CN112085740A - Tooth fast segmentation method based on three-dimensional tooth jaw model - Google Patents

Abstract

The invention provides a tooth rapid segmentation method based on a three-dimensional tooth jaw model, which comprises the following steps: s1, preprocessing the data of the input three-dimensional dental model; s2 re-modeling the shape on the three-dimensional dental model; s3 determining the grid maximum path search based on the Astar algorithm; s4, identifying the characteristics of the gingival margin line and the fusion area by using the convolutional neural network, and separating out a single tooth. The segmentation method can rapidly and accurately segment each tooth, and provides better reference value for doctors of subsequent false tooth repair.

Description

Fast segmentation method of teeth based on 3D dental model

technical field

The invention relates to a fast tooth segmentation method based on a three-dimensional dental and jaw model, and belongs to the technical field of tooth model image processing.

Background technique

With the development of computer technology, 3D dental and jaw models can be easily digitized by intraoral or extraoral measurement technology, so that CAD/CAM technology is introduced into the oral restoration system, and is used in clinical applications (such as orthodontics, oral cavity, etc.) and collar surgery, etc.) with great success.

Tooth segmentation is an important step in computer-aided orthodontic systems, and its main task is to accurately locate, identify and extract teeth from the patient's three-dimensional jaw model. However, in the process of digitizing the dental and jaw model, due to the influence of the overlapping interference between the teeth caused by the oral deformity, the accuracy of the measurement equipment and the low resolution of the surface reconstruction method, the adjacent teeth of the three-dimensional dental and jaw model are stuck together, and there is no Clear interdental gaps, resulting in partial shape loss of a single tooth segmented. In the CAD/CAM system for prosthodontics, from the production of inlays, veneers, full crowns, partial crowns, simple fixed bridges and even complete dentures, independent single tooth models with original shapes are required. However, the current tooth segmentation methods cannot accurately locate the gingival boundary of the teeth or rely too much on manual interaction, so they cannot guarantee the fast and accurate segmentation of the teeth.

SUMMARY OF THE INVENTION

In order to solve the problems existing in the prior art, the invention proposes a method for rapid tooth segmentation based on a three-dimensional dental and jaw model, which can quickly and accurately segment each tooth, which is convenient for dental workers to perform treatment.

In order to achieve the above object, the method for rapid tooth segmentation based on the three-dimensional dental model of the present invention includes the following steps:

S1 performs data preprocessing on the input 3D dental and jaw model;

S2 remodels the shape on the 3D dental and jaw model;

S3 determines the grid most-path search based on the Astar algorithm;

S4 uses a convolutional neural network to identify the gingival margin line and fusion area features, and separate a single tooth.

Further, step S1 includes:

S11 uses the local surface fitting method to perform discrete curvature analysis on the three-dimensional tooth model, and obtains the curvature value k(p) of each vertex p on the tooth-jaw model based on the principle of maximum principal curvature;

S12 performs a stretching transformation based on histogram equalization on the curvature value, and performs a threshold operation {k(p)>h (h is a threshold)} on the transformed curvature value to obtain an initial feature area;

Some noise points and breakpoints in the S13 area are processed by the three-dimensional morphological opening and closing operation method to obtain the final characteristic area of the boundary between the teeth and the gums.

Further, step S2 includes:

S21 Detects the interdental fusion area, and realizes the automatic identification of the feature line of the feature area by matching the branch points. According to the principle of the shortest distance, select another branch point Jets(i) in the branch point set Jets as the endpoint. An endpoint Jets(j) is used as a matching point to obtain the corresponding fusion area feature line, and automatically identify it, and perform three morphological expansion operations on the identified fusion area feature line until it covers the interdental fusion area, that is, to realize the tooth fusion area. Automatic identification of inter-fusion regions;

S22 delete the interdental fusion area, obtain interdental holes, retain the matched branch points, and realize automatic bridge repair of holes for subsequent tooth restoration;

S23 restores the tooth shape. Based on the vertex information of the boundary of the interdental cavity, the surface corresponding to the missing part is constructed by means of surface energy constraint, and the tooth shape with a high degree of approximation to the original tooth is obtained.

Further, in step S3, the Astar algorithm is used on the effectively restored surface of the three-dimensional dental and jaw model to store the information of each abstracted vertex, and before searching for the shortest distance from the starting point to the target point, two ListA and ListB are created. Collection, where the collection ListA is used to store the nodes that have not been processed, and the collection ListB is used to store the nodes that have been visited. Assuming that the starting node is P and the target node is Q, the specific steps of the algorithm are as follows:

S31 stores the starting node P in the ListA set;

S32 judges whether the ListA collection is empty, if the ListA collection is empty, it means that there is no next node that meets the screening conditions, that is, there is no path, and the search ends, and if there is a node in the ListA collection, proceed to the next step;

S33 takes the node with the smallest cost value from the ListA set as the current optimal node, and judges whether this node is the target node. If it is, it means that the shortest path has been found, and the algorithm ends, and if not, continue to the next step;

S34 checks the neighborhood point of the current node, if the neighborhood point cannot pass or the adjacent node is in the ListB set, skip and continue to process the next node in the neighborhood;

S35 calculates the f value of each adjacent node. If the current adjacent node is neither in the ListA set nor in the ListB set, record the f value of the adjacent node, then add the current node to the path stack, and then add the adjacent node to the path stack. The node is stored in the ListA collection. If the current adjacent node is already in the ListA collection, then compare the size of the newly calculated f value with the current f value. If the new value is smaller, replace the old with the new f value. f value, and then add the adjacent node to the path stack, if the new value is larger, it will not be processed, and the next adjacent node will be processed;

S36: The neighbor nodes of the current node have been processed, and the current node is stored in the ListB set, and then the current node is removed from the ListA set;

S37: Jump to S32 until the optimal path between points P and Q is found or there is no path.

Further, step S4 includes:

S41 inputs the dental and jaw data model provided in step S2 and step S3 into the segmentation network for training, and obtains three caffemodel models with optimal network weights with different parameter settings respectively;

S42 completes the segmentation of a single tooth through the obtained caffemodel model in turn, and uses the conditional random field model to optimize the boundary of the gingival margin area and the interdental contact area;

S43 completes the post-processing of the dental and jaw model through the back-projection ray intersection algorithm and point cloud reconstruction technology.

Further, in step S41, a tooth recognition network is constructed based on 2-level hierarchical feature learning, wherein the ReLU activation function is used to effectively alleviate the gradient dispersion phenomenon, and the three-dimensional convolution expression is:

为权重。Among them, f(*) is the activation function, P _i ×Q _i ×R _i is the weight vector of the convolution kernel at (p, q, r), H ^(m) is the feature vector of the mth channel, b _i,j are bias terms,

for weight.

The fast tooth segmentation method based on the three-dimensional dental and jaw model of the present invention has the following beneficial effects:

(1) The segmentation speed is fast, compared with the traditional interactive marking control algorithm, the tooth segmentation speed of the algorithm proposed by the present invention is significantly improved;

(2) The segmentation accuracy is high, and the traditional interactive labeling algorithm has under-segmentation, but the algorithm proposed by the present invention does not have under-segmentation;

(3) Strong adaptability, for most of the three-dimensional tooth mesh models, the method proposed in the present invention can perform accurate segmentation.

Description of drawings

The present invention will be further described and explained below in conjunction with the accompanying drawings.

FIG. 1 is a flow chart of a method for rapid tooth segmentation based on a three-dimensional dental model according to a preferred embodiment of the present invention.

Detailed ways

The technical solutions of the present invention will be more clearly and completely described below through the description of the preferred embodiments of the present invention with reference to the accompanying drawings.

As shown in FIG. 1 , the method for rapid tooth segmentation based on the three-dimensional dental model according to the preferred embodiment of the present invention includes the following steps:

S1 performs data preprocessing on the input 3D dental and jaw model.

Specifically, step S1 further includes the following steps:

S11 uses the local surface fitting method to perform discrete curvature analysis on the three-dimensional tooth model, and obtains the curvature value k(p) of each vertex p on the tooth-jaw model based on the principle of maximum principal curvature;

S12 performs a stretching transformation based on histogram equalization on the curvature value, and performs a threshold operation {k(p)>h (h is a threshold)} on the transformed curvature value to obtain an initial feature area;

Some noise points and breakpoints in the S13 area are processed by the three-dimensional morphological opening and closing operation method to obtain the final characteristic area of the boundary between the teeth and the gums.

Open operation:

Closing operation:

S2 remodels the shape on the 3D dental and jaw model.

In the process of digitization of the dental and jaw model, due to the influence of the overlapping interference between the teeth caused by the oral deformity and the accuracy of the measurement equipment, the adjacent teeth of the three-dimensional dental and jaw model are stuck together, and there is no clear characteristic area of the boundary between the teeth and the gums. , resulting in the partial shape loss of the segmented single tooth. Therefore, it is very important to remodel the tooth shape of the 3D dental and jaw model.

Specifically, step S2 further includes the following steps:

S21 Detects the interdental fusion area, and realizes the automatic identification of the feature line of the feature area by matching the branch points. According to the principle of the shortest distance, select another branch point Jets(i) in the branch point set Jets as the endpoint. An endpoint Jets(j) is used as a matching point to obtain the corresponding fusion area feature line, and automatically identify it, and perform three morphological expansion operations on the identified fusion area feature line until it covers the interdental fusion area, that is, to realize the tooth fusion area. The automatic identification of the fusion area between the two can achieve better results for most dental and jaw models;

S22 delete the interdental fusion area, obtain interdental holes, retain the matched branch points, and realize automatic bridge repair of holes for subsequent tooth restoration;

S23 restores the tooth shape. Based on the vertex information of the boundary of the interdental cavity, the surface corresponding to the missing part is constructed by means of surface energy constraint, and the tooth shape with a high degree of approximation to the original tooth is obtained.

Through the above method, a higher approximation degree to the original teeth can be obtained, and an accurate three-dimensional jaw model can be provided for oral restoration and orthodontic systems.

S3 determines the grid most-path search based on the Astar algorithm.

Specifically, step S3 uses the Astar algorithm on the effectively restored surface of the three-dimensional dental and jaw model, stores the information of each abstracted vertex, and creates two sets ListA and ListB before searching for the shortest distance from the starting point to the target point, in which The set ListA is used to store the nodes that have not been processed, and the set ListB is used to store the nodes that have been visited. Assuming that the starting node is P and the target node is Q, the specific steps of the algorithm are as follows:

S31 stores the starting node P in the ListA set;

S32 judges whether the ListA collection is empty, if the ListA collection is empty, it means that there is no next node that meets the screening conditions, that is, there is no path, and the search ends, and if there is a node in the ListA collection, proceed to the next step;

S33 takes the node with the smallest cost value from the ListA set as the current optimal node, and judges whether this node is the target node. If it is, it means that the shortest path has been found, and the algorithm ends, and if not, continue to the next step;

S34 checks the neighborhood point of the current node, if the neighborhood point cannot pass or the adjacent node is in the ListB set, skip and continue to process the next node in the neighborhood;

S35 calculates the f value of each adjacent node. If the current adjacent node is neither in the ListA set nor in the ListB set, record the f value of the adjacent node, then add the current node to the path stack, and then add the adjacent node to the path stack. The node is stored in the ListA collection. If the current adjacent node is already in the ListA collection, then compare the size of the newly calculated f value with the current f value. If the new value is smaller, replace the old with the new f value. f value, and then add the adjacent node to the path stack, if the new value is larger, it will not be processed, and the next adjacent node will be processed;

S36: The neighbor nodes of the current node have been processed, and the current node is stored in the ListB set, and then the current node is removed from the ListA set;

S37: Jump to S32 until the optimal path between points P and Q is found or there is no path.

Through the Astar algorithm, the shortest path search of the ideal gingival margin line can be obtained and the speed and accuracy of the algorithm can be guaranteed.

S4 uses a convolutional neural network to identify the gingival margin line and fusion area features, and separate a single tooth.

Specifically, step S4 further includes the following steps:

S41 inputs the dental and jaw data model provided in step S2 and step S3 into the segmentation network for training, and obtains three caffemodel models with optimal network weights with different parameter settings respectively;

S42 completes the segmentation of a single tooth through the obtained caffemodel model in turn, and uses the conditional random field model to optimize the boundary of the gingival margin area and the interdental contact area;

S43 completes the post-processing of the dental and jaw model through the back-projection ray intersection algorithm and point cloud reconstruction technology.

More specifically, in step S41, a tooth recognition network is constructed based on 2-level hierarchical feature learning, wherein the ReLU activation function is used to effectively alleviate the gradient dispersion phenomenon, and the three-dimensional convolution expression is:

为权重。Among them, f(*) is the activation function, P _i ×Q _i ×R _i is the weight vector of the convolution kernel at (p, q, r), H ^(m) is the feature vector of the mth channel, b _i,j are bias terms,

for weight.

Then, the classifier layers are connected in a max pooling and fully connected fashion. In order to avoid training overfitting and improve the generalization ability of the model, random inactivation technology is used in the fully connected layer, and finally the feature values passed to the output layer are input into the classifier for classification prediction, so as to realize the extraction of the tooth feature area. Construction of tooth recognition models.

The encoder-decoder structure is used to establish a CNN-based dental and jaw segmentation network, and a fully connected conditional random field is introduced into the dental segmentation network. At the same time, when calculating the energy function, by considering the correlation between any adjacent points in the dental and jaw model , the gingival margin area and the interdental contact area are optimized, and the local detail features of the segmented area are obtained. In order to improve the segmentation accuracy of a single tooth, in the constructed 3D CRF (conditional random field) model, a bilateral Gaussian filter is used to classify adjacent point clouds in space into the same label; a spatial smoothing Gaussian filter is used to remove isolated points in the segmentation results. Then, by optimizing the energy function of the conditional random field, the segmentation boundary is continuously smoothed, and at the same time, the backprojection ray intersection method is used to project the label result onto the original dental model, and point cloud reconstruction is performed on it. Based on this, a single tooth can be effectively segmented in the end, the accuracy and efficiency of tooth segmentation and recognition can be improved, and a better reference value can be provided for the subsequent denture restoration doctors.

The above-mentioned specific embodiments merely describe the preferred embodiments of the present invention, but do not limit the protection scope of the present invention. Without departing from the design concept and spirit scope of the present invention, various modifications, substitutions and improvements made to the technical solutions of the present invention by those of ordinary skill in the art according to the text description and drawings provided by the present invention shall belong to protection scope of the present invention. The protection scope of the present invention is determined by the claims.

Claims (6)

1. A tooth fast segmentation method based on a three-dimensional tooth jaw model is characterized by comprising the following steps:

s1, preprocessing the data of the input three-dimensional dental model;

s2 re-modeling the shape on the three-dimensional dental model;

s3 determining the grid maximum path search based on the Astar algorithm;

s4, identifying the characteristics of the gingival margin line and the fusion area by using the convolutional neural network, and separating out a single tooth.

2. The method for rapid tooth segmentation based on three-dimensional dental model according to claim 1, wherein the step S1 includes:

s11, discrete curvature analysis is carried out on the tooth three-dimensional model by adopting a local curved surface fitting method, and a curvature value k (p) of each vertex p on the dental jaw model is obtained based on the maximum principal curvature principle;

s12, stretching transformation based on histogram equalization is carried out on the curvature value, and threshold operation { k (p) > h (h is a threshold value) } is carried out on the transformed curvature value, so as to obtain an initial characteristic region;

some miscellaneous points and break points in the S13 area are processed by a three-dimensional morphological open-close operation method to obtain a final characteristic area of the boundary between the tooth and the gum.

3. The method for rapid tooth segmentation based on three-dimensional dental model according to claim 2, wherein the step S2 includes:

s21 detecting an interdental fusion area, realizing automatic identification of characteristic area characteristic lines by matching branch points, selecting another end point Jets (j) from the segmented characteristic lines taking any branch point Jets (i) in the branch point set as an end point according to the shortest distance principle to be used as a matching point to obtain a corresponding fusion area characteristic line, automatically identifying the corresponding fusion area characteristic line, and executing 3 times of morphological expansion operation on the identified fusion area characteristic line until the interdental fusion area is covered, namely realizing the automatic identification of the interdental fusion area;

s22, deleting the inter-dental fusion area to obtain inter-dental holes, and keeping the matched branch points to realize automatic bridging of hole repairing for subsequent tooth repair;

s23 restoring the tooth shape, constructing a curved surface corresponding to the missing part by adopting a curved surface energy constraint mode based on the vertex information of the boundary of the interdental cavity, and obtaining the tooth shape with higher approximation degree with the original tooth.

4. The method for tooth fast segmentation based on three-dimensional dental model as claimed in claim 3, wherein in step S3, the Astar algorithm is used to store the information of each abstracted vertex on the curved surface of the three-dimensional dental model for effective restoration, and two sets of ListA and ListB are created before searching the shortest distance from the starting point to the target point, wherein the set ListA is used to store the nodes that have not been processed, and the set ListB is used to store the nodes that have been visited, and assuming that the starting node is P and the target node is Q, the algorithm comprises the following steps:

s31 storing the starting node P into a ListA set;

s32 judges whether the ListA collection is empty, if the ListA collection is empty, it represents that there is no next node meeting the screening condition, i.e. there is no path, the search is finished, if there is node in the ListA collection, the next step is carried out;

s33, taking the node with the minimum cost value from the ListA set as the current optimal node, judging whether the node is the target node, if so, indicating that the shortest path is found, finishing the algorithm, and if not, continuing the next step;

s34 checking the neighborhood point of the current node, and if the neighborhood point fails or the neighboring node is in the ListB set, skipping to continue processing the next node in the neighborhood;

s35, calculating the f value of each adjacent node, if the current adjacent node is not in the ListA set or the ListB set, recording the f value of the adjacent node, then adding the current node into the path stack, then storing the adjacent node into the ListA set, if the current adjacent node is already in the ListA set, then comparing the newly calculated f value with the current f value, if the new value is smaller, replacing the old f value with the new f value, then adding the adjacent node into the path stack, if the new value is larger, not processing, processing the next adjacent node;

s36: all neighborhood nodes of the current node are processed, the current node is stored in a ListB set, and then the current node is removed from the ListA set;

s37: and jumping to S32 until an optimal path between P and Q points is found or no path exists.

5. The method for rapid tooth segmentation based on three-dimensional dental model according to claim 4, wherein the step S4 includes:

s41, inputting the dental data models provided in the steps S2 and S3 into a segmentation network for training, and respectively obtaining ca ffemodel models with network optimal weights set by 3 different parameters;

s42, completing the segmentation of the single tooth through the obtained ca ffemodel model in sequence, and performing boundary optimization processing on the gingival margin area and the interdental contact area by adopting a conditional random field model;

s43, finishing post-processing of the dental model through a back projection ray intersection algorithm and a point cloud reconstruction technology.

6. The method for tooth fast segmentation based on three-dimensional dental model according to claim 5, wherein in the step S41, a tooth recognition network is constructed based on 2-level feature learning, wherein a ReLU activation function is adopted to effectively alleviate the gradient diffusion phenomenon, and the three-dimensional convolution expression is:

wherein f () is an activation function, P_i×Q_i×R_iWeight vector, H, for the convolution kernel at (p, q, r)^(m)Is the feature vector of the mth channel, b_i，jIn order to be a term of the offset,

are weights.

Images