WTNet Pediatric Mandibular Wisdom Tooth Germ Segmentation Network Architecture Method

Abstract

This invention discloses a WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method, mainly comprising an input enhancement module with a region feature enhancement module and a bone feature separation module. Independent scale-specific feature fusion modules are provided in the tooth segmentation branch and the bone segmentation branch of the bone feature separation module, respectively. For an input and its corresponding ground truth label, the input is first fed into the input enhancement module to generate a mask and supervised using the ground truth label; then, the input and mask are fed into the region feature enhancement module to obtain enhanced input; in the bone feature separation module, the enhanced input is fed into the scale-specific feature fusion modules of the tooth segmentation branch and the bone segmentation branch respectively after passing through a shared encoder, and then fed into the decoders of the two branches. The tooth and bone ground truth labels are used to complete the supervision of the two branches respectively, obtaining tooth segmentation masks and bone segmentation masks. The two masks are fused to obtain the final segmentation result.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, specifically relating to a WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method. Background Technology

[0002] Lower wisdom tooth extraction often involves a large incision and requires a long recovery period for patients. If the immature lower wisdom tooth germ can be removed minimally invasively during adolescence, the pain and inconvenience caused by impacted lower wisdom teeth can be greatly reduced. Wisdom tooth germ resection is a common pediatric oral surgery. It involves removing the early-developing wisdom tooth germ using minimally invasive techniques, thus preventing the lower wisdom tooth from becoming impacted and significantly reducing the potential pain and inconvenience. However, this surgery faces several challenges in clinical practice: firstly, there is a lack of clinical consensus, leading doctors to rely on clinical experience to determine the necessity and optimal surgical location, resulting in significant subjectivity; secondly, due to a lack of strong reference points, some parents may adopt a wait-and-see attitude or harbor wishful thinking, delaying the optimal surgical time. Using artificial intelligence to assist doctors in determining surgical plans and providing suggestions would greatly improve these issues.

[0003] Patent CN114862771A proposes a method for wisdom tooth recognition and classification based on panoramic dental images using deep learning networks. This method first acquires panoramic X-ray images to construct an image dataset for preprocessing, then labels the wisdom teeth and divides the dataset to train a YOLO model to obtain local wisdom tooth feature maps. Next, based on the local wisdom tooth feature maps, the crown surface of the wisdom tooth is labeled, a set of labeled points is obtained, and linear regression is performed on the labeled point set to calculate the tooth axis direction. Classification is then performed based on a set threshold. Patent CN113139977A proposes a method for wisdom tooth segmentation based on transverse dental images using YOLO and U Net. This method preprocesses transverse dental images and performs positional labeling to obtain wisdom tooth location labels, which are then used to train a model to obtain the spatial location information of the wisdom teeth. Then, based on the spatial location information of the wisdom teeth, slice processing is performed. All slices containing wisdom teeth are preprocessed, and pixel-level category labeling is applied to the slices in the training set after preprocessing. The obtained data is then used to train a U Net model to finally complete the wisdom tooth segmentation.

[0004] Several common shortcomings exist in existing technologies. First, while CBCT images are widely used in dentistry, current research primarily focuses on panoramic dental images, lacking studies based on CBCT 3D images. Furthermore, existing technologies primarily study adult wisdom teeth in general dental clinics, neglecting research specifically on unerupted wisdom tooth germs in pediatric dentistry, creating a research gap in this area. Finally, existing technologies utilize open-source network models, without designing targeted deep learning network model architectures specific to the morphology and characteristics of wisdom teeth.

[0005] For wisdom tooth germ resection surgery, only a small portion of the mandibular wisdom tooth germ, the second molar, and the alveolar bone surrounding both needs to be dissected, rather than the entire alveolar bone. Therefore, guiding the model to dissect the mandibular wisdom tooth germ and the alveolar bone surrounding the second molar in the correct locations is also important. On the other hand, because tooth germs are uncalcified soft tissues, they are generally smaller in volume compared to the second molar, which in turn is smaller than the surrounding alveolar bone. These factors mean that if a model is to be used to dissect a patient's mandibular wisdom tooth germ, the second molar, and the alveolar bone surrounding both simultaneously, it must be able to address multi-scale issues. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method, which solves the problem of multi-scale segmentation by constructing a semantic separation scale-specific feature fusion network of WTNet.

[0007] This invention is achieved through the following technical solution:

[0008] The WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method includes the following steps:

[0009] S1: Construct a dataset, collect pediatric CBCT data and desensitize it, classify and annotate the mandibular wisdom tooth germ, second molar and alveolar bone surrounding the former two in each CBCT data, and divide it into training set, validation set and test set.

[0010] S2: Generate a mask for the region of interest in the original CBCT slice image using UNet, and convert the CBCT slice image annotation results into binary labels to supervise the mask generation process of UNet;

[0011] When the input is a 3D CBCT image data block, 3D-UNet is used to generate a mask for the region of interest, and binary labels from the CBCT 3D data annotation are used to supervise the mask generation process.

[0012] S3: Input the generated region of interest mask and the original input image into the region feature enhancement module to complete the region feature enhancement based on the original image, and obtain the enhanced input;

[0013] S4: The enhanced input image is fed into the shared encoder of the bone feature separation module. The shared encoder has four encoder layers to obtain shared multi-level image features, and a bottleneck layer to generate shared enhanced input hierarchical features.

[0014] S5: Input multi-level image features into the tooth segmentation branch and the bone segmentation branch respectively. Multi-scale feature learning and fusion are performed through the independent scale-specific feature fusion module in each branch. Channel attention mechanism and spatial attention mechanism are used to enhance the network feature learning ability.

[0015] S6: The features learned by each branch are fed into a specific decoder for decoding, and the tooth segmentation mask and bone segmentation mask are obtained by progressive upsampling;

[0016] S7: Fuse the tooth segmentation mask and the bone segmentation mask to obtain the final segmentation map;

[0017] S8: Iterative training of the S2-S7 process is performed using the training set, and the segmentation network is debugged and verified using the validation set and test set.

[0018] Furthermore, the original annotation is labeled with a label i, i∈{0,1,2,3}, representing the background, the mandibular wisdom tooth germ, the second molar, and the alveolar bone located around the mandibular wisdom tooth germ and the second molar, respectively. The binary label is generated by setting the label pixel value of the second molar and the alveolar bone located around the mandibular wisdom tooth germ and the second molar in the original annotation to 1.

[0019] Furthermore, the encoder used in the input enhancement module contains four encoder layers, and the number of feature channels of the four encoder layers sharing the encoder in the bone feature separation module is [64, 128, 256, 512].

[0020] Furthermore, the region feature enhancement module obtains the enhanced input X using the following formula:

[0021]

[0022] Where I is the input image; Y is the generated region of interest mask; Concat represents concatenation along the channel dimension; Conv 1×1(×1) This refers to a 1×1 (×1) convolutional layer with padding = 0; Indicates pixel-by-pixel multiplication; This indicates pixel-by-pixel addition.

[0023] Furthermore, the tooth segmentation branch is only responsible for segmenting the mandibular wisdom tooth germ and the second molar. By setting the alveolar bone label pixel value of the part surrounding the mandibular wisdom tooth germ and the second molar in the original annotation to 0, a ground truth label for the tooth segmentation branch is generated for supervision of the mask output of the tooth segmentation branch. The bone segmentation branch is only used to segment the alveolar bone surrounding the mandibular wisdom tooth germ and the second molar. By setting the pixel value of the mandibular wisdom tooth germ and the second molar label in the original annotation to 0, and setting the pixel value of the alveolar bone label pixel value of the part surrounding the mandibular wisdom tooth germ and the second molar to 1, a ground truth label for the bone segmentation branch is generated for supervision of the mask output of the bone segmentation branch.

[0024] Furthermore, the scale-specific feature fusion module concatenates the input multi-level features through a sampling layer to form intermediate feature 1. Intermediate feature 1 is then passed through a channel attention module to obtain intermediate feature 2. Intermediate feature 2 is then passed through a recovery layer to obtain an enhanced feature of the same size as the input multi-level features. In the recovery layer, four 1×1 (×1) convolutional layers are used to adjust the number of channels in intermediate feature 2 to align it with the input multi-level features. Various resampling layers are then used to restore the spatial resolution of intermediate feature 2 to match the spatial resolution of the multi-level features, thus obtaining enhanced features. Each enhanced feature is then added pixel-by-pixel to its corresponding multi-level feature. The result is then passed through four independent spatial attention layers to extract spatial features and obtain the final output feature.

[0025] Furthermore, the feature output process of the tooth branch scale-specific feature fusion module is as follows:

[0026] Intermediate feature 1 in the tooth branch scale-specific feature fusion module is represented by Z. T This can be obtained using the following formula:

[0027] Z T =Concat(AP k＝2 (F1),F2,UP k＝2 (F3),UP k＝4 (F4))

[0028] Where AP is the adaptive average pooling layer, UP is the bilinear upsampling layer, k is the scaling factor, and F1, F2, F3, and F4 are the multi-level features of the input.

[0029] Z T The image size is half the size of the input multi-level features;

[0030] The final output features of the tooth branch scale-specific feature fusion module It is obtained through the following formula:

[0031]

[0032] Where SA refers to the spatial attention module, Restore represents the recovery layer in the tooth branch scale-specific feature fusion module, and F... i For the hierarchical features of the input, The tooth branch intermediate feature 2 is the output of the channel attention module. This indicates pixel-by-pixel addition.

[0033] Furthermore, the feature output process of the skeletal branch scale-specific feature fusion module is as follows:

[0034] Intermediate feature 1 in the skeletal branch scale-specific feature fusion module is represented by Z. B This can be obtained using the following formula:

[0035] Z B =Concat(AP k＝2 (F1),UP k＝4 (F2), F3, UP k＝2 (F4))

[0036] Where AP is the adaptive average pooling layer, UP is the bilinear upsampling layer, k is the scaling factor, and F1, F2, F3, and F4 are the multi-level features of the input.

[0037] Z B The image size is one-quarter of the input multi-level features;

[0038] The final output features of the skeletal branch scale-specific feature fusion module It is obtained through the following formula:

[0039]

[0040] Where SA refers to the spatial attention module, Restore represents the recovery layer in the skeletal branch scale-specific feature fusion module, and F... i For the hierarchical features of the input, The tooth branch intermediate feature 2 is the output of the channel attention module. This indicates pixel-by-pixel addition.

[0041] Furthermore, each specific decoder in the tooth and bone feature separation module employs linear interpolation for upsampling and uses two 3×3 (×3) convolutional layers and ReLU activation for feature learning; specifically, the output D of each decoder in the tooth segmentation branch and the bone segmentation branch... i The following formula is used to calculate:

[0042]

[0043] Where Concat represents concatenation along the channel dimension; Conv3×3(×3) This refers to a 3×3 (×3) convolutional layer, where UP is a bilinear upsampling layer. D5 represents the output feature of the scale-specific feature fusion module, and D5 represents the output of the bottleneck layer in the shared encoder.

[0044] In the final output layer, the output feature maps in each branch are adjusted to the same resolution as the enhanced input image, and then a 1×1 (×1) convolutional layer is used to predict the segmentation result.

[0045] Furthermore, during the training phase, cross-entropy loss is employed. and Dice loss The combination of these is used as the loss function for each output layer:

[0046]

[0047]

[0048] Where P and G represent the predicted image and ground truth label of the input image, respectively, H×W provides the height and width of the output mask in pixels, and D... * τ represents the data depth when the input image is a 3D data block, and τ is the smoothing term.

[0049] The entire segmentation network model is trained using the following loss function:

[0050]

[0051] in and Let represent the loss functions for the binary mask, the tooth segmentation mask, and the bone segmentation mask, respectively.

[0052] Compared with the prior art, the beneficial effects of this invention are as follows:

[0053] 1. This invention is designed for the clinical characteristics of pediatric wisdom tooth germs and has powerful region of interest extraction capabilities and multi-scale feature learning and segmentation capabilities.

[0054] 2. Compared with other existing segmentation networks, this invention achieves the best results in the segmentation of pediatric wisdom tooth germs and can be used as an important prerequisite for preoperative auxiliary diagnosis in pediatric wisdom tooth germ resection.

[0055] Therefore, the WTNet deep learning architecture model provided by this invention can accurately segment CBCT images of the mandibular wisdom tooth germ region in pediatric oral surgery, which can lay an important foundation for artificial intelligence-assisted pediatric wisdom tooth germ resection, while further reducing the pain of children and contributing to improving the level of oral health in my country. Attached Figure Description

[0056] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0057] Figure 1 This is a schematic diagram of the segmentation network architecture of the present invention;

[0058] Figure 2 This is a schematic diagram of the segmentation network region feature enhancement module of the present invention;

[0059] Figure 3 This is a schematic diagram of the segmentation network scale-specific feature fusion module of the present invention;

[0060] Figure 4 This is a comparison diagram of the segmentation results of the segmentation network of this invention with those of other networks;

[0061] Figure 5 This is a comparison chart of network segmentation results with different configurations in the segmentation network ablation study of this invention. Detailed Implementation

[0062] Exemplary embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the invention to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0063] This invention uses a deep learning network to intelligently process and analyze CBCT images of the mandibular wisdom tooth region in pediatric oral cavity, and to segment and identify the mandibular wisdom tooth germ, the second molar, and a portion of the alveolar bone surrounding both.

[0064] This invention discloses a WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method, such as... Figures 1-3 As shown, the segmentation network mainly includes an Input Enhancement Block (IE module) and a Tooth Bone Features Separation Block (TBFS module).

[0065] The IE module aims to enhance the feature representation of the target segmentation region. It includes an encoder and a decoder, and also features a Regional Feature Enhancement Block (RFE) module to enhance the raw input, providing a stronger ROI functional representation and avoiding obfuscation. This module uses a binary ROI mask, allowing the shared encoder of the TBFS module to focus more on the target segmentation region.

[0066] The TBFS module utilizes a shared encoder and two independent decoders to predict segmentation masks for teeth and bones, respectively, thereby improving segmentation performance. To enable the independent branches in the TBFS module to learn appropriate multi-scale features, a Scale-Specific Feature Fusion Block (SFF) module is added to each of the two independent branches. This allows the tooth segmentation branch and the bone segmentation branch to adaptively select suitable features from the encoder, thus enhancing WTNet's multi-scale segmentation capabilities.

[0067] Specifically, given an image Where H, W, and C represent the image height, width, and channel number, respectively, and D... * Optional dimension: This field indicates the depth of the input data when the input is a 3D data block; it does not apply to 2D slices. Input image I to the IE module to obtain a binary ROI mask. Then input the image I and the binary mask Y into the RFE module to obtain enhanced input. In the TBFS module, the augmented input X is first sent to a shared encoder with four encoder layers and a bottleneck layer following the VGG16 approach (for 3D training, the shared encoder uses a 3D-UNet encoder) to generate hierarchical features for the augmented input X. The multi-level features of the encoder layers are represented as follows: Where i∈{1,2,3,4}. Then the multi-level features F i The data is sent separately to the SFF modules of the tooth segmentation branch and the bone segmentation branch, respectively, so that different branches can adaptively select multi-scale features and local spatial information for scale-specific feature fusion. Finally, the tooth segmentation mask is output. and skeletal segmentation mask The final segmentation result is obtained by fusing the tooth segmentation mask with the bone segmentation mask.

[0068] The implementation method is explained in detail step by step below:

[0069] S1: Collect large-scale pediatric CBCT data and complete data cleaning and desensitization. Classify and annotate the mandibular wisdom tooth germ, second molar, and alveolar bone surrounding the former two in each CBCT data set at the pixel level, and divide the data into training set, validation set, and test set to facilitate the training and validation of the segmentation network.

[0070] In the implementation of this invention, a CBCT dataset called NKUT was first constructed specifically for pediatric mandibular wisdom tooth germ segmentation tasks. This dataset contains 133 3D CBCT datasets, totaling over 53,000 slices. Patients ranged in age from 7 to 22 years, with a mean age of 13.2 years. All scans in the NKUT dataset were manually annotated in great detail by pediatric experts, encompassing three different pixel-level annotations: bilateral mandibular wisdom tooth germs (MWT), bilateral mandibular second molars (SM), and alveolar bone (AB) located around the bilateral mandibular wisdom tooth germs and second molars.

[0071] The NKUT dataset contains 34 cases aged 18-22 years. Firstly, to enhance the diversity of the NKUT dataset, including all developmental stages of the MWT, thereby improving the robustness of the NKUT dataset; secondly, to improve the generalization performance of the segmentation network model, so that the segmentation network model trained on the NKUT dataset can not only identify the MWT, but also identify all developmental stages of the mandibular wisdom tooth from tooth germ to complete calcification.

[0072] S2: For the labeled data, the feature representation of the region of interest is first enhanced in the input augmentation module to facilitate segmentation and recognition in the next stage of the network. Specifically, the labeled results are first converted into binary labels, and then a UNet or 3D-UNet is trained based on these labels. This enables the UNet to generate a mask for the region of interest of the input 2D slice image or 3D image patch. The binary labels are used to supervise the mask generation process.

[0073] The specific process is as follows: In the IE module, a UNet network with a VGG16 backbone is first used to obtain the binary ROI mask of the input CBCT slice image. For 3D image training, 3D-UNet is used. During this process, binary labels generated from the original annotations are used for supervision. Given an original annotation with label i, where i∈{0,1,2,3}, representing background, MWT, SM, and AB respectively. The binary labels used in the IE module can be generated by setting the label pixel values ​​of SM and AB in the original annotations to 1.

[0074] S3: Input the original image and the region of interest mask generated in S2 into the region feature enhancement module simultaneously to complete the region feature enhancement based on the original image.

[0075] The original input image I and the binary mask Y output by UNet are sent to the RFE module to obtain the augmented input X, which is obtained by the following formula (1):

[0076]

[0077] Where Concat represents concatenation along the channel dimension, which can effectively fuse the original image and the enhanced image [in the formula] and This approach leverages the characteristics of the ROI (Region of Interest) mask while also avoiding excessive influence of the ROI mask on the final segmentation result.

[0078] Conv 1×1(×1) This refers to a 1×1 (×1) convolutional layer with padding = 0, which can further fuse features and perform channel dimensionality reduction; Indicates pixel-by-pixel multiplication. This is pixel-by-pixel addition.

[0079] S4: The enhanced input image is fed into the shared encoder of the bone feature separation module to obtain shared multi-level image features. The shared encoder has four encoder layers and one bottleneck layer to generate the hierarchical features of the enhanced input.

[0080] A shared encoder is used to extract hierarchical features from enhanced slice images. Two independent, specific decoders are used to generate segmentation mask images for teeth and bones. Specifically, the tooth segmentation branch is only responsible for the segmentation of MWT and SM. By setting the AB label pixel values ​​in the original labels to 0, the ground truth label for the tooth segmentation branch is generated, which is used to supervise the output of the tooth segmentation masking process. The other bone segmentation branch is only used for AB segmentation. By setting the MWT and SM label pixel values ​​in the original labels to 0 and the AB label pixel values ​​to 1, the ground truth label for the bone segmentation branch is generated, which is used to supervise the output of the bone segmentation masking process. The multi-level feature map output by the shared encoder can be represented as F. i , i∈{1,2,3,4,5}, where F5 corresponds to the output of the bottleneck layer. Compared with traditional skip connections, which fuse features with the same spatial resolution only at a single scale, in the technical solution of this invention, {F1,F2,F3,F4} in WTNet will be sent to the SFF module for specific scale fusion.

[0081] S5: Input the multi-level image features obtained in the previous step into the tooth segmentation branch and the bone segmentation branch respectively. Perform multi-scale feature learning and fusion through the independent scale-specific feature fusion module in each branch, and use channel attention mechanism and spatial attention mechanism to enhance the network feature learning ability.

[0082] In the SFF module, the full-scale multi-level features {F1, F2, F3, F4} obtained from the shared encoder are used as input. Within each SFF module, the four outputs are resized to the same spatial resolution as the input multi-level features and then concatenated along the channel dimension. More specifically, in the tooth segmentation SFF module, {F1, F2, F3, F4} are first resized and concatenated to the intermediate feature 1 of the tooth segmentation branch, denoted as... Using formula (2), we get:

[0083] Z T =Concat(AP k＝2 (F1),F2,UP k＝2 (F3),UP k＝4 (F4)) (2)

[0084] Where AP, UP, and k represent the adaptive average pooling layer, the bilinear upsampling layer, and the scaling factor, respectively, and C f =C1+C2+C3+C4, where B represents the batch size.

[0085] Then, Z T The input channel attention (CA) module adaptively adjusts the attention scores for different feature scales. The CA module used in this invention is the open-source SENet (Squeeze-and-Excitation Networks). The intermediate feature 2 of the tooth segmentation branch output of the CA module is represented as... Size and Z T Same. The final output of the SFF module for tooth segmentation branches. The following formula (3) can be used to calculate:

[0086]

[0087] SA represents the spatial attention module. This indicates pixel-by-pixel addition.

[0088] The spatial attention (SA) module used in this invention is the open-source CBAM (Convolutional Block Attention Module) module.

[0089] The recovery layer in the SFF module is used to... Converted into four different enhanced features X i Each is associated with the hierarchical feature F i The spatial resolution and number of channels are matched. More specifically, in the recovery layer, four 1×1 (×1) convolutional layers are used to adjust... The number of channels, making it consistent with F iAlignment. Then, various resampling layers are used to recover. Spatial resolution, to match F i Spatial resolution. The entire process of the restoration layer of the tooth segmentation branch SFF module can be considered as the inverse operation of formula (2). Therefore, This will be transformed into four different X's. i Where i∈{1,2,3,4}, X i Size and F i Consistent. Then each X i All of them are their corresponding F i Pixel-by-pixel summation is performed. The result will then be processed through four independent SA layers to extract spatial features and obtain the final output. Used for skip connections at corresponding positions within the decoder.

[0090] Similarly, in the skeletal segmentation branch, the intermediate feature 1 of the SFF module is represented by Z. B express, The following formula (4) can be used to calculate:

[0091] Z B =Concat(AP k＝2 (F1),UP k＝4 (F2), F3, UP k＝2 (F4)) (4)

[0092] In the tooth segmentation branch, the resolution of intermediate feature 1 for all scales is adjusted to half the resolution of the input multi-level features, while in the bone segmentation branch, intermediate feature 1 is adjusted to one-quarter of the input feature resolution. This allows more low-level features to be retained in the tooth branch, while extracting more valuable information from the high-level features of the bone branch. Due to the divide-and-conquer approach of the TBFS module, both the tooth and bone branches can learn scale-specific features more effectively and better. The intermediate feature 2 output by the CA module in the bone branch is used... This indicates the final output of the SFF module in the skeleton segmentation branch. It can be calculated using formula (5). Here, Restore represents the recovery layer in the SFF module of the bone segmentation branch, which can be regarded as the inverse operation of formula (4).

[0093]

[0094] S6: The features learned by each branch are fed into a specific decoder for decoding, and then progressively upsampled to obtain the tooth segmentation mask and the bone segmentation mask.

[0095] S7: Merge the tooth segmentation mask and bone segmentation mask obtained from S6 to obtain the final segmentation map.

[0096] S8: Iteratively train the process from S2 to S7 using sliced ​​images from the training set, and debug and verify the segmentation network using the validation set and test set.

[0097] In each specific decoder during the 2D training and prediction phase, upsampling is performed using linear interpolation, followed by feature learning using two 3×3 (×3) convolutional layers and ReLU activation. Specifically, the output of each decoder in the tooth segmentation branch and the bone segmentation branch can be calculated using the following formula (6):

[0098]

[0099] Where D5 represents the output of the bottleneck layer F5 in the shared encoder. In the final output layer, the feature maps in each branch are adjusted to the same resolution as the augmented input X. Then, a 1×1 convolutional layer is used to predict the segmentation result, obtaining T∈ and Here, T and B represent the final segmentation results of the tooth segmentation branch and the bone segmentation branch, respectively. The 3D-UNet decoder is used directly as the specific decoder during the 3D training phase.

[0100] During the training phase, cross-entropy loss is used. and Dice loss The combination of these factors serves as the loss function for each output layer. and The definitions are as shown in formulas (7) and (8):

[0101]

[0102]

[0103] Where P and G represent the predicted image and ground truth label of the input image, respectively, H×W provides the number of pixels in the output mask, and τ is the smoothing term.

[0104] The entire architecture can be trained using the loss function of the following formula (9):

[0105]

[0106] in and Let represent the loss functions for the binary mask, the tooth segmentation mask, and the bone segmentation mask, respectively.

[0107] The NKUT dataset was preprocessed as follows for network training and testing: The window level and width of all CBCT images were adjusted to 800 and 2500, respectively. The HU values ​​of all CBCT images were normalized to the range [0, 255]. During the 2D training and testing phases, images and labels were extracted slice by slice along the horizontal plane from the original CBCT images. The input to the 2D network was a 3-channel RGB image of size 256×256. For the 3D training phase, due to GPU memory limitations, 150 sub-blocks of size 64×64×64 were randomly cut around the labeled region in each training set CBCT scan; therefore, the input to the 3D network was a single-channel grayscale CBCT image patch of size 64×64×64. During the testing phase of the 3D network, a non-overlapping sliding window was used to sequentially predict the 64×64×64 sub-blocks. In 2D and 3D testing, for pixels where there is ambiguity between tooth branches and bone branches, the label pixel value of the segmentation mask is directly set to 0 to generate the final segmentation mask.

[0108] To test the effectiveness of the WTNet segmentation network of this invention, it was used to segment the test set of the NKUT dataset against other existing segmentation networks, and the results were compared.

[0109] During implementation, data augmentation techniques including horizontal flipping, vertical flipping, and random rotation were employed to reduce overfitting. The Adam optimizer was used to train all models. The total number of training iterations was set to 200, with an initial learning rate of 0.0001. To ensure fairness, the same learning rate decay strategy was used for all models: if the training loss did not decrease within more than 3 training iterations, the learning rate was multiplied by 0.8 to reduce it. Furthermore, training was terminated if the validation loss did not decrease within more than 10 training iterations. The number of feature channels in the four encoder layers of the IE and TBFS modules were [64, 128, 256, 512] (the same settings were used for 2D and 3D).

[0110] The following metrics were used to evaluate system performance: mean intersection-to-union ratio (mIOU), pixel accuracy (Acc), dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and mean symmetric surface distance (ASSD).

[0111] Table 1: Performance comparison of different segmentation network models on the NKUT dataset

[0112]

[0113] Table 2: Comparison of segmentation results of different segmentation network models on the NKUT dataset

[0114]

[0115] In Table 2, |T| represents the critical value of T for the current test.

[0116] As can be seen from the segmentation results of these networks in Tables 1 and 2, the 2D version of WTNet of this invention achieves the best segmentation results.

[0117] Visual comparison results as follows Figure 4 As shown, WTNet-2D demonstrates excellent performance on the NKUT dataset. Compared to network models from other methods, it can segment complete teeth and bones with well-defined boundaries, and the segmentation results are most similar to the manual annotations of pediatric dentists. Thanks to the divide-and-conquer strategy and the powerful scale-specific feature fusion function in the TBFS module, WTNet can accurately segment teeth, tooth germs, and bones simultaneously. Compared to skip connections in UNet and its variants, the SFF module ensures scale-specific feature fusion, thus preserving the semantic integrity of the decoder. This is crucial for achieving higher segmentation accuracy and structural integrity, especially when identifying tiny tooth germs. The inclusion of the IE module enables WTNet to accurately localize without confusion with other adjacent teeth or bones.

[0118] To validate the effectiveness of the IE, TBFS, and SFF modules, ablation studies based on UNet were conducted to better understand the impact of each module component. The qualitative and quantitative segmentation results for each module are as follows: Figure 5 As shown. From Figure 5 This is more clearly visible in the visual results. Figure 5 The segmentation results shown in (f) reveal more complex details and clearer boundaries. The WTNet segmentation network of this invention no longer confuses MWT and SM, resulting in more comprehensive skeleton segmentation. The segmentation results of UNet+IE+TBFS+SFF (i.e., WTNet) are comparable to... Figure 5 (b) shows a very similar GT, highlighting the indispensable role played by each module in WTNet.

Claims

1. A WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method, characterized in that, Includes the following steps: S1: Construct a dataset, collect pediatric CBCT data and desensitize it, classify and annotate the mandibular wisdom tooth germ, second molar and alveolar bone surrounding the former two in each CBCT data, and divide it into training set, validation set and test set. S2: Generate a mask for the region of interest in the original CBCT slice image using UNet, and convert the CBCT slice image annotation results into binary labels to supervise the mask generation process of UNet; When the input is a 3D CBCT image data block, 3D-UNet is used to generate a mask for the region of interest, and binary labels from the CBCT 3D data annotation are used to supervise the mask generation process. S3: The generated region of interest mask and the original input image are fed into the region feature enhancement module to complete the region feature enhancement based on the original image, and the enhanced input is obtained; S4: The enhanced input image is fed into the shared encoder of the bone feature separation module. The shared encoder has four encoder layers to obtain shared multi-level image features, and a bottleneck layer to generate shared enhanced input hierarchical features. S5: Input multi-level image features into the tooth segmentation branch and the bone segmentation branch respectively. Multi-scale feature learning and fusion are performed through the independent scale-specific feature fusion module in each branch. Channel attention mechanism and spatial attention mechanism are used to enhance the network feature learning ability. The scale-specific feature fusion module concatenates the input multi-level features through a sampling layer to form intermediate feature 1. Intermediate feature 1 is then passed through a channel attention module to obtain intermediate feature 2. Intermediate feature 2 is then passed through a recovery layer to obtain an enhanced feature of the same size as the input multi-level features. In the recovery layer, four 1×1 2D convolutional layers or four 1×1×1 3D convolutional layers are used to adjust the number of channels of intermediate feature 2 to align it with the input multi-level features. Then, various resampling layers are used to restore the spatial resolution of intermediate feature 2 to match the spatial resolution of the multi-level features, thus obtaining enhanced features. Each enhanced feature is then added pixel by pixel to its corresponding multi-level features. The result will be processed by four independent spatial attention layers to extract spatial features and obtain the final output features. Intermediate feature 1 in the tooth branch scale-specific feature fusion module is used This can be obtained using the following formula: ； in AP For adaptive average pooling, UP Bilinear upsampling layer k As a scaling factor, The input consists of multi-level features; The image size is half the size of the input multi-level features; The final output features of the tooth branch scale-specific feature fusion module It is obtained through the following formula: ； in SA This refers to the spatial attention module. Restore This represents the recovery layer in the tooth branch scale-specific feature fusion module. For the hierarchical features of the input, The tooth branch intermediate feature 2 is the output of the channel attention module. This indicates pixel-by-pixel addition; Intermediate feature 1 in the skeletal branch scale-specific feature fusion module is used This can be obtained using the following formula: ； in AP For adaptive average pooling layer, UP It is a bilinear upsampling layer. k As a scaling factor, The input consists of multi-level features; The image size is one-quarter of the input multi-level features; The final output features of the skeletal branch scale-specific feature fusion module It is obtained through the following formula: ； in SA This refers to the spatial attention module. Restore This represents the recovery layer in the skeletal branch scale-specific feature fusion module. For the hierarchical features of the input, The skeletal branch intermediate feature 2 is the output of the channel attention module. This indicates pixel-by-pixel addition; S6: The features learned by each branch are fed into a specific decoder for decoding, and the tooth segmentation mask and bone segmentation mask are obtained by progressive upsampling; S7: Fuse the tooth segmentation mask and the bone segmentation mask to obtain the final segmentation map; S8: Iterative training of the S2-S7 process is performed using the training set, and the segmentation network is debugged and verified using the validation set and test set.

2. The WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method according to claim 1, characterized in that, Set labels for the original annotations , , respectively representing the background, mandibular wisdom tooth germ, second molar, and alveolar bone located around the mandibular wisdom tooth germ and the second molar. The binary label is generated by setting the label pixel value of the second molar and the alveolar bone located around the mandibular wisdom tooth germ and the second molar in the original annotation to 1.

3. The WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method according to claim 1, characterized in that, The encoder used in the UNet or 3D-UNet contains four encoder layers, and the number of feature channels of the four encoder layers that share the encoder in the bone feature separation module are [64, 128, 256, 512].

4. The WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method according to claim 1, characterized in that, The region feature enhancement module obtains the enhanced input X using the following formula: ； in Input image; A mask for the generated region of interest; Concat Indicates concatenation along the channel dimension; Conv 1×1（×1） This refers to a 1×1 2D convolutional layer or a 1×1×1 3D convolutional layer, with padding=0; Indicates pixel-by-pixel multiplication; This indicates pixel-by-pixel addition.

5. The WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method according to claim 2, characterized in that, The tooth segmentation branch is only responsible for segmenting the mandibular wisdom tooth germ and the second molar. By setting the alveolar bone label pixel value of the part of the mandibular wisdom tooth germ and the second molar in the original annotation to 0, the ground truth label of the tooth segmentation branch is generated and used for supervision of the tooth segmentation branch mask output. The bone segmentation branch is only used to segment the alveolar bone located around the mandibular wisdom tooth germ and the second molar. By setting the pixel values ​​of the mandibular wisdom tooth germ and the second molar label in the original annotation to 0, and setting the pixel values ​​of the alveolar bone label located around the mandibular wisdom tooth germ and the second molar to 1, the ground truth label of the bone segmentation branch is generated and used for supervision of the bone segmentation branch mask output.

6. The WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method according to claim 1, characterized in that, Each specific decoder in the tooth and bone feature separation module uses linear interpolation for up-sampling, uses two 3x3 2D convolution layers or two 3x3x3 3D convolution layers, and ReLU activation for feature learning; specifically, the output D i The following formula is used to calculate: ； in This indicates splicing along the channel dimension; This refers to a 3×3 2D convolutional layer or a 3×3×3 3D convolutional layer. UP It is a bilinear upsampling layer. The output features of the scale-specific feature fusion module This represents the output of the bottleneck layer in the shared encoder; In the final output layer, the output feature maps in each branch are adjusted to the same resolution as the enhanced input image, and then the segmentation results are predicted using a 1×1 2D convolutional layer or a 1×1×1 3D convolutional layer.

7. The WTNet pediatric mandibular wisdom tooth germ segmentation network architecture method according to claim 1, characterized in that, During the training phase, cross-entropy loss is used. and Dice loss The combination of these is used as the loss function for each output layer: ； in P and G These represent the predicted image and the ground truth label of the input image, respectively. H×W Provide the height and width of the output mask in pixels. The data depth when the input image is a 3D data block. τ For smoothing terms; The entire segmentation network model is trained using the following loss function: ； in , and Let represent the loss functions for the binary mask, the tooth segmentation mask, and the bone segmentation mask, respectively.

Images