A dental model registration and segmentation system based on deep learning algorithm

Abstract

The present invention discloses a dental mold registration and segmentation system based on a deep learning algorithm, and specifically relates to the technical field of dental mold registration and segmentation. The present invention constructs a deep learning model. Through the deep learning model, the system can automatically learn and recognize the shape and structure of the dental mold, which greatly improves the accuracy of cutting and positioning. Since the deep learning model has powerful learning and recognition capabilities, the present invention can adapt to various different dental mold shapes and sizes and has strong adaptability. A dental mold segmentation module and a dental mold registration module are provided to control the segmentation and registration processes of the dental mold, which is beneficial to improving the segmentation accuracy and the registration accuracy. The actual application quality index is calculated based on the recorded actual application related data, and the calculation result is evaluated to determine whether it meets the actual application requirements, which is beneficial to optimizing the system processing speed and improving the actual application quality of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of dental model registration and segmentation, and more specifically, to a dental model registration and segmentation system based on a deep learning algorithm. Background Art

[0002] In the field of stomatology, dental mold registration and segmentation is an important step, which involves the precise cutting and positioning of the tooth model. Currently, this process mainly relies on manual operation, which is not only inefficient but also prone to errors. The design of an automated system for dental mold registration and segmentation has become a hot topic.

[0003] The existing dental mold registration and segmentation system obtains the patient's oral structure data through a three-dimensional scanning device. After scanning, a professional doctor will export the data in polygonal files or stereolithography format as needed, use relevant software to mark the boundaries of individual teeth at an appropriate resolution, and use modeling software to build a three-dimensional model of the teeth based on the tooth boundary marking information, which greatly improves the efficiency of dental mold making and guarantees the accuracy of dental mold making to a certain extent.

[0004] However, the above systems still have some problems: although they can improve accuracy and efficiency to a certain extent, they require a lot of manual adjustment and optimization, and when dealing with complex dental mold structures, misjudgments or omissions may occur. In addition, these systems have limited learning and recognition capabilities and are difficult to adapt to various dental mold shapes and sizes. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a dental cast registration and segmentation system based on a deep learning algorithm to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a dental model registration and segmentation system based on a deep learning algorithm, comprising:

[0007] Original data set generation module: after obtaining the permission, the dental and maxillary image data are retrieved from various dental hospitals, and the dental and maxillary image data are numbered according to the receiving time as the original data set;

[0008] Raw data preprocessing module: used to perform image denoising and image enhancement on the raw data set;

[0009] Deep learning model construction module: Determine the model information based on the characteristics of the dataset used and the model construction requirements, use the preprocessed dataset to train and test the designed model, and put the model into use when the model recognition performance index meets the preset value, otherwise the model is tuned;

[0010] Dental model segmentation module: takes the output of the deep learning model as input, pre-processes the input image data, and then uses the image segmentation algorithm to segment the dental image data. The segmentation effect is evaluated. If the preset conditions are met, the segmentation scheme is used. If the preset conditions are not met, the segmentation scheme is optimized.

[0011] Dental model registration module: Retrieve the recognition results of the deep learning model and the segmented 3D coordinate point data of a single tooth to reconstruct the 3D tooth model and perform point cloud fusion of the reconstructed 3D tooth point cloud data with the real 3D tooth point cloud data to evaluate whether the current registration scheme can meet the needs. If it does, it will be put into use. If not, the registration scheme needs to be optimized.

[0012] Practical application module: used for the practical application of dental mold registration and segmentation, and records the relevant data of the actual application process;

[0013] Practical application quality comprehensive evaluation module: calculates the actual application quality index based on the recorded actual application related data, and evaluates the calculation results to determine whether they meet the actual application requirements;

[0014] Database: used to store data information of all modules in the system.

[0015] Preferably, the original data set generation module includes an original data acquisition unit, a data annotation unit, a data retrieval unit, and a data set generation unit, wherein the original data acquisition unit scans the patient's oral cavity with an intraoral scanner to acquire a patient's dental plane image, wherein the dental plane image includes front and side images of the teeth; the data annotation unit is used to annotate the tooth boundaries and contours of the acquired original data; the data retrieval unit retrieves the annotated dental image data from various dental hospitals after obtaining authorization; and the original data set generation unit numbers the retrieved annotated dental model image data in order of receipt time as the original data set.

[0016] Preferably, the deep learning model construction module includes a model selection unit, a data partitioning unit, a model training unit, a model evaluation unit, a model application determination unit, and a model recognition result display unit. The model selection unit selects a network architecture and designs a network structure and the number of layers based on the characteristics of the data set used and the model construction requirements, and imports relevant packages and modifies important global parameters after determining the model used; the data partitioning unit divides the preprocessed original data set into a training data set and a test data set after normalization; the model training unit uses the training data set to perform model training, and uses different optimization algorithms and learning strategies to improve the performance of the model during the training process; the model performance evaluation unit uses the test data set to evaluate the performance of the model, divides the test samples of the test data set into positive and negative cases, and respectively counts the true negative examples TN, false negative examples FN, true positive examples TP, and false positive examples FP tested, and calculates the correct recognition rate CRR and F1 value based on the statistical results. The specific formulas are as follows: , , based on the correct recognition rate and F1 value, the recognition model performance index Y is calculated MX , the specific formula is: , e is a natural constant; the model application determination unit determines that the model can be put into application when the calculated model recognition performance index is greater than a preset value, and determines that the model needs to be tuned when the calculated model recognition performance index is less than or equal to the preset value; the model recognition result display unit is used to display an image containing multiple areas recognized by the deep learning model, each area representing a tooth or a part of the oral structure.

[0017] Preferably, the dental model segmentation module includes a data input unit, a preprocessing unit, an image segmentation unit, and a segmentation effect evaluation unit. The data input unit takes the output result of the deep learning model as input; the preprocessing unit is used to filter and enhance the contrast of the input image data; the image segmentation unit segments the preprocessed tooth image data based on the seed point expansion tooth segmentation algorithm of YOLOv5; the segmentation effect evaluation unit is used to evaluate the image segmentation effect and record the three-dimensional coordinates P of the jth point on the i-th tooth in the k-th input sample after segmentation. fkij (x fkij ,y fkij ,z fkij ), retrieve the real three-dimensional coordinates P of the jth point in the same spatial coordinate system between the i-th tooth and the corresponding segmented tooth in the k-th input sample tkij (x tkij ,y tkij ,z tkij ), count the number of three-dimensional coordinate points m that are consistent between the two zbki 、The total number of non-repeated three-dimensional coordinate points m hbki, the number of pixels m that actually belong to teeth in the statistical segmentation results ytki , pixel m in the segmentation result yfki , and the pixel m that is actually a tooth pixel and is correctly segmented by the segmentation algorithm fzki , based on the number of three-dimensional coordinate points m that are consistent between the two zbki , the total number of three-dimensional coordinate points m hbki Calculate the similarity α of the segmentation result of the i-th tooth in the k-th input sample fki , the specific calculation formula is: , based on the pixel m that actually belongs to the tooth in the statistical segmentation result ytki , pixel m in the segmentation result yfki , and the pixel m that is actually a tooth pixel and is correctly segmented by the segmentation algorithm fzki Calculate the coordination coefficient α of the segmentation effect of the i-th tooth in the k-th input sample tki , the specific formula is: , based on the similarity α of the segmentation result of the i-th tooth in the k-th input sample fki , segmentation effect coordination value α tki , the number of teeth in the kth input sample m qyk , input sample size n sb The specific formula for calculating the average segmentation performance index YFG is: When the calculated segmentation performance index is greater than a preset value, it is determined that the image segmentation scheme is used; when the calculated segmentation performance index is less than or equal to the preset value, it is determined that the image segmentation scheme needs to be optimized.

[0018] Preferably, the dental mold registration module includes a dental mold image information retrieval unit, a three-dimensional reconstruction unit, a point cloud fusion unit, a registration accuracy calculation unit, and a registration scheme determination unit. The dental mold information retrieval unit is used to retrieve the deep learning model recognition result and the three-dimensional coordinate point data of the segmented single tooth; the three-dimensional reconstruction unit re-establishes the spatial rectangular coordinate system and integrates the three-dimensional coordinate point data of the single tooth and the distribution image accordingly; the point cloud fusion unit fuses the k-th reconstructed tooth three-dimensional model with the real tooth three-dimensional model, and counts the number of overlapping three-dimensional coordinate points m ryk , the number of non-overlapping three-dimensional coordinate points m rxk , the number of teeth that completely overlap m vyk , and the number of teeth with non-overlapping three-dimensional coordinate points m vxk The registration accuracy calculation unit is based on the number of overlapping three-dimensional coordinate points m ryk , the number of non-overlapping three-dimensional coordinate points m rxk , the number of teeth that completely overlap m vyk , and the number of teeth with non-overlapping three-dimensional coordinate points mvxk Calculate the average registration accuracy R Z The specific formula is: , e is a natural constant; the registration scheme determination unit compares the calculated registration accuracy with a preset value, and if it is greater than the preset value, the current registration scheme is applied, and if it is less than or equal to the preset value, the current registration scheme needs to be optimized.

[0019] Preferably, the practical application module includes a tooth information acquisition unit, a preprocessing unit, a tooth information recognition unit, a tooth model segmentation and registration unit, an application data recording unit, and an application data output unit. The tooth information acquisition unit is used to scan the patient's oral cavity through an intraoral scanner to acquire a patient's dental planar image; the preprocessing unit is used to denoise and enhance the acquired image data; the tooth information recognition unit inputs the processed tooth image data into a deep learning model for image recognition; the tooth model segmentation and registration unit segments and registers the recognized tooth image information according to the segmentation scheme and the registration scheme; the application data recording unit is used to record the recognition time T of the ith practical application si , segmentation processing time T fi , registration time T pi , and the recognition model performance index Y when applied MXi , average segmentation performance index Y FGi , average registration accuracy R Zi ; The application data output unit is used to output the recorded application data to the actual application quality comprehensive evaluation module.

[0020] Preferably, the actual application quality comprehensive evaluation module includes a data receiving unit, an actual application quality index calculation unit and an actual application quality evaluation unit, wherein the data receiving unit is used to receive the recorded application data; the actual application quality index calculation unit is based on the recognition time T of the ith actual application. si , segmentation processing time T fi , registration time T pi , recognition model performance index Y when applied MXi , average segmentation performance index Y FGi , average registration accuracy R Zi Calculate the actual application quality index W Ci The specific formula is: The actual application quality evaluation unit compares the calculated actual application quality index with a preset value. If the calculated value is greater than the preset value, no operation is performed. If the calculated value is less than or equal to the preset value, the system needs to be optimized.

[0021] Technical effects and advantages of the present invention:

[0022] 1. The present invention sets up an original data set generation module to retrieve the dental and maxillary image data from various dental hospitals after obtaining the authority, and numbers the dental and maxillary image data according to the receiving time as the original data set, sets up an original data preprocessing module to perform image denoising and image enhancement on the original data set, sets up a deep learning model construction module to determine the model information based on the characteristics of the data set used and the model construction requirements, uses the preprocessed data set to train and test the designed model, and puts the model into use when the model recognition performance index of the model meets the preset value, otherwise the model is tuned to ensure the model's ability to recognize teeth. Through the deep learning model, the system can automatically learn and recognize the shape and structure of the dental mold, greatly improving the accuracy of cutting and positioning. Since the deep learning model has powerful learning and recognition capabilities, the present invention can adapt to various dental mold shapes and sizes and has strong adaptability.

[0023] 2. The present invention sets up a dental model segmentation module to take the output result of the deep learning model as input, pre-processes the input image data, and then uses the image segmentation algorithm to segment the tooth image data, evaluates the segmentation effect, and puts this segmentation scheme into use if the preset conditions are met, and optimizes this segmentation scheme if the preset conditions are not met. A dental model registration module is set up to retrieve the recognition result of the deep learning model and the three-dimensional coordinate point data of the segmented single tooth, and then reconstructs the tooth three-dimensional model and performs point cloud fusion of the reconstructed tooth three-dimensional point cloud data with the real tooth three-dimensional point cloud data, evaluates whether the current registration scheme can meet the needs, and puts it into use if it meets the needs, and optimizes it if it does not meet the needs. If the requirements of the actual application are met, the registration scheme needs to be optimized. A comprehensive evaluation module for actual application quality is set up to calculate the actual application quality index based on the recorded actual application related data, and the calculation results are evaluated to determine whether they meet the actual application requirements. While improving the segmentation accuracy and registration accuracy, the processing time of different steps is also optimized, which is beneficial to improving the application quality of the system and realizing the iterative optimization of the system. At the same time, since the whole process is carried out automatically, the efficiency is greatly improved. The traditional dental mold registration and segmentation process requires a lot of manual adjustment, while the present invention greatly reduces this link, making the whole process more concise and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a system structure block diagram of the present invention.

[0025] Figure 2 The present invention is a system operation flow chart. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] like Figure 1 The present embodiment shown provides a dental cast registration and segmentation system based on a deep learning algorithm, including an original data set generation module, an original data preprocessing module, a deep learning model construction module, a dental cast segmentation module, a dental cast registration module, an actual application module, an actual application quality comprehensive evaluation module, and a database. The original data set generation module, the original data preprocessing module, the deep learning model construction module, the dental cast segmentation module, and the dental cast registration module are connected in sequence, the original data preprocessing module, the deep learning model construction module, the dental cast segmentation module, and the dental cast registration module are respectively connected to the actual application module, the actual application module is connected to the actual application quality comprehensive evaluation module, and all modules in the system are connected to the database.

[0028] The original data set generation module retrieves the dental and maxillary image data from various dental hospitals after obtaining the authority, and numbers the dental and maxillary image data according to the receiving time as the original data set.

[0029] Furthermore, the original data set generation module includes an original data acquisition unit, a data annotation unit, a data retrieval unit, and a data set generation unit. The original data acquisition unit scans the patient's mouth with an intraoral scanner to acquire a patient's dental plane image, and the dental plane image includes front and side images of the teeth; the data annotation unit is used to annotate the tooth boundaries and contours of the collected original data; the data retrieval unit retrieves the annotated dental image data from various dental hospitals after obtaining permission; the original data set generation unit numbers the retrieved annotated dental model image data in the order of reception time as the original data set.

[0030] The raw data preprocessing module is used to perform image denoising and image enhancement processing on the raw data set;

[0031] What needs to be specifically explained in this embodiment is that the raw data preprocessing module includes an image denoising unit and an image enhancement unit. The image denoising unit is used to perform denoising on the dental plane image; the image enhancement unit adopts a depth curve estimation method to enhance the image with poor lighting. The depth curve estimation method includes three parts: a light enhancement curve, a depth curve network, and a non-reference loss function. The light enhancement curve automatically maps the input image to the curve, and the parameters of the curve are completely derived from the pixels of the input image. The input of the depth curve network is an image with darker lighting, and the output is a light enhancement curve. The non-reference loss function is used to evaluate the quality of the enhanced image obtained. The selected loss functions are respectively a spatial consistency function L s , exposure control function L e , color constant function L c and the lighting smoothness function L t , the specific function expression is as follows: , where M K represents the number of local regions, P(i) represents the adjacent region centered on region i, Y represents the average value of the enhanced region, and I represents the average value of the input region. , where i is the order of non-overlapping local regions, M fc is the number of non-overlapping local regions, H is the grayscale in the RBG color space, , where (p,q) represents a pair of channels, J p , J q Represent the average area value of p and q channels respectively, , where nd is the total number of iterations, i is the iteration order, ▽ x , ▽ y Respectively represent the horizontal and vertical gradient evaluation of x and y, and the total loss function L z The expression is: , a1 and a2 are the weight coefficients of the loss function.

[0032] Specifically, in this embodiment, a method for designing a denoising model is provided, and the denoising process of the dental and maxillary plane image is completed by using the model, which specifically includes the following steps:

[0033] A1. Take the input noisy jaw plane image sample U as the output layer to complete the network model noise image T (U) prediction;

[0034] A2. The nonlinear feature mapping of the dental plane is completed through the Lecky ReLU function to extract image features. The network model layers are as follows: (1) The first layer: DConv+Leaky ReLU, DConv is the convolution kernel for generating sub-features, and Leaky ReLU is used to complete non-feature mapping; (2) The second to fifth layers: DConv+BN+Leaky ReLU, using the mode of increasing convolution kernel layer by layer, and completing the batch normalization operation of the dental plane image in the expanded convolution layer and Leaky ReLU; (3) The sixth to ninth layers: DConv+BN+Leaky ReLU, using convolution kernels with gradually decreasing layers; (4) DConv layer: Finally, the face image is reconstructed and output through a 3×3×32 convolution kernel;

[0035] A3. Predict the loss function Z of the noise-free image through the mapping function s , the specific formula is: , where m ys is the number of sample pixels, g is the actual noise-free image, c is the denoised image, d t is the image width, h t is the image height, o is the pixel particle;

[0036] A4. Train the mapping model through the residual learning method, and obtain the mean square error σ of the expected residual and predicted residual of the image through the original noise-free image t (n,c o , U o ), U o is a noisy pixel particle, and the specific formula is: , n is the network learning parameter, T e，n is the residual mapping, n z is the noise quantity;

[0037] A5. Use the gradient descent algorithm to optimize the model's parameter matrix β and parameter vector X to calculate the minimum mean square error. The specific formula is: .

[0038] The deep learning model construction module determines the model information based on the characteristics of the data set used and the model construction requirements, uses the preprocessed data set to train and test the designed model, and puts the model into application when the model recognition performance index of the model meets the preset value, otherwise the model is tuned.

[0039] Further, the deep learning model construction module includes a model selection unit, a data partitioning unit, a model training unit, a model evaluation unit, a model application determination unit, and a model recognition result display unit. The model selection unit selects a network architecture and designs a network structure and the number of layers based on the characteristics of the data set used and the model construction requirements, imports relevant packages and modifies important global parameters after determining the model used; the data partitioning unit divides the preprocessed original data set into a training data set and a test data set after normalization; the model training unit uses the training data set to perform model training, and uses different optimization algorithms and learning strategies to improve the performance of the model during the training process; the model performance evaluation unit uses the test data set to evaluate the performance of the model, divides the test samples of the test data set into positive and negative cases, and respectively counts the true negative examples TN, false negative examples FN, true positive examples TP, and false positive examples FP tested, and calculates the correct recognition rate CRR and F1 value based on the statistical results. The specific formulas are as follows: , , based on the correct recognition rate and F1 value, the recognition model performance index Y is calculated MX , the specific formula is: , e is a natural constant; the model application determination unit determines that the model can be put into application when the calculated model recognition performance index is greater than a preset value, and determines that the model needs to be tuned when the calculated model recognition performance index is less than or equal to the preset value; the model recognition result display unit is used to display an image containing multiple areas recognized by the deep learning model, each area representing a tooth or a part of the oral structure.

[0040] The dental model segmentation module takes the output result of the deep learning model as input, pre-processes the input image data, and then uses the image segmentation algorithm to segment the dental image data, evaluates the segmentation effect, and uses this segmentation scheme if it meets the preset conditions, and optimizes this segmentation scheme if it does not meet the preset conditions.

[0041] Furthermore, the dental model segmentation module includes a data input unit, a preprocessing unit, an image segmentation unit, and a segmentation effect evaluation unit. The data input unit takes the output result of the deep learning model as input; the preprocessing unit is used to filter and enhance the contrast of the input image data; the image segmentation unit segments the preprocessed dental image data based on the seed point expansion tooth segmentation algorithm of YOLOv5; the segmentation effect evaluation unit is used to evaluate the image segmentation effect and record the three-dimensional coordinates P of the jth point on the i-th tooth in the k-th input sample after segmentation. fkij (x fkij ,y fkij ,z fkij), retrieve the real three-dimensional coordinates P of the jth point in the same spatial coordinate system between the i-th tooth and the corresponding segmented tooth in the k-th input sample tkij (x tkij ,y tkij ,z tkij ), count the number of three-dimensional coordinate points m that are consistent between the two zbki 、The total number of non-repeated three-dimensional coordinate points m hbki , the number of pixels m that actually belong to teeth in the statistical segmentation results ytki , pixel m in the segmentation result yfki , and the pixel m that is actually a tooth pixel and is correctly segmented by the segmentation algorithm fzki , based on the number of three-dimensional coordinate points m that are consistent between the two zbki , the total number of three-dimensional coordinate points m hbki Calculate the similarity α of the segmentation result of the i-th tooth in the k-th input sample fki , the specific calculation formula is: , based on the pixel m that actually belongs to the tooth in the statistical segmentation result ytki , pixel m in the segmentation result yfki , and the pixel m that is actually a tooth pixel and is correctly segmented by the segmentation algorithm fzki Calculate the coordination coefficient α of the segmentation effect of the i-th tooth in the k-th input sample tki , the specific formula is: , based on the similarity α of the segmentation result of the i-th tooth in the k-th input sample fki , segmentation effect coordination value α tki , the number of teeth in the kth input sample m qyk , input sample size n sb The specific formula for calculating the average segmentation performance index YFG is: When the calculated segmentation performance index is greater than a preset value, it is determined that the image segmentation scheme is used; when the calculated segmentation performance index is less than or equal to the preset value, it is determined that the image segmentation scheme needs to be optimized.

[0042] It should be specifically noted in this embodiment that the YOLOv5 model can only be trained on labeled data, so it is necessary to use the labelImg tool to manually annotate the training set data, attach a range box to each tooth in the image, and determine the seed point of each tooth.

[0043] The dental mold registration module retrieves the recognition results of the deep learning model and the segmented three-dimensional coordinate point data of a single tooth, reconstructs the tooth three-dimensional model, and performs point cloud fusion of the reconstructed tooth three-dimensional point cloud data with the real tooth three-dimensional point cloud data to evaluate whether the current registration scheme can meet the needs. If it does, it will be put into use. If not, the registration scheme needs to be optimized.

[0044] Furthermore, the dental model registration module includes a dental model image information retrieval unit, a three-dimensional reconstruction unit, a point cloud fusion unit, a registration accuracy calculation unit, and a registration scheme determination unit. The dental model information retrieval unit is used to retrieve the deep learning model recognition result and the three-dimensional coordinate point data of the segmented single tooth; the three-dimensional reconstruction unit re-establishes the spatial rectangular coordinate system and integrates the three-dimensional coordinate point data and the distribution image of the single tooth in correspondence; the point cloud fusion unit fuses the k-th reconstructed tooth three-dimensional model with the real tooth three-dimensional model, and counts the number m of overlapping three-dimensional coordinate points. ryk , the number of non-overlapping three-dimensional coordinate points m rxk , the number of teeth that completely overlap m vyk , and the number of teeth with non-overlapping three-dimensional coordinate points m vxk The registration accuracy calculation unit is based on the number of overlapping three-dimensional coordinate points m ryk , the number of non-overlapping three-dimensional coordinate points m rxk , the number of teeth that completely overlap m vyk , and the number of teeth with non-overlapping three-dimensional coordinate points m vxk Calculate the average registration accuracy R Z The specific formula is: , e is a natural constant; the registration scheme determination unit compares the calculated registration accuracy with a preset value, and if it is greater than the preset value, the current registration scheme is applied, and if it is less than or equal to the preset value, the current registration scheme needs to be optimized.

[0045] The practical application module is used to perform practical application of dental model registration and segmentation, and record relevant data of the practical application process.

[0046] Furthermore, the practical application module includes a tooth information acquisition unit, a preprocessing unit, a tooth information recognition unit, a tooth model segmentation and registration unit, an application data recording unit, and an application data output unit. The tooth information acquisition unit is used to scan the patient's oral cavity through an intraoral scanner to acquire a patient's dental planar image; the preprocessing unit is used to denoise and enhance the acquired image data; the tooth information recognition unit inputs the processed tooth image data into a deep learning model for image recognition; the tooth model segmentation and registration unit segments and registers the recognized tooth image information according to the segmentation scheme and the registration scheme; the application data recording unit is used to record the recognition time T of the ith practical application si , segmentation processing time T fi , registration time T pi , and the recognition model performance index Y when applied MXi , average segmentation performance index Y FGi , average registration accuracy R Zi; The application data output unit is used to output the recorded application data to the actual application quality comprehensive evaluation module.

[0047] The actual application quality comprehensive evaluation module calculates the actual application quality index based on the recorded actual application related data, and evaluates the calculation result to determine whether it meets the actual application requirements.

[0048] Furthermore, the actual application quality comprehensive evaluation module includes a data receiving unit, an actual application quality index calculation unit and an actual application quality evaluation unit, wherein the data receiving unit is used to receive the recorded application data; the actual application quality index calculation unit is based on the recognition time T of the ith actual application. si , segmentation processing time T fi , registration time T pi , recognition model performance index Y when applied MXi , average segmentation performance index Y FGi , average registration accuracy R Zi Calculate the actual application quality index W Ci The specific formula is: The actual application quality evaluation unit compares the calculated actual application quality index with a preset value. If the calculated value is greater than the preset value, no operation is performed. If the calculated value is less than or equal to the preset value, the system needs to be optimized.

[0049] The database is used to store data information of all modules in the system.

[0050] It should be specifically noted in this embodiment that the preset values ​​and coefficients used are selected based on actual needs and are not limited to specific values ​​here.

[0051] like Figure 2 The present embodiment provides an operation process of a dental model registration and segmentation system based on a deep learning algorithm, comprising the following steps:

[0052] S1: After obtaining the permission, the dental image data are retrieved from various dental hospitals, and the dental image data are numbered according to the receiving time as the original data set;

[0053] S2: Perform image denoising and image enhancement on the original data set;

[0054] S3: Determine the model information based on the characteristics of the dataset used and the model building requirements, use the preprocessed dataset to train and test the designed model, and apply the model when the model recognition performance index meets the preset value, otherwise, tune the model;

[0055] S4: Taking the output result of the deep learning model as input, preprocessing the input image data, and then segmenting the tooth image data using the image segmentation algorithm, evaluating the segmentation effect, and using this segmentation scheme if it meets the preset conditions, and optimizing this segmentation scheme if it does not meet the preset conditions;

[0056] S5: Retrieve the recognition results of the deep learning model and the segmented 3D coordinate point data of the single tooth, reconstruct the 3D tooth model, and perform point cloud fusion of the reconstructed 3D tooth point cloud data with the real 3D tooth point cloud data to evaluate whether the current registration scheme can meet the needs. If it does, it will be put into use. If not, the registration scheme needs to be optimized.

[0057] S6: Perform practical application of dental mold registration and segmentation, and record relevant data of the actual application process;

[0058] S7: Calculate the actual application quality index based on the recorded actual application related data, and evaluate the calculation results to determine whether they meet the actual application requirements. If they meet the requirements, no operation is performed; if they do not meet the requirements, the system needs to be optimized.

[0059] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A dental model segmentation and registration system based on deep learning algorithm, characterized by: include: Original data set generation module: after obtaining the permission, the dental and maxillary image data are retrieved from various dental hospitals, and the dental and maxillary image data are numbered according to the receiving time as the original data set; Raw data preprocessing module: used to perform image denoising and image enhancement on the raw data set; Deep learning model construction module: Determine the model information based on the characteristics of the dataset used and the model construction requirements, use the preprocessed dataset to train and test the designed model, and put the model into use when the model recognition performance index meets the preset value, otherwise the model is tuned; The deep learning model construction module includes a model selection unit, a data partitioning unit, a model training unit, a model performance evaluation unit, a model application determination unit, and a model recognition result display unit. The model selection unit selects a network architecture and designs a network structure and the number of layers based on the characteristics of the data set used and the model construction requirements. After determining the model used, the relevant packages are imported and important global parameters are modified; the data partitioning unit divides the preprocessed original data set into a training data set and a test data set after normalization operation; the model training unit uses the training data set to perform model training, and uses different optimization algorithms and learning strategies to improve the performance of the model during the training process; the model performance evaluation unit uses the test data set to evaluate the performance of the model, divides the test samples of the test data set into positive and negative cases, and respectively counts the true negative examples TN, false negative examples FN, true positive examples TP, and false positive examples FP tested, and calculates the correct recognition rate CRR and F1 value based on the statistical results. The specific formulas are as follows: , , based on the correct recognition rate and F1 value, the recognition model performance index Y is calculated MX , the specific formula is: , e is a natural constant; the model application determination unit determines that the model can be put into application when the calculated model recognition performance index is greater than a preset value, and determines that the model needs to be tuned when the calculated model recognition performance index is less than or equal to the preset value; the model recognition result display unit is used to display an image containing multiple regions recognized by the deep learning model, each region representing a tooth or a part of the oral structure; Dental model segmentation module: takes the output of the deep learning model as input, pre-processes the input image data, and then uses the image segmentation algorithm to segment the dental image data. The segmentation effect is evaluated. If the preset conditions are met, the segmentation scheme is used. If the preset conditions are not met, the segmentation scheme is optimized. Dental model registration module: Retrieve the recognition results of the deep learning model and the segmented 3D coordinate point data of a single tooth to reconstruct the 3D tooth model and perform point cloud fusion of the reconstructed 3D tooth point cloud data with the real 3D tooth point cloud data to evaluate whether the current registration scheme can meet the needs. If it does, it will be put into use. If not, the registration scheme needs to be optimized. The dental model registration module includes a dental model image information retrieval unit, a three-dimensional reconstruction unit, a point cloud fusion unit, a registration accuracy calculation unit, and a registration scheme determination unit. The dental model information retrieval unit is used to retrieve the deep learning model recognition result and the three-dimensional coordinate point data of the segmented single tooth; the three-dimensional reconstruction unit re-establishes the spatial rectangular coordinate system and integrates the three-dimensional coordinate point data of the single tooth and the distribution image accordingly; the point cloud fusion unit fuses the k-th reconstructed tooth three-dimensional model with the real tooth three-dimensional model, and counts the number of overlapping three-dimensional coordinate points m ryk , the number of non-overlapping three-dimensional coordinate points m rxk , the number of teeth that completely overlap m vyk , and the number of teeth with non-overlapping three-dimensional coordinate points m vxk The registration accuracy calculation unit is based on the number of overlapping three-dimensional coordinate points m ryk , the number of non-overlapping three-dimensional coordinate points m rxk , the number of teeth that completely overlap m vyk , and the number of teeth with non-overlapping three-dimensional coordinate points m vxk Calculate the average registration accuracy R Z The specific formula is: , e is a natural constant, n sb The number of samples is input; the registration scheme determination unit compares the calculated registration accuracy with a preset value, and if it is greater than the preset value, the current registration scheme is applied; if it is less than or equal to the preset value, the current registration scheme needs to be optimized; Practical application module: used for the practical application of dental model segmentation and registration, and records the relevant data of the actual application process; The practical application module includes a tooth information acquisition unit, a preprocessing unit, a tooth information recognition unit, a tooth model segmentation and registration unit, an application data recording unit, and an application data output unit. The tooth information acquisition unit is used to scan the patient's oral cavity through an intraoral scanner to acquire a patient's dental and maxillary plane image; the preprocessing unit is used to perform denoising and image enhancement on the acquired image data; The tooth information recognition unit inputs the preprocessed tooth image data into the deep learning model for image recognition; the tooth model segmentation and registration unit segments and registers the recognized tooth image information according to the segmentation scheme and the registration scheme; the application data recording unit is used to record the recognition time T of the ith actual application si , segmentation processing time T fi , registration time T pi , and the recognition model performance index Y when applied MXi , average segmentation performance index Y FGi , average registration accuracy R Zi ; The application data output unit is used to output the recorded application data to the actual application quality comprehensive evaluation module; Practical application quality comprehensive evaluation module: Calculates the actual application quality index based on the recorded actual application related data, and evaluates the calculation results to determine whether they meet the actual application requirements.

2. The dental model segmentation and registration system based on deep learning algorithm according to claim 1, characterized in that: The raw data set generation module includes a raw data acquisition unit, a data annotation unit, a data retrieval unit, and a data set generation unit. The raw data acquisition unit scans the patient's oral cavity with an intraoral scanner to acquire a patient's dental plane image, wherein the dental plane image includes front and side images of the teeth; the data annotation unit is used to annotate the tooth boundaries of the collected raw data; the data retrieval unit retrieves the annotated dental image data from various dental hospitals after obtaining the authority; the raw data set generation unit numbers the retrieved annotated dental model image data in the order of receiving time as the raw data set.

3. The dental model segmentation and registration system based on deep learning algorithm according to claim 1, characterized in that: The dental model segmentation module includes a data input unit, a preprocessing unit, an image segmentation unit, and a segmentation effect evaluation unit. The data input unit takes the output result of the deep learning model as input; the preprocessing unit is used to filter and enhance the contrast of the input image data; the image segmentation unit segments the preprocessed tooth image data based on the seed point expansion tooth segmentation algorithm of YOLOv5; the segmentation effect evaluation unit is used to evaluate the image segmentation effect and record the three-dimensional coordinates P of the jth point on the i-th tooth in the k-th input sample after segmentation. fkij (x fkij ,y fkij ,z fkij ), retrieve the real three-dimensional coordinates P of the jth point in the same spatial coordinate system between the i-th tooth and the corresponding segmented tooth in the k-th input sample tkij (x tkij ,y tkij ,z tkij ), count the number of three-dimensional coordinate points m that are consistent between the two zbki 、The total number of non-repeated three-dimensional coordinate points m hbki , the number of pixels m that actually belong to teeth in the statistical segmentation results ytki , pixel m in the segmentation result yfki , and the pixel m that is actually a tooth pixel and is correctly segmented by the segmentation algorithm fzki , based on the number of three-dimensional coordinate points m that are consistent between the two zbki , the total number of three-dimensional coordinate points m hbki Calculate the similarity α of the segmentation result of the i-th tooth in the k-th input sample fki , the specific calculation formula is: , based on the pixel m that actually belongs to the tooth in the statistical segmentation result ytki , pixel m in the segmentation result yfki , and the pixel m that is actually a tooth pixel and is correctly segmented by the segmentation algorithm fzki Calculate the coordination coefficient α of the segmentation effect of the i-th tooth in the k-th input sample tki , the specific formula is: , based on the similarity α of the segmentation result of the i-th tooth in the k-th input sample fki , segmentation effect coordination value α tki , the number of teeth in the kth input sample m qyk , input sample size n sb The specific formula for calculating the average segmentation performance index YFG is: When the calculated segmentation performance index is greater than a preset value, it is determined that the image segmentation scheme is used; when the calculated segmentation performance index is less than or equal to the preset value, it is determined that the image segmentation scheme needs to be optimized.

4. The dental model segmentation and registration system based on deep learning algorithm according to claim 1, characterized in that: The actual application quality comprehensive evaluation module includes a data receiving unit, an actual application quality index calculation unit and an actual application quality evaluation unit. The data receiving unit is used to receive the recorded application data; the actual application quality index calculation unit is based on the recognition time length T of the ith actual application. si , segmentation processing time T fi , registration time T pi , recognition model performance index Y when applied MXi , average segmentation performance index Y FGi , average registration accuracy R Zi Calculate the actual application quality index W Ci The specific formula is: The actual application quality evaluation unit compares the calculated actual application quality index with a preset value. If the calculated value is greater than the preset value, no operation is performed. If the calculated value is less than or equal to the preset value, the system needs to be optimized.

Images