Oral cbct image segmentation method based on MRDUNet model of double boundary loss

Abstract

The application provides an oral CBCT image segmentation method based on a MRDUNet model of double-boundary loss, and belongs to the technical field of tooth detection; solves the problem that an image segmentation algorithm in the prior art often has prediction errors when predicting tooth boundaries and tooth and root canal boundary regions, leading to an unsatisfactory segmentation effect; comprises the following steps: coarse segmentation of a tooth segmentation region of interest is performed through ResUNet to obtain a tooth region label file; then, MRDUNet proposed in the application is combined with double-boundary loss to perform double segmentation of teeth and root canals on the labeled region of interest; finally, a plurality of sets of comparative experimental results are discussed and analyzed; the application is applied to oral CBCT image segmentation.

Description

Technical Field

[0001] This invention provides a method for segmenting oral CBCT images based on the MRDUNet model with dual boundary loss, which is used for segmenting CBCT three-dimensional tooth and root canal images, and belongs to the field of dental detection technology. Background Technology

[0002] Projecting images of the entire or part of the oral cavity onto a two-dimensional image plane has certain limitations. In the diagnosis or treatment planning of special cases, advanced CBCT three-dimensional imaging can provide dentists with more information. High-precision three-dimensional model segmentation of teeth and root canals within the oral cavity provides more detailed hierarchical structures and contour shapes, better assisting clinicians in clinical diagnosis and preoperative planning. However, manually annotating teeth and root canals on medical image slices is not only prone to manual errors at structural boundaries but also extremely time-consuming and labor-intensive.

[0003] Artificial intelligence is being increasingly applied in the field of oral medicine. Its integration with oral imaging diagnostics can help improve the accuracy of imaging diagnosis, facilitate doctor-patient communication, preoperative diagnosis, intraoperative guidance, and postoperative evaluation, thus having significant clinical implications. Tooth region segmentation includes the segmentation of tooth and root canal regions and their boundaries. However, image segmentation algorithms often make prediction errors when predicting tooth boundaries and the tooth-root canal boundary region, resulting in unsatisfactory segmentation results. Summary of the Invention

[0004] To address the problem that existing image segmentation algorithms often make prediction errors when predicting tooth boundaries and the boundary between teeth and root canals, resulting in unsatisfactory segmentation results, this invention proposes a CBCT image segmentation method for the oral cavity based on the MRDUNet model with dual boundary loss.

[0005] To solve the above-mentioned technical problems, the technical solution adopted in this invention is: a CBCT image segmentation method for the oral cavity based on the MRDUNet model with dual boundary loss, comprising the following steps:

[0006] S1: The patient's oral cavity CBCT images of teeth and root canals were obtained by cone-beam CT of the maxillofacial region as a dataset. The ResUNet model was used for preliminary segmentation to determine the region of interest. During training, the tooth and root canal regions were treated as the same sample label for training to extract the tooth structure region.

[0007] S2: The Adam algorithm is used to optimize network training and obtain a coarse segmentation training model;

[0008] S3: Constructing a fine segmentation model: Based on DenseUNet and combined with multi-branch dilated convolution, an improved MRDUNet network model is proposed. The MRDUNet network model is used as a fine segmentation model. The MRDUNet network model is based on multi-branch convolution and dilated convolution to obtain the MRBlock structure. The MRBlock structure is introduced into the DenseUNet model to obtain the MRDUNet network model.

[0009] S4: Image segmentation is performed using a segmentation method based on the 2.5D concept. Adjacent slices are stacked together as input to the network to generate a segmentation map of the center slice. For each layer of segmentation, the adjacent layers are also used as multi-channel input.

[0010] S5: A dual-boundary loss function is proposed to address the characteristics of tooth and root canal images. The dual-boundary loss function increases the network training cost for the pixels at the edges of the tooth and root canal respectively. MRDUNet is combined with the dual-boundary loss function to perform dual segmentation of the labeled region of interest for the tooth and root canal.

[0011] S6: The Adam algorithm is used for network training and optimization, and the corresponding training parameters are set to obtain a fine segmentation model;

[0012] S7: Input the coarsely segmented image into the trained fine segmentation model to obtain segmented images of oral teeth and root canals.

[0013] In step S1, the ResUNet model is obtained through the following steps: the convolutional block of the encoder in the U-Net network is replaced with a residual unit Res_Conv. At the beginning of the Res_ConvA module, a 1x1 convolution is designed. Then, the model feature matrix is ​​transformed in the channel dimension through the 1x1 convolution. Finally, it is input into the residual block for parameter training.

[0014] The MRBlock structure in step S3 is as follows:

[0015] There are four branches after the Previous layer. The first branch includes a 1x1 convolutional layer and a 3x3 convolutional layer with a dilation rate of 1. The second branch includes a 1x1 convolutional layer, a 1x3 convolutional layer, and a 3x3 convolutional layer with a dilation rate of 2. The third branch includes a 1x1 convolutional layer, a 3x1 convolutional layer, and a 3x3 convolutional layer with a dilation rate of 3. The fourth branch includes a 1x1 convolutional layer. The first three branches are connected by a 1x1 convolutional layer and then output together with the fourth branch.

[0016] The structure of the MRDUNet network model is as follows:

[0017] The first convolutional layer, conv1, uses a 7x7 convolutional kernel with a stride of 2, and employs BN and ReLU.

[0018] The first pooling layer, pool1, uses max pooling with a step size of 2.

[0019] First layer MRBlock;

[0020] The first transition layer uses a 1x1 convolution and a 2x2 AvgPooling;

[0021] Second layer MRBlock;

[0022] The second transition layer uses a 1x1 convolution and a 2x2 AvgPooling;

[0023] Third layer MRBlock;

[0024] The third transition layer uses a 1x1 convolution and a 2x2 AvgPooling;

[0025] Third layer MRBlock;

[0026] The second convolutional layer uses a 1x1 convolutional kernel and employs BN and ReLU.

[0027] The formula for calculating the double-boundary loss function in step S5 is as follows:

[0028] BDLoss = α1xGDLoss + α2xTeethBoundaryTyLoss + α3xRootBoundaryTyLoss, where BDLoss represents the double-boundary loss function, GDLoss represents the loss value of the overlapping region of multiple categories, and TeethBoundaryLoss and RootBoundaryLoss represent the pixel loss of the tooth boundary and root canal boundary, respectively. α1, α2 and α3 are the corresponding loss weight ratios, representing the ratio of category weight loss, the ratio of tooth region boundary loss weight, and the ratio of root canal region boundary loss weight, respectively.

[0029] The parameters set in step S6 include the learning rate, the exponential decay rate of the first moment estimate, the exponential decay rate of the second moment estimate, and a minimum number to prevent the denominator from being zero.

[0030] The evaluation index of the coarse segmentation model is the Dice coefficient, which is calculated using the following formula:

[0031]

[0032] In the above formula: TP: True Positive, the correctly predicted part; FP: False Positive, the false positive part; FN: False Negative, the false negative part.

[0033] Morphological algorithms are used to label the tooth boundaries and tooth and root canal boundaries in oral CBCT images, and boundary mask labels are automatically generated. During network training, in addition to reading the original image and label data, the model also reads the boundary mask data.

[0034] The formula for calculating GDLoss is as follows:

[0035]

[0036]

[0037] In the above formula: r ln Let p represent the true category of category l at the nth pixel position, and p ln w represents the predicted probability value of the current position. l This represents the weight of each category, where l represents the category and n represents the position.

[0038] After obtaining the coarse segmentation training model, the training dataset for fine segmentation is input into the model, resulting in a statistical table of coordinate data of the ROI region of the tooth in CBCT. In subsequent fine segmentation experiments, the error range will be taken into account when annotating and statistically analyzing the ROI region of the tooth. A certain number of pixels will be added to the coordinates of the tooth region in each direction in the slice image. Based on the coordinate labeling data, the fine segmentation model will only be trained on the segmented ROI region.

[0039] The advantages of this invention over the prior art are as follows: When using deep learning algorithms to segment oral CBCT images, this invention strengthens the attention to boundary regions to obtain more boundary information, designs a loss function that is more suitable for segmenting boundary regions, and improves the overall segmentation performance of the image. Attached Figure Description

[0040] The present invention will be further described below with reference to the accompanying drawings:

[0041] Figure 1 This is a diagram showing the results of the validation set in the coarse segmentation experiment provided by this invention;

[0042] Figure 2 This is a diagram showing the ROI coordinate partitioning results of the coarse segmentation network model provided by this invention;

[0043] Figure 3 The diagrams provided by this invention show the layer structure (a) in the original DenseNet and the layer structure (b) in the MRBlock module.

[0044] Figure 4 This is a structural diagram of the MRDUNet network model provided by the present invention;

[0045] Figure 5 This is a comparison chart of the segmentation results of MRDUNet with and without boundary loss provided by this invention;

[0046] Figure 6 This is a diagram showing some of the validation set results during the training process provided by the present invention;

[0047] Figure 7 This is a diagram showing the training and validation set results of the ResUNet network training provided by this invention. Detailed Implementation

[0048] like Figures 1 to 7 As shown, this invention provides a method for segmenting oral CBCT images based on the MRDUNet model with dual-boundary loss, comprising the following steps:

[0049] S1: Obtain image regions containing tooth structures. During training, treat teeth and root canal regions as the same sample label for training. Extract tooth structure regions and use the ResUNet model for preliminary segmentation to determine regions of interest (POIs). This reduces the computational resource consumption of the refined segmentation model experiment and improves segmentation accuracy.

[0050] S2: To address the issue in S1 where adding direct residual units to the decoder of the ResUNet network might cause the network training to lose convergence, the convolutional blocks of the encoder in U-Net are replaced with residual units Res_Conv, which can also solve the gradient divergence and network degradation problems in deep networks.

[0051] S3: To address the issue of maintaining consistency in channel dimensions before and after short connection operations in the ResUNet network in S2, a 1x1 convolution needs to be designed at the beginning of the Res_ConvA module. Then, the model feature matrix is ​​transformed in terms of channel dimensions through the 1x1 convolution and finally input into the residual block for parameter training.

[0052] S4: The Adam algorithm is used for network training optimization.

[0053] S5: After obtaining the coarse segmentation training model, the training dataset for fine segmentation is input into the model, resulting in a statistical table of coordinate data for the CBCT tooth ROI regions. Subsequent fine segmentation experiments will consider the error range when annotating and statistically analyzing the tooth ROI regions, adding 20 pixels to the coordinates of each direction in the sliced ​​images of the tooth regions. Based on the coordinate-labeled data, the fine segmentation model will only be trained on the segmented ROI regions.

[0054] S6: An improved MRDUNet network model based on DenseUNet combined with multi-branch dilated convolution is proposed for fine segmentation.

[0055] S7: To address the limitation of using only 2D slice information in 3D CBCT image segmentation, a segmentation method based on 2.5D concepts is proposed.

[0056] S8: A dual-boundary loss function is proposed to address the characteristics of tooth and root canal images.

[0057] S9: Use the Adam algorithm to optimize network training and set parameters.

[0058] S10: Experimentally verify the effectiveness of the proposed model and loss function.

[0059] The network optimization steps using the Adam algorithm in step S4 are as follows:

[0060] The learning rate decayed to 0.9, and the learning rate was set to 0.0001. The original slice image size was 640x640. To adapt to network training and reduce training resource consumption, the image input size was scaled to 512x512 in this embodiment. The labels of the input images in the coarse segmentation model are [0,1,1], where the background label is 0, and the labels for teeth and root canal regions are both set to 1. To achieve the best training effect, the training batch size was set to 4, and the training was iterated 50 times. Finally, the trained segmentation model was saved for the acquisition and labeling of ROIs in the tooth region. The evaluation metric for the coarse segmentation experiment is the Dice coefficient. Its calculation method is as follows:

[0061]

[0062] In the above formula: TP: True Positive, the correctly predicted part; FP: False Positive, the false positive part; FN: False Negative, the false negative part.

[0063] The guiding metrics for training the ResUNet network in step S1 include:

[0064] The ResUNet network uses Dice_Per_Case and Dice_Global as guiding metrics for network training. Dice_Per_Case refers to calculating the Dice coefficient for each test sample and then averaging the results across all samples. Dice_Global refers to calculating the Dice coefficient over all test samples and then averaging all the results as the metric.

[0065] The MRDUNet network model is designed in step S6 as follows:

[0066] To design a module with strong representational power and low computational cost to replace the convolutional structure in the DenseLayer of the Backbone, an MRBlock structure based on multi-branch convolution and dilated convolution is designed to fully extract multi-scale features. This structure is then introduced into the DenseUNet model, resulting in the MRDUNet network model. MRDUNet integrates the advantages of multi-scale dilated convolution, dense connections, and UNet network connections. The MRBlock module retains the original DenseLayer structure as much as possible while adding branches with different dilation ratios to expand the receptive field. To reduce computation, 1x3 and 3x1 convolutions are used instead of 3x3 convolutions. Unlike DenseNet, long-range skip connections are added to the MRDUNet model to support low-level spatial feature preservation for better intra-slice context exploration. Furthermore, a grouped image input method is designed, and a composite feature fusion layer effectively extracts intra-slice and inter-slice features, achieving feature optimization. Through long-range feature connections, multiple feature fusion and optimization are achieved, thereby extracting features more suitable for teeth and root canals.

[0067] Step S7, the segmentation method based on 2.5D principles, specifically includes the following ideas:

[0068] Adjacent slices are stacked together and used as input to the network to generate a segmentation map of the central slice. For each layer's segmentation, several adjacent layers (from layer i+n to layer in) are also used as multi-channel input, similar to RGB in natural images. This saves computational resources and fully utilizes the spatial information of CBCT image slices to extract more comprehensive and expressive features of the tooth and root canal structures. This results in more accurate and comprehensive extraction of tooth feature information. In slices where the root canal first appears, the pixel information is not obvious and may be ignored during algorithm segmentation. However, simultaneously reading the current slice and several adjacent subsequent layers can enhance the representation of root structure information. In the upper and lower molars, the tooth topology changes significantly, and the shape in the slice is irregular. Simultaneously reading multiple layers of images can more accurately extract tooth features, resulting in more diverse features. Therefore, the segmentation method based on the 2.5D concept can overcome some of the shortcomings of 2D and 3D methods, improving the accuracy and stability of tooth segmentation.

[0069] The specific details of step S8, the double-boundary loss function, are as follows:

[0070] The Generalized Dice Loss (GDLoss) function, based on the Dice coefficient, is used as the fundamental loss function for network training. Dice Loss calculates the loss for segmentation of a specific class. However, for multiple segmentation scenarios, the Dice Loss for each class needs to be calculated. GDLoss calculates the mean or weighted mean of the Dice Loss for each target class, and finally uses a representative index integrated from all classes as the quantitative indicator of the segmentation result. Its calculation formula is as follows:

[0071]

[0072]

[0073] In the above formula: r ln Let p represent the true category of category l at the nth pixel position, and p ln w represents the predicted probability value of the current position. l This represents the weight of each category.

[0074] To optimize segmentation boundaries in medical image segmentation, this invention proposes a dual-boundary loss for tooth and root canal images. This loss function increases the training cost of the network for each pixel at the edge of the tooth and root canal, thereby enhancing the importance of these boundary pixels in the network's classification decisions. In this way, the proposed model can achieve more accurate and smoother boundary regions, resulting in higher segmentation performance.

[0075] This invention uses a morphological algorithm to label tooth boundaries and tooth-root canal boundaries in oral CBCT images and automatically generates boundary mask labels. In the boundary mask, the boundary between the background and the tooth, as well as the inner boundary of the tooth, is set to 1, the inner boundary of the root canal is set to 2, and other areas are set to 0. During network training, in addition to reading the original image and label data, the model also reads the boundary mask data to help the network learn boundary information more effectively.

[0076] The final loss function, BDLoss, is defined as:

[0077] BDLoss = α1xGDLoss + α2xTeethBoundaryTyLoss + α3xRootBoundaryTyLoss, where GDLoss represents the loss value of the overlapping region of multiple classes, while TeethBoundaryLoss and RootBoundaryLoss represent the pixel loss of the tooth boundary and root canal boundary, respectively. α1, α2, and α3 are the corresponding loss weight ratios, representing the weight loss ratios of the class region, tooth region boundary, and root canal region boundary, respectively. During model training, the values ​​of these weight ratios are adjusted to achieve better segmentation results.

[0078] The specific parameters for step S9 are as follows:

[0079] (1) Learning rate: The learning rate controls the speed of weight updates. This invention can choose a smaller learning rate value (e.g., 1.0E-5) to help the training process tend towards better performance; or choose a larger learning rate value (e.g., 0.3) to perform faster initial learning before updates.

[0080] (2) Exponential decay rate F1 (Beta1) of the first moment estimate: The Adam algorithm uses the first moment estimate (i.e., the average gradient) to estimate the gradient change. F1 controls this exponential decay rate, which is typically set to 0.9.

[0081] (3) Exponential decay rate F2 (Beta2) of the second moment estimate: The Adam algorithm uses the second moment estimate (i.e., the mean squared gradient) to estimate the gradient change. F2 controls this exponential decay rate, which is usually close to 1, typically taking a value of 0.999.

[0082] (4) e (Epsilon): e is a very small number to prevent the denominator from being 0. Generally, a number close to 0 is chosen, such as 10E-8.

[0083] During the training of MRDUNet, the initial learning rate was set to 0.00001, the momentum parameter was 0.99, and the weight coefficients were 0.0005. Furthermore, to avoid overfitting in the generator, 1 / 4 of the nodes in the generator's dropout layer were randomly selected for deletion. Considering resource consumption and preventing overfitting in the deep learning network model, an early stopping strategy was implemented in the fine-grained segmentation experiment, and the initial number of iterations was set to 50.

[0084] The evaluation metrics for the experimental test data in step S10 are as follows:

[0085] The average segmentation value is calculated using multiple metrics, including volumetric overlap error (VOE), relative volume difference (RVD), average symmetric surface distance (ASSD), and maximum symmetric surface distance (MSSD).

[0086] Evaluation Indicators Introduction, V seg V represents the predicted segmentation result. gt The gold standard segmentation mask:

[0087] (1) The Dice coefficient is a set similarity measure function, commonly used to calculate the similarity or overlap between two samples, and it is widely used in medical image segmentation. The Dice coefficient ranges from [0,1], where 0 represents no overlapping regions and 1 represents completely overlapping segments. Its calculation formula is as follows:

[0088]

[0089] (2) Volume overlap error (VOE) is essentially similar to the Jaccard index (also known as Intersection over Union, IoU), but in the field of medical imaging, it is more commonly used to evaluate segmentation performance due to its greater sensitivity to segmentation error rate. Its calculation formula is as follows:

[0090]

[0091] (3) Voxel relative error is an index used to describe the difference between two volumes. It is calculated by dividing the difference between the segmented volume and the true volume by the true volume, thus obtaining the ratio of the relative error. The calculation formula is as follows:

[0092]

[0093] (4) Average Symmetrical Surface Distance is a terrain analysis technique commonly used to measure the distance between two surfaces in medical images. This metric utilizes symmetry to symmetrically match the distances between two surfaces and calculates their average value to obtain the final distance value. The calculation formula is as follows:

[0094]

[0095] In the above formula: S(V gt ) represents V gt Surface voxels in the set. S(V) seg ) represents Vseg The surface voxels in the set. d(v,S(V) gt )) represents any voxel to S(V gt The shortest distance d(s) is. Vgt ,S(V seg )) represents V gt voxels in the set to S(V) seg The shortest distance d(s) is. Vseg ,S(V gt )) represents V seg voxels in the set to S(V) gt The shortest distance.

[0096] (5) The maximum surface distance at symmetrical locations, also known as the Hausdorff Distance, is a terrain analysis technique commonly used to measure the maximum distance between two surfaces in medical images. This metric utilizes symmetry to symmetrically match the distances between two surfaces and finds the maximum distance value, i.e., MSSD. Its calculation formula is as follows:

[0097]

[0098] In the above formula: S(V gt ) represents V gt Surface voxels in the set. S(V) seg ) represents V seg The surface voxels in the set. d(v,S(V) gt )) represents any voxel to S(V gt The shortest distance d(s) is. Vgt ,S(V seg )) represents V gt voxels in the set to S(V) seg The shortest distance d(s) is. Vseg ,S(V gt )) represents V seg voxels in the set to S(V) gt The shortest distance.

[0099] The proposed MRDUNet network structure is as follows:

[0100] The first convolutional layer, conv1, uses a 7x7 convolutional kernel with a stride of 2, and employs BN and ReLU.

[0101] The first pooling layer, pool1, uses max pooling with a step size of 2.

[0102] First layer MRBlock;

[0103] The first transition layer uses a 1x1 convolution and a 2x2 AvgPooling;

[0104] Second layer MRBlock;

[0105] The second transition layer uses a 1x1 convolution and a 2x2 AvgPooling;

[0106] Third layer MRBlock;

[0107] The third transition layer uses a 1x1 convolution and a 2x2 AvgPooling;

[0108] Third layer MRBlock;

[0109] The second convolutional layer uses a 1x1 convolutional kernel and employs BN and ReLU;

[0110] The input and output of the MRDUNet network structure are shown in Table 1 below:

[0111]

[0112] Table 1.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for segmenting oral CBCT images based on the MRDUNet model with dual-boundary loss, characterized in that: Includes the following steps: S1: The patient's oral cavity CBCT images of teeth and root canals were obtained by cone-beam CT of the maxillofacial region as a dataset. The ResUNet model was used for preliminary segmentation to determine the region of interest. During training, the tooth and root canal regions were treated as the same sample label for training to extract the tooth structure region. S2: The Adam algorithm is used to optimize network training and obtain a coarse segmentation training model; S3: Constructing a fine segmentation model: Based on DenseUNet and combined with multi-branch dilated convolution, an improved MRDUNet network model is proposed. The MRDUNet network model is used as a fine segmentation model. The MRDUNet network model is based on multi-branch convolution and dilated convolution to obtain the MRBlock structure. The MRBlock structure is introduced into the DenseUNet model to obtain the MRDUNet network model. The structure of the MRDUNet network model is as follows: The first convolutional layer, conv1, uses a 7x7 convolutional kernel with a stride of 2, and employs BN and ReLU. The first pooling layer, pool1, uses max pooling with a step size of 2. First layer MRBlock; The first transition layer uses a 1x1 convolution and a 2x2 AvgPooling; Second layer MRBlock; The second transition layer uses a 1x1 convolution and a 2x2 AvgPooling; Third layer MRBlock; The third transition layer uses a 1x1 convolution and a 2x2 AvgPooling; Third layer MRBlock; The second convolutional layer uses a 1x1 convolutional kernel and employs BN and ReLU; S4: Image segmentation is performed using a segmentation method based on the 2.5D concept. Adjacent slices are stacked together as input to the network to generate a segmentation map of the center slice. For each layer of segmentation, the adjacent layers are also used as multi-channel input. S5: A dual-boundary loss function is proposed to address the characteristics of tooth and root canal images. The dual-boundary loss function increases the network training cost for the pixels at the edges of the tooth and root canal respectively. MRDUNet is combined with the dual-boundary loss function to perform dual segmentation of the labeled region of interest for the tooth and root canal. S6: The Adam algorithm is used for network training and optimization, and the corresponding training parameters are set to obtain a fine segmentation model; S7: Input the coarsely segmented image into the trained fine segmentation model to obtain segmented images of oral teeth and root canals.

2. The oral CBCT image segmentation method based on the MRDUNet model with dual boundary loss according to claim 1, characterized in that: In step S1, the ResUNet model is obtained through the following steps: the convolutional block of the encoder in the U-Net network is replaced with a residual unit Res_Conv. At the beginning of the Res_ConvA module, a 1x1 convolution is designed. Then, the model feature matrix is ​​transformed in the channel dimension through the 1x1 convolution. Finally, it is input into the residual block for parameter training.

3. The oral CBCT image segmentation method based on the MRDUNet model with dual boundary loss according to claim 1, characterized in that: The MRBlock structure in step S3 is as follows: There are four branches after the Previous layer. The first branch includes a 1x1 convolutional layer and a 3x3 convolutional layer with a dilation rate of 1. The second branch includes a 1x1 convolutional layer, a 1x3 convolutional layer, and a 3x3 convolutional layer with a dilation rate of 2. The third branch includes a 1x1 convolutional layer, a 3x1 convolutional layer, and a 3x3 convolutional layer with a dilation rate of 3. The fourth branch includes a 1x1 convolutional layer. The first three branches are connected by a 1x1 convolutional layer and then output together with the fourth branch.

4. The oral CBCT image segmentation method based on the MRDUNet model with dual boundary loss according to claim 1, characterized in that: The formula for calculating the double-boundary loss function in step S5 is as follows: BDLoss = α1xGDLoss + α2xTeethBoundaryLoss + α3xRootBoundaryLoss, where BDLoss represents the double-boundary loss function, GDLoss represents the loss value of the overlapping region of multiple categories, and TeethBoundaryLoss and RootBoundaryLoss represent the pixel loss of the tooth boundary and root canal boundary, respectively. α1, α2 and α3 are the corresponding loss weight ratios, representing the ratio of category weight loss, the ratio of tooth region boundary loss weight, and the ratio of root canal region boundary loss weight, respectively.

5. The oral CBCT image segmentation method based on the MRDUNet model with dual boundary loss according to claim 1, characterized in that: The parameters set in step S6 include the learning rate, the exponential decay rate of the first moment estimate, the exponential decay rate of the second moment estimate, and a minimum number to prevent the denominator from being zero.

6. The oral CBCT image segmentation method based on the MRDUNet model with dual boundary loss according to claim 1, characterized in that: The evaluation metric for the coarse segmentation training model is the Dice coefficient, which is calculated using the following formula: ； In the above formula: TP: True Positive, the correctly predicted part; FP: False Positive, the falsely detected part; FN: False Negative, the falsely detected part.

7. The oral CBCT image segmentation method based on the MRDUNet model with dual boundary loss according to claim 1, characterized in that: Morphological algorithms are used to label the tooth boundaries and tooth and root canal boundaries in oral CBCT images, and boundary mask labels are automatically generated. During network training, in addition to reading the original image and label data, the model also reads the boundary mask data.

8. The oral CBCT image segmentation method based on the MRDUNet model with dual boundary loss according to claim 4, characterized in that: The formula for calculating GDLoss is as follows: ； ； In the above formula: r ln Indicate category l In the n The true category of each pixel location, and p ln This represents the predicted probability value for the current location. w l This represents the weight of each category. l Indicates category, n Indicates location.

9. The oral CBCT image segmentation method based on the MRDUNet model with dual boundary loss according to claim 1, characterized in that: After obtaining the coarse segmentation training model, the training dataset for fine segmentation is input into the model, resulting in a statistical table of coordinate data of the ROI region of the tooth in CBCT. In subsequent fine segmentation experiments, the error range will be taken into account when annotating and statistically analyzing the ROI region of the tooth. A certain number of pixels will be added to the coordinates of the tooth region in each direction in the slice image. Based on the coordinate labeling data, the fine segmentation model will only be trained on the segmented ROI region.

Images