Tooth point cloud segmentation method based on deep learning
By optimizing tooth point cloud segmentation using the Point Transformer V3 model and an interactive segmentation framework, the problems of reliance on manual labeling and insufficient segmentation accuracy in existing tooth segmentation technologies are solved, achieving efficient and accurate tooth segmentation.
Patent Information
- Application Number
- CN202511168216.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2025-07-23
- Filing Date
- 2025-08-20
- Publication Date
- 2025-11-28
AI Technical Summary
Existing tooth segmentation methods rely on manual labeling by doctors, resulting in insufficient segmentation accuracy and low efficiency in complex scenarios. Traditional methods are sensitive to noise and struggle to handle situations with large differences in tooth shape, irregularities, and blurred boundaries. Multi-stage model training is complex and the parameter tuning process is cumbersome.
The Point Transformer V3 model is used for preprocessing, initial segmentation, annotation, and correction of tooth point clouds. The model parameters are optimized through an interactive segmentation framework. By combining sparse tensor data and skip connections with attention mechanisms, end-to-end tooth segmentation is achieved.
It improves the accuracy and efficiency of tooth point cloud segmentation, reduces the amount of manual correction work, can more accurately distinguish between different tissues such as teeth and gums, and simplifies the segmentation process.
Smart Images
Figure CN121033424A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of oral medical image processing, and in particular to a tooth point cloud segmentation method based on deep learning. BACKGROUND
[0002] With the digital transformation of modern dental technology, three-dimensional point cloud data has become an important technical means for tooth diagnosis, treatment planning, and oral health research. Three-dimensional point cloud data can comprehensively present the three-dimensional morphology of teeth, including the specific position, shape and surface features of teeth, and is widely used in orthodontic, dental implant design and tooth restoration scenarios.
[0003] However, robust segmentation of individual teeth from intraoral scanning data still faces the following key problems: first, manual segmentation of each tooth by doctors is a time-consuming and difficult process due to the special nature of 3D data, especially involving the boundary part of the tooth; second, traditional segmentation methods based on geometric features are very sensitive to noise and difficult to handle cases with large differences in tooth shape, irregularity and blurred boundaries; in addition, although existing deep learning models can automatically segment teeth to some extent, the segmentation accuracy in complex scenarios is insufficient, especially for special cases such as tooth defects and dense arrangement, and the multi-stage model training is complex, requiring a complex parameter tuning process. SUMMARY
[0004] In view of the above deficiencies in the prior art, the present application provides a tooth point cloud segmentation method based on deep learning, which solves the problems of manual marking by doctors, insufficient segmentation accuracy in complex scenarios and low segmentation efficiency in the prior art.
[0005] In order to achieve the above-mentioned purposes, the technical solution adopted by the present application is: The present application provides a tooth point cloud segmentation method based on deep learning, comprising: S1: preprocessing the initial point cloud data of the input tooth, the preprocessing including down-sampling, format standardization and principal component analysis pose normalization of the initial point cloud data; S2: performing initial segmentation on the preprocessed point cloud by a Point Transformer V3 model to generate an initial segmentation result; S3: performing labeling and correction operations on the initial segmentation result to obtain a labeled point set and a corrected label; S4: optimizing the parameters of the Point Transformer V3 model based on the labeled point set and the corrected label; S5: based on the optimized Point Transformer V3 model, the preprocessed point cloud is segmented, if the segmentation result reaches the preset threshold, the segmentation result is obtained based on the optimized Point Transformer V3 model, otherwise, S3 is returned, and the optimization is continued until the segmentation result reaches the preset threshold.
[0006] Further, the initial segmentation of the preprocessed point cloud by the Point Transformer V3 model generates an initial segmentation result, comprising: S201: serializing the preprocessed point cloud data; S202: constructing sparse tensor data based on the serialized point cloud data; S203: based on the constructed sparse tensor data, input the encoder-decoder network structure for feature extraction and restoration, enhance the spatial semantic expression through the jump connection and attention mechanism, and output the initial point-level segmentation result.
[0007] Further, the initial segmentation of the preprocessed point cloud by the Point Transformer V3 model generates an initial segmentation result, comprising: A1: input the sparse tensor data into the encoder for down-sampling processing to extract the point cloud features; A2: the decoder up-samples the extracted point cloud features, gradually improves the spatial resolution of the point cloud features, and fuses the point cloud features through the jump connection to obtain point cloud data close to the initial point cloud data spatial resolution; A3: the segmentation head module receives the point cloud data output by the encoder to obtain the label category of each point cloud data, and further obtains the initial segmentation result.
[0008] Further, the formula of the jump connection is:
[0009] wherein, is the deep feature after up-sampling, is the feature of the corresponding stage of the encoder, is a learnable weight matrix, is an activation function.
[0010] Further, the encoder adopts a five-stage hierarchical design, and the feature dimension of each stage increases while the spatial resolution decreases; the decoder adopts a symmetric four-stage design, and gradually restores the spatial resolution; the encoder and the decoder of each layer adopt a serialized attention mechanism, and the attention calculation formula of the serialized attention mechanism is:
[0011] In the formula, , and These are query, key, and value matrices, respectively. This is a relative position code calculated based on the positional relationships after serialization. The dimension of the key vector, superscript For transpose; The formula for calculating the relative position code is as follows:
[0012] In the formula, and This is the one-dimensional position index after serialization. For point Relative to point Location encoding, This represents the index of the current query point or target point. Point The neighboring points.
[0013] Furthermore, the step of labeling and correcting the initial segmentation results to obtain a set of labeled points and corrected labels includes: S301: Connect the initial segmentation results to the interactive segmentation framework and calculate the segmentation index of the label category for each point cloud data in the initial segmentation results; S302: If the segmentation index is lower than the preset index threshold, load the visualized point cloud of label category information, observe the visualized point cloud, select the points with segmentation errors through interactive tools, the interactive segmentation framework receives the selected points, and uses the adjacency search algorithm to select all points with the same label within the specified radius, while accepting the correct label category input, to obtain the set of labeled points that need to be corrected and the correct correction label.
[0014] Furthermore, the optimization of the Point Transformer V3 model parameters based on the labeled point set and corrected labels includes: S401: Optimize the loss function based on the labeled point set and corrected labels; S402: Use the optimized loss function as the new loss function for the Point Transformer V3 model to obtain the optimized Point Transformer V3 model.
[0015] Furthermore, the loss function is a weighted sum of the corrected energy and the stable energy:
[0016] in, To correct the energy, To stabilize energy, This is a weighting factor.
[0017] Furthermore, the correction energy is defined as the sparse cross-entropy loss of the labeled point set:
[0018] In the formula, The number of labeled point sets, For point The one-hot encoding of the tag, For point Predicted as category The probability, where M is the number of categories; The stable energy is defined as a weighted sum of global points:
[0019] In the formula, For point Filtering score, This represents the number of point clouds after preprocessing and downsampling. The filtering score is dynamically updated based on entropy changes:
[0020] In the formula, For point exist Entropy change in round optimization For the first The filtering score of the wheel, For the first The filtering score of the wheel, and For a predefined threshold, No. The stable energy of the wheel For the first The stable energy of the wheel Optimize the number of rounds after each interaction.
[0021] The beneficial effects of this application are: This application provides a deep learning-based tooth point cloud segmentation method that effectively captures the features and details of tooth point clouds by employing the PointTransformer V3 model. Furthermore, the interactive correction mechanism allows for annotation and correction of the initial segmentation results. The model dynamically adjusts its parameters based on these corrections, enabling further optimization of the final segmentation results based on deep learning. This results in more accurate differentiation between teeth and different tissues such as gums, reducing the workload of manual correction and improving the accuracy of tooth point cloud segmentation. Attached Figure Description
[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the accompanying drawings in the following description only represent some of the embodiments of the present application, and other embodiments can also be obtained by those skilled in the art based on these drawings.
[0023] Figure 1 A flowchart of a tooth point cloud segmentation method based on deep learning provided by an embodiment of the present application. DETAILED DESCRIPTION
[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art based on the present application belong to the scope of protection of the present application.
[0025] Prior art one discloses a 3D tooth segmentation and classification method based on deep learning. The model framework used by the prior art is divided into four stages. Stage one: sample data collection and preparation; stage two: deep network model building and training based on the two segmentation between teeth and gums; stage three: deep network model building and training based on the sixteen segmentation between teeth; stage four: integration of the results of stage two and stage three. The prior art has the technical effects of "based on deep convolutional network, through two classification tasks to segment teeth and gums, to separate the target teeth, then perform sixteen segmentation tasks between the teeth, due to the difference in the number of tooth surfaces between single teeth, use different loss weights to train the deep learning model, then combine the results of teeth and gums and apply graph cut algorithm for post-processing, finally use support vector machine algorithm to upsample back to the original 3D tooth data, to achieve the purpose of accurate labeling of actual 3D teeth". However, this method needs to train two stage models, one is two classification to distinguish teeth and gums, the other is multi-classification to segment sixteen teeth, the training process is complex and cannot achieve end-to-end segmentation, in addition, it needs complex parameter tuning.
[0026] The prior art two discloses a region growth algorithm, which can also segment the dental crown. Region growth refers to starting from a certain pixel, gradually adding adjacent pixels that meet the similarity criterion according to certain criteria, and terminating region growth when the growth stopping condition is met. The similarity criterion can be image information such as pixel gray value, color, texture feature, etc. (Another region growth is to first segment the image into many small regions with strong consistency, such as small regions with the same pixel gray value in the region, and then fuse the small regions into large regions according to certain rules to achieve the purpose of segmenting the image).
[0027] The region growth algorithm generally includes three steps: determining a growth seed point, defining a growth criterion, and determining a growth stopping condition.
[0028] The steps of implementing the region growth algorithm are as follows: 1. Scan the image in order, find the first pixel that has not been assigned, and set the pixel as (x0, y0); 2. Take (x0, y0) as the center, consider the 4-neighborhood or 8-neighborhood pixels (x, y) of (x0, y0), if (x0, y0) meets the growth criterion, merge (x, y) with (x0, y0) (in the same region), and push (x, y) into the stack at the same time; 3. Take a pixel from the stack, treat it as (x0, y0), and return to step 2; 4. When the stack is empty, return to step 1; 5. Repeat steps 1-4 until each point in the image has been assigned, and the growth is complete.
[0029] Currently, the region growth algorithm is used to segment the dental crown, Loop subdivision is used to smooth the mesh and smooth the segmentation path, and then an initial seed point is manually interacted to each tooth, two bending degree evaluation functions H1 and H2 are used to replace the accurate calculation of curvature, and the region growth algorithm grows from the seed point to the boundary of the dental crown. However, the region growth algorithm needs to manually interact an initial seed point to each tooth, and over-segmentation is prone to occur.
[0030] Based on this, the embodiment of the application provides a tooth point cloud segmentation method based on deep learning, which can refer to Figure 1 , Figure 1 as shown in the flowchart of the tooth point cloud segmentation method based on deep learning provided by the embodiment of the application, which includes: S1: pre-processing the initial point cloud data of the input tooth, the pre-processing includes down-sampling the initial point cloud data to 10000 points, format standardization, and principal component analysis pose normalization.
[0031] In an embodiment of the present application, the spatial coordinate information of the input initial point cloud data is normalized to [-1, 1], the number of points of the original point cloud data is uniformly down-sampled to 10,000 points using the farthest point sampling method, and the spatial pose offset of the data is adjusted using the principal component analysis method to ensure that the poses of all data are approximately the same.
[0032] S2: performing initial segmentation on the preprocessed point cloud by using a Point Transformer V3 model to generate an initial segmentation result.
[0033] The specific segmentation steps include: S201: performing serialization processing on the preprocessed point cloud data.
[0034] In an embodiment of the present application, the serialization processing is performed on the defined three-dimensional point cloud where the point cloud By using a space-filling curve such as a Z-order and a Hilbert curve to traverse points in a three-dimensional space, the curve can map the points to a high-dimensional discrete space while maintaining the spatial proximity between the points. Mathematically, the space-filling curve can be defined as a bijective function φ: where is the dimension of the space, is the dimension of the mapped high-dimensional space.
[0035] S202: constructing sparse tensor data based on the point cloud data after serialization processing.
[0036] In an embodiment of the present application, the serialized point cloud data is mapped into a sparse tensor format acceptable to a sparse convolutional network (SpConv) based on a voxelization method, and a high-efficiency data structure is provided for subsequent operations by calculating grid coordinates and dynamic sparse space shapes.
[0037] S203: inputting the constructed sparse tensor data into an encoder-decoder network structure for feature extraction and restoration, enhancing spatial semantic expression through a skip connection and an attention mechanism, and outputting an initial point-level segmentation result.
[0038] In an embodiment of the present application, the encoding and decoding parts adopt a conventional U-Net structure, the point cloud features are extracted through continuous down-sampling of the encoder, and then the decoder gradually recovers the point cloud features through a skip connection and feature fusion to generate an accurate point-level segmentation prediction, which is specifically as follows: A1: inputting the sparse tensor data into an encoder for down-sampling processing to extract point cloud features.
[0039] In an embodiment of the present application, the encoder adopts a five-stage hierarchical design, with the feature dimension increasing and the spatial resolution decreasing at each stage.
[0040] Stage configuration: Stage 0: [N, 16] - initial feature embedding, keeping original resolution Stage 1: [N / 4, 32] - first downsampling, doubling feature dimension Stage 2: [N / 16, 64] - mid-level feature extraction, expanding receptive field Stage 3: [N / 64, 128] - deep semantic feature, enhanced processing Stage 4: [N / 256, 256] - deepest global feature, largest receptive field wherein N represents the number of point clouds after pre-processing and downsampling, and the downsampling ratios of each stage are 1, 1 / 4, 1 / 16, 1 / 64 and 1 / 256, respectively.
[0041] A2: The decoder upsamples the extracted point cloud features, gradually improves the spatial resolution of the point cloud features, and fuses the point cloud features through skip connection to obtain point cloud data close to the spatial resolution of the initial point cloud data.
[0042] In an embodiment of the present application, the decoder adopts a symmetric four-stage design, gradually recovering the spatial resolution.
[0043] Decoding stage mapping: Dec Stage 3: [N / 256, 256]→[N / 64, 128]+Skip[N / 64, 128] Dec Stage 2: [N / 64, 128]→[N / 16, 64]+Skip[N / 16, 64] Dec Stage 1: [N / 16, 64]→[N / 4, 32]+Skip[N / 4, 32] Dec Stage 0: [N / 4, 32]→[N, 16]+Skip[N, 16] wherein Skip represents a residual connection adopted by the skip connection.
[0044] The formula of the skip connection is:
[0045] wherein, is the deep feature after upsampling, is the feature of the corresponding stage of the encoder, is a learnable weight matrix, is an activation function.
[0046] The encoder and the decoder of each layer adopt a serialized attention mechanism with linear complexity compared with the standard self-attention, and the attention calculation formula of the serialized attention mechanism is as follows:
[0047] In the formula, , and are query, key and value matrices respectively, is a relative position encoding calculated based on the serialized position relationship, is a key vector dimension, and the superscript is a transpose.
[0048] wherein the calculation formula of the relative position encoding is as follows:
[0049] wherein, and are one-dimensional position indexes after serialization, is a position encoding of a point relative to a point , represents an index of a current query point or a target point, represents a neighbor point of the point .
[0050] A3: The segmentation head module receives the point cloud data output by the encoder, obtains the label category of each point cloud data, and further obtains an initial segmentation result.
[0051] In an embodiment of the present application, the segmentation head first performs dimension transformation on the input high-dimensional feature vector through a multi-layer perception (MLP), gradually compresses and maps to a preset category space. Then, the Softmax function is used to normalize the class score of each point to obtain its probability distribution belonging to each category. Finally, the system selects the class with the maximum probability value of each point as its predicted label, thereby completing the initial semantic segmentation of the entire point cloud.
[0052] S3: Perform labeling and correction operations on the initial segmentation result to obtain a labeled point set and a corrected label.
[0053] In an embodiment of the present application, the initial segmentation result is input into an interactive segmentation framework, which can be a general term for inference using the point cloud segmentation model obtained in the previous step and then combining the subsequent interactive segmentation, the segmentation indicators of the label categories of each point cloud data in the initial segmentation result are calculated, if the segmentation indicators are lower than the preset indicator threshold, the label category information of the visualized point cloud is loaded, the user observes the visualized point cloud, selects a point with segmentation error through the interactive tool shift+left mouse button click, the interactive segmentation framework receives the selected point, and uses an adjacent search algorithm to select all points within a specified radius around the point and with the same label, while accepting the correct label category input by the user, to obtain a set of labeled points requiring correction and a correct corrected label.
[0054] S4: optimizing the parameters of the Point Transformer V3 model based on the set of labeled points and the corrected label.
[0055] Further, the optimizing the parameters of the Point Transformer V3 model based on the set of labeled points and the corrected label comprises: S401: optimizing a loss function based on the set of labeled points and the corrected label.
[0056] In an embodiment of the present application, the set of labeled points is used to optimize the initial segmentation by optimizing an objective function, the objective function is a loss function, and includes a weighted sum of a correction energy and a stability energy, wherein the correction energy is a sparse cross-entropy loss of the set of labeled points, and the stability energy is an entropy weighted loss of global points.
[0057] The correction energy reflects the degree of correct segmentation of the clicked point (the set of labeled points), and is defined as a sparse cross-entropy loss of the clicked point (the set of labeled points):
[0058] In the formula, is the number of the set of labeled points that are also clicked points, is the label of the point in one-hot encoding, is the probability that the point is predicted to be of the category , and is the number of categories. The stability energy reflects the overall stability of the segmentation result, and is defined as a weighted sum of global points:
[0059] In the formula, is the filtering score of the point , and is the number of point clouds after pre-processing and down-sampling.
[0060] The filtering score is dynamically updated by the entropy change, and the specific rules are as follows:
[0061] Wherein, is the point In the entropy change in the optimization round, is the filtering score of the first round, is the filtering score of the first round, and is a predefined threshold, is the stable energy of the first round, is the stable energy of the first round, is the number of optimization rounds after each interaction.
[0062] The final test loss function is defined as the weighted sum of the correction energy and the stabilization energy:
[0063] Wherein, is the correction energy, is the stabilization energy, is the weighting factor.
[0064] S402: The optimized loss function is used as the new loss function of the Point Transformer V3 model, and the optimized Point Transformer V3 model is obtained.
[0065] S5: Based on the optimized Point Transformer V3 model, the preprocessed point cloud is segmented, if the segmentation result reaches the preset threshold, the segmentation is performed on the to-be-segmented data based on the optimized Point Transformer V3 model, otherwise, return to S3, continue to optimize until the segmentation result reaches the preset threshold.
[0066] In an embodiment of the present application, the loss function obtained above is used as the new loss function of the previous model, and the parameters of the original model are automatically updated by normal deep learning training for 5 rounds (after network back propagation) to optimize the model parameters for specific data. The point cloud data is again inferred through the updated model to obtain new label data for each point, visualize the new results, and calculate new index information. If the index is still below, repeat this step (from selecting points by the user), until the index exceeds the set threshold, or the maximum number of iterations is exceeded. If the optimization process always fails to meet the set threshold, a maximum number of iterations is set to exit. For results that meet the standard or do not need to be modified, the mapped original point cloud data is restored to its original size. The mapping method is a feature propagation method based on tri-linear interpolation. Specifically, for each point in the original point cloud, the Euclidean distance between the point and the down-sampled point set is calculated, the three closest points are selected, and the feature interpolation result is calculated according to the distance weighting.
[0067] The tooth point cloud segmentation method based on deep learning provided by the present application improves the accuracy of tooth point cloud segmentation by using the PointTransformer V3 model. The multi-layer Transformer architecture can effectively capture the features and details of the tooth point cloud, and is more convenient and effective than traditional methods and previous two-stage deep learning for complex situations such as dense tooth arrangement and fuzzy boundaries. The interactive correction mechanism allows users to label and correct the initial segmentation result, and the model dynamically adjusts the parameters based on these corrections, so that the final segmentation result can be further optimized to accurately distinguish between teeth and gums and other different tissues based on deep learning. The interactive optimization design reduces the workload of manual modification; traditional methods require professional personnel to manually adjust a large number of incorrect segmentation points, while the present scheme only requires users to label a small amount of data in key error areas, and the model can automatically complete the subsequent optimization process. This semi-automatic processing method retains the key judgments of professionals while avoiding tedious point-by-point modification, making the entire segmentation process more efficient and convenient.
[0068] It should be noted that those skilled in the art will realize that the embodiments described herein are intended to aid the reader in understanding the principles of the present application, and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations of the present application without departing from the spirit of the present application, and these modifications and combinations are still within the scope of protection of the present application.
Claims
1. A tooth point cloud segmentation method based on deep learning, characterized in that, include: S1: Preprocess the initial point cloud data of the input teeth. The preprocessing includes downsampling, format standardization, and principal component analysis pose normalization of the initial point cloud data. S2: Perform initial segmentation on the preprocessed point cloud using the Point Transformer V3 model to generate initial segmentation results; S3: Perform annotation and correction operations on the initial segmentation results to obtain the annotation point set and correction labels; S4: Optimize the parameters of the Point Transformer V3 model based on the labeled point set and corrected labels; S5: Segment the preprocessed point cloud based on the optimized Point Transformer V3 model. If the segmentation result reaches the preset threshold, segment the data to be segmented based on the optimized Point Transformer V3 model. Otherwise, return to S3 and continue optimization until the segmentation result reaches the preset threshold.
2. The tooth point cloud segmentation method based on deep learning according to claim 1, characterized in that, The initial segmentation of the preprocessed point cloud using the Point Transformer V3 model to generate initial segmentation results includes: S201: Perform serialization processing on the preprocessed point cloud data; S202: Constructing sparse tensor data based on serialized point cloud data; S203: Based on the constructed sparse tensor data, feature extraction and restoration are performed in the encoder-decoder network structure. Spatial semantic expression is enhanced through skip connections and attention mechanisms, and the initial point-level segmentation result is output.
3. The tooth point cloud segmentation method based on deep learning according to claim 2, characterized in that, The constructed sparse tensor data is input into the encoder-decoder network structure for feature extraction and reconstruction. Spatial semantic representation is enhanced through skip connections and attention mechanisms, and the initial point-level segmentation results are output, including: A1: Input the sparse tensor data into the encoder for downsampling to extract point cloud features; A2: The decoder upsamples the extracted point cloud features to gradually improve the spatial resolution of the point cloud features, and fuses the point cloud features through skip connections to obtain point cloud data with a spatial resolution close to that of the initial point cloud data. A3: The segmentation head module receives the point cloud data output by the encoder, obtains the label category of each point cloud data, and then obtains the initial segmentation result.
4. The tooth point cloud segmentation method based on deep learning according to claim 3, characterized in that, The formula for the jump connection is: in, These are the deep features after upsampling. Features corresponding to the encoder stage, The weight matrix is a learnable matrix. This is the activation function.
5. The tooth point cloud segmentation method based on deep learning according to claim 3, characterized in that, The encoder employs a five-stage hierarchical design, with the feature dimension increasing and spatial resolution decreasing in each stage; the decoder employs a symmetrical four-stage design, gradually restoring spatial resolution; each layer's encoder and decoder utilize a serialized attention mechanism, the attention calculation formula of which is: In the formula, , and These are query, key, and value matrices, respectively. This is a relative position code calculated based on the positional relationships after serialization. The dimension of the key vector, superscript For transpose; The formula for calculating the relative position code is as follows: In the formula, and This is the one-dimensional position index after serialization. For point Relative to point Location encoding, This represents the index of the current query point or target point. Point The neighboring points.
6. The tooth point cloud segmentation method based on deep learning according to claim 1, characterized in that, The step of labeling and correcting the initial segmentation results to obtain a set of labeled points and corrected labels includes: S301: Connect the initial segmentation results to the interactive segmentation framework and calculate the segmentation index of the label category for each point cloud data in the initial segmentation results; S302: If the segmentation index is lower than the preset index threshold, load the visualized point cloud of label category information, observe the visualized point cloud, select the points with segmentation errors through interactive tools, the interactive segmentation framework receives the selected points, and uses the adjacency search algorithm to select all points with the same label within the specified radius, while accepting the correct label category input, to obtain the set of labeled points that need to be corrected and the correct correction label.
7. The tooth point cloud segmentation method based on deep learning according to claim 6, characterized in that, The parameters of the Point Transformer V3 model optimized based on the labeled point set and corrected labels include: S401: Optimize the loss function based on the labeled point set and corrected labels; S402: Use the optimized loss function as the new loss function for the Point Transformer V3 model to obtain the optimized Point Transformer V3 model.
8. The tooth point cloud segmentation method based on deep learning according to claim 7, characterized in that, The loss function is a weighted sum of the corrected energy and the stable energy: in, To correct the energy, To stabilize energy, This is a weighting factor.
9. The tooth point cloud segmentation method based on deep learning according to claim 8, characterized in that, The correction energy is defined as the sparse cross-entropy loss of the labeled point set: In the formula, The number of labeled point sets, For point The one-hot encoding of the tag, For point Predicted as category The probability, Number of categories; The stable energy is defined as a weighted sum of global points: In the formula, For point Filtering score, This represents the number of point clouds after preprocessing and downsampling. The filtering score is dynamically updated based on entropy changes: In the formula, For point exist Entropy change in round optimization For the first The filtering score of the wheel, For the first The filtering score of the wheel, and For a predefined threshold, No. The stable energy of the wheel For the first The stable energy of the wheel Optimize the number of rounds after each interaction.