Tooth generation method for missing tooth CBCT images based on missing tooth detection and probability diffusion model

Abstract

The present invention discloses a method for generating teeth in edentulous CBCT images based on edentulous detection and a probability diffusion model. The present invention uses an artificial intelligence algorithm to automatically identify edentulous areas in input CBCT images to ensure accurate positioning of the edentulous areas. The identified edentulous areas are then cropped, and the tooth images are generated using an improved probability diffusion model, which can generate high-resolution, morphologically accurate tooth images and effectively complete the edentulous areas. Finally, the generated tooth images are accurately transplanted back to the original images according to the original cropping coordinates to form a complete CBCT image. The present invention is suitable for surgical planning of implant placement, and can also serve as a reference for denture modeling to help patients better understand the expected preoperative effects. The present invention realizes automatic detection of edentulous areas in CBCT images and rapid generation of tooth images through the collaborative work of edentulous detection and a probability diffusion model, greatly reducing human intervention and manual operations, and improving work efficiency and accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of medical image detection and generation, and specifically relates to a method for generating teeth from missing tooth CBCT images based on missing tooth detection and a probability diffusion model. This method realizes the entire process from missing tooth area detection to missing tooth image generation through an automated process, significantly improving the efficiency and accuracy of dental diagnosis and treatment, and has important clinical significance and practical application value. Background Art

[0002] Missing teeth refers to the situation where one or more teeth are lost or missing from the mouth due to various reasons. This situation may affect a person's chewing function, language ability and facial aesthetics, and may also cause oral health problems. Common causes of missing teeth include caries (cavities), periodontal disease, trauma, etc. Dental implant technology is considered to be the best solution in the field of modern dentistry to replace missing teeth and restore oral function and aesthetics. This technology works by implanting implants (usually made of titanium or titanium alloy) into the alveolar bone of the edentulous area as artificial roots, and then installing artificial crowns on them, thereby restoring the function and aesthetics of the teeth. The implantation of implants requires doctors to evaluate the patient's oral health and alveolar bone quality through clinical examinations and imaging, and to develop a detailed implant plan. Therefore, designing a method that can quickly find the location of missing teeth and generate information about healthy tooth crowns will effectively guide the implantation of implants and assist doctors in treatment.

[0003] As an advanced three-dimensional imaging technology, cone-beam computed tomography (CBCT) has become an indispensable tool in modern dentistry and related medical fields. It acquires three-dimensional image data of the entire scanned area in a single scan. Its high resolution, low radiation dose, and wide clinical application make it a vital tool in dental diagnosis and treatment. It can provide detailed three-dimensional images that more accurately display anatomical structures and pathological conditions than traditional two-dimensional X-rays. Simultaneously, the application of artificial intelligence in dental implant procedures is increasing. Direct analysis of three-dimensional data will yield more in-depth information, but storing and processing this data requires high-performance computer equipment and specialized software, which increases operational complexity and cost. Therefore, it is necessary to improve existing image detection and generation models, identify oral information through radiographic images, predict the generation of three-dimensional teeth, and optimize the aesthetic design of teeth.

[0004] Surgical planning of implant placement is crucial to avoid damage to important anatomical structures. However, manually measuring edentulous bone on cone beam computed tomography images is very time-consuming and prone to human error. Automated processes have the potential to reduce human error, saving time and costs. Therefore, the first phase of the present invention uses an artificial intelligence solution to automatically detect the location of missing teeth in CBCT images. Since there is no obvious tooth structure in the edentulous area, only alveolar bone or possible gingival tissue, these areas have fewer features than intact teeth and are more difficult to identify using traditional image detection methods and deep learning models. It is very easy for humans to find the location of missing teeth, but it is a difficult point for artificial intelligence.

[0005] Denture production is a relatively complex process. Traditional methods require the creation of impressions and models. If missing teeth could be quickly generated from edentulous CBCT images, it would improve implant surgery planning, accelerate the denture production process, facilitate personalized treatment plans, and enhance patient communication and compliance. Therefore, the second phase of the present invention proposes the use of a probability diffusion model as a tooth generation model to generate dentures. The probability diffusion model is capable of generating high-quality, realistic data samples. Compared to some other generation models, the probability diffusion model has the ability to generate diverse data. By introducing noise through an iterative process, the generated sample images achieve a certain degree of diversity. It also has a certain degree of stability during training, making it easier to train. It also has the ability to handle incomplete data, such as missing image areas, which is very important in the restoration and repair of missing teeth. Currently, CBCT image generation still has limitations such as low contrast resolution, noise, and artifacts. Therefore, the present invention improves the generation effect of the probability diffusion model by improving the network architecture, enhancing data preprocessing, and refining the loss function method. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this paper proposes a tooth generation method for edentulous CBCT images based on tooth loss detection and a probability diffusion model. This method, primarily based on a deep learning artificial intelligence system, comprises a tooth loss detection network for detecting edentulous areas and a tooth generation model for generating three-dimensional CBCT image data. This process enables efficient, accurate, and fully automatic segmentation and generation of CBCT images from real patients. It achieves a Dice (a measure of the spatial overlap between segmentation results and physician annotations) of over 90% for tooth loss detection, while the generated tooth generation model accurately reflects the shape, size, and orientation of normal teeth. The reconstruction error (MSE) between the generated image and the original image is maintained between 0.0003 and 0.0007, and the peak signal-to-noise ratio (PSNR) is maintained between 30 and 40, making the difference between the reconstructed and original images imperceptible to the human eye.

[0007] The technical solutions adopted to solve the technical problems are as follows:

[0008] The tooth generation method for missing tooth CBCT images based on missing tooth detection and probability diffusion model has the following steps:

[0009] Step 1: Acquire and process a dataset: Two different training networks are used, one trained on missing tooth detection data and the other trained on single tooth instance data. The dataset is derived from 115 cone-beam computed tomography (CBCT) images from the Department of Stomatology at Zhejiang Provincial People's Hospital. The dataset includes annotation of missing tooth regions and tooth instance segmentation.

[0010] Step 2: Preprocess the processed dataset. Based on the balance between computational efficiency and segmentation accuracy, the physical resolution of the CBCT images in the collected dataset is normalized to an isotropic resolution of 0.4 × 0.4 × 0.4 mm3. In addition, to reduce the influence of extreme values, especially in the metal artifact area, the intensity values ​​of each CBCT image scan are clipped to [0, 3096] before intensity normalization.

[0011] Then, each tooth instance in the single tooth instance data is segmented, and the segmented area is cropped with a size of 64×64×64 pixels around its center as the anchor point to obtain the segmented tooth instance dataset.

[0012] Step 3: Define a missing tooth detection network. In this invention, a Unet convolutional neural network is used as the first-stage missing tooth detection network and trained using supervised learning. Of the 115 cases with annotated missing areas obtained in Step 2, 85 are used as the training dataset, 10 as the validation dataset, and the remaining 20 as the test dataset to test the model's performance.

[0013] Step 4: Define a missing tooth generation network. In this invention, the second stage of the missing tooth generation network uses self-supervised learning to train by automatically generating damaged and repaired pairs. The missing tooth area of ​​the input missing tooth image of the probability diffusion model is masked. The probability diffusion model controls the generation and removal of noise based on the surrounding unmasked information, repairs the occluded portion of the missing tooth area, generates a 3D image, and predicts the shape, size, and orientation of the tooth at the masked location.

[0014] Step 5: Crop the missing tooth detection area. The CBCT image, which has been cropped by intensity value in step 2, is center-cropped according to the missing tooth detection result in step 3. This removes information in the CBCT image that is irrelevant to the target teeth. This aims to improve training efficiency, reduce the risk of overfitting, and improve the robustness of the model.

[0015] Step 6: Input the cropped data from Step 5 into the missing tooth generation network. The probabilistic diffusion model for generating teeth gradually adds noise to the 3D mask data through a forward diffusion process. Then, through a reverse diffusion process, it learns information about the surrounding oral cavity and restores the original image from the noisy image, thus completing the mask image restoration task. By controlling the noise generation and removal processes, the mask information can be used to restore the occluded portions of the missing tooth image, generating a 3D image that reflects the healthy tooth features learned during model training.

[0016] Step 7: Blend the original image and fill the newly generated tooth image of the missing tooth area into the cropped area of ​​the original image in step 5. By comparing the generated tooth image with the surrounding important anatomical structure information, the implant surgery planning and denture production are designed.

[0017] Furthermore, the single tooth instance data described in step 1 is obtained by preprocessing the original image data in step 2, then cropping the segmented region of each tooth instance to a size of 64×64×64 pixels around its center as an anchor point. The cropped image contains not only the central tooth but also the surrounding oral information. Through this method, the present invention obtained data samples of 3552 healthy teeth, covering all tooth categories including incisors, canines, premolars, and molars; 2623 of these samples were used as training datasets for the tooth generation model, and 929 were used as test datasets to test the model's performance.

[0018] The oral information includes important anatomical structure information and missing tooth information.

[0019] The tooth-missing feature extraction described in step 3 is a key component of image processing in traditional computer vision research. Feature extraction requires extracting meaningful features from the original image and converting the image information into a compact, information-rich representation for further analysis, recognition, classification, or other tasks. However, the difficulty of tooth-missing detection lies in the lack of clear feature information. In the patient's CBCT image representation, the tooth-missing region lacks obvious tooth structure, consisting only of alveolar bone or possible gingival tissue and surrounding air. Finding feature information for this region is challenging. Therefore, the present invention employs a segmentation model trained using supervised learning to segment the tooth-missing region in CBCT images. Unlike traditional artificial intelligence object detection models that identify objects present in an image, this model can identify missing objects. The primary implementation method is through a convolutional neural network (Unet), trained from scratch using collected original images and labeled data from the tooth-missing region in a supervised learning mode. Unet is an encoder-decoder network architecture, where the encoder captures feature information in the image and the decoder converts the feature map into pixel-level segmentation results. This structure facilitates simultaneous local feature extraction and global information recovery. It also introduces skip connections, connecting the encoder's feature map with the decoder's feature map to preserve and transfer multi-scale information, which helps improve the accuracy of missing tooth detection and segmentation. Based on this, the present invention targets missing tooth areas for segmentation. The segmentation network loss function uses a combination of binary cross entropy loss and Dice loss to address class imbalance and improve the accuracy of segmentation results.

[0020] The binary cross entropy loss is defined as follows:

[0021]

[0022] The Dice coefficient loss function is defined as follows:

[0023]

[0024] Where N is the number of samples, y i is the true label of the i-th sample, p i is the probability that the model's prediction is true. The final loss function is designed as a weighted sum of the binary cross entropy loss function and the Dice coefficient loss function. The smaller the loss function, the higher the overlap between the model's prediction and the true label, and the better the model's segmentation accuracy.

[0025] The tooth generation model described in step 4 uses a probabilistic diffusion generation model. In the diffusion model, the process of generating data is modeled as a process that gradually diffuses over time: at the beginning, the input data is a noisy or low-quality version. By gradually "diffusing" this data, the model gradually generates samples of higher quality that are closer to the real data. The diffusion process can be defined as:

[0026]

[0027] where x t and x t-1 Represent the data at time step t and t-1 respectively. is a conditional probability distribution where is its mean, β t I represents the covariance matrix, and I is the identity matrix.

[0028] The opposite of generating data is the reverse process, which starts with high-quality data and gradually evolves backward to a noisy or low-quality version. The reverse process is a parameterized Gaussian distribution:

[0029]

[0030] in and The mean and covariance matrices are controlled by the model parameters θ, which are parameterized functions of the model and learned through training. The loss function of the model is calculated as follows:

[0031]

[0032] where q(x t |x0) is the data distribution generated at time step t when the initial data x0 is given in the forward process, Given x0, the random variable x t About the expected value of the distribution q. In the reverse process, the model generates t The probability distribution to x0. is the predicted mean of the model at time step t, with parameters θ. σ t is the standard deviation of the noise at time step t. The purpose of this loss function is to minimize the mean of the model predictions and the actual reverse process distribution The error between them is reduced, so that the training model can more accurately recover the original data from the noise. By minimizing this loss function, the model learns how to start from any intermediate state and gradually denoise to restore the real data. The diffusion generation model uses neural networks to control the diffusion and reverse processes, and trains the parameters of these neural networks to generate samples that match the actual data distribution. The present invention uses it to achieve free-form image restoration, that is, to generate image data of any missing part through the model. First, a single tooth instance image is trained in the three-dimensional tooth generation network to let the network learn how to unconditionally generate teeth, and use this as the basic network. Then, in the toothless CBCT image, the damaged and missing tooth areas that need to be repaired are masked (that is, the first stage of toothless detection). The basic generation model samples the unmasked surrounding teeth and oral information to change the reverse iteration of the model, that is, it can generate denture images that meet oral aesthetic evaluation in the original CBCT image in the toothless area it masks. The model uses an improved three-dimensional U-Net model structure. The entire training process is actually training the parameters of the U-Net network. The training process of the model is as follows:

[0033] Extract a sample from the data set, randomly select a time t from 1-T, pass x0 and t to GaussionDiffusion, GaussionDiffusion samples a random noise, and adds it to x0 to form x t , then x t and t are put into U-net, U-net generates sinusoidal position encoding and x according to t t Combined, U-net predicts the added noise and returns the noise. GaussionDiffusion calculates the loss of this noise and random noise. The noise predicted by the neural network U-net is combined with the random noise sampled by GaussionDiffusion to calculate the loss, calculate the gradient, and update the weights. The loss function is calculated as follows:

[0034]

[0035] Where x0 represents the initial tooth instance data, ∈ represents the noise term adopted from the normal distribution. The goal of this loss function is to minimize the difference between the noise ∈ sampled from the standard normal distribution and the noise ∈ generated by the model. θ (x t ,t) to train the probability diffusion generation model. The model uncertainty under different initial tooth instance data x0 and noise ∈ is considered, making the loss function more comprehensive and taking into account the situation of generating data under different conditions.

[0036] Repeat the above steps until the probability diffusion model is trained. Advantages of this method include generating restorations that more closely resemble the distribution of real tooth data and can handle different restoration tasks, such as hole filling and object removal. It also maintains image consistency and smoothness to a certain extent, allowing the restored area to blend naturally with the surrounding image.

[0037] The mixed original image described in step 7 realizes the entire process from missing tooth area detection to missing tooth image generation through an automated process. The collaborative work of missing tooth detection and the probability diffusion model realizes automatic detection of missing tooth areas in CBCT images and rapid generation of tooth images, greatly reducing human intervention and manual operations, and improving work efficiency and accuracy.

[0038] The beneficial effects of the present invention are as follows:

[0039] Based on the work of predecessors, this invention collects dental data sets and optimizes and improves the neural network model, allowing the model to automatically detect the location of missing teeth and automatically generate dentures in the area of ​​missing teeth. The predicted size of the missing tooth location and the true label have a Dice of more than 90%, and the reconstruction error (MSE) of the generated teeth in the structural morphology and size direction compared with the original image can be between 0.0003 and 0.0007. The peak signal-to-noise ratio (PSNR) can be maintained between 30 and 40, and the difference between the reconstructed image and the original image is difficult for the human eye to detect. This proves that this invention can achieve better detection of missing tooth areas and generation of missing tooth images, and can be used to assist in implant surgery planning and crown design, and better provide diagnosis and treatment plans for dentists. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of a specific process of the tooth generation method of missing tooth CBCT images based on missing tooth detection and probability diffusion model of the present invention.

[0041] Figure 2 This is an interface display of the tooth generation method of missing tooth CBCT images based on missing tooth detection and probability diffusion model.

[0042] Figure 3 This is the network structure of the missing teeth detection model in the first stage of the present invention.

[0043] Figure 4 This is the probability diffusion model structure of the tooth generation in the second stage of the present invention.

[0044] Figure 5 It is an effect diagram of missing tooth detection prediction of the present invention.

[0045] Figure 6 This is a three-dimensional CBCT image generation effect diagram of the present invention.

[0046] Figure 7 This is a schematic diagram of the process of implant placement based on generated teeth. DETAILED DESCRIPTION

[0047] The present invention will be further described in detail below with reference to the accompanying drawings.

[0048] The present invention provides a method for generating teeth from CBCT images based on missing tooth detection and a probability diffusion model. This method first uses an artificial intelligence algorithm to automatically identify missing tooth areas in the input CBCT image, ensuring precise localization of these edentulous areas. Subsequently, the identified missing tooth areas are cropped, and a modified probability diffusion model is applied to generate tooth images. This model generates high-resolution, morphologically accurate tooth images, effectively completing the missing tooth areas. Finally, the generated tooth images are precisely transplanted back to the original image based on the original cropped coordinates, forming a complete CBCT image. This method is not only suitable for surgical planning of implant placement but can also serve as a reference for denture modeling and help patients better understand their expected preoperative outcomes. By integrating missing tooth detection with a probability diffusion model, the present invention achieves automatic detection of missing tooth areas in CBCT images and rapid generation of tooth images, significantly reducing human intervention and manual operations, thereby improving efficiency and accuracy. Furthermore, the system features standardized processes, effectively reducing human error and enhancing treatment safety. Furthermore, implementation of the present invention shortens patient wait times and improves the patient experience.

[0049] like Figure 1 As shown, the present invention provides a method for generating teeth from missing tooth CBCT images based on missing tooth detection and probability diffusion model. The specific method interface is shown in FIG. Figure 2 As shown, it mainly consists of two stages. The two-stage training model structure can be referred to Figure 3 and Figure 4 The specific steps are as follows:

[0050] Step 1: Dataset acquisition. First, the missing tooth areas in each image are recorded using medical annotation software to construct a new missing tooth CBCT image dataset for the first stage of missing tooth detection task.

[0051] Step 2: Preprocess the input edentulous CBCT image data and annotation information to obtain a resolution of 0.4×0.4×0.4mm. 3 , a collection of 3D toothless data blocks of uniform size of 200×200×200 pixels. Next, each 3D toothless data block is cropped according to its tooth segmentation instance results to construct a dataset of healthy tooth instance images of the same resolution and uniform size of 64×64×64 pixels for the second stage of tooth generation task.

[0052] Step 3: Input the three-dimensional missing tooth data block obtained in step 2 and its corresponding annotation information as data into the missing tooth detection network model for supervised training. The model used this time is the standard Unet segmentation model. Because the CBCT image data is a single-channel grayscale image, the number of input image channels of the network is 1, and the number of output channels is also 1. The input image size of the model is 192*192*192 pixels, the feature map size is 32, the hidden layer size is set to 768, the dimension of the multi-layer perceptron is 3072, 12 attention heads are used, and residual blocks are used to improve the performance and training speed of the model. After model inference, a mask with the same size and resolution as the original image will be obtained, and the mask position will be extracted in the original image in the form of Figure 5 At the same time, the model has good scalability and can mark different missing tooth numbers in different missing tooth areas, and further classify the missing tooth types of missing tooth data to meet a wider range of needs.

[0053] Step 4 trains the second stage of the tooth image generation model. Define the probability diffusion network architecture, using a three-dimensional U-net as the base network. The healthy tooth image dataset obtained in Step 2 is used as input. During each round of training, a random time step t is selected for each training example, and Gaussian noise corresponding to t is applied to the image. The U-net is trained to predict the noise, and the loss is calculated using the predicted noise and the actual noise. Backpropagation of gradients is performed to optimize the model parameters. These steps are repeated until the U-net training is complete.

[0054] In step 5, after obtaining the mask of the missing tooth area output by the tooth missing detection model in step 3, a 64×64×64 cropping is performed on the mask and the original image around the missing tooth area according to its array coordinates. Sometimes there is more than one missing tooth area, and all cropped missing tooth areas will be traversed.

[0055] In step 6, the cropped image is input to the tooth image generative model trained in step 4. The model now uses two sampling methods: the denoising diffusion probabilistic model (DDPM) and the denoising diffusion implicit model (DDIM). Both are generative models based on the probability diffusion model. Their key idea is to gradually transition from a simple distribution to the target data distribution through a series of iterative steps. In each iterative step, noise is added or a transformation is performed to denoise the data, thereby gradually generating samples that are closer to the target data distribution. DDPM emphasizes probabilistic modeling of probability diffusion, which can be viewed as a time-varying probability distribution, with a conditional probability density function guiding the data generation process. DDIM emphasizes implicit modeling. That is, it does not explicitly model the conditional probability distribution at each iterative step, but instead generates data through an implicit generator function. DDIM introduces a noise network to add noise to the potential input of the generator function, which is then trained through backpropagation. Unlike DDPM, DDIM does not require explicit modeling of the conditional distribution at each iterative step, making it theoretically more flexible. Compared with DDPM, DDIM generates images faster and with better quality. Figure 6 As shown, the sampling strategy can be selected according to different needs. The present invention can flexibly use two methods to generate the required images, and the images generated by the generative model are diverse, so in clinical practice, the optimal surgical strategy can be adopted according to different images.

[0056] Step 7: Replace the image of the missing tooth area in the original CBCT image with the generated image according to the array coordinates on the original mask, and the operation of automatically detecting and generating missing teeth on the CBCT image can be completed. The invention generates missing teeth on the CBCT image and can more accurately simulate the final position and shape of the implant. This helps dentists to make detailed plans before the operation and determine the best implant position and angle (such as Figure 7 The surgical plan and expected results are presented in detail to increase patient confidence in treatment, improve treatment compliance, and reduce the likelihood of postoperative dissatisfaction and disputes.

Claims

1. A tooth generation method for missing teeth CBCT images based on missing teeth detection and probability diffusion model, characterized by The following steps are included: Step 1: Acquire the data set and process the acquired data set; Step 2: Preprocess the processed dataset. Based on the balance between computational efficiency and segmentation accuracy, the physical resolution of the CBCT images in the collected dataset is normalized to a uniform isotropic resolution. Step 3: Define the missing teeth detection network, use the Unet convolutional neural network as the first stage missing teeth detection network, and train it through supervised learning method; Step 4: Define the missing teeth generation network. The missing teeth generation network in the second stage adopts a self-supervised learning method and is trained by automatically generating damaged and repaired pairs. Step 5: Crop the missing tooth detection area. The CBCT image, which has been cropped by intensity value in step 2, is centrally cropped according to the missing tooth detection result in step 3 to remove information in the CBCT image that is not related to the target teeth. Step 6: Input the cropped data from step 5 into the missing tooth generation network. The probabilistic diffusion model for generating teeth gradually adds noise to the 3D mask data through a forward diffusion process. Then, through a reverse diffusion process, it learns information about the surrounding oral cavity to restore the original image from the noisy image, thereby achieving the mask image restoration task. By controlling the noise generation and removal process, the occluded portion of the missing tooth image can be restored based on the mask information, generating a 3D image of the healthy tooth features learned during the model training process. Step 7: Blend the original image and fill the newly generated tooth image of the missing tooth area into the cropped area of ​​the original image in step 5. By comparing the generated tooth image with the surrounding important anatomical structure information, the implant surgery planning and denture production are designed.

2. The tooth generation method for missing tooth CBCT images based on missing tooth detection and probability diffusion model according to claim 1, characterized in that: Step 1 uses two different training networks, one for missing tooth detection data and the other for single tooth instance data.

3. The tooth generation method for missing tooth CBCT images based on missing tooth detection and probability diffusion model according to claim 1, characterized in that: In step 2, the physical resolution of the CBCT images in the collected dataset is normalized to an isotropic resolution of 0.4 × 0.4 × 0.4 mm3, and the intensity value of each CBCT image scan is clipped to [0, 3096] before intensity normalization.

4. The tooth generation method for missing tooth CBCT images based on missing tooth detection and probability diffusion model according to claim 1, characterized in that: The single tooth instance data is obtained by preprocessing the original oral image data in step 2, and then cropping the area of ​​each tooth instance segmentation with a size of 64×64×64 pixels around its center as the anchor point to obtain the segmented tooth instance dataset. The cropped image contains not only the central tooth, but also the surrounding oral information, including all tooth categories such as incisors, canines, premolars and molars.

5. The tooth generation method for missing tooth CBCT images based on missing tooth detection and probability diffusion model according to claim 1, characterized in that: The tooth-missing area of ​​the input tooth-missing image of the probability diffusion model will be covered by a mask. The probability diffusion model will control the generation and removal of noise based on the surrounding unmasked information, repair the occluded part in the tooth-missing area and generate a three-dimensional image, and predict the shape, size and direction of the teeth at the mask position.

6. The method for generating teeth from missing tooth CBCT images based on missing tooth detection and probability diffusion model according to claim 4, characterized in that: A supervised learning segmentation model was used to segment the edentulous areas on CBCT images. The convolutional neural network (Unet) was trained from scratch in a supervised learning mode using the collected original images and labeled data of the edentulous areas. On this basis, the edentulous areas were segmented as the target. The segmentation network loss function used a combination of binary cross entropy loss and Dice loss to address the class imbalance problem and improve the accuracy of the segmentation results.

7. The method for generating missing tooth CBCT images based on missing tooth detection and probability diffusion model according to claim 1, characterized in that: The tooth generation model described in step 4 adopts a probability diffusion generation model to achieve free-form image restoration, that is, the image data of any missing part is generated by the model. First, a single tooth instance image is trained in a three-dimensional missing tooth generation network to let the network learn how to unconditionally generate teeth, and this is used as the basic generation model. Then, in the CBCT image of the missing tooth, the damaged and missing tooth areas that need to be repaired are masked, which is the first stage of missing tooth detection; the basic generation model changes the reverse iteration of the model by sampling the unmasked surrounding teeth and oral information, that is, it can generate a denture image that meets the requirements in the missing tooth area masked in the original CBCT image.

8. The method for generating teeth from missing tooth CBCT images based on missing tooth detection and probability diffusion model according to claim 7, characterized in that: The tooth generation model uses an improved three-dimensional U-Net model structure. The model training process is as follows: Extract a sample from the data set, randomly select a time t from 1-T, pass x0 and t to GaussionDiffusion, GaussionDiffusion samples a random noise, and adds it to x0 to form x t , then x t and t are put into U-net, which generates sinusoidal position encoding and x according to time t. t Combined, U-net predicts the added noise and returns the noise, and GaussionDiffusion calculates the loss of this noise and random noise; The noise predicted by U-net is compared with the random noise sampled by GaussionDiffusion before to calculate the loss, calculate the gradient, and update the weight; the loss function is calculated as follows: Where x0 represents the initial tooth instance data, ∈ represents the noise term adopted from the normal distribution; the goal of this loss function is to minimize the difference between the noise ∈ sampled from the standard normal distribution and the noise ∈ generated by the model θ (x t , the mean square error between t, to train the probability diffusion generation model; the expected operation The model uncertainty under different initial conditions x0 and noise ∈ is considered, and the above steps are repeated until the probability diffusion model training is completed.

Images