A Digital Dental Implant Classification Method Based on Deep Learning

Abstract

The present invention belongs to the field of computer vision, and particularly relates to a digital dental implant classification method based on deep learning, including: obtaining real-time CBCT images and preprocessing the CBCT images; inputting the preprocessed CBCT images into the segmentation branch of a trained multi-task dental implant classification model for binary segmentation; cropping the implant area according to the segmentation result to obtain a cropped image and a Mask area; connecting the cropped image and the Mask area in channels, and inputting the connected image into the classification branch of the multi-task dental implant classification model to obtain a classification result; the multi-task dental implant classification model includes an encoder, a fully associated feature enhancement module, a decoder, a segmentation layer, 3 convolutional attention mixing modules, and a classifier; the present invention uses progressive network training to extract multi-granularity features, which can effectively improve the classification accuracy of the implant system.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and particularly relates to a method for classifying digital dental implants based on deep learning. Background Art

[0002] With the continuous development of digital technology, the field of oral medicine is transforming from the traditional diagnostic mode to computer-aided diagnosis and treatment. Dental implants have been used for more than half a century and are currently considered the most reliable long-term (10 years and above) treatment for replacing missing teeth. However, biological complications (such as peri-implant mucositis and peri-implantitis) and mechanical complications (such as fractures of implant prostheses or fixtures and screw loosening) are often associated with implant systems, and the cumulative complication rate in patients with implants is 48.03%. Therefore, for clinicians, being able to accurately classify the brands and systems of implants in a patient's oral cavity from CBCT images is a key condition for subsequent treatment.

[0003] There are the following difficulties in manually identifying implants: There are up to more than a thousand types of implant systems on the market currently, and their shapes are different. Even for doctors with relatively rich clinical experience, it is quite difficult to visually identify the implant system.

[0004] Traditional implant classification methods use the K-Nearest Neighbor algorithm as a classifier. In the field of deep learning, methods generally use models pre-trained on large natural image datasets and then fine-tune them on implant datasets. The former relies on prior knowledge and requires manual interaction, while the latter achieves automatic classification but does not design a network specifically for the characteristics of implants. These methods have certain effects, but they have not yet well solved the following problems: 1) Although the overall shapes of implant systems are similar, there are significant differences in some details (such as taper, presence or absence of threads, shape of threads, etc.), and there is no specific method to analyze them. 2) Only relying on two-dimensional films for feature learning and not using three-dimensional imaging features cannot comprehensively grasp the feature information. These problems result in low accuracy of implant classification. Therefore, there is an urgent need for a new method that can automatically and accurately classify implants in three-dimensional CBCT images. Summary of the Invention

[0005] To solve the problems existing in the above prior art, the present invention proposes a digital dental implant classification method based on deep learning, which includes: acquiring real-time CBCT images and preprocessing the CBCT images; inputting the preprocessed CBCT images into the segmentation branch of a trained multi-task dental implant classification model for binary segmentation; cropping the implant area according to the segmentation result to obtain a cropped image and a Mask area; connecting the cropped image and the Mask area in channels, and inputting the connected image into the classification branch of the multi-task dental implant classification model to obtain a classification result; the multi-task dental implant classification model includes an encoder, a fully associated feature enhancement module, a decoder, a segmentation layer, 3 convolutional attention hybrid modules, and a classifier; among them, the encoder, the fully associated feature enhancement module, the decoder, and the segmentation layer serve as the segmentation branch, and the encoder, the 3 convolutional attention hybrid modules, and the classifier serve as the classification branch;

[0006] The process of training the multi-task dental implant classification model includes:

[0007] S1: Acquire a CBCT image dataset, and preprocess the images in the dataset to obtain a training set and a test set;

[0008] S2: Input the CBCT images in the training set into the segmentation branch of the multi-task dental implant classification model for training to obtain the optimal parameters and segmentation results in the segmentation branch;

[0009] S3: Input the segmentation results output by the segmentation branch into the classification branch of the multi-task dental implant classification model for training to obtain the optimal parameters and classification results in the classification branch; complete the training of the model;

[0010] S4: Input the data in the test set into the trained multi-task dental implant classification model for testing.

[0011] Preferably, the process of training the segmentation branch of the multi-task dental implant classification model includes:

[0012] S21: Input the CBCT images in the training set into the encoder for multi-resolution feature extraction to obtain multiple feature maps with different resolutions;

[0013] S22: Perform skip connections on the other feature maps of the feature maps output by the encoder except the last layer;

[0014] S23: Input the feature map output by the last layer of the encoder into the fully associated feature enhancement module for feature enhancement to obtain an enhanced feature map;

[0015] S24: Input the skip-connected feature map and the enhanced feature map into the decoder and perform upsampling to obtain a restored full-size feature map;

[0016] S25: Input the full-size feature map into the segmentation output layer, and process the output image using the Softmax function to obtain the segmentation result;

[0017] S26: Calculate the segmentation branch loss function of the model based on the segmentation result. Through backpropagation of gradients, continuously update the parameters of the encoder, decoder, and fully associated feature enhancement module. When the loss function converges to the minimum value, complete the training of the model to obtain the parameters of the segmentation part in the trained multi-task implant classification model.

[0018] Further, the process of using the fully associated feature enhancement module to perform feature enhancement processing on the features output by the encoder includes:

[0019] S231: Input the feature layer X output by the last layer of the encoder into the fully associated feature enhancement module. After a dilated convolution with a dilation coefficient of 1, save its output feature map Y1;

[0020] S232: Add the output feature map Y1 and X, and perform a dilated convolution with a dilation coefficient of 2, and save the output feature map Y2;

[0021] S233: Add Y2 and X, perform a dilated convolution with a dilation coefficient of 4, and save the output feature map Y3. Finally, perform multi-path feature fusion on X, Y1, Y2, and Y3 to obtain the enhanced output feature map.

[0022] Preferably, the process of training the classification branch includes:

[0023] S31: Crop the input CBCT image for a single implant area according to the segmentation result, connect the cropped image with the Mask area into 2 channels, and use the connected image as the input of the classification branch;

[0024] S32: Input the image of a single implant into the encoder of the model for feature extraction, and save the feature map at this resolution before each downsampling to obtain multiple feature maps with different resolutions;

[0025] S33: Input the feature maps of the last three layers of the encoder into 3 convolutional attention mixing modules (CAMM) respectively to obtain 3 feature-enhanced feature maps;

[0026] S34: Fuse the 3 enhanced feature maps, input the fused feature map into the classifier to obtain the classification result;

[0027] S35: Calculate the classification branch loss function of the model based on the classification results. Update the parameters of the encoder, CAMM, and classifier through backpropagation of gradients. When the loss function converges to the minimum value, complete the training of the model to obtain the parameters of the classification part in the trained multi-task implant classification model.

[0028] Further, the process of the convolutional attention mixing module CAMM processing the input feature map includes:

[0029] S331: Map the input feature map using three 1×1×1 convolutions and reshape it into N blocks to obtain intermediate features containing 3×N feature maps;

[0030] S332: For the convolutional branch, if the size of its convolutional kernel is K, then use a fully connected layer to obtain K 3 feature maps, where one feature map indicates that it can be shifted to obtain N convolutional output feature maps;

[0031] S333: For the self-attention branch, divide the intermediate features into N groups, each group contains three parts of features, which act as query, key, and value respectively. The output layer feature map of the self-attention mechanism is obtained by using the multi-head self-attention module, and then the aggregated features are obtained by feature concatenation;

[0032] S334: Set two weight coefficients, multiply the N convolutional output feature maps and the aggregated features by the corresponding weight coefficients respectively and then add them to obtain the final output result of a CAMM.

[0033] Advantages of the present invention:

[0034] 1. The present invention uses an improved multi-task 3D convolutional neural network model to segment and classify CBCT images, adopts progressive network training to extract multi-granularity features, and uses convolutional and attention mechanisms to enhance the multi-granularity features, which can better extract the edge information of the implant system, make the classification results more accurate, and promote the segmentation network part;

[0035] 2. The present invention uses a residual fully associated dilated convolution module at the end of the encoder in the segmentation part to adjust the effective kernel function of the feature point receptive field without reducing the resolution of the feature map, and uses a residual structure to combine the feature information under different receptive fields, which can effectively restore the global information of the feature map and improve the segmentation accuracy. Description of the Drawings

[0036] Figure 1 is the overall flowchart of the present invention;

[0037] Figure 2Block diagram of the multi-task dental implant classification model of the present invention;

[0038] Figure 3 Block diagram of each convolutional module of the present invention;

[0039] Figure 4 Block diagram of the structure of the full association module of the present invention;

[0040] Figure 5 Block diagram of the structure of the convolutional attention mixing module of the present invention. Detailed implementation manner

[0041] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0042] A digital dental implant classification method based on deep learning, as Figure 1 shown, the method includes: obtaining CBCT images in real time and preprocessing the data; inputting the preprocessed CBCT image data into the segmentation branch of the trained multi-task dental implant classification model for binary segmentation; cropping the single implant area according to the segmentation result, connecting the original image and the Mask obtained by cropping in channels, and then sending them into the classification branch of the multi-task dental implant classification model to classify the implant; the multi-task dental implant classification model is an improved 3D convolutional neural network model, including an encoder, a full association feature enhancement module FLM, a decoder, 3 convolutional attention mixing modules CAMM, and a classifier.

[0043] A specific real-time manner of a digital dental implant classification method based on deep learning, the method includes: obtaining CBCT images, preprocessing the CBCT images; inputting the preprocessed CBCT images into the segmentation branch of the trained multi-task dental implant classification model for binary segmentation; cropping the implant area according to the segmentation result to obtain a cropped image and a Mask area; connecting the cropped image and the Mask area in channels, and inputting the connected image into the classification branch of the multi-task dental implant classification model to obtain a classification result; the multi-task dental implant classification model includes an encoder, a full association feature enhancement module, a decoder, a segmentation layer, 3 convolutional attention mixing modules, and a classifier; wherein the encoder, the full association feature enhancement module, the decoder, and the segmentation layer serve as the segmentation branch, and the encoder, 3 convolutional attention mixing modules, and the classifier serve as the classification branch

[0044] In this embodiment, the framework structure diagram of the multi-task dental implant classification model is asFigure 2 As shown, the original image preprocessed at the encoder input of the network is connected to the end of the encoder at the input of the fully associated feature enhancement module FLM in the segmentation part, the input of the decoder is connected to the output of the fully associated feature enhancement module FLM, and the output of the decoder is connected to the output layer; in the classification part, the inputs of the three convolutional attention mixing modules CAMM are respectively connected to the outputs of the last three layers of the encoder, and the features at the outputs of the three convolutional attention mixing modules CAMM are concatenated and then connected to the input of the classifier to form a multi-task implant classification model.

[0045] The training process of the multi-task implant classification model includes but is not limited to the following steps:

[0046] S1: Obtain the CBCT image dataset, preprocess the images in the dataset, and allocate the preprocessed data according to a ratio of 7:3 to obtain a training set and a test set; since the data is from multiple centers, it is necessary to preprocess the original CBCT image data containing implants. The preprocessing includes image intensity value standardization and normalization. The voxel intensity value after preprocessing is:

[0047]

[0048] where P mean is the average value of the intensity values in the data, P std is the standard deviation of the intensity values; P min is the minimum value of the intensity values, P max is the maximum value of the intensity values.

[0049] S2: Input the CBCT image data into the encoder of the model for feature extraction, and save the feature map at this resolution before each downsampling to obtain multiple feature maps with different resolutions. As Figure 3 shown, the feature extraction includes 1 basic convolutional block and 4 residual convolutional modules. There is 1 maxpooling module for downsampling after each convolutional module; each basic convolutional module and residual convolutional module includes 2 3×3×3 convolutional layers, 2 batch normalization layers, and 2 ReLU activation functions; each time the data extracts features, the size of the feature map is reduced to half of the original:

[0050]

[0051] where M represents the feature map before downsampling, C represents the number of channels of the feature map, represents the number field where the image is located, H, W, and L respectively represent the height, width, and length of the image, and M * represents the feature map after downsampling.

[0052] S3: Perform skip connections on the feature maps other than the feature map output by the last layer of the encoder, and fuse them with the feature maps of the corresponding resolution in the decoder.

[0053] S4: Input the feature map output by the last layer of the encoder into the FLM for feature enhancement; the dilated convolution in the FLM structure is an effective kernel function for adjusting the receptive field of feature points without reducing the resolution of the feature map. Utilizing this feature, add the feature map after each dilation to the feature map input in the previous stage, and perform feature enhancement again, which can show a strong ability to improve the segmentation accuracy.

[0054] In this embodiment, the fully associated module FLM is as Figure 4 shown, and specifically includes the following process:

[0055] S41: Input the feature layer X output by the last layer of the encoder into the FLM, perform a dilated convolution with a dilation coefficient of 1 once, and save its output feature map Y1;

[0056] S42: Add the output feature map Y11 of S41 and X, perform a dilated convolution with a dilation coefficient of 2 once, and save the output feature map Y2;

[0057] S43: Add Y2 and X, perform a dilated convolution with a dilation coefficient of 4 once, and save the output feature map Y3. To better fuse features with inconsistent semantics, perform attention-based multi-path feature fusion on X, Y1, Y2, and Y3 to obtain the final output. Denote X = Y0, then the output of one layer of the FLM is expressed as:

[0058]

[0059] The total output result is:

[0060]

[0061]

[0062]

[0063] L(Y′) = B(PWConv1(δ(B(PWConv2(Y′)))))

[0064] where denotes element-wise multiplication of the corresponding pixel values, denotes the addition of corresponding pixel values, g(Y′) denotes global average pooling, σ denotes the Sigmoid function, δ denotes the ReLU activation function, where n is the number of dilated convolutions in the FLM module, PWConv is a convolution with a kernel size of 1, and B denotes batch normalization. During the process of attention-based multi-path feature fusion, the sum of the four features is used to extract channel attention weights through two branches. One branch uses global average pooling to extract the attention of global features, and the other branch directly uses a convolution with a kernel size of 1 to extract the channel attention of local features, so that the finally fused feature map not only has global features but also includes local features, making the final result more accurate; compared with the fusion of other feature maps, the more complex attention-based multi-path feature fusion method for feature fusion has great potential to produce better results.

[0065] S5: Input the feature map output by the FLM and the feature map for skip connection into the decoder, and obtain the restored full-size feature map through the upsampling module. The structure diagram of the upsampling module is as Figure 2 shown.

[0066] S6: Input the full-size feature map into the segmentation output layer, and use the Softmax function to process the output image to obtain the segmentation result Y out :

[0067]

[0068] where X represents each pixel output by the decoder, X C represents the output corresponding to the Cth category semantics in X, n represents the number of categories, and i represents the semantics of the ith category.

[0069] S7: Calculate the loss function of the model according to the segmentation result. Since the implant target accounts for a relatively small proportion compared to the entire oral CT during the segmentation stage, the Focal loss function (Focal Loss, FL) adapted to the size of the implant target and the Dice loss (Dice Loss, DL) function for measuring the similarity between the predicted label and the true label are used; sum the Focal loss and the Dice loss to obtain the total model loss function for segmentation:

[0070] L Totle = λ FL L FL + λ Dice L Dice

[0071]

[0072]

[0073] Among them, N represents the total number of voxels, n represents the number of categories, i represents the semantics of the i-th category, represents the predicted category of the i-th voxel, represents the true category of the i-th voxel, γ is the hyperparameter of Focal Loss, and experiments have shown that the best effect is achieved when γ takes 2; λ FL and λ Dice are two hyperparameters for balancing the loss.

[0074] After calculating the loss function, through backpropagation of the gradient, the parameters of the encoder, decoder, and FLM are continuously updated. When the loss function converges to the minimum value, the training of the model is completed, and the parameters of the segmentation part in the trained multi-task implant classification model are obtained.

[0075] S8: After the segmentation part is trained, the classification part is trained. The data is passed through the segmentation part to output the segmentation result. According to the segmentation result, the original image and the binary mask are cropped for each implant and connected into a 2-channel image as the input of the classification part.

[0076] S9: The image of a single implant is input into the encoder of the model for feature extraction. After the last three downsamplings, the feature maps at this resolution are saved to obtain feature maps with different resolutions.

[0077] S10: The feature maps of the last three layers of the encoder are respectively input into 3 CAMM networks to obtain 3 feature maps with enhanced features.

[0078] In this embodiment, the convolutional attention mixing module CAMM is as Figure 5 shown. Taking one layer as an example, the specific process includes:

[0079] S101: The input feature is mapped by three 1×1×1 convolutions and then reshaped into N blocks to obtain intermediate features containing 3×N feature maps;

[0080] S102: For the convolutional branch, assuming the convolutional kernel size is K, a fully connected layer is used to obtain K 3 feature maps, and one of the feature maps can be Then, N convolutional output feature maps are obtained through shifting. The shifting operation is defined as:

[0081]

[0082] where Δx, Δy, and Δz represent the displacements in the Figure 3 feature directions, f is the input feature map, and is the output feature map;

[0083] The complete convolutional output result after shifting is:

[0084]

[0085] Among them, p, q, and r represent the coordinates of the weights in the convolutional kernel, and the aggregated features are obtained by means of feature concatenation.

[0086] S103: For the self-attention branch, the intermediate features are divided into N groups, and each group contains three parts of features, which serve as query, key, and value respectively. Then, through the traditional multi-head self-attention module, the output layer feature map of the self-attention mechanism is obtained, and the aggregated features are obtained by means of feature concatenation. The expression for obtaining the aggregated features is:

[0087]

[0088] Among them, H represents the number of heads of the attention mechanism. respectively represent the query, key, and value in head l, and N k (i, j, k) represents the pixels in the local area centered on i, j, and k, and A(*) represents the attention weight of the corresponding area related to the feature.

[0089] S104: Introduce two weight coefficients, multiply the aggregated features output by S102 and S103 by the weight coefficients respectively and then add them to obtain the final output result of a CAMM:

[0090] F out = αF att + βF conv

[0091] Among them, F conv represents the feature map result through the attention branch, and F att represents the feature map result through the convolutional displacement branch, and α and β respectively represent the weight coefficients of the two branches.

[0092] In this embodiment, the two weight coefficients α and β are learnable scalars. During the training process of the model, the two weight coefficients are first initialized and then updated during the network iteration process. The outputs of the two branches are combined according to the updated weight coefficients α and β, which reflects the bias of the model towards convolution or self-attention at different depths.

[0093] S11: Perform feature fusion on the 3 feature maps and send them into the classifier, and use the Softmax function to process the results to obtain the final classification output result:

[0094]

[0095] S12: Calculate the loss function of the model according to the classification result, and use the CE loss function:

[0096]

[0097] Among them, N is the number of categories, and p i represents the predicted category, and g i represents the true category.

[0098] Update the parameters of the encoder, CAMM, and classifier through backpropagation of gradients. When the loss function converges to the minimum value, the training of the model is completed, and the parameters of the classification part in the trained multi-task implant classification model are obtained.

[0099] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium. The storage medium can include: ROM, RAM, magnetic disk, optical disc, etc.

[0100] The above-mentioned embodiments further elaborate on the purpose, technical solution, and advantages of the present invention. It should be understood that the above-mentioned embodiments are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made to the present invention within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A digital dental implant classification method based on deep learning, characterized in that, Including: CBCT images obtained in real time, and preprocessing the CBCT images; Inputting the preprocessed CBCT images into the segmentation branch of the trained multi-task dental implant classification model for binary segmentation; Cropping the implant area according to the segmentation result to obtain a cropped image and a Mask area; connecting the cropped image and the Mask area in channels, and inputting the connected image into the classification branch of the multi-task dental implant classification model to obtain a classification result; the multi-task dental implant classification model includes an encoder, a fully associated feature enhancement module, a decoder, a segmentation layer, 3 convolutional attention mixing modules, and a classifier; Among them, the encoder, the fully associated feature enhancement module, the decoder, and the segmentation layer serve as the segmentation branch, and the encoder, 3 convolutional attention mixing modules, and the classifier serve as the classification branch; The process of training the multi-task dental implant classification model includes: S1: Obtaining a CBCT image dataset, and preprocessing the images in the dataset to obtain a training set and a test set; S2: Inputting the CBCT images in the training set into the segmentation branch of the multi-task dental implant classification model for training to obtain the optimal parameters and segmentation results in the segmentation branch; In the segmentation branch, the process of processing the features output by the decoder by the fully associated feature enhancement module includes: Step 1: Inputting the feature layer X output by the last layer of the encoder into the fully associated feature enhancement module, performing a dilated convolution with a dilation coefficient of 1, and saving its output feature map Y1; Step 2: Adding the output feature map Y1 and X and performing a dilated convolution with a dilation coefficient of 2, and saving the output feature map Y2; Step 3: Adding Y2 and X and performing a dilated convolution with a dilation coefficient of 4 to save the output feature map Y3, and fusing the features of X, Y1, Y2, and Y3 to obtain an enhanced feature map; S3: Inputting the segmentation results output by the segmentation branch into the classification branch of the multi-task dental implant classification model for training to obtain the optimal parameters and classification results of the classification branch; completing the training of the model; In the classification branch, processing the input feature map through the convolutional attention mixing module CAMM, specifically including: Step 1: Mapping the input feature map using three 1×1×1 convolutions and reshaping it into N blocks to obtain an intermediate feature containing 3×N feature maps; Step 2: For the convolutional branch, if the size of its convolutional kernel is K, a fully connected layer is used to obtain K 3 feature maps, where one feature map is denoted as and N convolutional output feature maps are obtained by shifting; denotes the feature map, p, q, r denote the coordinates of the weights in the convolutional kernel, and i, j, k denote the coordinates of the pixels in the feature map; Step 3: For the self-attention branch, dividing the intermediate feature into N groups, each group containing three parts of features, which serve as query, key, and value respectively, using the multi-head self-attention module to obtain the output layer feature map of the self-attention mechanism, and then obtaining the aggregated feature by means of feature splicing; Step 4: Setting two weight coefficients, multiplying the N convolutional output feature maps and the aggregated feature by the corresponding weight coefficients respectively and then adding them to obtain a final output result of a CAMM; S4: Inputting the data in the test set into the trained multi-task dental implant classification model for testing.

2. The digital dental implant classification method based on deep learning according to claim 1, characterized in that The process of training the segmentation branch of the multi-task dental implant classification model includes: S21: Input the CBCT images in the training set into the encoder for multi-resolution feature extraction to obtain multiple feature maps with different resolutions; S22: Perform skip connections on the feature maps output by the encoder except the last layer; S23: Input the feature map output by the last layer of the encoder into the fully associated feature enhancement module for feature enhancement to obtain the enhanced feature map; S24: Input the feature map after skip connection and the enhanced feature map into the decoder, and perform upsampling to obtain the restored full-size feature map; S25: Input the full-size feature map into the segmentation output layer, and use the Softmax function to process the output image to obtain the segmentation result; S26: Calculate the segmentation branch loss function of the model. Through backpropagation of gradients, continuously update the parameters of the encoder, decoder, and fully associated feature enhancement module. When the loss function converges to the minimum value, complete the training of the model to obtain the parameters of the segmentation part in the trained multi-task implant classification model.

3. The digital dental implant classification method based on deep learning according to claim 2, wherein, The expression for calculating the segmentation branch loss function is: L Totle = λ FL L FL + λ Dice L Dice Among them, N represents the total number of voxels, n represents the number of categories, and i represents the semantics of the i-th category. represents the predicted category of the i-th voxel. represents the true category of the i-th voxel, and γ is the hyperparameter of FocalLoss, λ FL and λ Dice are two hyperparameters for balancing the loss, and L FL is the expression of the Focal loss function, and L Dice is the expression of the Dice loss function.

4. A digital dental implant classification method based on deep learning according to claim 2, characterized in that The expression for the segmentation result is: where X represents each pixel of the output of the decoder, X C represents the output corresponding to the semantic of the C-th category in X, n represents the number of categories, and i represents the semantic of the i-th category.

5. A digital dental implant classification method based on deep learning according to claim 1, characterized in that, The process of training the classification branch includes: S31: Crop the input CBCT image for a single implant area according to the segmentation result, connect the cropped image with the Mask area into 2 channels, and use the connected image as the input of the classification branch; S32: Input the image of a single implant into the encoder of the model for feature extraction, and save the feature map of this resolution before each downsampling to obtain multiple feature maps with different resolutions; S33: Input the feature maps of the last three layers of the encoder into 3 Convolutional Attention Mixing Modules (CAMM) respectively to obtain 3 enhanced feature maps; S34: Fuse the 3 enhanced feature maps, input the fused feature map into the classifier to obtain the classification result; S35: Calculate the classification branch loss function of the model. Through backpropagation of gradients, update the parameters of the encoder, CAMM, and classifier. When the loss function converges to the minimum value, complete the training of the model to obtain the parameters of the classification part in the trained multi-task implant classification model.

6. The digital dental implant classification method based on deep learning according to claim 1, characterized in that, The expression for N convolutional output feature maps is: Among them, represents the feature vector, [.] represents rounding down, and K represents the convolutional kernel parameter.

7. A digital dental implant classification method based on deep learning according to claim 1, characterized in that, The formula for aggregating features is: Among them, H represents the number of heads of the attention mechanism, respectively represent the query, key, value, and N in head l k (i, j, k) represents the pixels within the local area centered on i, j, k, and A(*) represents the attention weight of the corresponding area related to the feature.

8. A digital dental implant classification method based on deep learning according to claim 1, characterized in that The expression for the classification branch loss function of the model is: Among them, N is the number of categories, and p i represents the predicted category, and g i represents the true category.

Images