CN114387259A

CN114387259A - Method and device for predicting missing tooth coordinates and training method of recognition model

Info

Publication number: CN114387259A
Application number: CN202210040685.7A
Authority: CN
Inventors: 钱坤; 黄志俊; 刘金勇; 张有健
Original assignee: Lancet Robotics Co Ltd
Current assignee: Lancet Robotics Co Ltd
Priority date: 2022-01-14
Filing date: 2022-01-14
Publication date: 2022-04-22

Abstract

A training method for identifying a missing tooth model, a method and a device for predicting missing tooth coordinates, an electronic device and a storage medium are provided. For CBCT data of a deficient missing tooth patient, missing tooth data are generated through coverage of pixel values, missing teeth are segmented through a neural network, and therefore coordinate positioning can be carried out on the segmented missing tooth data. During training, according to the priori knowledge, a relatively obvious boundary point existing between the missing tooth and the normal tooth and soft tissue is converted into a thermodynamic diagram through Gaussian transformation, and the characteristic diagram is amplified or reduced through a corresponding convolution layer and a corresponding pooling layer and is added into a down-sampling channel corresponding to the Unet network. The coordinates of 8 points of the upper surface and the lower surface of the tooth model are obtained from the divided tooth model, so that a 3D rectangular frame surrounding the tooth model is obtained, the center point of each surface of the rectangular frame is obtained respectively and connected to obtain the intersection point of the central axes of the four surfaces as the coordinate of the center point, and the coordinate can be used as the predicted coordinate of the planting pit point.

Description

Method and device for predicting missing tooth coordinates and training method of recognition model

Technical Field

The invention relates to the technical field of medical image processing, in particular to a method and a device for predicting missing tooth coordinates, a training method for training a missing tooth recognition model, a computer-readable storage medium and electronic equipment.

Background

In recent years, with the progress of medical imaging technology and computer technology, especially artificial intelligence technology, medical image processing and analysis, especially deep learning-based, for example, deep convolutional neural networks, have attracted attention, and can automatically segment and extract implicit physiological or diagnostic features from medical image big data.

Common clinical medical images include, for example, MRI images, CT images, X-ray images, ultrasound images, PET images, and CBCT (Cone Beam Computed Tomography) images. Among them, CBCT is widely used by oral physicians due to its advantages of small radiation dose, short scanning time, low shooting cost, high image spatial resolution, and the like.

However, since the dentist reads a large amount of medical optical sheet data every day, such work may cause recognition deviation of tooth position coordinates, especially deviation of implant planning, due to personal factors such as fatigue, emotion, low experience level, and the like. Because of the complexity of the oral structure, planning of dental implants is a cumbersome task for the dentist of the mouth, often requiring the dentist to spend a great deal of time in planning placement of the dental implant.

Therefore, in order to improve the accuracy of implant planning, a scheme for predicting tooth position coordinates or missing tooth coordinates is developed, and the method has great significance for reducing the workload of dentists, reducing the labor cost and reducing the occurrence of misdiagnosis.

Disclosure of Invention

However, due to the privacy protection of the patient with the missing tooth, the CBCT scan data of the patient with the missing tooth is relatively lacking and is not enough to support the training of the deep neural network. Therefore, although there is a need for predicting missing tooth coordinates using deep learning, overfitting is severe due to the lack of the number of training pictures, and therefore a further corresponding solution is needed for technical implementation of predicting missing tooth coordinates using a deep learning method.

To solve the above problems, an object of the present invention is to provide a technique capable of predicting missing tooth coordinates using a deep learning method.

(1) According to an aspect of the present invention, there is provided a training method for identifying a missing tooth model, comprising the steps of:

step 201: preprocessing a tooth image as original image data by overlaying pixel values to generate an image including missing tooth data;

step 202: the generated image containing the missing tooth data and the artificially marked tooth label form a training set, and the training set is trained by using a Unet neural network,

wherein: converting a feature map containing the coordinates of the missing tooth area into a thermodynamic diagram through a Gaussian variation function H, amplifying or reducing the feature map through a corresponding convolution layer and a corresponding pooling layer, and adding the feature map into a corresponding down-sampling channel of the Unet neural network to train the neural network.

More specifically, the following method and apparatus are also provided according to the present invention.

(2) Preferably, step 200 is further included before step 201: and acquiring the tooth image in the CBCT data format as original image data.

(3) Preferably, the step 201 comprises: and manually labeling the CBCT data to mark a single target tooth, generating a corresponding labeled STL file, and overlaying the label data of the labeled STL file on the original image data.

(4) Preferably, the Hu value of the overlay data is set to 200.

(5) Preferably, the gaussian variation function H is:

where φ is the scaling of the feature map, d is the center coordinate of the missing tooth region, b is the edge coordinate point of the missing tooth, p is any coordinate point in the feature map, δ_bIs the standard deviation of the coordinates of the center and the edge of the missing tooth area.

(6) Preferably, the uet neural network comprises:

a CPLD operation module for feature extraction, the CPLD operating to: generating a characteristic diagram for an input image file through 2 convolution layers of 3-by-3 convolution kernels, performing maximum pooling layer processing, outputting by using LeakyReLU as an activation function, and performing spatial dropout processing to obtain an output characteristic diagram; and

a decCONV operation module for an upsample operation, the decCONV operation being to: after deconvolution operation is carried out on the output characteristic graph of each layer of CPLD operation, Concat characteristic fusion input is carried out on the output characteristic graph of the last layer of CPLD operation to obtain an output characteristic graph,

the neural network is configured to:

the feature diagram CPLD1 with one time enlarged feature channel obtained after the input original image data passes through the first layer of CPLD operation module is used as the input of the next layer of CPLD operation module, and the feature diagrams CPLD 2-CPLD 5 with one time enlarged feature channel are sequentially obtained;

inputting a feature diagram CPLD5 into an output feature diagram decCONV1 obtained by the decCONV operation module, and taking the output feature diagram decCONV1 as the input of the next layer of decCONV operation module to sequentially obtain decCONV 1-decCONV 4;

and then using the sigmoid activation function to obtain an output image group.

(7) Preferably, the neural network is configured to: inputting an original image into a first CPLD layer to obtain a characteristic diagram CPLD1 with one time of characteristic channel expansion, inputting CPLD1 serving as input into a second CPLD layer to obtain a characteristic diagram CPLD2 with one time of characteristic channel expansion, and performing 5 CPLD layer operations in the same way to obtain a characteristic diagram CPLD5 with 1024 characteristic channels;

the method comprises the steps of performing 2-time deconvolution on a feature graph CPLD5 serving as input, performing Concat feature fusion input on the feature graph CPLD4 to obtain an output feature graph decCONV1, enabling a convolution kernel of the deconvolution to be 2 x 2, then performing Concat feature fusion on the decCONV1 serving as input and CPLD3 through 4-time deconvolution to obtain an output feature graph decCONV2, then performing 8-time deconvolution on the decCONV2 serving as input, performing Concat feature fusion on the output feature graph decCONV3 and CPLD2, and then performing 16-time deconvolution to obtain decCONv 4.

(8) Preferably, the labeled training set is converted from dicom format to a 2-dimensional image in jpg format, and the window width is adjusted to 1312hu and the window level is adjusted to 213 hu.

(9) Training device for identifying missing tooth models, comprising the following modules:

a preprocessing module configured to preprocess a tooth image, which is original image data, by coverage of pixel values to generate an image including missing tooth data;

a training module configured to combine the generated image including the missing tooth data and the artificially labeled tooth label into a training set and train the training set using a Unet neural network,

the training module is configured to: and converting a feature map containing the coordinates of the missing tooth area into a thermodynamic diagram through a Gaussian variation function H, amplifying or reducing the feature map through a corresponding convolution layer and a corresponding pooling layer, and adding the feature map into a corresponding down-sampling channel of the Unet neural network to train the neural network.

(10) A method for predicting missing tooth coordinates using a neural network determined by the above training as a neural network for segmentation, comprising the steps of:

step 101, acquiring a tooth image to be processed of a patient with missing teeth as original image data,

step 102, the original image data is predicted through a neural network for division, the divided missing tooth data is stored,

and 103, calculating the coordinates of the missing tooth model based on the segmented missing tooth data.

(11) Preferably, step 103 further comprises: the matrix data of the jpg image of the divided missing tooth data is read, and the jpg image is converted into a 3D STL file format.

(12) Preferably, the 8 point coordinates of the upper and lower surfaces of the tooth model in the 3D STL file format are obtained to obtain a 3D rectangular frame surrounding the tooth model, and the center point of each surface of the rectangular frame is obtained and connected to obtain the intersection point of the central axes of the four surfaces as the center point coordinates of the tooth model.

(13) An apparatus for predicting missing tooth coordinates, comprising the modules of:

an original image data acquisition module configured to acquire a tooth image to be processed of a missing tooth patient as original image data,

a target tooth model segmentation module configured to predict original image data through the neural network determined by the training and store the segmented missing tooth data,

and the missing tooth coordinate positioning module is used for calculating the coordinates of the missing tooth model based on the segmented missing tooth data.

(14) A computer-readable storage medium storing a computer program for performing any of the above methods.

(15) An electronic device, comprising: a processor and a memory for storing instructions executable by the processor, the processor being configured to read the instructions from the memory and execute the instructions to implement any of the methods described above.

Thus, according to the present invention, it is possible to perform deep neural network training to automatically segment missing teeth based on deep learning while sufficiently protecting the privacy of a missing tooth patient, thereby reducing the subjectivity of a doctor, avoiding misdiagnosis, and the like caused by eye fatigue of the doctor, rapidly obtaining a diagnosis result by calculating nest data, and giving a reasonable planting suggestion.

Drawings

Fig. 1 schematically shows a flow chart of a method for predicting missing tooth coordinates according to an embodiment of the present invention.

FIG. 2 illustrates a CBCT missing dental data image made in accordance with an embodiment of the present invention.

Fig. 3 schematically illustrates a network infrastructure for training a predictive missing tooth coordinate recognition model according to an embodiment of the present invention.

FIG. 4 illustrates a segmented 3D dental model according to an embodiment of the present invention.

Fig. 5 schematically shows a 3D rectangular box for obtaining coordinates of a center point of a tooth model.

Fig. 6 schematically shows the demarcation points between the missing tooth and the other parts.

Fig. 7 is a schematic structural diagram of a training apparatus for identifying a missing tooth model according to an embodiment of the present invention.

Fig. 8 is a schematic structural diagram illustrating an apparatus for predicting missing tooth coordinates based on deep learning according to an embodiment of the present invention.

Detailed Description

Exemplary embodiments of the present invention are described in detail below with reference to the accompanying drawings. The exemplary embodiments described below and illustrated in the figures are intended to teach the principles of the present invention and enable one skilled in the art to implement and use the invention in several different environments and for several different applications. The scope of the invention is, therefore, indicated by the appended claims, and the exemplary embodiments are not intended to, and should not be considered as, limiting the scope of the invention.

Segmentation methods based on deep learning, such as convolutional neural networks, have been the research focus of medical image analysis, but in the field of focusing on privacy protection, such as oral medicine, the scale of training set data also imposes a great limit on the segmentation methods. Even if a network based on small sample training data, such as a U _ Net network, is adopted, the image edge features still have the problem of unclear segmentation.

The following description of at least one exemplary embodiment of the invention is merely illustrative in nature and is in no way intended to limit the invention, its application, or uses.

As a development environment, for example, a software environment can be set: programming languages such as python3.6, C + +, integrated framework tensrflow2.5, cmake3.20, other components such as cuda11.1, vtk 9.0.0 +; the experimental environment is, for example, Windows10, RTX3090 graphics card.

< exemplary method >

Fig. 1 schematically shows a flow diagram of a method 200 for predicting missing tooth coordinates according to an embodiment of the invention. The embodiment can be applied to an electronic device, as shown in fig. 1, and includes the following steps:

in step 101, a tooth image to be processed of a missing tooth patient is acquired as original image data,

in step 102, the original image data is predicted by the neural network for segmentation, and the segmented missing tooth data is saved,

in step 103, coordinates of the missing tooth model are calculated based on the segmented missing tooth data.

Therefore, the image containing the missing tooth can be automatically segmented by utilizing the neural network to predict the missing tooth model, so that the dentist is assisted to quickly obtain the predicted coordinate for planting the pit point.

And training the neural network before segmenting the original image data through the neural network.

According to one embodiment, the training of the neural network comprises:

step 201: the tooth image as the original image data is preprocessed by the overlay of pixel values to generate an image containing missing tooth data.

Step 202: the generated image containing missing tooth data and the artificially labeled tooth labels may be combined into a training set and used as a segmentation network using a Unet neural network.

Preferably, step 200 is further included before step 201: the tooth image is acquired as original image data, and the tooth image in the CBCT data format is used as the original image data.

Therefore, the images containing missing tooth data based on the CBCT data format and the artificially labeled tooth labels can form a training set, and the Unet neural network is used as a segmentation network.

Due to the privacy protection of the patient with the missing teeth, the CBCT data of the patient with the missing teeth is relatively lack and is not enough to support the training of a deep neural network. Therefore, the CBCT data of the missing tooth are self-generated by the pixel value covering method. In this way not only the data set can be increased but also the privacy of the patient can be protected, since fewer samples of missing dental data are available.

More specifically, when it is desired to obtain missing tooth data about any one tooth, even if CBCT data of a patient missing the tooth (hereinafter referred to as a missing tooth) is relatively missing, a robust CBCT data sample of the tooth may be used to perform pixel value coverage on an image of the tooth, for example, to cover the tooth with black or white pixels, and when a missing tooth pixel set at the tooth position is formed, missing tooth data formed after the pixel value coverage of the tooth is acquired and saved, so as to provide pixel information for the subsequent coordinate positioning of the missing tooth.

According to one embodiment, the preprocessing of the original image data by the overlay of pixel values to generate an image containing missing tooth data includes:

firstly, a data annotating person uses the Segmentation function of the mics software to select a single tooth (data needing to cover pixels) on oral cavity CBCT data to draw a label and generate a corresponding STL (Standard Template Library) file. Then, untagged DICOM (Digital Imaging and Communications in Medicine) data and a tagged STL file are read through a Simple ITK library, the STL formatted file is parsed into Uint8 format by a sit k.cast function, label data is overlaid on the original data using sit k.label overlay function, and Hu value of the overlaid data is set to 200.

The pixel coverage covers all the hierarchical data of the CBCT because the used STL file is in a 3-dimensional format, and can simulate the data of the missing teeth relatively truly on a 3D level. Pixel filling may be performed by copying pixel values of other vacuum regions on the CT image as overlaid pixel values. This is mainly done because CT data itself has noise, and the data can be made more realistic by copying pixels in the noise region for filling.

Here, the Hu (hounsfiled unit) value is a dimensionless unit commonly used in Computed Tomography (CT), and generally, the higher the density is, the higher the Hu value is, and here, the Hu value of the region covering pixels is set to 200, which is close to the Hu value of soft tissue, so that missing tooth data can be closer to reality.

As described above, the original data of the missing tooth may be obtained by using the raw data of the entire teeth and covering the tooth at the missing tooth position with another pixel (for example, black or white pixel) for the target missing tooth position at any desired position.

Accordingly, the later-described label is also a tooth model in which the missing tooth is used, and the center point may be the center point of the missing tooth or the center point of the rectangular frame, as described later in detail, with reference to fig. 4 and 5.

According to an embodiment of the present invention, for the missing tooth data as shown in fig. 2, a total of, for example, 20 CBCTs are made. The 20 sets of the manufactured missing tooth CBCT data and the manually drawn tooth label are combined into a training set, and an improved Unet neural network is used as a segmentation network, so that a better segmentation effect can be obtained.

Here, since the edge of the created missing dental data is unclear and the data amount is small, and it is difficult to achieve a good segmentation effect by using a conventional neural network such as the Unet network, the present inventors will add prior knowledge to the Unet network to achieve a good segmentation effect. Here, the prior knowledge refers to: there is a relatively sharp demarcation point between the missing tooth and other parts (i.e., the surrounding normal teeth and soft tissue). As shown in fig. 6, the white dotted line is the demarcation point of the missing tooth and other parts of the surroundings and is input as a priori knowledge.

The inventor adds a neural network channel by using a thermodynamic diagram conversion mode as a priori knowledge, and particularly converts the boundary points into a thermodynamic diagram by performing Gaussian transformation on the boundary points according to the existence of obvious boundary points between the missing tooth and surrounding normal teeth and soft tissues, wherein the formula of the Gaussian change is as follows:

Step 203: through the Gaussian variation function H, the feature map containing the missing tooth area coordinates can be converted into a thermodynamic diagram, and the generated thermodynamic diagram is amplified or reduced through the corresponding convolution layer and pooling layer and added to a down-sampling channel corresponding to the Unet network to train the network. As shown in particular in figure 3.

More specifically, because the thermodynamic value of the edge region of the missing tooth in the generated thermodynamic diagram is required to be different from the thermodynamic values of other regions, a specific conversion method is to firstly determine the central point of the missing tooth region, determine the corresponding thermodynamic value by calculating the distance value between the central point and each point of the edge region, thereby generating a corresponding matrix, and because OpenCV library processing is used, the dtype format of the matrix is set to Uint8 for convenient display, and the above gaussian transformation formula can be referred to specifically. And finally, displaying a corresponding thermodynamic diagram through a COLORMAP _ HOT function in OpenCV.

After 20000 epochs are trained on the Unet network added with the priori knowledge, the network obtains a good segmentation effect.

According to one embodiment, the training network adds a priori knowledge to a neural network architecture based on the classical Unet, namely, the training network is divided into three parts: respectively a feature extraction (down-sampling), up-sampling, prior region.

Firstly, a training set with a label converts CT data in a DICOM format into a 2-dimensional image of jpg through a pydicom library, and the inventor notices that the window width is adjusted to 1312hu and the window level is adjusted to 213hu through a plurality of tests in consideration of the fact that the hu values of CT in different regions are different greatly. The pictures generated by the training set are input into a neural network configured as follows for training.

1. Feature extraction:

the image generates a feature map by 2 convolution layers of 3-by-3 convolution kernels, and the size of the feature map is reduced by a maximum pooling layer, so that the field of view of the convolution kernels is enlarged. And using LeakyReLU as an activation function for outputting, and reducing the parameter number through Spatial dropout, wherein the series of operations are abbreviated as CPLD (Conv-Pool-LeakyReLU-Spatial Droupout), when an original image is input and passes through a CPLD layer, a characteristic diagram with one time of characteristic channel expansion is obtained and is marked as CPLD1, and when CPLD1 is input and passes through a 2 nd CPLD layer, a characteristic diagram with one time of characteristic channel expansion is obtained and is marked as CPLD 2. By analogy, through 5-layer CPLD layer operation, a feature diagram CPLD5 with a feature channel number of 1024 is obtained.

The feature extraction network structure is shown in the left half of fig. 3.

2. And (3) upsampling:

after feature extraction, the number of output feature channels is continuously increased after 5-layer CPLD layer operation, and a feature map needs to be restored and output through deconvolution operation. Therefore, the input feature map CPLD5 is subjected to 2-fold deconvolution operation, and then is subjected to Concat feature fusion input with CPLD4 to obtain an output feature map, which is denoted as decCONV1, and the convolution kernel of the deconvolution is 2 × 2. DecCONV1 is used as input, and is subjected to deconvolution by 4 times and is subjected to Concat characteristic fusion with CPLD3 to obtain an output characteristic diagram, which is recorded as decCONV2, then DecCONV2 is used as input, and is subjected to deconvolution by 8 times and is subjected to Concat characteristic fusion with CPLD2 to obtain an output characteristic diagram, which is recorded as decCONV3, and then is subjected to deconvolution by 16 times to obtain decCONv 4. The upsampling operation is shown in the right half of fig. 3.

3. A prior layer: the thermodynamic diagram converted from the demarcation points of the soft tissues at the missing tooth and other soft tissues (of the tooth position adjacent to the missing tooth) is taken as an a priori input, and is converted into corresponding characteristic diagrams through a convolution layer and a pooling layer, and is respectively Concat into the corresponding CPLD1-CPLD5 layers.

And then using the sigmoid activation function to obtain an output image group.

In this way, a trained neural network is obtained to determine a neural network model for predicting the missing tooth model.

Inputting the test set into a trained neural network to obtain a jpg image predicted to be output, reading matrix data of the jpg image through a numpy library, converting the jpg image into a 3D STL model through a numpy _ to _ vtk function, and obtaining a segmentation result as shown in FIG. 4.

< coordinates of planting Point >

After the 3D tooth model is obtained, in order to obtain the coordinates of the implant point, the implant pit point is often positioned at the center of the missing tooth according to the knowledge of the implant tooth, so the coordinates of the implant point can be obtained by obtaining the central axis of the missing tooth area,

the specific scheme includes that 8 point coordinates (corresponding to coordinate information of upper and lower four corner points of an external rectangular cuboid frame of the tooth model, namely, the leftmost two points and the rightmost two points of the upper surface, the leftmost two points and the rightmost two points of the lower surface, namely, 4 points of one surface and 8 points in total) of the upper surface and the lower surface of the tooth model are obtained through a mesh function in an stl library, a 3D rectangular frame surrounding the tooth model is obtained, and central points of all surfaces of the rectangular frame are obtained respectively and connected to obtain a central axis of the surface. The intersection point of the central axes of the four surfaces is the coordinate of the center point of the tooth model, and is also the predicted coordinate of the implant pit point, as shown in fig. 5.

Of course, the method for obtaining the center point coordinates is not limited to this, and may also be determined by using the midpoint information of each side of the rectangular parallelepiped frame.

Therefore, planning time can be shortened by planning the implant on the coordinate in the later period, and labor cost is reduced.

Moreover, the predictive segmentation can predict the tooth extraction result before the extraction of the affected tooth, thereby bringing great convenience for a doctor to explain the operation plan to a patient.

Although the original data may be without missing teeth, the data of missing teeth can be formed by arbitrarily selecting one tooth and overlaying the pixel value of the tooth according to the invention. In this way, data of missing teeth can be formed after one tooth is arbitrarily selected and covered by the pixel value of the tooth, so that after the prediction models for all the teeth are accumulated, an estimated image of any tooth with pertinence can be provided, such as a cavity situation after tooth extraction and a possible planting situation on the basis of the estimated image.

Further, an electronic device according to one embodiment includes: a processor and a memory; wherein the memory is to store processor-executable instructions; the processor is used for reading the executable instructions from the memory and executing the instructions to realize the method.

Further, a computer-readable storage medium according to an embodiment stores a computer program for executing the above-described method.

The electronic device may be either or both of the first device and the second device, or a stand-alone device separate from them, which stand-alone device may communicate with the first device and the second device to receive the acquired input signals therefrom. The processor may be a Central Processing Unit (CPU) or other form of processing unit having data processing capabilities and/or instruction execution capabilities, and may control other components in the electronic device to achieve desired functionality. The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and/or non-volatile memory.

In the present invention, "first" and "second", "S101", "S102", etc. are used only for the purpose of distinguishing between descriptions of two different object features, and are not used to indicate an order of arrangement, relative importance, or to implicitly indicate the number of indicated technical features.

The method steps, modules and other components of each example described in the embodiments disclosed in the present invention can be implemented by electronic hardware, computer software or a combination of the two, and some or all of them can be selected according to actual needs to implement the purpose of the embodiment.

While the invention has been described with reference to various specific embodiments, it should be understood that changes can be made within the spirit and scope of the inventive concepts described. Accordingly, it is intended that the invention not be limited to the described embodiments, but that it will have the full scope defined by the language of the following claims.

Claims

1. A training method for identifying missing tooth models, comprising the steps of:

wherein: converting a feature map containing the missing tooth area coordinates into a thermodynamic diagram through a Gaussian variation function H, amplifying or reducing the feature map through a corresponding convolution layer and a corresponding pooling layer by the generated thermodynamic diagram, and adding the feature map into a down-sampling channel corresponding to the Unet neural network to train the neural network.

2. Training method for identifying a missing dental model according to claim 1,

before the step 201, the method further comprises the step 200: and acquiring the tooth image in the CBCT data format as original image data.

3. Training method for identifying a missing dental model according to claim 1,

the step 201 comprises: and manually labeling the CBCT data to mark a single target tooth, generating a corresponding labeled STL file, and overlaying the label data of the labeled STL file on the original image data.

4. Training method for identifying a missing dental model according to any of claims 1 to 3,

the Hu value of the overlay data is set to 200.

5. Training method for identifying a missing dental model according to any of claims 1 to 3,

the gaussian variation function H is:

6. Training method for identifying a missing dental model according to claim 1,

the Unet neural network includes:

the neural network is configured to:

and then using the sigmoid activation function to obtain an output image group.

7. Training method for identifying a missing dental model according to claim 6,

the neural network is configured to:

inputting an original image into a first CPLD layer to obtain a characteristic diagram CPLD1 with characteristic channels being enlarged by one time, inputting CPLD1 serving as input into a second CPLD layer to obtain the characteristic diagram CPLD2 with characteristic channels being enlarged by one time, and performing operation on 5 CPLD layers in the same way to obtain the characteristic diagram CPLD5 with the characteristic channel number of 1024;

performing 2-time deconvolution on the feature map CPLD5 serving as input, performing Concat feature fusion input on the feature map and CPLD4 to obtain the output feature map decCONV1, wherein a convolution kernel of the deconvolution is 2 x 2, then performing Concat feature fusion on the decCONV1 serving as input, performing 4-time deconvolution and on CPLD3 to obtain the output feature map decCONV2, then performing 8-time deconvolution on the decCONV 35567 serving as input, performing Concat feature fusion on the input and CPLD 8645 to obtain the output feature map decCONV3, and performing 16-time deconvolution on the output feature map decCONV 4.

8. Training method for identifying a missing dental model according to claim 1,

the labeled training set is converted from dicom format to a 2-dimensional image in jpg format and the window width is adjusted to 1312hu and the window level to 213 hu.

9. Training device for identifying missing tooth models, characterized in that it comprises the following modules:

10. A method for predicting missing tooth coordinates, characterized by using the neural network determined by the training method of any one of claims 1 to 8 as a neural network for segmentation, comprising the steps of:

11. The method for predicting missing tooth coordinates of claim 10,

step 103 further comprises: the matrix data of the jpg image of the divided missing tooth data is read, and the jpg image is converted into a 3D STL file format.

12. The method for predicting missing tooth coordinates of claim 11,

and obtaining 8 point coordinates of each of the upper surface and the lower surface from the tooth model in the 3D STL file format to obtain a 3D rectangular frame surrounding the tooth model, and obtaining a central point of each surface of the rectangular frame and connecting to obtain an intersection point of central axes of the four surfaces as the central point coordinate of the tooth model.

13. An apparatus for predicting missing tooth coordinates, comprising:

the original image data acquisition module is used for acquiring a tooth image to be processed of the patient with the missing tooth as original image data;

a target tooth model segmentation module configured to predict original image data by the neural network determined by the training method according to any one of claims 1 to 8, and store the segmented missing tooth data; and

14. A computer-readable storage medium, characterized in that a computer program is stored for executing the training method for identifying a missing tooth model according to any one of claims 1 to 8 or the method for predicting missing tooth coordinates according to any one of claims 10 to 12.

15. An electronic device, comprising: a processor and a memory for storing instructions executable by the processor, the processor being configured to read the instructions from the memory and execute the instructions to implement the training method for identifying a missing tooth model according to any one of claims 1 to 8 or the method for predicting missing tooth coordinates according to any one of claims 10 to 12.