CBCT image tooth segmentation method based on deep learning

Abstract

The invention relates to a CBCT image tooth segmentation method. The method comprises the following steps: generating a two-dimensional oral cavity panoramic film from a CBCT image; carrying out tooth two-dimensional segmentation and tooth ID classification on the two-dimensional oral cavity panoramic film by adopting a deep learning algorithm; calculating a key point coordinate of a two-dimensional single tooth and mapping the key point coordinate to a three-dimensional CBCT image coordinate; respectively cutting the CBCT image to obtain an area of each tooth; segmenting each three-dimensional tooth region by adopting a deep learning algorithm; and assigning ID to all the tooth segmentation results, and splicing to obtain a three-dimensional tooth segmentation result. According to the method, accurate segmentation of the tooth area in the CBCT image can be realized, complex conditions such as tight tooth arrangement, similar shapes and artifact interference can be effectively handled, the problems of tooth segmentation result fusion, ID confusion and metal artifact interference can be effectively solved, and resource consumption is remarkably reduced by determining the tooth key points on the panoramic film, so that the tooth segmentation efficiency is improved. The overall segmentation efficiency is improved, and the segmentation time is shortened, so that the accuracy of the segmentation result and the processing speed are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of oral image processing, and in particular to a CBCT image tooth segmentation method based on deep learning. Background Art

[0002] As an advanced imaging technology, cone beam computed tomography (CBCT) has been widely used in the field of stomatology with its advantages of high resolution and low radiation dose. Through CBCT, the three-dimensional structural information of the patient's teeth, bones and soft tissues in the mouth can be accurately obtained, providing clinicians with a clear basis for diagnosis. Accurate segmentation of each tooth in CBCT images and the establishment of a three-dimensional model can not only help clinicians with orthodontic treatment planning and the design and placement of dental implants, but also promote communication between doctors and patients, making the presentation of surgical plans more intuitive and vivid.

[0003] However, due to factors such as the close arrangement of teeth, similar shapes of adjacent teeth, and changes in tooth morphology caused by oral diseases, tooth segmentation poses great challenges. In addition, low-quality imaging (such as metal artifacts) further increases the complexity of segmentation. In this case, existing deep learning segmentation algorithms are prone to problems such as tooth fusion and confusion of tooth IDs, making it difficult to ensure the accuracy of segmentation results. Manual labeling is not only time-consuming and requires high professionalism, but also difficult to apply in practice. Summary of the invention

[0004] The technical problem to be solved by the present invention is to design a CBCT image tooth segmentation method based on deep learning, which can realize accurate segmentation of the tooth area in the CBCT image, effectively deal with complex situations such as close arrangement of teeth, similar morphology, and artifact interference, thereby solving the existing technical problems.

[0005] In order to solve the above technical problems, the present invention provides a CBCT image tooth segmentation method based on deep learning, which specifically includes the following steps:

[0006] Step S1: Generate a two-dimensional oral panoramic film from the CBCT image;

[0007] Step S2: using a deep learning algorithm to perform two-dimensional tooth segmentation and tooth ID classification on the two-dimensional oral panoramic film;

[0008] Step S3: Calculate the key point coordinates of a single tooth in two dimensions and map them to three-dimensional CBCT image coordinates;

[0009] Step S4: cropping the CBCT image to obtain the area of ​​each tooth according to the three-dimensional CBCT image coordinates;

[0010] Step S5: using a deep learning algorithm to segment each tooth region in three dimensions;

[0011] Step S6: assign IDs to all tooth segmentation results and then splice them to obtain the three-dimensional tooth segmentation result of the CBCT image.

[0012] Furthermore, in step S1, according to the tooth region of interest of the CBCT image, the slice plane and slice direction of the three-dimensional CBCT image are defined, and each layer of slices is obtained and superimposed to generate a two-dimensional oral panoramic film.

[0013] Furthermore, step S1 specifically includes the following steps:

[0014] Step S11: extracting dental arch lines from CBCT three-dimensional images;

[0015] Step S12: interpolating the dental arch line so that the distance between every two adjacent points on the dental arch line is one pixel;

[0016] Step S13: traverse the point set on the dental arch line, and for each point, calculate the direction of the corresponding perpendicular bisector;

[0017] Step S14: starting from the nth point on the dental arch line, sampling is performed in the mth layer of the CBCT image along the direction of the corresponding perpendicular bisector according to a predetermined step length. The multiple pixel values ​​collected in this direction are summed to obtain the pixel value corresponding to the mth row and nth column in the panoramic image. This process traverses all points on the dental arch line and finally forms a complete two-dimensional panoramic image.

[0018] Furthermore, in step S2, a two-dimensional tooth segmentation model is used to process the two-dimensional oral panoramic film. The model performs single tooth segmentation according to the morphological characteristics of the teeth, obtains the tooth contour, and assigns a unique ID to each tooth.

[0019] Furthermore, in step S2, the two-dimensional tooth segmentation model adopts a tooth segmentation model based on a UNet neural network, and the training process of the two-dimensional tooth segmentation model specifically includes the following steps:

[0020] Collect 2D oral panoramic film samples generated by CBCT and annotate single teeth with labels and IDs;

[0021] Perform sample preprocessing and use image enhancement to expand the panoramic film;

[0022] The preprocessed 2D oral panoramic film samples were divided into training set, validation set and test set in a ratio of 7:2:1;

[0023] Construct a UNet-based 2D oral panoramic film tooth segmentation model for training. Specifically, set the model training parameters: batch size, initial learning rate lr 0The model uses the Adam optimizer and combines the first-order moment estimation and second-order moment estimation of the gradient in the parameter update process. The learning rate lr is dynamically adjusted using a polynomial decay strategy, which is defined as follows:

[0024]

[0025] Among them, lr 0 is the initial learning rate, t is the current number of iterations, T max is the total number of iterations, p is the decay exponent;

[0026] When the performance of the model on the validation set reaches stability, stop training and save the best model;

[0027] By inputting the 2D oral panoramic film into the UNet-based tooth segmentation model, the 2D tooth segmentation results and tooth ID values ​​can be obtained.

[0028] Furthermore, in step S3, a key point coordinate generation algorithm is used to extract the two-dimensional key point coordinates of each tooth, and according to the original three-dimensional coordinate system of the CBCT image, the extracted two-dimensional key point coordinates are mapped back to the three-dimensional space through a back projection algorithm to restore the actual position of the key point of each tooth in the three-dimensional space.

[0029] Furthermore, step S3 specifically includes the following steps:

[0030] Step S31: extracting the contour information of each tooth from the two-dimensional oral panoramic film;

[0031] Step S32: Process the contour of each tooth to obtain a bounding box of the contour;

[0032] Step S33: taking the Y value of the center of the bounding box to determine the Y coordinate of the key point, ensuring that the key point is at the longitudinal center of the tooth;

[0033] Step S34: at the determined Y coordinate, traverse all points along the X axis in the tooth contour, filter out the X coordinate values ​​corresponding to these points, then sum and average all the X coordinate values ​​to obtain the X coordinate of the key point of the tooth;

[0034] Step S35: starting from the dental arch point of the X coordinate, searching the tooth area along the perpendicular direction, calculating the center point to obtain the three-dimensional key point coordinates.

[0035] Furthermore, in step S4, the spatial range of the corresponding tooth in the CBCT image is determined using the three-dimensional key point coordinates of each tooth, and then a three-dimensional cropping operation is performed to crop the area of ​​a single tooth from the original CBCT image, ensuring that the cropped area contains the entire structure of the tooth and the tooth is in the center.

[0036] Furthermore, in step S5, a three-dimensional tooth segmentation model is used to extract features from the cropped CBCT image to remove the influence of artifacts and noise, and a tooth segmentation result is obtained according to the three-dimensional morphology of the teeth. The segmentation result is post-processed to ensure the accuracy of the segmentation result.

[0037] Furthermore, in step S5, a three-dimensional tooth segmentation model based on the nnUNet neural network is used, and the training process of the three-dimensional tooth segmentation model specifically includes the following steps:

[0038] Collect CBCT image samples and annotate single tooth labels;

[0039] Get the three-dimensional key point coordinates of a single tooth;

[0040] Crop samples, perform preprocessing, and use data enhancement to expand samples;

[0041] The samples were divided into training set, validation set and test set in a ratio of 7:2:1;

[0042] A 3D tooth segmentation model based on the nnUNet neural network was constructed for training. Specifically, the training set was normalized to the maximum and minimum values ​​and then input into the network according to the batch size. The network used DC_and_BCE_loss as the loss function to compare the gap between the segmentation result and the true label. The stochastic gradient descent (SGD) optimizer was used for parameter optimization and the polynomial learning rate scheduler was used to adjust the learning rate. In each round of training, the network calculated the gradient and updated the weight parameters through back propagation to gradually reduce the loss value and save the weight parameter model with the minimum loss.

[0043] The cropped CBCT tooth area is input into the three-dimensional tooth segmentation model based on the nnUNet neural network. After the single tooth segmentation result is obtained, the maximum connected domain processing is performed to ensure the accuracy of the segmentation result.

[0044] Furthermore, step S6 specifically includes the following steps:

[0045] Step S61: Calculate the position of the tooth region in the original CBCT image according to the key point coordinates of each tooth;

[0046] Step S62: Calculate the actual filling amount of each tooth region in the original image to ensure the alignment of the tooth region in the original image, then extract the valid part of the tooth region and exclude the invalid area caused by the alignment;

[0047] Step S63: assigning pixel values ​​to the tooth area according to the tooth ID;

[0048] Step S64: stitching all tooth regions to obtain a three-dimensional tooth segmentation result.

[0049] Beneficial effects of the present invention:

[0050] The present invention provides a CBCT image tooth segmentation method based on deep learning, which can achieve accurate segmentation of tooth areas in CBCT images, effectively deal with complex situations such as close arrangement of teeth, similar morphology, and artifact interference, ensure the independent segmentation and identification of each tooth, and provide reliable three-dimensional data support for subsequent clinical applications, such as orthodontic treatment planning and dental implant design. In addition, it also promotes communication between doctors and patients, enhances patients' understanding and cooperation of treatment plans, and improves the overall treatment effect. Compared with existing tooth segmentation algorithms, the present invention can effectively solve problems such as tooth segmentation result fusion, ID confusion, and metal artifact interference. By determining the key points of teeth on panoramic films, it significantly reduces resource consumption, improves the overall segmentation efficiency, and shortens the segmentation time, thereby improving the accuracy of the segmentation results and the processing speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The specific implementation of the present invention will be further explained below in conjunction with the accompanying drawings.

[0052] Figure 1 It is a flow chart of the CBCT image tooth segmentation method based on deep learning of the present invention;

[0053] Figure 2 A flow chart of the method for generating a two-dimensional oral panoramic film in the present invention;

[0054] Figure 3 A flow chart of a method for mapping two-dimensional single tooth key point coordinates to three-dimensional CBCT image coordinates in the present invention;

[0055] Figure 4 This is a flow chart of the method for splicing teeth after three-dimensional tooth segmentation in the present invention. DETAILED DESCRIPTION

[0056] Combination Figure 1 In this embodiment, a CBCT image tooth segmentation method based on deep learning specifically includes the following steps:

[0057] Step S1: Generate a two-dimensional oral panoramic film from the CBCT image;

[0058] Step S2: using a deep learning algorithm to perform two-dimensional tooth segmentation and tooth ID classification on the two-dimensional oral panoramic film;

[0059] Step S3: Calculate the key point coordinates of a single tooth in two dimensions and map them to three-dimensional CBCT image coordinates;

[0060] Step S4: cropping the CBCT image to obtain the area of ​​each tooth according to the three-dimensional CBCT image coordinates;

[0061] Step S5: using a deep learning algorithm to segment each tooth region in three dimensions;

[0062] Step S6: assign IDs to all tooth segmentation results and then splice them to obtain the three-dimensional tooth segmentation result of the CBCT image.

[0063] Preferably, in this embodiment, in step S1, the slice plane and slice direction of the three-dimensional CBCT image are defined according to the tooth region of interest of the CBCT image, and each layer of slices is obtained and superimposed to generate a two-dimensional oral panoramic image.

[0064] In this embodiment, preferably, Figure 2 , step S1 specifically includes the following steps:

[0065] Step S11: extracting a dental arch line from the CBCT three-dimensional image, where the dental arch line represents the arrangement trajectory of teeth in the oral cavity;

[0066] Step S12: interpolating the dental arch line so that the distance between every two adjacent points on the dental arch line is one pixel;

[0067] Step S13: traverse the point set on the dental arch line, and for each point, calculate the corresponding perpendicular midline direction; the perpendicular midline direction is the direction perpendicular to the dental arch line from the dental arch point outward, which represents the orthogonal direction along the tooth arrangement direction in the CBCT image.

[0068] Step S14: starting from the nth point on the dental arch line, sampling is performed in the mth layer of the CBCT image along the direction of the corresponding perpendicular bisector according to a predetermined step length. The multiple pixel values ​​collected in this direction are summed to obtain the pixel value corresponding to the mth row and nth column in the panoramic image. This process traverses all points on the dental arch line and finally forms a complete two-dimensional panoramic image.

[0069] Preferably, in step S2 of this embodiment, a two-dimensional tooth segmentation model is used to process the two-dimensional oral panoramic film, and the model performs single tooth segmentation according to the morphological characteristics of the teeth to obtain tooth contours, and assigns a unique ID to each tooth.

[0070] In this embodiment, preferably, in step S2, the two-dimensional tooth segmentation model adopts a tooth segmentation model based on a UNet neural network, and the training process of the two-dimensional tooth segmentation model specifically includes the following steps:

[0071] Collect 2D oral panoramic film samples generated by CBCT and annotate single teeth with labels and IDs;

[0072] Specifically, in this embodiment, 1000 two-dimensional oral panoramic films were collected, and two dental experts performed tooth contour and ID annotation.

[0073] Perform sample preprocessing and use image enhancement to expand the panoramic film;

[0074] Specifically, in this embodiment, in order to avoid overfitting of the model, data enhancement technology is used to expand the sample, and the enhancement method includes random cropping of the top, bottom, left, and right, with a cropping amount of 0% to 15%. Horizontal flipping and random rotation are performed, and the rotation angle is -5° to 5°. Considering the limitations of computing efficiency and hardware performance, the size of the panoramic film is uniformly adjusted to 640×256, and normalized by using a mean of 0 and a variance of 1.

[0075] The preprocessed 2D oral panoramic film samples were divided into training set, validation set and test set in a ratio of 7:2:1;

[0076] Specifically, in this embodiment, the preprocessed two-dimensional oral panoramic film samples are divided into a training set of 700 images, a validation set of 200 images, and a test set of 100 images.

[0077] Construct a UNet-based 2D oral panoramic film tooth segmentation model for training. Specifically, set the model training parameters: batch size, initial learning rate lr 0 The model uses the Adam optimizer and combines the first-order moment estimation and second-order moment estimation of the gradient in the parameter update process. The learning rate lr is dynamically adjusted using a polynomial decay strategy, which is defined as follows:

[0078]

[0079] Among them, lr 0 is the initial learning rate, t is the current number of iterations, T max is the total number of iterations, p is the decay exponent;

[0080] Specifically, in this embodiment, 12 images are processed in each batch (Batch size is 12), and the initial learning rate lr 0 It is set to 0.001, the number of training rounds is set to 1500, and the decay exponent p is usually set to 2.

[0081] The UNet model consists of two parts: an encoder and a decoder. The encoder gradually extracts high-level features of the panoramic image, while the decoder gradually restores these features to outputs of the same size as the input to generate pixel-level segmentation results. The core structure of UNet is as follows:

[0082] f out =g(d(e(f in )));

[0083] Among them, f in is the normalized panoramic image of the input, e(·) is the encoder, d(·) is the decoder, and g(·) is the activation function of the final output. The Softmax function is used here to ensure the probability that each pixel of the output belongs to a specific category. For each convolutional layer and downsampling layer within the model, ReLU is usually used as the activation function to introduce nonlinearity and accelerate convergence. It is defined as:

[0084] ReLU(x)=max(0,x)

[0085] Applying the ReLU activation function after each convolutional layer can effectively prevent the gradient vanishing problem and facilitate feature extraction.

[0086] When the performance of the model on the validation set reaches stability, stop training and save the best model;

[0087] By inputting the 2D oral panoramic film into the UNet-based tooth segmentation model, the 2D tooth segmentation results and tooth ID values ​​can be obtained.

[0088] Preferably, in step S3 of the present embodiment, a key point coordinate generation algorithm is used to extract the key point coordinates of each tooth in two dimensions, and according to the original three-dimensional coordinate system of the CBCT image, the extracted two-dimensional key point coordinates are mapped back to the three-dimensional space through a back projection algorithm to restore the actual position of the key point of each tooth in the three-dimensional space.

[0089] In this embodiment, preferably, Figure 3 , step S3 specifically includes the following steps:

[0090] Step S31: extracting the contour information of each tooth from the two-dimensional oral panoramic film; these contours are composed of the boundary pixel points of the teeth and are the basic data for key point calculation.

[0091] Step S32: Process the contour of each tooth to obtain a bounding box of the contour; the bounding box is used to define the position and size of the tooth contour in the image.

[0092] Step S33: taking the Y value of the center of the bounding box to determine the Y coordinate of the key point, ensuring that the key point is at the longitudinal center of the tooth;

[0093] Step S34: At the determined Y coordinate, traverse all points along the X axis in the tooth contour, filter out the X coordinate values ​​corresponding to these points, then sum and average all the X coordinate values ​​to obtain the X coordinate of the key point of the tooth; this coordinate is the horizontal center position of the tooth in the panoramic film.

[0094] Step S35: After obtaining the coordinates of the two-dimensional key points, considering the resolution of the CBCT image and the correspondence with the two-dimensional oral panoramic film, the dental arch point of the X coordinate is used as the starting point, the tooth area is searched along the midpoint direction, and the center point is calculated to obtain the coordinates of the three-dimensional key points.

[0095] Preferably, in step S4 of the present embodiment, a three-dimensional cropping operation is performed after the spatial range of the corresponding tooth in the CBCT image is determined using the three-dimensional key point coordinates of each tooth, and the region of a single tooth is cropped from the original CBCT image to ensure that the cropped region contains the entire structure of the tooth and the tooth is in the center position.

[0096] Specifically, in this embodiment, step S4 includes the following steps:

[0097] Step S41: Calculate the starting coordinates of the cropping area according to the patch size (96,96,96) and the three-dimensional key points.

[0098] Step S42: Determine whether the cropping area exceeds the image boundary. If the starting coordinate exceeds the size range of the image, calculate the number of pixels that need to be filled in the top, bottom, left, right, front, and back after cropping.

[0099] Step S43: cropping the CBCT image, and filling the part beyond the image according to the -1000Hu value to obtain a cropped single tooth area.

[0100] Preferably, in step S5 of the present embodiment, a three-dimensional tooth segmentation model is used to extract features from the cropped CBCT image to remove the influence of artifacts and noise, and a tooth segmentation result is obtained according to the three-dimensional morphology of the teeth. The segmentation result is post-processed to ensure the accuracy of the segmentation result.

[0101] In this embodiment, preferably, in step S5, a three-dimensional tooth segmentation model based on the nnUNet neural network is used, and the training process of the three-dimensional tooth segmentation model specifically includes the following steps:

[0102] Collect CBCT image samples and annotate single tooth labels;

[0103] Specifically, in this embodiment, 300 CBCT images were collected, covering different tooth conditions, including implants, metal artifacts, etc., to ensure sample diversity. Two dental experts used annotation tools to accurately mark each tooth to ensure that the contour and position of each tooth were correctly identified and marked.

[0104] Get the three-dimensional key point coordinates of a single tooth;

[0105] Specifically, in this embodiment, all pixel coordinates and pixel values ​​of each single tooth label are extracted, and teeth with too few pixels are removed to ensure the validity of key points. The mean of the pixel coordinates is calculated to obtain the key point coordinates, and the pixel value of the coordinate is the tooth ID.

[0106] Crop samples, perform preprocessing, and use data enhancement to expand samples;

[0107] Specifically, in this embodiment, the coordinates of the key points of the teeth are randomly adjusted in the range of -4 to 4 pixels, which helps to improve the generalization ability of the model. According to the adjusted coordinates, the resolution of the CBCT image and the label image is downsampled to a spacing of 0.4 to eliminate the too dark and too bright values, and then cropped to a size of 96×96×96 pixels. If the cropped area exceeds the image boundary, it is filled. The value of the filled area of ​​the CBCT image is set to -1000Hu, and the value of the filled area of ​​the label image is set to 0 to prevent invalid information outside the cropped area from affecting the model training.

[0108] For the cropped label image, only the label value pixels corresponding to the key points are retained and set to 1, and the remaining label pixel values ​​are set to 0 to highlight the label information of the current tooth.

[0109] Due to the limited number of data samples, data enhancement technology is used to expand the samples. The enhancement methods include random rotation with a rotation angle of -15° to 15°; adding Gaussian noise to simulate the noise in the acquisition process; applying texture filters to enhance the texture features of the image, etc. These data enhancement operations help improve the robustness of the model.

[0110] The samples were divided into training set, validation set and test set in a ratio of 7:2:1;

[0111] Specifically, in this embodiment, the samples are divided into 210 training sets, 60 validation sets and 30 test sets.

[0112] A 3D tooth segmentation model based on the nnUNet neural network was constructed for training. Specifically, the training set was normalized to the maximum and minimum values ​​and then input into the network according to the batch size. The network used DC_and_BCE_loss as the loss function to compare the gap between the segmentation result and the true label. The stochastic gradient descent (SGD) optimizer was used for parameter optimization and the polynomial learning rate scheduler was used to adjust the learning rate. In each round of training, the network calculated the gradient and updated the weight parameters through back propagation to gradually reduce the loss value and save the weight parameter model with the minimum loss.

[0113] Specifically, in this embodiment, the batch size is 18 and the network is input, and the number of training rounds is set to 1500.

[0114] Among them, the nnUNet model inherits the encoder-decoder structure of UNet. The encoder part downsamples layer by layer to extract features, and the decoder part upsamples layer by layer to restore spatial resolution. This structure helps the network effectively capture local and global information of the image. And use jump connections between the encoder and decoder to pass high-resolution features to the decoder part, thereby improving the accuracy of segmentation and detail retention.

[0115] Specifically, in this embodiment, DC_and_BCE_loss is a loss function that combines Dice loss and cross-entropy loss, which can simultaneously optimize the pixel-level classification accuracy and overall structural matching of the segmentation result.

[0116] Dice loss is an indicator to measure the similarity between two samples. It can handle the problem of class imbalance and is defined as follows:

[0117]

[0118] Among them, A is the predicted segmentation area, B is the real labeled area, |A∩B| represents the overlapping part of the predicted and real areas, and |A| and |B| are the areas of the predicted and real areas respectively.

[0119] Cross entropy loss is used to measure the difference between the predicted probability distribution and the true label. For each pixel, the cross entropy loss calculates the negative log-likelihood of the predicted probability distribution and the true label, and the formula is:

[0120]

[0121] Among them, y is the true value, is the predicted value.

[0122] DC_and_BCE_loss combines the above two loss functions to simultaneously optimize the pixel-level classification accuracy and overall structural matching of the segmentation results. The definition is as follows:

[0123] DC_and_BCE_loss=α·Dice Loss+β·Cross_Entropy Loss;

[0124] Among them, α and β are balance coefficients, which are used to adjust the weights of the two losses in the final loss.

[0125] The cropped CBCT tooth area is input into the three-dimensional tooth segmentation model based on the nnUNet neural network. After the single tooth segmentation result is obtained, the maximum connected domain processing is performed to ensure the accuracy of the segmentation result.

[0126] Specifically, in step S6, each tooth in all the tooth segmentation results is assigned a unique ID so as to distinguish each tooth during the splicing process. All the tooth segmentation results are spliced ​​back into the three-dimensional CBCT coordinate system according to their original positions in the CBCT image to ensure that each tooth is accurately positioned in the three-dimensional space. The overall three-dimensional tooth segmentation results after splicing are optimized and corrected to eliminate possible splicing gaps or overlapping areas.

[0127] In this embodiment, preferably, Figure 4 , step S6 specifically includes the following steps:

[0128] Step S61: Calculate the position of the tooth region in the original CBCT image according to the key point coordinates of each tooth;

[0129] Step S62: Calculate the actual filling amount of each tooth region in the original image to ensure the alignment of the tooth region in the original image, then extract the valid part of the tooth region and exclude the invalid area caused by the alignment;

[0130] Step S63: assigning pixel values ​​to the tooth area according to the tooth ID;

[0131] Step S64: stitching all tooth regions to obtain a three-dimensional tooth segmentation result.

[0132] Many specific details are described in the above description to facilitate a full understanding of the present invention. However, the above description is only a preferred embodiment of the present invention. The present invention can be implemented in many other ways different from those described herein, so the present invention is not limited to the specific implementation disclosed above. At the same time, any person familiar with the art can make many possible changes and modifications to the technical solution of the present invention using the methods and technical contents disclosed above without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A CBCT image tooth segmentation method based on deep learning, characterized by: The steps include: Step S1: Generate a two-dimensional oral panoramic film from the CBCT image; Step S2: using a deep learning algorithm to perform two-dimensional tooth segmentation and tooth ID classification on the two-dimensional oral panoramic film; Step S3: Calculate the key point coordinates of a single tooth in two dimensions and map them to three-dimensional CBCT image coordinates; Step S4: cropping the CBCT image to obtain the area of ​​each tooth according to the three-dimensional CBCT image coordinates; Step S5: using a deep learning algorithm to segment each tooth region in three dimensions; Step S6: assign IDs to all tooth segmentation results and then splice them to obtain the three-dimensional tooth segmentation result of the CBCT image.

2. The CBCT image tooth segmentation method based on deep learning according to claim 1, characterized in that: In step S1, according to the tooth region of interest of the CBCT image, the slice plane and slice direction of the three-dimensional CBCT image are defined, and each layer of slices is obtained and superimposed to generate a two-dimensional oral panoramic film.

3. The CBCT image tooth segmentation method based on deep learning according to claim 2, characterized in that: Step S1 specifically includes the following steps: Step S11: extracting dental arch lines from CBCT three-dimensional images; Step S12: interpolating the dental arch line so that the distance between every two adjacent points on the dental arch line is one pixel; Step S13: traverse the point set on the dental arch line, and for each point, calculate the direction of the corresponding perpendicular bisector; Step S14: starting from the nth point on the dental arch line, sampling is performed in the mth layer of the CBCT image along the direction of the corresponding perpendicular bisector according to a predetermined step length. The multiple pixel values ​​collected in this direction are summed to obtain the pixel value corresponding to the mth row and nth column in the panoramic image. This process traverses all points on the dental arch line and finally forms a complete two-dimensional panoramic image.

4. The CBCT image tooth segmentation method based on deep learning according to claim 1, characterized in that: In step S2, a two-dimensional tooth segmentation model is used to process the two-dimensional oral panoramic film. The model performs single tooth segmentation according to the morphological characteristics of the teeth, obtains the tooth contour, and assigns a unique ID to each tooth.

5. The method for tooth segmentation based on CBCT images according to claim 4, characterized in that: In step S2, the two-dimensional tooth segmentation model adopts a tooth segmentation model based on a UNet neural network. The training process of the two-dimensional tooth segmentation model specifically includes the following steps: Collect 2D oral panoramic film samples generated by CBCT and annotate single teeth with labels and IDs; Perform sample preprocessing and use image enhancement to expand the panoramic film; The preprocessed 2D oral panoramic film samples were divided into training set, validation set and test set in a ratio of 7:2:1; A UNet-based 2D oral panoramic film tooth segmentation model was constructed for training. Specifically, the parameters of model training were set: the number of images processed in each batch, the initial learning rate lr0, and the number of training rounds. The model used the Adam optimizer, and combined the first-order moment estimation and second-order moment estimation of the gradient in the parameter update process. The learning rate lr was dynamically adjusted using a polynomial attenuation strategy, which is defined as follows: Among them, lr0 is the initial learning rate, t is the current number of iterations, and T max is the total number of iterations, p is the decay exponent; When the performance of the model on the validation set reaches stability, stop training and save the best model; By inputting the 2D oral panoramic film into the UNet-based tooth segmentation model, the 2D tooth segmentation results and tooth ID values ​​can be obtained.

6. The CBCT image tooth segmentation method based on deep learning according to claim 1, characterized in that: In step S3, a key point coordinate generation algorithm is used to extract the two-dimensional key point coordinates of each tooth. According to the original three-dimensional coordinate system of the CBCT image, the extracted two-dimensional key point coordinates are mapped back to the three-dimensional space through a back projection algorithm to restore the actual position of the key point of each tooth in the three-dimensional space.

7. The method for tooth segmentation based on CBCT images according to claim 6, characterized in that: Step S3 specifically includes the following steps: Step S31: extracting the contour information of each tooth from the two-dimensional oral panoramic film; Step S32: Process the contour of each tooth to obtain a bounding box of the contour; Step S33: taking the Y value of the center of the bounding box to determine the Y coordinate of the key point, ensuring that the key point is at the longitudinal center of the tooth; Step S34: at the determined Y coordinate, traverse all points along the X axis in the tooth contour, filter out the X coordinate values ​​corresponding to these points, then sum and average all the X coordinate values ​​to obtain the X coordinate of the key point of the tooth; Step S35: starting from the dental arch point of the X coordinate, searching the tooth area along the perpendicular direction, calculating the center point to obtain the three-dimensional key point coordinates.

8. The CBCT image tooth segmentation method based on deep learning according to claim 1, characterized in that: In step S4, the spatial range of the corresponding tooth in the CBCT image is determined using the three-dimensional key point coordinates of each tooth, and then a three-dimensional cropping operation is performed to crop the area of ​​a single tooth from the original CBCT image, ensuring that the cropped area contains the entire structure of the tooth and the tooth is in the center.

9. The CBCT image tooth segmentation method based on deep learning according to claim 1, characterized in that: In step S5, a three-dimensional tooth segmentation model is used to extract features from the cropped CBCT image to remove the influence of artifacts and noise, and a tooth segmentation result is obtained according to the three-dimensional morphology of the teeth. The segmentation result is post-processed to ensure the accuracy of the segmentation result.

10. The method for tooth segmentation based on CBCT images according to claim 9, characterized in that: In step S5, a three-dimensional tooth segmentation model based on the nnUNet neural network is used. The training process of the three-dimensional tooth segmentation model specifically includes the following steps: Collect CBCT image samples and annotate single tooth labels; Get the three-dimensional key point coordinates of a single tooth; Crop samples, perform preprocessing, and use data enhancement to expand samples; The samples were divided into training set, validation set and test set in a ratio of 7:2:1; A 3D tooth segmentation model based on the nnUNet neural network was constructed for training. Specifically, the training set was normalized to the maximum and minimum values ​​and then input into the network according to the batch size. The network used DC_and_BCE_loss as the loss function to compare the gap between the segmentation result and the true label. The stochastic gradient descent optimizer was used for parameter optimization and the polynomial learning rate scheduler was used to adjust the learning rate. In each round of training, the network calculated the gradient and updated the weight parameters through back propagation to gradually reduce the loss value and save the weight parameter model with the minimum loss. The cropped CBCT tooth area is input into the three-dimensional tooth segmentation model based on the nnUNet neural network. After the single tooth segmentation result is obtained, the maximum connected domain processing is performed to ensure the accuracy of the segmentation result.

11. The CBCT image tooth segmentation method based on deep learning according to claim 1, characterized in that: Step S6 specifically includes the following steps: Step S61: Calculate the position of the tooth region in the original CBCT image according to the key point coordinates of each tooth; Step S62: Calculate the actual filling amount of each tooth region in the original image to ensure the alignment of the tooth region in the original image, then extract the valid part of the tooth region and exclude the invalid area caused by the alignment; Step S63: assigning pixel values ​​to the tooth area according to the tooth ID; Step S64: stitching all tooth regions to obtain a three-dimensional tooth segmentation result.

Images