The aggregation method is as follows: each rearranged point cloud feature sequence is first normalized and then passed through a fully connected layer and a SiLU (Sigmoid Linear Array). The Unit activation function obtains the weight of each point, resulting in four weight sequences. The four SSM-processed point cloud feature sequences are then multiplied element-wise with their corresponding weight sequences. Finally, the four weighted feature sequences are summed element-wise to obtain the sequence feature aggregation result. Finally, the point features of the same tooth are aggregated to obtain tooth-level features.

[0054] Specifically, the above-mentioned point cloud feature sequence operations based on Mamba, such as Figure 4 As shown, the point cloud feature sequence F needs to be... i / F t The sequences are rearranged into four distinct sequences S1, S2, S3, and S4, and then subjected to the following operations:

[0055]

[0056] in, represents an auxiliary branch used to adjust the output; SILU represents the Sigmoid linear unit activation function; Linear represents a fully connected layer; LN represents layer normalization; k=1, 2, 3, 4; S k and S k ´ represent the point cloud feature sequences before and after processing by the Mamba network, respectively, S k The original point feature sequence is rearranged into four point cloud feature sequences after being scanned in four different orders. Each feature sequence is first normalized, and then passed through a fully connected layer and an activation layer to obtain activation values. S k ´ represents embedded features; Merge represents aggregation operations; This indicates element-wise multiplication.

[0057] The Merge operation reorders the four consecutive inputs unidirectionally and integrates the directional features into a unified representation, as follows: Figure 4 As shown in (a). Figure 4 As shown in (b), the Mamba network architecture includes a fully connected layer, a deep convolutional layer (DWConv), an activation function (SILU), a selective SSM layer, and a normalization layer connected in sequence. The output of the normalization layer is added to the input of the Mamba network and the output is used as the output of the Mamba network.

[0058] After obtaining the tooth-level feature sequence, the above scanning operation is repeated again, and the interdental relationship is extracted and the four feature sequences are aggregated through a network with the same structure. Finally, the obtained tooth-level feature sequence is input into a Transformer-based tooth transformation predictor for subsequent tooth transformation prediction.

[0059] 6. Using the tooth point cloud and the tooth poses of all existing tooth movement steps (including initial position, target position, and all predicted intermediate steps) as described in steps 3-4 above, obtain the tooth-level feature sequences of all existing tooth movement steps, and input them into a Transformer-based tooth transformation predictor. Specifically, taking a total of 50 steps as an example, in the first prediction step, input the tooth-level feature sequences corresponding to the initial and target tooth poses into a Transformer-based bidirectional tooth transformation prediction network to obtain the tooth transformations for steps 1 and 49. Then, calculate the tooth poses for steps 1 and 49 based on the initial and target poses. Next, input the tooth-level feature sequences corresponding to steps 0-1 and 49-50 into the bidirectional tooth transformation prediction network to obtain the tooth transformations for steps 2 and 48. Repeat this process until a complete orthodontic path is obtained.

[0060] Specifically, the Transformer-based bidirectional tooth transformation prediction network includes a bidirectional self-attention block and an action distribution decoder. The role of the bidirectional self-attention block is to fuse information between each tooth movement step, allowing the model to combine information from all predicted tooth movement steps to make the next prediction. Figure 5 (a) and Figure 5 Figure (b) illustrates the difference between the bidirectional self-attention mechanism and the traditional self-attention mechanism. The bidirectional self-attention mechanism used in this invention employs a unique bidirectional mask, simultaneously considering both forward and backward tooth sequence predictions, which better suits the bidirectional prediction design of this invention. Taking a total of 50 steps as an example, when predicting tooth movement in step 6, the traditional self-attention mechanism uses a look-ahead mask, allowing the model to only see tooth information from steps 0-5. The bidirectional self-attention mechanism, however, uses a bidirectional mask, enabling the model to combine tooth information from steps 0-5 and 45-50 to predict steps 6 and 44.

[0061] The action distribution decoder generates the distribution of tooth movement actions. Compared to traditional Transformer decoders, this invention decodes not a fixed action, but a distribution of actions. Specifically, the tooth transformation used is a 6-DOF action, including translation and rotation of the teeth in three axes. The output of the action distribution decoder is 6 means and 6 variances, representing a Gaussian action distribution with 6 degrees of freedom. The tooth action is obtained by sampling from these 6 distributions, as follows:

[0062]

[0063] Where, μ i and μ L-i H represents the mean of the action distribution at step i and step Li, respectively; N This indicates the hidden state of the last transformer layer; and Let represent the variances of the action distributions at step i and step Li, respectively. and These are the preset upper and lower bounds of the variance. and Let represent the 6-DOF motion parameters of each tooth at step i and step Li, respectively. 'Sample' represents the sampling method from the motion distribution, including probability sampling and averaging.

[0064] After sampling the translation and rotation amounts of the teeth, the tooth movements are post-processed according to certain rules. Specifically, a minimum translation amount tmin and a minimum rotation amount rmin are first set. If the sampled translation amount t and rotation amount r are less than the set values, it is considered that the tooth has not been moved or rotated in this step.

[0065] After obtaining the two new tooth poses, a termination condition needs to be determined. Specifically, taking a set number of steps of 50 as an example, if the current number of steps is less than 25, the new tooth pose needs to be concatenated with the previously predicted poses and steps 4-6 above needs to be repeated. For example, if the current prediction is the 6th time, obtaining the tooth poses at steps 6 and 44, then concatenating the tooth poses from steps 0-6 and 44-50 and repeating steps 4-6, until the 25th time, all tooth poses from steps 0-50 are obtained. It should be noted that the 25th prediction will obtain two 25th-step poses; the average of the two poses can be taken as the final 25th-step tooth pose.

[0066] All networks in this invention are trained using collected datasets. Specifically, the geometric feature extraction network uses point cloud reconstruction loss. L cdOptimization was performed using a tooth transformation prediction network based on Mamba and Transformer, employing a negative log-likelihood loss function. L p The optimization is performed using the following calculation method:

[0067]

[0068] Among them, L p Represents the movement loss function; and These represent the hyperparameters of the equilibrium rotation and translation components, respectively; T represents the number of teeth, and L represents the number of steps; and Let represent the actual rotation and displacement vectors of the v-th tooth at step i, respectively; This indicates that the mean is The variance is The probability of sampling rotation from a Gaussian distribution.

[0069] This invention evaluates the results on a predefined test set. The evaluation of the generated orthodontic path includes the following metrics:

[0070] Point cloud average distance (ADD): Calculates the average L2 distance between the model's predicted pose and the ground truth for each tooth at each step, in mm.

[0071] Translation error ME trans : Calculate the average L2 distance between the model's predicted movement and the true value for each tooth at each step, in mm.

[0072] Rotational error ME rot : Calculate the average L2 distance between the model's predicted rotation amount and its true value for each tooth at each step, in degrees.

[0073] like Figure 6 of (a) Figure 6 of (b) Figure 7 (a) and Figure 7 As shown in (b), there are four examples with slightly different total number of steps, covering common clinical conditions such as dental arch crowding, missing teeth, overbite, and underbite. Each example includes four key steps of tooth pose at four angles. The test results are shown in Table 1.

[0074] Table 1

[0075]

[0076] like Figure 8 As shown, it demonstrates Figure 7The complete orthodontic path for the case shown in (a) is illustrated, where the "Predicted" column displays the orthodontic path predicted by the model, and the "True Value" column displays the orthodontic path designed by a professional dentist. From four perspectives—maxillary, mandibular, frontal, and lateral—the model's predictions closely match the path designed by the dentist, with an average point cloud distance (ADD) of 0.23 mm. trans It is 0.18mm, ME rot It is 1.29°.

[0077] The results show that the network structure of the present invention can learn doctors' planning patterns from a large-scale dataset containing various clinical orthodontic cases, and the trained model can generate orthodontic path planning results that are more in line with clinical practice more quickly.

[0078] It should be understood that the embodiments described in this invention are exemplary, and the specific parameters used in the embodiment description are only for the purpose of describing this invention and are not intended to limit this invention.

Claims

1. A method for orthodontic path planning based on artificial intelligence, characterized in that, include: Step 1: Scan the entire set of teeth to be orthodontized to construct dental model data, and obtain the initial pose of the teeth based on the dental model data; Step 2: Extract the initial point cloud of teeth from the tooth model data and perform preprocessing to obtain the point cloud of the entire set of teeth; Step 3: Construct an artificial intelligence model based on bidirectional autoregressive prediction and train it to obtain a dental orthodontic path planning model. Input the initial tooth pose, the preset target pose, and the point cloud of the whole set of teeth into the dental orthodontic path planning model for processing. After processing, the dental orthodontic path planning model outputs the initial tooth movement amount for each step, and combines heuristic methods to post-process the initial tooth movement amount to construct the final tooth movement path, thereby realizing dental orthodontic path planning. In step three, the artificial intelligence model includes a tooth point cloud perceptron and a tooth transformation predictor connected in sequence. The tooth point cloud perceptron includes a pre-trained point cloud feature extractor and a Mamba-based hierarchical sequence modeling module connected in sequence. The artificial intelligence model is pre-set with a preset number of movement steps L. The initial pose and preset target pose of the teeth input to the artificial intelligence model are processed by a fully connected network to extract the initial pose code and target pose code. The entire tooth point cloud is processed by the pre-trained point cloud feature extractor to extract the point cloud feature code sequence. The initial pose code and the point cloud feature code sequence are concatenated to obtain the initial concatenation result, which is then input into the Mamba-based hierarchical sequence modeling module. The target pose code and the point cloud feature code sequence are then concatenated. The target stitching result is then input into the Mamba-based hierarchical sequence modeling module. The point information between the initial stitching result and the target stitching result is extracted and then input into the tooth transformation predictor for bidirectional autoregressive prediction. The initial tooth pose is used as the 0th step of the movement path, and the preset target pose is used as the Lth step of the movement path. In the i-th prediction, the model receives the tooth poses from steps 0 to i-1 and from steps L-i+1 to L as input. The tooth transformation predictor outputs the tooth displacement and rotation amounts of steps i and Li, thereby obtaining the tooth poses of steps i and Li. These poses are then returned to the tooth point cloud perceptron and the tooth transformation predictor for processing until the tooth displacement and rotation amounts of all steps are obtained as the initial tooth movement amounts.

2. The artificial intelligence-based orthodontic path planning method according to claim 1, characterized in that: In step two, the preprocessing specifically involves first downsampling the initial point cloud of each tooth using the farthest point sampling (FPS) method, then normalizing the initial point cloud of the entire set of teeth, and finally obtaining the point cloud of the entire set of teeth.

3. The artificial intelligence-based orthodontic path planning method according to claim 1, characterized in that: The point cloud feature extractor specifically employs a PointNet deep learning model, a multi-level PointNet++ deep learning model, or a Dilated Gate Convolutional Neural Network (DGCNN) model.

4. The artificial intelligence-based orthodontic path planning method according to claim 1, characterized in that: The Mamba-based hierarchical sequence modeling module includes two sequentially connected sequence modules. Each sequence module includes a position encoding processing, interdental scanning processing, a point-scale state space model (SSM), and an aggregation layer, all connected sequentially. The aggregation layer of the first sequence module includes a sequence feature aggregation layer and a point cloud feature aggregation layer, all connected sequentially. The aggregation layer of the second sequence module only includes a sequence feature aggregation layer. For each point cloud feature input to the Mamba-based hierarchical sequence modeling module, it is first processed by the position encoding of the first sequence module, and then interdental scanning is performed in the scanning order of top to bottom, bottom to top, left to right, and right to left to obtain four rearranged point cloud sequences. Each rearranged point cloud sequence is then processed by the point-scale state space model (SSM) to extract inter-point feature sequences. Then, the four inter-point features are aggregated by the first sequence feature aggregation layer. During aggregation, each rearranged point cloud sequence is first normalized and then passes through a fully connected layer and an activation function (SiLU) to obtain the weight of each point, resulting in four weight sequences. The four inter-point feature sequences and their corresponding weight sequences are multiplied element-wise to obtain four weighted feature sequences. Finally, the four weighted feature sequences are added element-wise to obtain the sequence feature aggregation result. The first sequence module also goes through a point cloud feature aggregation layer, which takes the average of the features of all points of the same tooth to aggregate and obtain a tooth-level feature sequence. The tooth-level feature sequence is then input into the second sequence module for processing and outputs a feature sequence that integrates information between teeth.

5. The artificial intelligence-based orthodontic path planning method according to claim 1, characterized in that: The described tooth transformation predictor is a Transformer-based bidirectional autoregressive predictor, comprising layer normalization, several layers of bidirectional self-attention BiMHSA, and an action distribution decoder connected in sequence. The tooth transformation predictor predicts from the initial tooth pose to a preset target pose, and simultaneously predicts from the preset target pose to the initial pose. Each prediction outputs the tooth displacement and rotation at step i and step Li, thereby obtaining the tooth pose at steps i and Li. The tooth poses at steps i and Li are returned to the tooth point cloud perceptron for processing, and the feature sequences at steps i and Li are output. These feature sequences are then concatenated with the previously obtained feature sequences from steps 0 to i-1 and from L-i+1 to L, and then input back into the tooth transformation predictor for processing, outputting the tooth displacement and rotation at steps i+1 and Li-1, until the tooth displacement and rotation of all steps are obtained as the initial tooth movement.

6. The artificial intelligence-based orthodontic path planning method according to claim 5, characterized in that: The aforementioned bidirectional self-attention BiMHSA is specifically a multi-head attention mechanism MHSA that employs a bidirectional masking mechanism. The output and input of the first-layer bidirectional self-attention BiMHSA are residually concatenated and then input into the second-layer bidirectional self-attention BiMHSA for processing. The output of each subsequent layer of bidirectional self-attention BiMHSA is residually concatenated with the output of the previous layer of bidirectional self-attention BiMHSA before being input into the next layer of bidirectional self-attention BiMHSA for processing.

7. The artificial intelligence-based orthodontic path planning method according to claim 5, characterized in that: The motion distribution decoder specifically includes a Transformer decoder and an upsampling operation connected in sequence. The input of the motion distribution decoder is processed by the Transformer decoder and outputs a six-degree-of-freedom Gaussian motion distribution of the tooth. Specifically, the six-degree-of-freedom motion of the tooth is first obtained, including the displacement and rotation of the tooth in three axes. Then, the six means and six variances of the six-degree-of-freedom motion are obtained, which are then used as the six-degree-of-freedom Gaussian motion distribution. Then, the next displacement and rotation of the tooth are selected by sampling.

8. The artificial intelligence-based orthodontic path planning method according to claim 7, characterized in that: In step three, during training, the artificial intelligence model uses a moving loss function to train the tooth point cloud perceptron and the tooth transformation predictor. The moving loss function is specifically composed of the weighted translation loss and rotation loss. Both the translation loss and rotation loss use the negative log-likelihood loss function. The input to the moving loss function is each mean and each variance of the six degrees of freedom action obtained by the action distribution decoder. Training is completed when the moving loss function converges.

9. The artificial intelligence-based orthodontic path planning method according to claim 1, characterized in that: In step three, the displacement and rotation of the teeth output in each step are processed using a heuristic method. Specifically, when the displacement of the teeth is less than a preset displacement threshold, the displacement is set to 0, and when the rotation of the teeth is less than their respective preset rotation thresholds, the rotation is set to 0.

Images