Intelligent tooth arrangement method and system based on data enhancement and Transform

Abstract

The invention belongs to the technical field of tooth correction, and particularly relates to an intelligent tooth arrangement method, device and system, and the method comprises the steps: carrying out automatic data enhancement processing on original intraoral scanning data, and obtaining standardized tooth point cloud data; the standardized tooth point cloud data is input to an intelligent tooth arrangement model, and the intelligent tooth arrangement model is a neural network constructed based on a Transform architecture and is used for modeling a dependency relationship between local geometric features and global space of teeth and outputting a prediction result of target tooth arrangement; the intelligent tooth arrangement model is optimized by adopting a joint supervision loss function, the joint supervision loss function comprehensively constrains error distribution, space structure consistency and prediction uncertainty of a prediction result, and the method adopts an STU automatic data enhancement strategy to simulate various position changes of teeth before orthodontic treatment; layered tooth information can be effectively extracted, and modeling of a complex spatial relationship can be enhanced.

Description

Technical Field

[0001] This invention belongs to the field of dental correction technology, specifically relating to an intelligent tooth alignment method and system based on data augmentation and Transformer. Background Technology

[0002] Malocclusion refers to misaligned teeth, abnormal relationships between the upper and lower dental arches, or abnormalities in the size, shape, or position of the jawbone. A significant portion of the global population is affected by malocclusion, with prevalence varying considerably across different regions and ethnicities, ranging from 39% to 93%. Orthodontic treatment is the primary method for correcting malocclusion. By applying multidirectional forces using various orthodontic appliances, teeth can be adjusted to their predetermined positions, thereby achieving an aesthetically pleasing, balanced, and stable dental arch shape and occlusal relationship.

[0003] However, current deep learning methods are often limited by the insufficient scale and diversity of clinical data, which presents a series of challenges. High-quality oral datasets require precise annotation by professional clinicians, which is costly and subject to strict privacy regulations, resulting in slow sample accumulation. Teeth are small and morphologically diverse, and data scarcity makes it difficult for models to capture fine-grained features at different levels, such as crowns and roots. The three-dimensional spatial relationships of teeth in malocclusion are complex, and limited data is insufficient to build accurate spatial relationship models. These challenges directly lead to weak model generalization ability, making them prone to diagnostic errors when faced with highly variable and complex malocclusions, thus failing to meet clinical needs.

[0004] Furthermore, tooth overlap is a significant technical bottleneck in digital dental diagnosis. This phenomenon is particularly common in malocclusion dental models. Physical interference between teeth leads to geometric distortion in 3D reconstruction models and directly affects the accuracy of subsequent tooth segmentation, position prediction, and orthodontic treatment planning—for example, in automatic tooth alignment algorithms, overlapping teeth prevent the model from accurately determining natural spatial relationships, resulting in alignments that do not conform to physiological structures.

[0005] With the advancement of artificial intelligence, machine learning-based tooth alignment methods are rapidly emerging, gradually overcoming the limitations of traditional techniques. These methods aim to achieve fully automated and more precise intelligent tooth alignment, providing a more efficient and reliable solution for orthodontic treatment. TPPMAGO designed a geometrically adaptive optimization and physiologically adaptive reconstruction strategy, integrating the physiological features and anatomical structure of teeth to adapt to the complexity of different clinical environments in real time and enhance pose invariance. TAlignDiff introduced two key components—a pointwise regression network (PRN) and a diffusion transform matrix denoising module (DTMD)—to further improve model performance. Li et al. proposed a deep learning-based automated orthodontic treatment planning method that achieves high-precision automatic alignment by fusing local and global spatial features of teeth to predict 3D transformations. iOrthoPredictor utilizes the 3D geometric latent space encoded by the unsupervised generative model StyleGAN to generate alignment processes in the image domain by finding semantic paths along the normal vector direction. Chen et al. detected markers on each tooth to identify key locations, constructing feature map structures for the jawbone and teeth. Li et al. proposed TEANet, which uses a dual-coordinate classifier to make orthodontic extraction decisions and utilizes an adaptive graph structure for feature propagation to handle crowded dental cases requiring extraction.

[0006] Despite these advances in tooth position prediction, most methods still have limitations. First, existing models are highly sensitive to dataset size and often rely on local features (such as point clouds or meshes), making it difficult to accurately represent the complex spatial topological relationships between teeth. Second, existing methods often fail to explicitly model the inherent physiological connections, alignment constraints, and pose consistency of teeth, resulting in significant room for improvement in generalization ability. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent tooth alignment method, device, and system based on data augmentation and Transformer to address the problems existing in the prior art. It adopts the STU automatic data augmentation strategy to simulate the diverse positional changes of teeth before orthodontic treatment, and at the same time constructs a two-level feature tooth position prediction network, which can effectively extract layered tooth information and enhance the modeling of complex spatial relationships.

[0008] The technical solution of this invention is: On the one hand, the present invention provides an intelligent tooth alignment method based on data augmentation and Transformer, comprising the following steps: The raw intraoral scan data is preprocessed using the STU automated data augmentation strategy (segmentation-transformation-uniform sampling) to enhance data diversity and unify the representation format; A two-level Transformer intelligent tooth alignment model is constructed, which predicts the target tooth alignment from a standardized point cloud by encoding local features and modeling global relationships. We design a joint supervised loss function to jointly optimize the model from multiple perspectives, including location error, geometric consistency, and prediction uncertainty, thereby improving the accuracy and rationality of the alignment results.

[0009] Preferably, the STU data enhancement strategy specifically includes: Segmentation: Based on geometric features, teeth and alveolar bone are separated from intraoral scan data to extract independent tooth models; Transformation: A random rigid transformation is applied to each tooth to simulate the diverse initial tooth postures in clinical settings; Uniform sampling: Sampling the tooth point cloud to obtain a standardized input with a uniform number of points.

[0010] Preferably, the two-level Transformer intelligent tooth alignment model includes: The local feature encoder uses the Transformer self-attention mechanism to extract the geometric features of a single tooth; A global relation encoder models spatial constraints between teeth in the entire mouth through a cross-attention mechanism; The feature fusion and prediction module, combined with residual connections, outputs the target position or transformation parameters for each tooth.

[0011] Preferably, the joint supervision loss function includes: Location error loss, such as Chamfer distance or point-to-point error; Geometric consistency loss is used to penalize inter-tooth collisions and non-physiological alignments; Uncertainty perception loss is used to adjust sample weights based on prediction variance, thereby enhancing model robustness.

[0012] The present invention also provides an intelligent tooth alignment device, including a data augmentation unit, an alignment model unit, and a training optimization unit, for performing the above-described method.

[0013] In addition, the present invention provides an intelligent tooth alignment system, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the program to implement the above-mentioned method.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0015] The intelligent tooth alignment method based on data augmentation and Transformer proposed in this invention provides a systematic solution to the shortcomings of existing methods described in the background section, and brings significant and verifiable performance improvements. Specific beneficial effects are compared below: 1. This invention introduces a structured STU automated data augmentation strategy, resulting in the following specific effects: Expanded data distribution: Through the "segmentation-transformation" step, a large number of training samples simulating the initial state of different malocclusions can be automatically generated, significantly expanding the pose distribution range of the training data and reducing the model's dependence on the original data size and perfect annotation; Improved standardization and robustness: Through the "uniform sampling" step, the format and size of the input data are unified, which not only reduces computational overhead but also allows the model to focus on learning general geometric and spatial relationships rather than specific point cloud densities or noise patterns, thereby improving the model's adaptability and robustness to scan data from different sources; In comparative experiments in data-scarce scenarios, compared with the baseline model without the STU strategy, the alignment error of the method of this invention is reduced by an average of about 25% on unseen complex cases, demonstrating its excellent generalization performance.

[0016] 2. This invention constructs a two-level Transformer intelligent tooth alignment model, resulting in the following specific effects: Fine-grained local geometry capture: The local feature encoder (first-level Transformer) utilizes a self-attention mechanism to effectively understand the complex three-dimensional morphology of each tooth, providing accurate initial geometric features for subsequent arrangement.

[0017] Global spatial relationship modeling: The global relationship encoder (second-level Transformer) explicitly calculates and learns the mutual influence between all tooth pairs through a cross-attention mechanism, thereby actively modeling global information such as arch curves, adjacency relationships, and occlusal constraints. The method of this invention reduces the collision volume between teeth by more than 30% on average compared to existing mainstream CNN methods, and the arranged dental arch morphology is more in line with physiological laws, resulting in higher scores from clinicians for visual rationality.

[0018] 3. This invention designs a joint supervised loss function, the specific effects of which are: Multi-dimensional constraint optimization: This loss function integrates supervision signals from three dimensions: positional error, geometric consistency, and prediction uncertainty. This guides the model to meet the physical and clinical constraints of tooth alignment while pursuing accurate localization.

[0019] Active collision and misalignment suppression: The geometric consistency loss term can explicitly penalize tooth penetration and alignment that violates the natural dental arch trend in the prediction, thus proactively reducing the risk of irrational results from the optimization objective level.

[0020] Adaptive learning and reliability enhancement: Uncertainty regularization loss enables the model to identify and carefully handle unpredictable samples (such as severely crowded cases), improving the overall reliability and safety of the model's predictions. In comprehensive clinical rationale assessments, the proportion of outputs from models trained using the joint loss function that were judged by experts to be "directly usable for clinical protocol design" was more than 40% higher than that of models using only location error loss.

[0021] In summary, this invention systematically solves the key bottlenecks in existing automatic tooth alignment technologies through a three-pronged innovation: data augmentation (STU strategy), model architecture (two-level Transformer), and optimization objective (joint loss function). This invention not only achieves advanced levels in standard error metrics but also makes significant breakthroughs in key dimensions such as generalization ability, geometric rationality, and clinical usability. It provides strong technical support for the automation, intelligence, and precision of digital orthodontic treatment, possessing broad clinical application prospects and market value. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of the intelligent tooth alignment method provided by the present invention; Figure 2 This is a schematic diagram of the intelligent tooth alignment model in this invention; Figure 3 This is a network structure diagram of the SWTB module of the intelligent tooth alignment model of the present invention; Figure 4 This is a network structure diagram of the PT module of the intelligent tooth alignment model of the present invention; Figure 5 Schematic diagram of the initial state of inter-tooth collision; Figure 6 This is a diagram showing the result after the tooth collision problem was resolved; Figure 7 These are the cumulative distribution curves of the average point-by-point distance for different methods; Figure 8 This is a comparison of tooth position prediction results from different methods. Detailed Implementation

[0023] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0024] Example 1: Intelligent Tooth Alignment Method Flowchart like Figure 1 As shown, the intelligent tooth alignment method provided in this embodiment includes the following steps: Step S101: Data Augmentation Preprocessing Obtain raw tooth point cloud or mesh data output from an intraoral scanner.

[0025] First of all, let This represents the original dental scan obtained from an STL mesh or point cloud. Since the original scan contains both teeth and gums, it is necessary to extract the individual geometry of each tooth. We employ a 3D instance segmentation method based on geometric feature analysis to independently extract each tooth from the input raw data.

[0026] Xori: Raw dental scan data (in STL mesh format) containing complete intraoral scan geometry information of teeth and gums.

[0027] Xend: Geometric data of a single tooth obtained after segmentation (initial data of the tooth in the target state). (However, no change in posture was made).

[0028] Next, a random rigid transformation is applied to each individual tooth, including translation within the range of [-3mm, 3mm] and rotation within the range of [-15°, 15°].

[0029] To fully simulate the diverse tooth postures that may occur in the early stages of orthodontic assessment, we applied a rigid transformation operator to each segmented tooth. This operator generates multiple augmented samples through controlled rotations and translations, thereby expanding the distribution range of tooth poses and exposing the model to more possible initial alignment scenarios.

[0030] Xstart: Geometric data of a single tooth after posture transformation (simulating the initial tooth posture at the beginning of orthodontic assessment).

[0031] Finally, the farthest point sampling algorithm was used to uniformly sample the point cloud of each tooth to 1024 points to form a standardized input.

[0032] Downsampling of point cloud data: Using the uniform sampling operator Uts, the point cloud data of each tooth is uniformly sampled to a fixed number of points. This step is consistently applied to the tooth model in both the initial and target states, aiming to reduce computational and training costs in subsequent stages.

[0033] : Point cloud data of a single tooth after uniform sampling and pose transformation (normalized initial state tooth point cloud).

[0034] : Point cloud data of a single tooth after uniform sampling and segmentation (normalized target state tooth point cloud).

[0035] Step S102: Two-level Transformer alignment prediction The processed tooth point cloud is input into... Figure 2 The two-level Transformer network shown.

[0036] The local encoder extracts features from each tooth separately and outputs a 128-dimensional feature vector for each tooth.

[0037] The global encoder receives all tooth features, calculates inter-tooth correlations through cross-attention, and outputs enhanced features with spatial context information.

[0038] The feature fusion module merges local and global features through residual connections and regresses the 6-DOF transformation parameters (3 translations and 3 rotations) of each tooth through a fully connected layer.

[0039] The specific principle of this model is as follows: it takes a normalized point cloud as input and predicts the clinically desired tooth alignment by learning the spatial mapping relationship of teeth. .

[0040] Xpre: Clinically expected tooth alignment geometry data (point cloud format) output by the model prediction.

[0041] The proposed two-level Transformer aligned prediction architecture is as follows: Figure 2 As shown. Given an input set of tooth points. The network first embeds the data into two parallel feature spaces through linear projection: The upper branch uses sinusoidal position coding (SPE) to generate The encoding features of the upper branch are refined through four spatial direction Transformer blocks (SWTB) to generate spatial descriptors. The lower branch uses a four-layer Point Transformer to generate geometry descriptors. The outputs of the two branches are concatenated into a unified hierarchical feature: F1: The unified hierarchical feature obtained after splicing (which integrates the spatial and geometric features of the teeth, and is a multi-dimensional fusion feature).

[0042] In order to model the relationship between teeth Obtained through another four-layer Point Transformer: , Among them, PT×4(·): another set of 4 consecutively stacked Point Transformer operators. Unlike the PT module in the lower branch, this module is specifically used to model the spatial association and interaction between different teeth (such as the relative position of adjacent teeth, occlusal relationship, etc.). F2: The feature vector output after modeling the relationships between teeth (capturing global correlation information between teeth). A residual connection merges the original features. (recorded as) The output of the relation module is passed to the relation module, and the two are combined to form the final relation feature: , FR: Reference features for residual connections, i.e., the original fusion features F1 without modeling the inter-tooth relationships. Residual connections are introduced to preserve the initial hierarchical features and avoid the gradient vanishing problem during deep network training. F3: Final relational features (containing both global relational features between teeth and original spatial-geometric fusion features) Then, fusion representation The data is fed into a multi-head regression module to predict the 6-DOF pose of each tooth: , Wherein, FR: Reference feature of residual connection, i.e., the original fusion feature before inter-tooth relationship modeling. F1 introduces residual connections to preserve the initial hierarchical features and avoid the gradient vanishing problem during deep network training; F3: final relational features (which simultaneously include global relational features between teeth and the original spatial-geometric fusion features).

[0043] Step S103: Joint Loss Supervised Training Total loss Defined as a weighted combination of multiple sub-losses, it jointly constrains the accuracy of tooth position prediction, the spatial relationships between teeth, and related uncertainties. For example... Figure 5 As shown, collisions between teeth severely interfere with the accurate localization of relative positions, thus reducing prediction accuracy. To address this issue, the proposed loss function effectively prevents tooth collisions and misalignments, such as... Figure 6 As shown, this ensures that the final tooth alignment is anatomically sound and globally consistent. The overall loss function is defined as: here, These are the balance coefficients for each loss term. L: The total loss function of the model, a weighted combination of multiple sub-losses, used to constrain the model training process and ensure the accuracy, rationality, and global consistency of tooth alignment prediction; λ1, λ2, ..., λ6: Balance coefficients (learnable or manually set hyperparameters) corresponding to the 6 sub-loss terms, used to adjust the contribution weight of each sub-loss in the total loss and balance the priorities of different training objectives; Ld: Interaction loss, its core function is to penalize penetration collisions between teeth and encourage teeth to maintain anatomically reasonable spacing; Lcore: Core loss term (presumably the basic loss for tooth position / morphology prediction, such as point cloud matching loss); LC: Collision-related auxiliary loss (or constraint loss, supplementing and strengthening the anti-collision effect); Lbuild: Tooth alignment structure loss (presumably used to constrain the overall structural consistency of tooth alignment, conforming to oral anatomy); Ldof: 6-DOF pose loss (for teeth, 6DoF). Error loss in pose prediction, such as regression loss for translation and rotation parameters); Lre: regularization loss (or residual loss, used to prevent model overfitting and improve generalization ability); +: weighted summation of loss values, integrating multi-dimensional constraint objectives.

[0044] Interactive loss penetrates between teeth and encourages anatomically efficient spacing by employing the Lennard-Jones potential function: Where: RMSE(⋅): Root Mean Square Error operator, used to calculate the error between the predicted value and the true value, measuring the degree of loss quantification; Pt: True tooth point set, i.e., the three-dimensional point cloud coordinates of the teeth in their anatomically reasonable positions; P^t: Predicted tooth point set (the original formula is written as Pt, which is presumably a typo and should be the predicted value); E: Tooth pair set, containing all tooth pairs that may come into contact or collide (such as adjacent teeth, maxillary and mandibular occlusal teeth); Pe: Euclidean error, referring to the three-dimensional Euclidean distance between the predicted tooth point set P^t and the true tooth point set Pt, representing the position prediction deviation of a single tooth; Pc: Distance threshold (hyperparameter), which is the critical distance for determining whether a tooth collision has occurred, usually set based on oral anatomy.

[0045] Step S104: Model Inference and Result Output During the inference phase, the patient's intraoral scan data is processed through step S101 and then input into the trained model. The target transformation parameters for each tooth are output and visualized or transmitted to the orthodontic design system.

[0046] In this embodiment, the Spatial Orientation Transformer Block (SWTB) in the aforementioned two-level Transformer alignment prediction is a key component of the two-level feature learning architecture, specifically designed to capture complex spatial dependencies within the tooth point cloud, such as... Figure 3 As shown, let the input be... The input is first divided into blocks, and then passed through a weighted matrix. Linear projection is performed on the fully connected (FC) layer, followed by... Space Window Block (SWB) stack processing: .

[0047] Where X: the input feature tensor of SWTB, corresponding to the pre-processed tooth point cloud features; ℝ^(B×T×N×C): Definition of the dimension of a 4-dimensional tensor in the real space: B: Batch Size, which is the number of dental point cloud samples input into the model in a single batch (e.g., processing the dental point clouds of 8 / 16 patients in a batch). T: Teeth number (the number of teeth in the tensor), the number of teeth contained in the tensor (e.g., 28 / 32 teeth in the whole mouth). N: Number of points / sequence length (number of points per tooth / length of feature sequence); C: Channel dimension (number of feature channels), the feature dimension of each point / sequence element (e.g., 128 / 256 dimensions); PatchPartition(·): The block segmentation (patch partitioning) operator, whose core function is to divide the point cloud / feature sequence of a single tooth into multiple local spatial windows (patch) according to spatial location, focusing on local spatial dependencies; Wfc: The weight matrix of the fully connected (FC) layer, with dimensions ℝ^(C×C), where the number of input / output channels is C, ensuring that the feature dimension remains unchanged; Wfc⋅: indicates linear projection (matrix multiplication) of the features output by PatchPartition onto a fully connected layer. SWB: Spatial Window Block, the core submodule of SWTB, is used to capture local spatial dependencies within a single spatial window; SWB×K0(·): K0 consecutively stacked SWB combination operators, where K0=2 indicates that 2 SWBs are stacked in the initial stage; X0: The intermediate feature tensor output after block partitioning, linear projection, and two SWB processing.

[0048] Subsequently, in three stratified stages ( In each of the three feature refinement stages (representing SWTB), the spatial sequence length is halved through the PatchMerging operation (the core function of which is to merge adjacent local windows (patches) along the spatial dimension, halving the feature sequence length N of a single tooth while maintaining the number of feature channels C unchanged through linear transformation). This operation merges adjacent elements along the spatial dimension. The elements are then subjected to a linear transformation. The resulting features are then... SWB refinement: ,in ( K1=2: 2 SWBs are stacked in stage 1; K2=6: 6 SWBs are stacked in stage 2 (core refinement stage, capturing more complex spatial dependencies); K3=2: 2 SWBs are stacked in stage 3), SWB × (·): Continuously stacked in stage l A SWB combination operator is used to refine the merged features; X1: The feature tensor output after stage l processing; X2: The final output feature tensor of SWTB; ℝ^(B×T×(N / 8)×C): The dimension of the final output tensor: N / 8: After 3 PatchMerging operations (each time the sequence length is halved: N→N / 2→N / 4→N / 8), the feature sequence length of a single tooth is compressed to 1 / 8 of the initial length; B / T / C: The batch size, number of teeth, and number of feature channels remain consistent with the input. Consistent with the architecture design. The final output of SWTB is .

[0049] In this embodiment, the Point Transformer (PT) layer in the two-level Transformer alignment prediction learns basic building blocks of local geometric features from the point cloud. For example... Figure 4 As shown, each PT layer undergoes a series of multi-branch feature fusion, neighborhood aggregation, and residual refinement operations.

[0050] Input features First, the process is handled by a multi-branch transform module, which contains two linear projections and one MLP branch. The outputs of these three branches are then fused element-wise to produce branch-enhanced features: in and The learnable weights represent the linear branches. This represents a nonlinear transformation implemented by the MLP branch.

[0051] Then, the branched fusion features obtained by aggregating across local neighborhoods are used using an attention-based aggregation operator: , Among them, Aggregation(·) is a neighborhood aggregation operator based on an attention mechanism. Its core function is to exchange feature information within the local neighborhood (such as the set of k nearest neighbors) of the i-th point, capturing the local geometric context of the tooth point cloud (such as the curvature of the tooth surface and the relative positions of adjacent points); fᵢʰʳᵃnᶜʰ is the branch enhancement feature of the i-th point (the input of the aggregation operator); fᵢᵃᵍᵍ is the aggregated feature of the i-th point after neighborhood aggregation (it incorporates the geometric information of neighboring points and is no longer a single isolated feature). This enables effective information exchange between adjacent points while preserving the local geometric context.

[0052] Then, residual connections are applied to combine the aggregated features with the original input features, followed by layer normalization to stabilize the training. The normalized features are further refined by an MLP submodule, which consists of two linear layers with GELU activation in between: in and The learnable weights represent the linear branches. This represents the Gaussian error linear unit activation function.

[0053] Finally, a second residual connection and layer normalization are used to generate the output features of the Point Transformer block: LayerNorm(·): Second layer normalization to ensure stable distribution of output features; fᵢᵒᵘᵗ: The final output feature of the i-th point, which is also the output result of the entire Point Transformer layer for the i-th point.

[0054] In this embodiment, the effectiveness of the loss function is verified through ablation results, as shown in Table 1 below.

[0055] Table 1 As shown in the table above, both error metrics exhibit a consistent downward trend as additional loss components are gradually added, indicating that each proposed loss term contributes positively to improving the accuracy of rotation and translation estimations. These findings further demonstrate the rationality and effectiveness of the multi-task loss design.

[0056] Example 2 This embodiment uses a publicly available high-quality 3D dental arch dataset as the experimental basis to conduct experiments to comprehensively evaluate the accuracy and robustness of the method proposed in Embodiment 1.

[0057] The dataset used in this embodiment is derived from clinical oral scans, covering a variety of common and complex cases. The samples exhibit significant structural and individual differences, providing sufficient clinical diversity for evaluating the robustness and generalization performance of the automated intelligent tooth alignment algorithm. Furthermore, the 3D models in the dataset were acquired using professional scanning equipment, possessing high geometric accuracy and surface detail quality, meeting the research requirements of geometry-based deep learning methods. During the data preprocessing stage, we performed systematic manual data cleaning on the original dental model scans, including: removing samples with significant noise, large-area scanning defects, or severe reconstruction errors; unifying the mesh resolution and coordinate system; and checking and repairing common geometric problems such as component separation and inverted facets. Simultaneously, scans with incomplete annotations, severely missing teeth, or unidentifiable individual tooth numbers were excluded to ensure data consistency and label reliability during model training. Through the above rigorous screening and quality control procedures, a total of 288 structurally complete and qualified dental model samples were obtained. Based on this, data augmentation was performed using the STU operation, expanding the total number of samples to 864. Finally, the dataset was randomly divided into training and testing sets in an 8:2 ratio, with 692 samples used for model training and validation, and the remaining 172 samples reserved for performance evaluation.

[0058] All experiments in this embodiment were conducted in a Python environment using the PyTorch deep learning framework (version 2.7.0 + cu128). The experiments were run on a Windows 11 system with hardware including an NVIDIA RTX 5070 Ti GPU and an AMD R7 9800X3D processor. The Adam optimizer was used during training, with the number of point samples set to N=512 and an initial learning rate of... The batch size is 16. The network is trained for 500 epochs, and the model is saved every 50 epochs.

[0059] This embodiment uses four quantitative indicators: the average distance of displacement points (ADD), the area under the ADD curve (ADD / AUC), and the rotational error (…). Translation error ).

[0060] The ADD index measures the average Euclidean distance between the predicted tooth posture and the corresponding points of the actual tooth posture, and is defined as: , in This represents the 3D point set representing teeth. and These represent the actual and predicted rotations and translations, respectively. A lower ADD value indicates a more accurate attitude prediction.

[0061] We further quantify the robustness of the model using ADD / AUC, which is defined as the robustness from 0 to a preset threshold. The integral of the ADD success rate. Specifically, ADD / AUC is defined as: in It is an indicator function. A higher ADD / AUC value reflects better overall performance and stability.

[0062] Rotational error Evaluation of predicted rotation With real rotation The angular difference between them is calculated using the following formula: This indicator directly reflects the accuracy of the direction estimation; the smaller the value, the better the alignment.

[0063] Translation error The Euclidean distance between the predicted translation vector and the true translation vector is defined as: , The lower the value, the more accurate the translation prediction.

[0064] Example 3 In order to further verify the effectiveness of the method provided by the present invention, this embodiment compares the method provided in Embodiment 1 with the SOTA method.

[0065] Table 2 below shows a quantitative comparison of the proposed DATNet with several state-of-the-art methods on four evaluation metrics.

[0066] Table 2 As shown in Table 1, DATNet achieves the lowest ADD error (0.7394), indicating a significant improvement in alignment accuracy compared to all competing methods. Simultaneously, our method achieves the highest ADD / AUC score (0.9132), demonstrating superior robustness across different alignment thresholds. In terms of geometric consistency, DATNet also achieves the lowest rotation error. (2.3119) and the lowest translation error (1.2478), significantly outperforming existing baseline methods. These results collectively validate the effectiveness of the proposed framework in capturing the spatial dependence of dental structural hierarchies and highlight its strong potential for clinically reliable tooth alignment.

[0067] In addition, such as Figure 7The distribution of the average point-by-point distance between the predicted tooth mesh and the corresponding ground truth vertices is shown. The horizontal axis represents the threshold of the average point-by-point distance, and the vertical axis represents the proportion of samples with prediction errors below that threshold. As shown in the figure, compared with all baseline methods, the curve of the proposed method rises faster and achieves a higher level of accuracy at a lower distance threshold. This indicates that the tooth arrangement generated by DATNet more closely resembles the anatomical geometry of the real data, reflecting its superior accuracy and consistency in capturing fine-grained spatial structure.

[0068] Figure 8 This paper presents a qualitative comparison of the proposed DATNet with several representative baseline methods (including PSTN, TANet, TADPM, TPPMAGO, and SWTBS) on multiple dental arch samples. For each method, areas with significant errors are marked with red borders. As shown in the figure, baseline methods often encounter difficulties in handling complex tooth arrangements, especially in areas of close occlusion, irregular spacing, or blurred geometric cues. For example, PSTN and TANet frequently exhibit local misalignment or incorrect offsets in the molar region, while TADPM, although an improvement, still shows instability in predicting the precise centroid position of teeth in challenging areas. In contrast, DATNet (column 7) consistently produces predictions that closely match reality, demonstrating superior structural consistency across various cases. The red-bordered areas indicate that DATNet not only achieves more accurate individual tooth localization but also better preserves the overall dental arch morphology. This improvement is attributed to the synergistic design of the SPE, SWTB, PT, and MH modules, enabling DATNet to effectively capture long-range dependencies and fine local geometric patterns.

[0069] Example 4: This embodiment provides an intelligent tooth alignment system, including a processor, a memory, and an input / output interface. The memory stores a computer program that, when executed, implements the method described in Embodiment 1. The system can be integrated into a dental diagnostic workstation or a cloud service platform, supporting real-time or batch intelligent tooth alignment analysis. It achieves high-precision and highly rational automatic tooth alignment, possessing significant clinical application value.

[0070] Example 5 This embodiment provides an intelligent tooth alignment device, including: The data enhancement unit is configured to perform the data enhancement processing described in the method provided in Embodiment 1 above; The alignment model unit includes the aforementioned two-level Transformer intelligent tooth alignment model, which is used to receive standardized tooth point cloud data and predict the target alignment. The training optimization unit is configured to train the alignment model unit using the joint supervised loss function.

[0071] Example 6 In this embodiment, to investigate the contribution of each component in the proposed DATNet, a comprehensive ablation study was conducted by progressively enabling the SPE, SWTB, PT, and MH modules. The quantitative results are summarized in Table 3.

[0072] Table 3 As shown in Table 2, removing any critical component leads to a significant performance degradation, while integrating all four modules yields the best overall accuracy. Specifically, the baseline setting without SPE and PT but with SWTB and MH produces an ADD of 0.8364 and an ADD / ACU of 0.8716. Introducing SPE or PT alone improves geometric estimation accuracy, but the gain remains limited due to the lack of complementary information modeling. The complete model equipped with all four modules achieves the best results across all metrics, including the lowest rotation and translation errors. ), These findings validate that each module makes a unique contribution to DATNet, and their combined effect leads to the most robust and accurate pose estimation.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A smart tooth alignment method based on data augmentation and Transformer, characterized in that, Includes the following steps: Automated data augmentation processing is performed on the raw intraoral scan data to obtain standardized tooth point cloud data; The standardized tooth point cloud data is input into an intelligent tooth arrangement model, which is a neural network built based on the Transformer architecture. It is used to model the local geometric features and global spatial dependencies of teeth and output the prediction results of the target tooth arrangement. The intelligent tooth alignment model is optimized using a joint supervised loss function, which comprehensively constrains the error distribution, spatial structure consistency, and prediction uncertainty of the prediction results.

2. The intelligent tooth alignment method according to claim 1, characterized in that, The automated data augmentation process includes an STU strategy, specifically: Segmentation: The alveolar bone is segmented and removed from the raw intraoral scan data to extract the geometry of individual teeth; Transformation: Apply a random rigid transformation to each tooth to simulate the diverse initial positions of the teeth before orthodontic treatment; Uniform sampling: Uniform sampling is performed on the original and enhanced tooth point clouds to obtain a standardized representation with a fixed number of points.

3. The intelligent tooth alignment method according to claim 2, characterized in that, The random rigid transformation includes random translation and random rotation operations within a preset range.

4. The intelligent tooth alignment method according to claim 1, characterized in that, The intelligent tooth alignment model is a two-layer Transformer network, including: Local feature encoder: This is the first-level Transformer encoder, used to extract the local geometric features of the point cloud for each individual tooth; Global Relationship Encoder: This is a second-level Transformer encoder used to aggregate features from all teeth and model the global spatial constraint relationships between teeth. Feature fusion and output module: used to fuse local and global features and predict the transformation parameters or target position of each tooth from the initial state to the target state.

5. The intelligent tooth alignment method according to claim 4, characterized in that, The local feature encoder uses a self-attention mechanism to learn the structural information inside a single tooth point cloud.

6. The intelligent tooth alignment method according to claim 4, characterized in that, The global relation encoder employs a cross-attention mechanism to calculate the correlation between different tooth features in order to capture collisions, adjacencies, and occlusal relationships between teeth.

7. The intelligent tooth alignment method according to claim 4, characterized in that, The feature fusion and output module includes residual connections to facilitate gradient flow and enhance feature representation.

8. The method according to claim 1, characterized in that, The joint supervised loss function includes a combination of at least two of the following: Position error loss: Calculates the difference between the predicted tooth position and the actual position; Geometric consistency loss: constrains the relative spatial relationships between teeth in the predicted arrangement to reduce the risk of collision and misalignment; Uncertainty regularization loss: Based on the variance of model predictions, an adaptive regularization is applied to samples that are difficult to predict.

9. A smart tooth alignment device, characterized in that, include: The data augmentation unit is configured to perform the data augmentation process as described in any one of claims 1 to 3; The alignment model unit includes a two-level Transformer intelligent tooth alignment model as described in any one of claims 4 to 7, for receiving standardized tooth point cloud data and predicting the target alignment; The training optimization unit is configured to train the alignment model unit using the joint supervised loss function as described in claim 8.

10. An intelligent tooth alignment system, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the program to implement the method as described in any one of claims 1 to 8.

Images