A tooth instance segmentation method and device based on anatomical morphology prior

Abstract

The application discloses a tooth instance segmentation method and device based on anatomical morphology prior, and belongs to the field of computer vision and artificial intelligence. The method trains a model on a CBCT image, processes a real segmentation mask into a three-class segmentation mask with enamel as a boundary as a supervision signal, enables the network to directly output a predicted segmentation mask containing tooth boundary information in an inference stage, and generates a detection frame according to the predicted segmentation mask. The detection frame is scaled by the same proportional coefficient and projected back to the original image, and each tooth is cut to obtain a single tooth CBCT image. In the segmentation training of the single tooth CBCT image, a loss function based on anatomical morphology prior is introduced, the network is constrained by an adaptively extracted morphological skeleton, the continuity of the predicted segmentation mask at the root tip part is promoted, and the tooth root segmentation fracture problem is effectively alleviated. The method realizes accurate positioning and high-quality segmentation of tooth instances, and makes the tooth instance segmentation in a full-mouth CBCT image more rapid and accurate.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and artificial intelligence, and specifically relates to a method and apparatus for segmenting tooth instances based on anatomical morphology priors. Background Technology

[0002] In the field of oral medical image analysis, tooth instance segmentation technology based on cone-beam computed tomography (CBCT) images plays a crucial role in clinical diagnosis, treatment planning, and digital oral medicine.

[0003] However, existing deep learning methods have significant shortcomings in handling tooth instance segmentation. Traditional instance segmentation networks are primarily designed for natural scene images, lacking consideration for the specific characteristics of medical images. More importantly, currently widely used loss functions, such as multi-class cross-entropy loss, focal loss, and multi-class Dice loss, mainly focus on pixel-level classification accuracy and region overlap, neglecting the unique anatomical features of teeth. As an anatomical unit with specific morphological structures, the slender and curved characteristics of teeth, as well as the continuity between the crown and root, are not effectively utilized by existing loss constraints. This deficiency prevents the network from fully learning the intrinsic morphological rules of teeth during the training phase, especially when dealing with detailed structures such as the root, which easily leads to segmentation fragmentation problems. Summary of the Invention

[0004] To address the shortcomings of existing technologies in tooth instance segmentation, in a first aspect, this invention proposes a tooth instance segmentation method based on anatomical morphology prior, comprising the following steps:

[0005] Step 1: Collect raw CBCT images containing various different fields of view and clinical conditions. Clean, preprocess, and annotate the teeth and pulp of the raw CBCT images. The processing yields a single tooth CBCT image containing at most one complete tooth and the corresponding tooth and pulp annotations.

[0006] Step 2: Take the labeled original CBCT images for tooth localization module training. Process the original segmentation mask according to the anatomical morphology prior, and divide it into three categories: background, enamel, and dentin, with the enamel as the boundary. Downsample the three-category segmentation mask and the corresponding original CBCT image at the same ratio coefficient to obtain the training image and its corresponding real segmentation mask that match the network input size. This image is used to perform supervised training on the tooth localization module of the segmentation model.

[0007] Step 3: In the inference phase of the tooth localization module, the input image is downsampled using the same scaling factor as in the training phase. The predicted segmentation mask is obtained through forward propagation inference using the training weights, and the detection box is extracted from the predicted segmentation mask.

[0008] Step 4: Project the detection box onto the original CBCT image to complete tooth localization and crop to obtain a single tooth CBCT image and the corresponding tooth and pulp true segmentation mask;

[0009] Step 5: Use the CBCT image of a single tooth obtained by cropping, which contains at most one complete tooth, to train the single tooth segmentation module of the segmentation model;

[0010] Step 6: Construct morphological constraint loss based on anatomical priors : From the accuracy of the predicted skeleton True skeleton sensitivity The calculation yielded, where To determine the proportion of the intersection between the predicted skeleton and the actual segmentation mask to the predicted skeleton, The proportion of the intersection of the real skeleton and the predicted probability map to the real skeleton; based on pixels. In category The actual segmentation mask below ,category Predicted probability Constructing multi-class cross-entropy loss and Dice loss The above losses are weighted and calculated to form a joint loss function, which is used to train the single tooth segmentation module of the segmentation model.

[0011] Step 7: The trained segmentation model's single-tooth segmentation module is used for inference on the single-tooth CBCT image obtained in Step 4. Softmax is used to output the probability distributions of teeth, pulp, and background, and thresholding is applied to generate the final single-tooth and pulp prediction segmentation mask. Based on the prediction mask of each instance tooth and its corresponding pulp, connected component analysis or morphological processing methods are used for post-processing to remove noise regions and ensure that each instance contains only one tooth. Combining the detection box information provided by the tooth localization module, the localization detection boxes obtained from the prediction segmentation mask are projected onto the coordinate system of the original CBCT image to achieve the final tooth instance segmentation output of the original CBCT image. Each detection box is loaded sequentially on the original CBCT image, segmentation inference is performed on each tooth, and the prediction segmentation masks of all teeth and pulp are integrated to construct a complete tooth and pulp prediction segmentation mask.

[0012] Furthermore, the specific calculation process of the joint loss function described in step 6 is as follows:

[0013] ,

[0014] ,

[0015] ,

[0016] in, For pixels In category The actual segmentation mask below; This represents the predicted probability for the corresponding category; These are weighting coefficients used to control the degree of influence of the three types of loss in the total loss; It is a smoothing constant; It is a joint loss function; It is a multi-class cross-entropy loss; It is a multi-class Dice loss; This is a morphological constraint loss based on anatomical priors;

[0017] ,

[0018] in, To improve the accuracy of the predicted skeleton, we define it as the proportion of the intersection between the predicted skeleton and the true segmentation mask to the predicted skeleton. The true skeleton sensitivity is defined as the proportion of the intersection of the true skeleton and the predicted probability map to the true skeleton.

[0019] ,

[0020] ,

[0021] For category The predicted probability map For the true segmentation mask, For adaptive morphological skeleton extraction operators, This indicates element-wise multiplication.

[0022] Furthermore, the scaling factor in step 2 is set within the range of [0.5, 0.75].

[0023] Furthermore, the adaptive morphological skeleton extraction operator The specific calculation formula is as follows:

[0024] ,

[0025] in, Given an input image tensor, The preset number of iterations, The skeleton increment for the k-th iteration is defined as:

[0026] ,

[0027] in, , This is the input data used to calculate the skeleton increment during the k-th iteration. , It was before The cumulative sum of the skeleton in each iteration This indicates that the tensors of the two are multiplied element-wise;

[0028] It is an adaptive opening operation.

[0029] ,

[0030] in For adaptive corrosion operation, For adaptive expansion operation;

[0031] ,

[0032] These are the max pooling operations performed with the sign reversed along different axes:

[0033] like ,

[0034] , For the longest axis index, the dilation kernel parameter is:

[0035]

[0036]

[0037]

[0038] For the Kronecker function: when hour, ,otherwise .

[0039] Furthermore, the method also includes: calculating the gradient value of the prior loss function based on anatomical morphology; updating the network parameters based on the gradient value; and repeatedly training using the updated network parameters until the loss function converges or a preset number of iterations is reached.

[0040] Secondly, the present invention proposes a tooth instance segmentation device based on anatomical morphology prior, the device comprising:

[0041] The tooth localization training module is used to train tooth localization on input CBCT images. It adopts a method based on boundary band segmentation and converts it into detection boxes. The real segmentation mask is divided into three categories: background, enamel and dentin, with the enamel as the boundary.

[0042] The tooth localization module is used to downsample the input CBCT image using the same scaling factor during the inference phase. The training weights obtained from the tooth localization training module are used for forward propagation inference to obtain the predicted segmentation mask, and the detection box is extracted from the predicted segmentation mask.

[0043] The image cropping module is used to project the detection box onto the original image to complete tooth localization and crop to obtain a single tooth CBCT image;

[0044] The single-tooth segmentation training module is used to train the segmentation model using cropped single-tooth CBCT images.

[0045] The joint loss calculation module is used to introduce joint loss during the training of the single-tooth segmentation module of the segmentation model, which is a morphological constraint loss based on anatomical morphology priors. From the accuracy of the predicted skeleton True skeleton sensitivity The calculation yielded, where To determine the proportion of the intersection between the predicted skeleton and the actual segmentation mask to the predicted skeleton, The proportion of the intersection of the true skeleton and the predicted probability map to the true skeleton; multi-class cross-entropy loss. and Dice loss From pixels In category The actual segmentation mask below ,category Predicted probability The calculated loss values ​​are weighted to form a joint loss. This module includes: a morphological skeleton extraction submodule, used for adaptive extraction of the morphological skeleton; and a loss constraint submodule, used for comprehensively calculating the morphological skeleton and the predicted segmentation mask to impose loss constraints on the segmentation model.

[0046] The single-tooth segmentation module is used to obtain a predicted segmentation mask by forward propagation inference using the CBCT image of a single tooth obtained by the image cropping module during the inference stage, and extract the pulp segmentation result of the single tooth from the predicted segmentation mask.

[0047] The gradient calculation module is used to calculate the gradient value based on the anatomical morphology prior loss function;

[0048] The parameter update module is used to update the network parameters based on the gradient values;

[0049] The iterative training module is used to repeatedly train the network using the updated parameters until the loss function converges or the preset number of iterations is reached.

[0050] This invention achieves a balance between speed and accuracy in tooth localization. Compared to simple boundary-band-based tooth localization methods, it can save 90% of inference time, enabling localization of all teeth within one minute. It also solves the problem of segmentation breakage in CBCT images, particularly at the ends of teeth and pulp, and in areas with unclear images. This results in more accurate segmentation masks for tooth and pulp segmentation prediction, achieving an average improvement of up to 1% in the Dice coefficient evaluation metric. Furthermore, the segmentation length at the pulp end is extended by 2-3 mm, effectively improving the model's segmentation performance. Attached Figure Description

[0051] Figure 1 This is an overall flowchart of the method of the present invention. Detailed Implementation

[0052] See Figure 1 This invention proposes a tooth instance segmentation method based on anatomical morphology priors, comprising the following steps:

[0053] Step 1: Collect raw CBCT images containing various different fields of view and clinical conditions. Clean, preprocess, and annotate the teeth and pulp of the raw CBCT images. The processing yields a single tooth CBCT image containing at most one complete tooth and the corresponding tooth and pulp annotations.

[0054] Step 2: Take the labeled original CBCT images for tooth localization module training. Process the original segmentation mask according to the anatomical morphology prior, and divide it into three categories: background, enamel, and dentin, with the enamel as the boundary. Downsample the three-category segmentation mask and the corresponding original CBCT image at the same ratio coefficient to obtain the training image and its corresponding real segmentation mask that match the network input size. This image is used to perform supervised training on the tooth localization module of the segmentation model.

[0055] Step 3: In the inference phase of the tooth localization module, the input image is downsampled using the same scaling factor as in the training phase. The predicted segmentation mask is obtained through forward propagation inference using the training weights, and the detection box is extracted from the predicted segmentation mask.

[0056] Step 4: Project the detection box onto the original CBCT image to complete tooth localization and crop to obtain a single tooth CBCT image and the corresponding tooth and pulp true segmentation mask;

[0057] Step 5: Use the CBCT image of a single tooth obtained by cropping, which contains at most one complete tooth, to train the single tooth segmentation module of the segmentation model;

[0058] Step 6: Construct morphological constraint loss based on anatomical priors : From the accuracy of the predicted skeleton True skeleton sensitivity The calculation yielded, where To determine the proportion of the intersection between the predicted skeleton and the actual segmentation mask to the predicted skeleton, The proportion of the intersection of the real skeleton and the predicted probability map to the real skeleton; based on pixels. In category The actual segmentation mask below ,category Predicted probability Constructing multi-class cross-entropy loss and Dice loss The above losses are weighted and calculated to form a joint loss function, which is used to train the single tooth segmentation module of the segmentation model.

[0059] Step 7: The trained segmentation model's single-tooth segmentation module is used for inference on the single-tooth CBCT image obtained in Step 4. Softmax is used to output the probability distributions of teeth, pulp, and background, and thresholding is applied to generate the final single-tooth and pulp prediction segmentation mask. Based on the prediction mask of each instance tooth and its corresponding pulp, connected component analysis or morphological processing methods are used for post-processing to remove noise regions and ensure that each instance contains only one tooth. Combining the detection box information provided by the tooth localization module, the localization detection boxes obtained from the prediction segmentation mask are projected onto the coordinate system of the original CBCT image to achieve the final tooth instance segmentation output of the original CBCT image. Each detection box is loaded sequentially on the original CBCT image, segmentation inference is performed on each tooth, and the prediction segmentation masks of all teeth and pulp are integrated to construct a complete tooth and pulp prediction segmentation mask.

[0060] Furthermore, the specific calculation process of the joint loss function described in step 6 is as follows:

[0061] ,

[0062] ,

[0063] ,

[0064] in, For pixels In category The actual segmentation mask below; This represents the predicted probability for the corresponding category; These are weighting coefficients used to control the degree of influence of the three types of loss in the total loss; It is a smoothing constant; It is a joint loss function; It is a multi-class cross-entropy loss; It is a multi-class Dice loss; This is a morphological constraint loss based on anatomical priors;

[0065] ,

[0066] in, To improve the accuracy of the predicted skeleton, we define it as the proportion of the intersection between the predicted skeleton and the true segmentation mask to the predicted skeleton. The true skeleton sensitivity is defined as the proportion of the intersection of the true skeleton and the predicted probability map to the true skeleton.

[0067] ,

[0068] ,

[0069] For category The predicted probability map For the true segmentation mask, For adaptive morphological skeleton extraction operators, This indicates element-wise multiplication.

[0070] Furthermore, the scaling factor in step 2 is set within the range of [0.5, 0.75].

[0071] Furthermore, the adaptive morphological skeleton extraction operator The specific calculation formula is as follows:

[0072] ,

[0073] in, Given an input image tensor, The preset number of iterations, The skeleton increment for the k-th iteration is defined as:

[0074] ,

[0075] in, , This is the input data used to calculate the skeleton increment during the k-th iteration. , It was before The cumulative sum of the skeleton in each iteration This indicates that the tensors of the two are multiplied element-wise;

[0076] It is an adaptive opening operation.

[0077] ,

[0078] in For adaptive corrosion operation, For adaptive expansion operation;

[0079] ,

[0080] These are the max pooling operations performed with the sign reversed along different axes:

[0081] like ,

[0082] , For the longest axis index, the dilation kernel parameter is:

[0083]

[0084]

[0085]

[0086] For the Kronecker function: when hour, ,otherwise .

[0087] Furthermore, the method also includes: calculating the gradient value of the prior loss function based on anatomical morphology; updating the network parameters based on the gradient value; and repeatedly training using the updated network parameters until the loss function converges or a preset number of iterations is reached.

[0088] The present invention also proposes a tooth instance segmentation device based on anatomical morphology prior, the device comprising:

[0089] The tooth localization training module is used to train tooth localization on input CBCT images. It adopts a method based on boundary band segmentation and converts it into detection boxes. The real segmentation mask is divided into three categories: background, enamel and dentin, with the enamel as the boundary.

[0090] The tooth localization module is used to downsample the input CBCT image using the same scaling factor during the inference phase. The training weights obtained from the tooth localization training module are used for forward propagation inference to obtain the predicted segmentation mask, and the detection box is extracted from the predicted segmentation mask.

[0091] The image cropping module is used to project the detection box onto the original image to complete tooth localization and crop to obtain a single tooth CBCT image;

[0092] The single-tooth segmentation training module is used to train the segmentation model using cropped single-tooth CBCT images.

[0093] The joint loss calculation module is used to introduce joint loss during the training of the single-tooth segmentation module of the segmentation model, which is a morphological constraint loss based on anatomical morphology priors. From the accuracy of the predicted skeleton True skeleton sensitivity The calculation yielded, where To determine the proportion of the intersection between the predicted skeleton and the actual segmentation mask to the predicted skeleton, The proportion of the intersection of the true skeleton and the predicted probability map to the true skeleton; multi-class cross-entropy loss. and Dice loss From pixels In category The actual segmentation mask below ,category Predicted probability The calculated loss values ​​are weighted to form a joint loss. This module includes: a morphological skeleton extraction submodule, used for adaptive extraction of the morphological skeleton; and a loss constraint submodule, used for comprehensively calculating the morphological skeleton and the predicted segmentation mask to impose loss constraints on the segmentation model.

[0094] The single-tooth segmentation module is used to obtain a predicted segmentation mask by forward propagation inference using the CBCT image of a single tooth obtained by the image cropping module during the inference stage, and extract the pulp segmentation result of the single tooth from the predicted segmentation mask.

[0095] The gradient calculation module is used to calculate the gradient value based on the anatomical morphology prior loss function;

[0096] The parameter update module is used to update the network parameters based on the gradient values;

[0097] The iterative training module is used to repeatedly train the network using the updated parameters until the loss function converges or the preset number of iterations is reached.

[0098] This invention achieves a balance between speed and accuracy in tooth localization. Compared to simple boundary-band-based tooth localization methods, it can save 90% of inference time, enabling localization of all teeth within one minute. It also solves the problem of segmentation breakage in CBCT images, particularly at the ends of teeth and pulp, and in areas with unclear images. This results in more accurate segmentation masks for tooth and pulp segmentation prediction, achieving an average improvement of up to 1% in the Dice coefficient evaluation metric. Furthermore, the segmentation length at the pulp end is extended by 2-3 mm, effectively improving the model's segmentation performance.

[0099] Example 1:

[0100] First, the model is trained by collecting at least 50 full-mouth or half-mouth CBCT images and marking them with realistic segmentation masks for the teeth and pulp regions. These masks are then input into a computational module based on anatomical morphology priors to extract tooth segmentation labels. Morphological erosion processing is then performed to obtain three-class segmentation masks for teeth, enamel, and dentin. The length, width, and height of these masks are downsampled by 0.5 times before training the tooth localization model based on anatomical morphology priors. For each cropped tooth image, single-tooth segmentation training is performed. During the training phase, the anatomical morphology prior loss uses an adaptive skeleton extraction method to obtain a consistency index between the predicted and real skeletons. In the specific calculation, the number of morphological erosion iterations k is set to 4, and the weights of each part in the joint loss function are... , and The values ​​are set to 0.5, 0.25, and 0.25. The gradient values ​​of the loss are calculated and backpropagation is performed to update the network parameters, so that the model can maintain the linear continuity and internal structure rationality of the predicted segmentation mask for teeth and pulp during subsequent inference segmentation.

[0101] Using a full-mouth CBCT image as input, with an image size of 512×512×300, the tooth instance segmentation method based on anatomical morphology prior of this invention is used to complete automatic segmentation:

[0102] The input CBCT image is first standardized and downsampled to accelerate initial localization. A pre-trained tooth localization network is used to obtain a predicted segmentation mask for the tooth region. Detection boxes are extracted from the dentin region within this mask. For each detection box, its length, width, and height are empirically expanded by a factor of 2, but not less than 1.2 times the minimum dimensions in the training dataset. The expanded detection boxes are then used to crop the original image, resulting in a series of single-tooth images, each containing at most one complete tooth. These single-tooth images are then fed into the single-tooth segmentation module of a segmentation model trained on anatomical morphology prior loss for inference, yielding predicted segmentation masks for the tooth and pulp.

[0103] After completing inference for all single-tooth regions, the system remaps the predicted segmentation masks for each tooth and pulp back to the original coordinate system, integrating them into a single instance predicted segmentation mask for all teeth and pulp. In the final output, each tooth and pulp structure is presented as an independent mask with clear boundaries and well-preserved structure, which can be directly used for clinical auxiliary diagnostic tasks such as orthodontic analysis and root canal identification.

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A tooth instance segmentation method based on anatomical morphology prior, characterized in that, Includes the following steps: Step 1: Collect raw CBCT images containing various different fields of view and clinical conditions. Clean, preprocess, and annotate the teeth and pulp of the raw CBCT images. The processing yields a single tooth CBCT image containing at most one complete tooth and the corresponding tooth and pulp annotations. Step 2: Take the labeled original CBCT images for tooth localization module training. Process the original segmentation mask according to the anatomical morphology prior, and divide it into three categories: background, enamel, and dentin, with the enamel as the boundary. Downsample the three-category segmentation mask and the corresponding original CBCT image at the same ratio coefficient to obtain the training image and its corresponding real segmentation mask that match the network input size. This image is used to perform supervised training on the tooth localization module of the segmentation model. Step 3: In the inference phase of the tooth localization module, the input image is downsampled using the same scaling factor as in the training phase. The predicted segmentation mask is obtained through forward propagation inference using the training weights, and the detection box is extracted from the predicted segmentation mask. Step 4: Project the detection box onto the original CBCT image to complete tooth localization and crop to obtain a single tooth CBCT image and the corresponding tooth and pulp true segmentation mask; Step 5: Use the CBCT image of a single tooth obtained by cropping, which contains at most one complete tooth, to train the single tooth segmentation module of the segmentation model; Step 6: Construct morphological constraint loss based on anatomical priors : From the accuracy of the predicted skeleton True skeleton sensitivity The calculation yielded, where To determine the proportion of the intersection between the predicted skeleton and the actual segmentation mask to the predicted skeleton, The proportion of the intersection of the true skeleton and the predicted probability map to the true skeleton; Based on pixels In category The actual segmentation mask below ,category Predicted probability Constructing multi-class cross-entropy loss and Dice loss The above losses are weighted and calculated to form a joint loss function, which is used to train the single tooth segmentation module of the segmentation model. Step 7: Use the single tooth segmentation module of the trained segmentation model for inference of the single tooth CBCT image obtained in Step 4. Use softmax to output the probability distribution of each category of tooth, pulp, and background, and generate the final single tooth and pulp prediction segmentation mask through thresholding. Based on the predicted mask of each instance tooth and its corresponding pulp, post-processing is performed using connected component analysis or morphological processing methods to remove noise regions and ensure that each instance contains only one tooth. Combining the detection box information provided by the tooth localization module, the localized detection boxes obtained from the predicted segmentation mask are projected onto the coordinate system of the original CBCT image to achieve the final tooth instance segmentation output of the original CBCT image. Each detection box is loaded sequentially on the original CBCT image, and segmentation inference is performed on each tooth. The predicted segmentation masks of all teeth and pulp are integrated to construct a complete tooth and pulp predicted segmentation mask.

2. The tooth instance segmentation method based on anatomical morphology prior as described in claim 1, characterized in that, The specific calculation process of the joint loss function described in step 6 is as follows: ， ， ， in, For pixels In category The actual segmentation mask below; This represents the predicted probability for the corresponding category; These are weighting coefficients used to control the degree of influence of the three types of loss in the total loss; It is a smoothing constant; It is a joint loss function; It is a multi-class cross-entropy loss; It is a multi-class Dice loss; This is a morphological constraint loss based on anatomical priors; ， in, To improve the accuracy of the predicted skeleton, we define it as the proportion of the intersection between the predicted skeleton and the true segmentation mask to the predicted skeleton. The true skeleton sensitivity is defined as the proportion of the intersection of the true skeleton and the predicted probability map to the true skeleton. ， ， For category The predicted probability map For the true segmentation mask, For adaptive morphological skeleton extraction operators, This indicates element-wise multiplication.

3. The tooth instance segmentation method based on anatomical morphology prior as described in claim 1, characterized in that, The scaling factor in step 2 is set within the range of [0.5, 0.75].

4. The tooth instance segmentation method based on anatomical morphology prior as described in claim 2, characterized in that, The adaptive morphological skeleton extraction operator The specific calculation formula is as follows: ， in, Given an input image tensor, The preset number of iterations, The skeleton increment for the k-th iteration is defined as: ， in, , This is the input data used to calculate the skeleton increment during the k-th iteration. , It was before The cumulative sum of the skeleton in each iteration This indicates that the tensors of the two are multiplied element-wise; It is an adaptive opening operation. ， in For adaptive corrosion operation, For adaptive expansion operation; , These are the max pooling operations performed with the sign reversed along different axes: ， , For the longest axis index, the dilation kernel parameter is: For the Kronecker function: when hour, ,otherwise .

5. The tooth instance segmentation method based on anatomical morphology prior as described in claim 1, characterized in that, The method further includes: calculating the gradient value of the prior loss function based on anatomical morphology; updating the network parameters based on the gradient value; and repeating the training using the updated network parameters until the loss function converges or a preset number of iterations is reached.

6. A tooth instance segmentation device based on anatomical morphology prior, characterized in that, The device includes: The tooth localization training module is used to train tooth localization on input CBCT images. It adopts a method based on boundary band segmentation and converts it into detection boxes. The real segmentation mask is divided into three categories: background, enamel and dentin, with the enamel as the boundary. The tooth localization module is used to downsample the input CBCT image using the same scaling factor during the inference phase. The training weights obtained from the tooth localization training module are used for forward propagation inference to obtain the predicted segmentation mask, and the detection box is extracted from the predicted segmentation mask. The image cropping module is used to project the detection box onto the original image to complete tooth localization and crop to obtain a single tooth CBCT image; The single-tooth segmentation training module is used to train the segmentation model using cropped single-tooth CBCT images. The joint loss calculation module is used to introduce joint loss during the training of the single-tooth segmentation module of the segmentation model, which is a morphological constraint loss based on anatomical morphology priors. From the accuracy of the predicted skeleton True skeleton sensitivity The calculation yielded, where To determine the proportion of the intersection between the predicted skeleton and the actual segmentation mask to the predicted skeleton, The proportion of the intersection of the true skeleton and the predicted probability map to the true skeleton; multi-class cross-entropy loss. and Dice loss From pixels In category The actual segmentation mask below ,category Predicted probability The calculated loss values ​​are weighted to form a joint loss. This module includes: a morphological skeleton extraction submodule, used for adaptive extraction of the morphological skeleton; and a loss constraint submodule, used for comprehensively calculating the morphological skeleton and the predicted segmentation mask to impose loss constraints on the segmentation model. The single-tooth segmentation module is used to obtain a predicted segmentation mask by forward propagation inference using the CBCT image of a single tooth obtained by the image cropping module during the inference stage, and extract the pulp segmentation result of the single tooth from the predicted segmentation mask. The gradient calculation module is used to calculate the gradient value based on the anatomical morphology prior loss function; The parameter update module is used to update the network parameters based on the gradient values; The iterative training module is used to repeatedly train the network using the updated parameters until the loss function converges or the preset number of iterations is reached.

Images