Auxiliary dental implant generation method based on diffusion model

Abstract

The invention provides an auxiliary dental implant generation method based on a diffusion model, and relates to the technical field of stomatology, and the method comprises the steps: obtaining historical patient oral cavity CBCT image data, carrying out the preprocessing of the historical patient oral cavity CBCT image data, obtaining a CBCT image data set, training a multi-task segmentation network through the CBCT image data set, and obtaining a CBCT image data set; obtaining an intra-oral tissue segmentation result in the segmentation network; respectively training a detection network from three dimensions of a cross section, a coronal plane and a sagittal plane by using an intraoral tissue segmentation result; obtaining detection results in three directions of a cross section, a coronal plane and a sagittal plane by using a detection network; fusing the detection results in the three directions of the cross section, the coronal plane and the sagittal plane, constructing a three-dimensional bounding box by using a majority voting algorithm, and obtaining an edentulous region; the intraoral segmentation result and the edentulous area serve as prompt information, the network is guided through the iteration process to generate the post-implantation effect, the implantation effect obtained through the method is high in accuracy, and a more accurate auxiliary tool can be provided for stomatology.

Description

Technical Field

[0001] The present invention relates to the field of oral medical technology, and in particular to a method for generating an auxiliary dental implant based on a diffusion model. Background Art

[0002] With the continuous development of stomatology, digital technology is changing the design and implementation of dental treatment plans, and has been widely used to assist dentists in designing plans. Among them, dental implant surgery is a key means to make up for missing teeth, reshape chewing function and aesthetic appearance, and its intelligent and precise development needs are becoming more and more urgent. In order to meet the needs of modern dentistry, exploring an intelligent, fast and efficient dental implant generation method to assist doctors in designing surgical plans can not only greatly improve the convenience and accuracy of surgical design, but also provide strong technical support for subsequent diagnosis and treatment tasks, thereby effectively ensuring the efficient execution and success rate of implant surgery.

[0003] The traditional method of creating dental implants uses materials such as alginate and silicone to make precise impressions of teeth and surrounding tissues in the patient's mouth. Plaster is poured into the impression to form a solid tooth model, and the doctor then designs the expected dental implant position and shape on the model. However, the intraoral impression can cause severe discomfort to the patient, and the design process is heavily dependent on the doctor's experience and technical level. There are problems such as low accuracy, low efficiency, and long time consumption.

[0004] The existing digital implant design requires the coordination of multiple steps, and the limitations of each step limit the accuracy of the implant position. First, the contrast between different tissues in CBCT images is low and closely arranged. Traditional segmentation methods are difficult to distinguish the boundary information between intraoral tissues, which directly leads to poor quality of segmented images and affects subsequent position determination. At the same time, traditional methods are also limited in determining the position of missing teeth in three-dimensional images. It is usually necessary to project the three-dimensional image into a two-dimensional depth map for detection, and the located position lacks spatial information. For image generation, traditional methods often use generative adversarial networks. The training process is unstable, prone to mode collapse, and consumes large amounts of computing resources and time for high-resolution images.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0006] The purpose of the present invention is to provide a method for generating an auxiliary dental implant based on a diffusion model to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for generating an auxiliary dental implant based on a diffusion model, the specific steps comprising:

[0009] S1: Acquire historical patient oral CBCT image data, preprocess the historical patient oral CBCT image data to obtain a CBCT image set, and determine a segmentation result set, a cross-sectional detection result set, a coronal plane detection result set, and a sagittal plane detection result set based on the historical patient oral CBCT image data, wherein the segmentation result set includes an edge information set and a mask information set;

[0010] S2: Take the CBCT image set as input and the segmentation result set as label to train the multi-task segmentation network, obtain the oral CBCT image data of the object to be predicted and preprocess it, input the CBCT image to be tested obtained after preprocessing into the trained segmentation network, and obtain the intraoral segmentation result of the intraoral tissue of the object to be predicted, which includes edge information and mask information;

[0011] S3: using the mask information set as input, and using the cross-sectional detection result set, the coronal plane detection result set, and the sagittal plane detection result set as labels, respectively training corresponding detection networks, and inputting the mask information of the oral tissue of the object to be predicted into the trained detection network to obtain detection results, wherein the detection results include cross-sectional direction detection results, coronal plane direction detection results, and sagittal plane direction detection results;

[0012] S4: Fusing the cross-sectional direction detection results, the coronal plane direction detection results, and the sagittal plane direction detection results, and using the majority voting algorithm to construct a three-dimensional bounding box to obtain the missing tooth area;

[0013] S5: The CBCT image to be tested and the random noise map are input into the trained diffusion model. The noise map is continuously diffused and denoised through the encoder, intermediate layer and decoder. The intraoral segmentation results and the edentulous area are used as prompt information. The network is guided to generate the effect after implantation through an iterative process.

[0014] Furthermore, the process of network training in the segmentation network includes:

[0015] S21: Perform image enhancement on the preprocessed CBCT image using the Haar filter algorithm;

[0016] S22: splicing the preprocessed CBCT image with the enhanced image to form a feature map, and sending the feature map to the encoder for feature extraction to obtain a first feature map, specifically: dividing the feature map into 4 groups of features along the channel dimension, the first group uses a 1×1×1 convolution to maintain the current ratio; the second and third groups use 3×3×3 depth convolutions with different expansion rates to capture information at different levels, where the expansion rate and the filling rate are set to 1 and 2 respectively, and the fourth group uses an image pooling operation to count the global spatial information; the features after layered processing are spliced ​​and then enter batch normalization integration and ReLU activation function to form output features, and the output features and feature maps are superimposed through residual connections to form the first feature map;

[0017] S23: Enhance the first feature map through a cross attention network, specifically: adjust the dimension of the channel in the first feature map by a 1×1×1 convolution, and further extract context information in the horizontal and vertical directions by two parallel convolution layers; adaptively enhance the importance of features through a multi-head attention module, and fuse the enhanced features with information in different channels and spatial positions through a multi-layer perceptron;

[0018] S24: The decoder gradually restores the spatial resolution of the cross-attention network output features through upsampling and multi-scale convolution, and connects the cross-attention network output features and the first feature map using a jump connection to obtain a second feature map;

[0019] S25: sending the second feature map to the output layer, processing it with an activation function, and outputting the final intra-oral segmentation result;

[0020] S26: The training process uses a weighted loss that combines Dice and cross entropy to optimize the boundary shape.

[0021] Furthermore, the expression of Haar filtering is:

[0022]

[0023] Among them, H N is the Haar filter matrix, h N-1 (N-1 / N) is the N-1 level Haar basis function, and N×N is the order of the Haar filter matrix.

[0024] Furthermore, the output of multi-scale convolution is expressed as:

[0025] F=δ(X(Concat(W1(S′1),W2(S′2),W3(S′3),P(S′4))))+S(1)

[0026] Where F is the first feature map, W1, W2, W3, and W4 are the weights of the four feature sets after corresponding processing, Concat represents the concatenation operation, P represents the global pooling operation, X represents the normalization operation, δ represents the ReLU activation function operation, S′1, S′2, S′3, and S′4 are the results of the four feature sets after corresponding processing, and S(1) represents the feature map.

[0027] Furthermore, the output of the segmentation network is calculated as:

[0028]

[0029] Where P represents the final segmentation result, Softmax(·) represents the normalized exponential function, M represents the pixel value output by the decoder, and C represents the number of categories; c Represents the prediction result corresponding to the cth category in M, where c is the category index;

[0030] The loss function expression of the segmentation network is:

[0031]

[0032]

[0033] in, represents the D ice loss function, K represents the total number of voxels, i represents the voxel index, represents the cross entropy loss function, represents the true label of the cth category of the i-th voxel, represents the overall loss function of the segmentation network; λ1 and λ2 represent the balanced loss parameters, and the tooth categories considered in the segmentation result specifically include: central incisor, lateral incisor, canine, first premolar, second premolar, first molar, and second molar.

[0034] Furthermore, the process of training the detection network includes:

[0035] S31: Input the intraoral tissue segmentation result set into the backbone network and use multiple convolutional layers to extract image features;

[0036] S32: Input the extracted feature map into the global attention module for feature enhancement;

[0037] S33: The last layer of the backbone network uses spatial pyramid pooling to integrate features from different regions;

[0038] S34: The output of the backbone network enters the fusion network, and the feature information from different scales is fused through the pyramid network;

[0039] S35: Input the fused feature map into the detector for regression prediction, and output a multi-scale prediction result including a bounding box, confidence, and category information;

[0040] S36: Input the multi-scale prediction results into the loss function and the flexible non-maximum suppression algorithm for post-processing, remove the overlapping bounding boxes by Gaussian confidence decay, and obtain the final detection result;

[0041] S37: The network completes the entire training process and obtains the optimal weight through continuous iterative updates and parameter adjustments to ensure that the loss parameter converges to the minimum value; the loss function expression of the detection network is:

[0042] Loss = a × loss obj +b×loss rect +c×loss cls

[0043] Among them, a, b, c represent weight coefficients, loss obj Represents confidence loss, loss rect Represents the rectangular box loss, loss cls represents the classification loss, and Loss represents the total loss of the detection network.

[0044] Furthermore, the specific steps of inputting the extracted feature map into the global attention module for feature enhancement include:

[0045] S321: Enhance the features extracted by the convolutional layer by inputting the channel attention; learn the relationship between channels through dimension exchange and multi-layer perceptron;

[0046] S322: swap the dimensions of the output of the multilayer perceptron and generate a channel attention map through a Sigmoid activation function;

[0047] S323: concatenate the output of the channel attention module with the output features extracted by the convolution layer to obtain an intermediate state, and send the intermediate state to the spatial attention module to further extract spatial features through two 7×7 convolutions;

[0048] S324: The convolution output is activated by the Sigmoid function to obtain the channel attention map;

[0049] S325: The channel attention weight, spatial attention weight and output features extracted by the convolutional layer are fused to make the network focus on the most important features in the image. The output feature F3 is used in subsequent network layers. The output of the intermediate state and GAM is expressed as:

[0050]

[0051] Among them, F1 represents the input feature, F2 represents the intermediate state, F3 represents the output feature, and M P is the channel attention map, M S is the spatial attention map, Represents the product of pixels; M P (F1) represents the channel attention map M P Applied to the input feature F1, the spatial attention map M S Applied to the intermediate state F2.

[0052] Furthermore, the specific steps for obtaining the missing teeth area are:

[0053] S41: CBCT images were used to detect missing tooth locations in the cross-sectional, coronal, and sagittal planes using a two-dimensional convolutional network;

[0054] S42: The candidate boxes output by the detection network in the three directions of cross section, coronal plane and sagittal plane are spliced, and a three-dimensional bounding box is obtained using a majority voting algorithm to obtain the tooth-missing area.

[0055] Furthermore, the specific steps of step 5 are:

[0056] S51: The input target image x0 is mapped to the implicit space through the encoder to obtain the latent variable, and the latent variable expression is obtained as:

[0057] z0=E(x0)

[0058] Among them, z0 is the hidden variable, x0 is the input target image, and E is the encoder encoding function;

[0059] S52: Define the forward diffusion process, by adding noise to the hidden variables z~q(z0), simulating the data degradation under real conditions, and gradually evolving the target image into a Gaussian noise distribution z T ; Forward diffusion process q(z 1:T |z0) is expressed as:

[0060]

[0061] Among them, β t Hyperparameters representing the variance of the noise from 0 to 1, z1,…,z T is a latent variable with the same dimension as the data z~q(z0); q(z 1:T |z0) is the forward diffusion process, t is the index of the diffusion step, represents the total number of diffusion steps, The representative mean is The variance is β t Gaussian distribution of; I represents the identity matrix;

[0062] S53: Send the located missing tooth position, the segmentation result of the intraoral tissue and the original image to the conditional encoder to extract the position feature, morphological feature and image feature, and fuse them to obtain the fusion result. The fusion result is expressed as:

[0063]

[0064] Among them, W is the fusion result, represents the embedded position feature, represents the embedded morphological features, Represents the feature embedding of the current step;

[0065] S54: Define the back diffusion process, from the noise image z T At the beginning, the fusion result is used as conditional information to guide the model to iteratively denoise. After T steps, the noise-free feature map z0 is obtained;

[0066] Specifically, the reverse diffusion process is expressed as:

[0067]

[0068] Among them, p θ (z 0:T |W) is the reverse diffusion process, and its mean is The variance is β t Gaussian distribution of θ represents the random noise of the sampling, α t :=1-β t ,

[0069] S55: The feature map z0 obtained above is decoded through the convolution layer and the upsampling layer, and the latent space feature is mapped back to the explicit space to generate the final target image. The decoding process is expressed as:

[0070] x′=D(z0)

[0071] Wherein, x′ is the final target image, and D is the decoder decoding function. Compared with the prior art, the beneficial effects of the present invention are:

[0072] The present invention uses an improved multi-task three-dimensional convolutional neural network to segment intraoral tissues in CBCT images, uses Haar filtering to enhance boundary contrast, improves the network's ability to perceive edge information, and uses a multi-scale convolution module to extract multi-granularity feature information. A cross-attention module is introduced at the connection between the encoder and the decoder to improve the network's ability to represent image features and understand context. This makes the segmentation results more accurate and the guidance of the generation network more effective.

[0073] The present invention uses a two-dimensional detection network to detect the missing tooth position from the CBCT image. By detecting the three directions of the cross section, coronal plane and sagittal plane, the detection results of the three branches are spliced, and a three-dimensional bounding box is obtained by using a majority voting algorithm to locate the missing tooth area. Compared with projecting the CBCT image into a two-dimensional panoramic film, more spatial information can be retained, which helps to locate the missing tooth area more accurately.

[0074] The present invention uses a conditional diffusion model in the generative network part to generate images in the step-by-step denoising process. Compared with the generative adversarial network, it is more stable and can generate images with high resolution and fine detail features. The diffusion process is guided by position information and structural information, making the generated image more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a schematic diagram of the overall method flow of the present invention.

[0076] Figure 2 Schematic diagram of the network segmentation of the present invention.

[0077] Figure 3 Schematic diagram of the multi-scale convolution module of the present invention.

[0078] Figure 4 Schematic diagram of missing tooth position detection according to the present invention.

[0079] Figure 5 It is a structural schematic diagram of the diffusion model of the present invention. DETAILED DESCRIPTION

[0080] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0081] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0082] Example:

[0083] See also Figure 1 , the present invention provides a technical solution:

[0084] A method for generating an auxiliary dental implant based on a diffusion model, the specific steps comprising:

[0085] S1: Acquire historical patient oral CBCT image data, preprocess the historical patient oral CBCT image data to obtain a CBCT image set, and determine a segmentation result set, a cross-sectional detection result set, a coronal plane detection result set, and a sagittal plane detection result set based on the historical patient oral CBCT image data, wherein the segmentation result set includes an edge information set and a mask information set;

[0086] Obtain a CBCT data set, including images of missing teeth cases, and use a preprocessing unit to process the data set. Since the data is collected from different equipment in different hospitals, the grayscale information and sampling interval of the same tissue are different. It is necessary to normalize the pixel value and sampling interval, and divide the data set according to the standard ratio of 7:3. The normalized pixel value is calculated as:

[0087]

[0088] Among them, X represents the original pixel value, X mean represents the average value, X sd represents the standard deviation, X nor Represents the normalized pixel value.

[0089] S2: Take the CBCT image set as input and the segmentation result set as label to train the multi-task segmentation network, obtain the oral CBCT image data of the object to be predicted and preprocess it, input the CBCT image to be tested obtained after preprocessing into the trained segmentation network, and obtain the intraoral segmentation result of the intraoral tissue of the object to be predicted, which includes edge information and mask information;

[0090] The tooth types considered in the segmentation result specifically include: central incisor, lateral incisor, canine, first premolar, second premolar, first molar, and second molar.

[0091] See also Figure 2 , the segmentation network structure of the dental implant generation model is as follows Figure 2 As shown;

[0092] The process of training the network in the segmentation branch of the dental implant generation model includes:

[0093] S21: Perform image enhancement on the preprocessed CBCT image using the Haar filter algorithm;

[0094] The expression of Haar filtering is:

[0095]

[0096] Among them, H N is the Haar filter matrix, h N-1 (N-1 / N) is the N-1 level Haar basis function, and N×N is the order of the Haar filter matrix.

[0097] S22: splicing the preprocessed CBCT image with the enhanced image to form a feature map, and sending the feature map to the encoder for feature extraction to obtain a first feature map, specifically: dividing the feature map into 4 groups of features along the channel dimension, the first group uses a 1×1×1 convolution to maintain the current ratio; the second and third groups use 3×3×3 depth convolutions with different expansion rates to capture information at different levels, where the expansion rate and the filling rate are set to 1 and 2 respectively, and the fourth group uses an image pooling operation to count the global spatial information; the features after layered processing are spliced ​​and then enter batch normalization integration and ReLU activation function to form output features, and the output features and feature maps are superimposed through residual connections to form the first feature map;

[0098] The filtered image is concatenated with the original CBCT image and sent to the encoder for feature extraction. The encoder includes three multi-scale convolution layers, three maximum pooling layers, and three downsampling layers. The multi-scale convolution layers and pooling layers are performed successively, and downsampling is performed after each pooling layer. The downsampling ratio is set to 2. The feature map after sampling is represented as follows:

[0099]

[0100] Where S′ represents the feature map after sampling, S represents the feature map before sampling, and J represents the number of feature map channels. It represents the number domain where the image is located, H represents the height of the feature map, and W represents the width of the feature map.

[0101] See also Figure 3 ,Multi-scale convolutional modules such as Figure 3 As shown, the specific process includes the following:

[0102] The input image is evenly divided into 4 groups along the channel dimension. The first group uses 1×1×1 convolution to maintain the current ratio. The second and third groups use 3×3×3 depth convolution with different expansion rates to capture information at different levels, where the expansion rate and filling rate are set to 1 and 2 respectively. The fourth group uses image global pooling operation to count the global spatial information;

[0103] The layered data is concatenated and integrated, and then output through batch normalization and ReLU activation function. The residual connection superimposes the processed features with the original input to obtain the final output feature F.

[0104] The output of multi-scale convolution is expressed as:

[0105] F=δ(X(Concat(W1(s′1),W2(s′2),W3(S′3),P(S′4))))+s(1)

[0106] Where F is the first feature map, W1, W2, W3, and W4 are the weights of the four feature sets after corresponding processing, Concat represents the concatenation operation, P represents the global pooling operation, X represents the normalization operation, δ represents the ReLU activation function operation, S′1, S′2, S′3, and S′4 are the results of the four feature sets after corresponding processing, and S(1) represents the feature map.

[0107] S23: Enhance the first feature map through a cross attention network, specifically: adjust the dimension of the channel in the first feature map by a 1×1×1 convolution, and further extract context information in the horizontal and vertical directions by two parallel convolution layers; adaptively enhance the importance of features through a multi-head attention module, and fuse the enhanced features with information in different channels and spatial positions through a multi-layer perceptron;

[0108] S231: The encoder output is adjusted to channel dimension by 1×1×1 convolution;

[0109] S232: Two parallel convolutions extract context information in the horizontal and vertical directions respectively;

[0110] S232: Multi-head attention module adaptively enhances the importance of relevant features and reduces the interference of irrelevant pixels;

[0111] S233: The enhanced features are fused with information from different channels and spatial positions through a multi-layer perceptron, and finally sent to the decoder through a residual connection with the output of the encoder.

[0112] S24: The decoder gradually restores the spatial resolution of the cross-attention network output features through upsampling and multi-scale convolution, and connects the cross-attention network output features and the first feature map using a jump connection to obtain a second feature map;

[0113] S25: sending the second feature map to the output layer, processing it with an activation function, and outputting the final intra-oral segmentation result;

[0114] The output of the segmentation network is calculated as:

[0115]

[0116] Where P represents the final segmentation result, Softmax(·) represents the normalized exponential function, M represents the pixel value output by the decoder, and C represents the number of categories; cRepresents the prediction result corresponding to the cth category in M, where c is the category index;

[0117] S26: The training process uses a weighted loss that combines Dice and cross entropy to optimize the boundary shape.

[0118] The loss function expression of the segmentation network is:

[0119]

[0120]

[0121] in, represents the D ice loss function, K represents the total number of voxels, i represents the voxel index, represents the cross entropy loss function, represents the true label of the cth category of the i-th voxel, represents the overall loss function of the segmentation network; λ1 and λ2 represent the balanced loss parameters. The tooth categories considered in the segmentation result include: central incisor, lateral incisor, canine, first premolar, second premolar, first molar, second molar. λ1>λ2, because the Dice loss function performs better in unbalanced data sets, especially when the target category occupies a small volume, because it pays more attention to the quality of the overlapping area. The cross entropy loss function performs well in terms of overall accuracy. In this example, the target volume is small, and the overlapping area does not play an important role in the subsequent step of determining the missing tooth area, so the balanced loss parameter is set to λ1>λ2.

[0122] S3: using the mask information set as input, and using the cross-sectional detection result set, the coronal plane detection result set, and the sagittal plane detection result set as labels, respectively training corresponding detection networks, and inputting the mask information of the oral tissue of the object to be predicted into the trained detection network to obtain detection results, wherein the detection results include cross-sectional direction detection results, coronal plane direction detection results, and sagittal plane direction detection results;

[0123] The mask image of the segmentation network is used to detect missing teeth from three anatomical directions: cross-section, coronal, and sagittal. A detection network is designed and trained for each direction.

[0124] The process of training the detection network includes:

[0125] S31: Input the intraoral tissue segmentation result set into the backbone network and use multiple convolutional layers to extract image features;

[0126] S32: The extracted feature map is input into the global attention module for feature enhancement;

[0127] S321: Enhance the features extracted by the convolutional layer by inputting the channel attention; learn the relationship between channels through dimension exchange and multi-layer perceptron;

[0128] S322: swap the dimensions of the output of the multilayer perceptron and generate a channel attention map through a Sigmoid activation function;

[0129] S323: concatenate the output of the channel attention module with the output features extracted by the convolution layer to obtain an intermediate state, and send the intermediate state to the spatial attention module to further extract spatial features through two 7×7 convolutions;

[0130] S324: The convolution output is activated by the Sigmoid function to obtain the channel attention map;

[0131] S325: The channel attention weight, spatial attention weight and output features extracted by the convolutional layer are fused to make the network focus on the most important features in the image. The output feature F3 is used in subsequent network layers. The output of the intermediate state and GAM is expressed as:

[0132]

[0133]

[0134] Among them, F1 represents the input feature, F2 represents the intermediate state, F3 represents the output feature, and M C is the channel attention map, M S is the spatial attention map, Represents the product of pixels; M C (F1) represents the channel attention map M C Applied to the input feature F1, the spatial attention map M S Applied to the intermediate state F2.

[0135] S33: The last layer of the backbone network uses spatial pyramid pooling to integrate features from different regions;

[0136] S34: The output of the backbone network enters the fusion network, and the feature information from different scales is fused through the pyramid network;

[0137] S35: Input the fused feature map into the detector for regression prediction, and output a multi-scale prediction result including a bounding box, confidence, and category information;

[0138] S36: Input the multi-scale prediction results into the loss function and the flexible non-maximum suppression algorithm for post-processing, remove the overlapping bounding boxes by Gaussian confidence decay, and obtain the final detection result;

[0139] S37: Through continuous iterative updates and parameter adjustments, the network ensures that the loss parameter converges to a minimum value to complete the entire training process and obtain the optimal weights. The expression of the loss function of the detection network is:

[0140] Loss = a × loss obj + b × loss rect + c × loss cls

[0141] where a, b, and c represent weight coefficients, loss obj represents the confidence loss, loss rect represents the bounding box loss, loss cls represents the classification loss, and Loss represents the total loss of the detection network. The weight coefficients satisfy a < b < c and a + b + c = 1. The classification loss is usually the most important part in object detection because it directly affects the model's ability to identify target classes; therefore, giving a larger weight can improve the classification accuracy; the bounding box loss is responsible for the accuracy of locating the target. Although important, it usually has a slightly lower priority than classification; a moderate weight can help the model learn more accurate localization; the confidence loss is mainly used to evaluate the model's confidence in the presence or absence of the target; the impact of this part of the loss is relatively small, so the weight is set to the minimum.

[0142] S4: Fuse the detection results in the cross-sectional direction, coronal direction, and sagittal direction, and use the majority voting algorithm to construct a three-dimensional bounding box to obtain the toothless area;

[0143] Please refer to Figure 4 , the toothless position detection is as Figure 4 shown,

[0144] Input the segmentation mask image into the trained detection network, and predict the toothless position in the cross-sectional, coronal, and sagittal directions;

[0145] Merge the detection results of the three branches in the cross-sectional, coronal, and sagittal directions, and use the majority voting algorithm to obtain the three-dimensional bounding box of the toothless position to locate the toothless position.

[0146] S51: The input target image x0 is mapped to the implicit space by the encoder to obtain a latent variable, and the expression for obtaining the latent variable is:

[0147] z0 = E(x0)

[0148] where z0 is the latent variable, x0 is the input target image, and E is the encoder encoding function;

[0149] S52: Define the forward diffusion process, by adding noise to the hidden variables z~q(z0), simulating the data degradation under real conditions, and gradually evolving the target image into a Gaussian noise distribution z T ; Forward diffusion process q(z 1:T |z0) is expressed as:

[0150]

[0151] Among them, β t Hyperparameters representing the amount of noise, z1,…,z T is a latent variable with the same dimension as the data z~q(z0); q(z 1:T |z0) is the forward diffusion process, t is the index of the diffusion step, represents the total number of diffusion steps, The representative mean is The variance is β t Gaussian distribution of; I represents the identity matrix;

[0152] The specific method of adding noise is: adding V noises to the latent variables, Add noise to satisfy 0<σ β <1, where is the rounding operation for T-1, σ β is the variance of the added noise.

[0153] S53: Send the located missing tooth position, the segmentation result of the intraoral tissue and the original image to the conditional encoder to extract the position feature, morphological feature and image feature, and fuse them to obtain the fusion result. The fusion result is expressed as:

[0154]

[0155] Among them, W is the fusion result, represents the embedded position feature, represents the embedded morphological features, Represents the feature embedding of the current step;

[0156] S54: Define the back diffusion process, from the noise image z T At the beginning, the fusion result is used as conditional information to guide the model to iteratively denoise. After T steps, the noise-free feature map z0 is obtained;

[0157] Specifically, the reverse diffusion process is expressed as:

[0158]

[0159] Among them, p θ (z 0:T|W) is the reverse diffusion process, and its mean is The variance is β t Gaussian distribution of θ represents the random noise of the sampling, α t :=1-β t ,

[0160] S55: The feature map z0 obtained above is decoded through the convolution layer and the upsampling layer, and the latent space feature is mapped back to the explicit space to generate the final target image. The decoding process is expressed as:

[0161] x′=D(z0)

[0162] Among them, x′ is the final target image and D is the decoder decoding function.

[0163] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0164] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0165] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0166] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A method for generating auxiliary dental implants based on a diffusion model, characterized in that: The specific steps include: S1: Acquire historical patient oral CBCT image data, preprocess the historical patient oral CBCT image data to obtain a CBCT image set, and determine a segmentation result set, a cross-sectional detection result set, a coronal plane detection result set, and a sagittal plane detection result set based on the historical patient oral CBCT image data, wherein the segmentation result set includes an edge information set and a mask information set; S2: Take the CBCT image set as input and the segmentation result set as label to train the multi-task segmentation network, obtain the oral CBCT image data of the object to be predicted and preprocess it, input the CBCT image to be tested obtained after preprocessing into the trained segmentation network, and obtain the intraoral segmentation result of the intraoral tissue of the object to be predicted, which includes edge information and mask information; S3: using the mask information set as input, and using the cross-sectional detection result set, the coronal plane detection result set, and the sagittal plane detection result set as labels, respectively training corresponding detection networks, and inputting the mask information of the oral tissue of the object to be predicted into the trained detection network to obtain detection results, wherein the detection results include cross-sectional direction detection results, coronal plane direction detection results, and sagittal plane direction detection results; S4: Fusing the cross-sectional direction detection results, the coronal plane direction detection results, and the sagittal plane direction detection results, and using the majority voting algorithm to construct a three-dimensional bounding box to obtain the missing tooth area; S5: The CBCT image to be tested and the random noise map are input into the trained diffusion model. The noise map is continuously diffused and denoised through the encoder, intermediate layer and decoder. The intraoral segmentation results and the edentulous area are used as prompt information. The network is guided to generate the effect after implantation through an iterative process.

2. The method for generating an auxiliary dental implant based on a diffusion model according to claim 1, characterized in that: The process of network training in the segmentation network includes: S21: Perform image enhancement on the preprocessed CBCT image using the Haar filter algorithm; S22: splicing the preprocessed CBCT image with the enhanced image to form a feature map, and sending the feature map to the encoder for feature extraction to obtain a first feature map, specifically: dividing the feature map into 4 groups of features along the channel dimension, the first group uses a 1×1×1 convolution to maintain the current ratio; the second and third groups use 3×3×3 depth convolutions with different expansion rates to capture information at different levels, where the expansion rate and the filling rate are set to 1 and 2 respectively, and the fourth group uses an image pooling operation to count the global spatial information; the features after layered processing are spliced ​​and then enter batch normalization integration and ReLU activation function to form output features, and the output features and feature maps are superimposed through residual connections to form the first feature map; S23: Enhance the first feature map through a cross attention network, specifically: adjust the dimension of the channel in the first feature map by a 1×1×1 convolution, and further extract context information in the horizontal and vertical directions by two parallel convolution layers; adaptively enhance the importance of features through a multi-head attention module, and fuse the enhanced features with information in different channels and spatial positions through a multi-layer perceptron; S24: The decoder gradually restores the spatial resolution of the cross-attention network output features through upsampling and multi-scale convolution, and connects the cross-attention network output features and the first feature map using a jump connection to obtain a second feature map; S25: sending the second feature map to the output layer, processing it with an activation function, and outputting the final intra-oral segmentation result; S26: The training process uses a weighted loss that combines Dice and cross entropy to optimize the boundary shape.

3. A method for generating an auxiliary dental implant based on a diffusion model according to claim 2, characterized in that: The expression of Haar filtering is: Among them, H N is the Haar filter matrix, h N-1 (N-1 / N) is the N-1 level Haar basis function, and N×N is the order of the Haar filter matrix.

4. The method for generating an auxiliary dental implant based on a diffusion model according to claim 3, characterized in that: The output of multi-scale convolution is expressed as: F=δ(X(Concat(W1(S′1),W2(S′2),W3(S′3),P(S′4))))+S(1) Where F is the first feature map, W1, W2, W3, and W4 are the weights of the four feature sets after corresponding processing, Concat represents the concatenation operation, P represents the global pooling operation, X represents the normalization operation, δ represents the ReLU activation function operation, S′1, S′2, S′3, and S′4 are the results of the four feature sets after corresponding processing, and S(1) represents the feature map.

5. The method for generating an auxiliary dental implant based on a diffusion model according to claim 3, characterized in that: The output of the segmentation network is calculated as: Where P represents the final segmentation result, Softmax(·) represents the normalized exponential function, M represents the pixel value output by the decoder, and C represents the number of categories; c Represents the prediction result corresponding to the c-th category in M, where c is the category index; The loss function expression of the segmentation network is: in, represents the Dice loss function, K represents the total number of voxels, i represents the voxel index, represents the cross entropy loss function, represents the true label of the cth category of the i-th voxel, represents the overall loss function of the segmentation network; λ1 and λ2 represent the balanced loss parameters, and the tooth categories considered in the segmentation result specifically include: central incisor, lateral incisor, canine, first premolar, second premolar, first molar, and second molar.

6. The method for generating an auxiliary dental implant based on a diffusion model according to claim 1, characterized in that: The process of training the detection network includes: S31: Input the intraoral tissue segmentation result set into the backbone network and use multiple convolutional layers to extract image features; S32: Input the extracted feature map into the global attention module for feature enhancement; S33: The last layer of the backbone network uses spatial pyramid pooling to integrate features from different regions; S34: The output of the backbone network enters the fusion network, and the feature information from different scales is fused through the pyramid network; S35: Input the fused feature map into the detector for regression prediction, and output a multi-scale prediction result including a bounding box, confidence, and category information; S36: Input the multi-scale prediction results into the loss function and the flexible non-maximum suppression algorithm for post-processing, remove the overlapping bounding boxes by Gaussian confidence decay, and obtain the final detection result; S37: The network completes the entire training process and obtains the optimal weight through continuous iterative updates and parameter adjustments to ensure that the loss parameter converges to the minimum value; the loss function expression of the detection network is: Loss=a×loss obj +b×loss rect +c×loss cls Among them, a, b, c represent weight coefficients, loss obj Represents confidence loss, loss rect Represents the rectangular box loss, loss cls represents the classification loss, and Loss represents the total loss of the detection network.

7. The method for generating an auxiliary dental implant based on a diffusion model according to claim 1, characterized in that: The specific steps of inputting the extracted feature map into the global attention module for feature enhancement include: S321: Enhance the features extracted by the convolutional layer by inputting the channel attention; learn the relationship between channels through dimension exchange and multi-layer perceptron; S322: swap the dimensions of the output of the multilayer perceptron and generate a channel attention map through a Sigmoid activation function; S323: concatenate the output of the channel attention module with the output features extracted by the convolution layer to obtain an intermediate state, and send the intermediate state to the spatial attention module to further extract spatial features through two 7×7 convolutions; S324: The convolution output is activated by the Sigmoid function to obtain the channel attention map; S325: The channel attention weight, spatial attention weight and output features extracted by the convolutional layer are fused to make the network focus on the most important features in the image. The output feature F3 is used in subsequent network layers. The output of the intermediate state and GAM is expressed as: Among them, F1 represents the input feature, F2 represents the intermediate state, F3 represents the output feature, and M P is the channel attention map, M S is the spatial attention map, Represents the product of pixels; M P (F1) represents the channel attention map M P Applied to the input feature F1, the spatial attention map M S Applied to the intermediate state F2.

8. The method for generating an auxiliary dental implant based on a diffusion model according to claim 1, characterized in that: The specific steps to obtain the edentulous area are: S41: CBCT images were used to detect missing tooth locations in the cross-sectional, coronal, and sagittal planes using a two-dimensional convolutional network; S42: The candidate boxes output by the detection network in the three directions of cross section, coronal plane and sagittal plane are spliced, and a three-dimensional bounding box is obtained using a majority voting algorithm to obtain the tooth-missing area.

9. The method for generating an auxiliary dental implant based on a diffusion model according to claim 1, characterized in that: The specific steps of step 5 are: S51: The input target image x0 is mapped to the implicit space through the encoder to obtain the latent variable, and the latent variable expression is obtained as: z0=E(x0) Among them, z0 is the hidden variable, x0 is the input target image, and E is the encoder encoding function; S52: Define the forward diffusion process, by adding noise to the hidden variables z~q(z0), simulating the data degradation under real conditions, and gradually evolving the target image into a Gaussian noise distribution z T ; Forward diffusion process q(z1: T |z0) is expressed as: Among them, β t Hyperparameters representing the amount of noise, z1,...,z T is a latent variable with the same dimension as the data z~q(z0); q(z 1：T |z0) is the forward diffusion process, t is the index of the diffusion step, T represents the total number of diffusion steps, The representative mean is The variance is β t Gaussian distribution of; I represents the identity matrix; S53: Send the located missing tooth position, the segmentation result of the intraoral tissue and the original image to the conditional encoder to extract the position feature, morphological feature and image feature, and fuse them to obtain the fusion result. The fusion result is expressed as: Among them, W is the fusion result, represents the embedded position feature, represents the embedded morphological features, Represents the feature embedding of the current step; S54: Define the back diffusion process, from the noise image z T At the beginning, the fusion result is used as conditional information to guide the model to iteratively denoise. After T steps, the noise-free feature map z0 is obtained; Specifically, the reverse diffusion process is expressed as: Among them, p θ (z 0：T |W) is the reverse diffusion process, and its mean is The variance is β t Gaussian distribution of θ represents the random noise of the sampling, α t : =1-β t , S55: The feature map z0 obtained above is decoded through the convolution layer and the upsampling layer, and the latent space feature is mapped back to the explicit space to generate the final target image. The decoding process is expressed as: x′=D(z0) Among them, x′ is the final target image and D is the decoder decoding function.

Images