Tooth CBCT image three-dimensional segmentation method and system

Abstract

The invention discloses a tooth CBCT image three-dimensional segmentation method and system, and relates to the technical field of image segmentation, and the method comprises the following steps: data collection and processing, image preprocessing, image registration, feature extraction and fusion, three-dimensional segmentation model construction and training, and segmentation effect evaluation. According to the method, dynamic MRI examination and generative adversarial network are introduced to generate virtual data, multi-task learning is combined, multiple modes and multi-time-point data are comprehensively utilized, analysis of soft tissue peripheral information is indirectly enhanced, and soft tissue lesion diagnosis is assisted; the dynamic change of teeth and surrounding tissues can be better captured through the attention to dynamic features in multi-time-point data acquisition, dynamic MRI image preprocessing, feature extraction and fusion and the synergistic effect of model training and evaluation links, and the limitation of CBCT as a static imaging technology in the aspect of evaluating dynamic related disease diagnosis and treatment effects is overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a method and system for three-dimensional segmentation of dental CBCT images. Background Art

[0002] The structure of teeth and their surrounding tissues is relatively complex. Through the three-dimensional segmentation of CBCT images, different tissues such as teeth, alveolar bone, pulp cavity, periodontal membrane, etc. can be clearly distinguished. Doctors can more accurately observe the morphology, position, and range of lesions of teeth, thereby making more accurate diagnoses of various diseases such as caries, apical periodontitis, alveolar bone absorption, impacted teeth, etc., providing a basis for formulating reasonable treatment plans.

[0003] However, CBCT mainly focuses on imaging hard tissues. Its imaging principle and resolution limitations result in relatively weak resolution of soft tissue details. Although the soft tissues around teeth can be distinguished from hard tissues through three-dimensional segmentation, it is difficult to clearly display and accurately segment the fine structures inside the soft tissues, such as the microvessels of the gums and the lesions of the mucosa. In some diagnosis and treatment involving soft tissue lesions, such as the early diagnosis of gingival cancer and the evaluation of oral mucosal diseases, CBCT image three-dimensional segmentation cannot provide sufficient information, which may affect the doctor's comprehensive understanding and accurate judgment of the disease. Moreover, CBCT is a static imaging technology that can only obtain images of teeth and surrounding tissues at a certain moment, and cannot capture the dynamic physiological processes of teeth and surrounding tissues like dynamic imaging examinations (such as fluorescence angiography, dynamic magnetic resonance imaging, etc.). For example, in tooth vitality testing and temporomandibular joint movement, CBCT image three-dimensional segmentation cannot directly reflect these dynamic changes. For the diagnosis of some diseases and evaluation of treatment effects related to dynamic physiological processes, such as pulp vitality recovery after pulp revascularization surgery and functional evaluation of temporomandibular joint disorders, the limitations of three-dimensional segmentation of CBCT images will make it impossible to fully and accurately evaluate the condition and treatment effects.

[0004] In view of this, a method and system for three-dimensional segmentation of dental CBCT images are provided to overcome the above problems. Summary of the invention

[0005] The object of the present invention is to provide a method and system for three-dimensional segmentation of dental CBCT images to solve the problems raised in the above background technology.

[0006] In order to solve the above technical problems, the present invention provides a method for three-dimensional segmentation of tooth CBCT images, comprising the following steps: Data collection and processing: CBCT scans were performed on the patients before and at the sixth month of orthodontic treatment, and a dynamic MRI examination was performed before the treatment began. The generator of the generative adversarial network was used to learn the distribution law of the data before and at the sixth month to generate virtual data for the third and ninth months. At the same time, a discriminator based on the physiological characteristics of teeth and surrounding tissues and the logic of orthodontic treatment changes was designed to evaluate the rationality of the virtual data. Multi-task learning was used to jointly train the time series prediction task and the tooth tissue segmentation task, and the spatial relationship between tissues in the segmentation task and the change pattern of orthodontic treatment were used to provide constraint information for the prediction model. Image preprocessing: Use median filtering algorithm to remove noise points and grayscale normalization on the acquired CBCT images, perform motion correction on dynamic MRI images, adjust contrast and brightness to highlight target structures, including but not limited to articular discs and muscles; Image registration: The nasal spine of the maxilla and the condyle of the mandible were selected as anatomical landmarks, and the CBCT images and dynamic MRI images were registered and calibrated using image registration software; Feature extraction and fusion: The attention mechanism in deep learning is used to extract the displacement, density, and angle change characteristics of key structures such as teeth, alveolar bones, and temporomandibular joints in CBCT images at different time points, including virtual data time points, as well as the displacement characteristics of articular discs and muscle contraction and relaxation characteristics in dynamic MRI images. The features are then weighted and averaged to form feature vectors based on their importance to segmentation and orthodontic treatment analysis. 3D segmentation model construction and training: A 3D convolutional neural network model is constructed based on the fused feature vectors. The dental CBCT image dataset with labeled information of different tissues is used as the training set to train the model. The validation set is used to monitor the model performance indicators to adjust the model hyperparameters. Segmentation effect evaluation: The CBCT image of the tooth to be segmented and the corresponding dynamic MRI image are input into the trained model to output the three-dimensional segmentation results of the teeth and surrounding tissues, which are then compared with the manually annotated gold standard segmentation results and evaluated using the Dice similarity coefficient and Hausdorff distance.

[0007] Furthermore, in data acquisition, specifically, when performing CBCT scanning, a real-time adaptive scanning resolution adjustment method is used. The scanning is performed at a conventional resolution at the beginning. As the scanning progresses, a preliminary analysis is performed on the acquired images. If a suspected microlesion area is found, the system automatically increases the scanning resolution of the area; for other areas, the original resolution is maintained or adjusted appropriately; At the same time, a deep learning model is used to preliminarily screen areas that may have tiny lesions in low-resolution images, so as to focus on these areas in subsequent high-resolution scanning and segmentation processes.

[0008] Furthermore, in data preprocessing, specifically when processing high-resolution CBCT images, local contrast enhancement processing and texture analysis algorithms are combined in one step; including, in the process of enhancing the contrast of areas prone to microlesions such as tooth surfaces, directly extracting texture features of these areas.

[0009] Furthermore, for areas where micro-lesions may exist, including but not limited to the enamel surface and near the pulp cavity, we focus on the newly extracted texture features and utilize the lesion possibility features output by the deep learning model.

[0010] Furthermore, when the weighted averaging method is used for feature fusion, an adaptive weight adjustment mechanism is designed; the weights are automatically adjusted according to the importance of the microlesion features at different stages, and the newly extracted microlesion-related features are weighted averaged together with other conventional features to form an updated fusion feature vector.

[0011] Furthermore, in the improvement and training of the 3D segmentation model: adding network layers to process the features related to small lesions, including introducing parameter adjustment mechanisms for small lesion features in the convolutional layer or the fully connected layer; The model was retrained using a dental CBCT image dataset containing known minor lesions, and during the training process, a high level of weight was given to the error in segmenting the minor lesion area.

[0012] Furthermore, in the evaluation of the segmentation effect: the tooth CBCT image to be segmented and the corresponding dynamic MRI image are input into the retrained 3D-CNN model, and the model outputs the three-dimensional segmentation results of the teeth and surrounding tissues, including the segmentation of teeth and alveolar bones, and also includes but is not limited to the separate labeling and segmentation of early caries areas and enamel microcracks areas.

[0013] Furthermore, the graph convolutional network is used in combination with the spatial information of the image to regard the CBCT image as a graph structure, in which each pixel is a node and the relationship between adjacent pixels is an edge. The local and global relationships between different regions in the image are learned through GCN. In the image preprocessing stage, GCN is used to enhance the image features and highlight the characteristics of the tiny lesion area.

[0014] A dental CBCT image three-dimensional segmentation system comprises: a data acquisition module, an image preprocessing module, an image registration module, a feature extraction and fusion module, a three-dimensional segmentation model construction and training module, a segmentation effect evaluation module, a microlesion targeted processing module, and an image space information utilization module.

[0015] Compared with the prior art, the present invention has the following beneficial effects: Solve the problem of data integration and dynamic change: Traditional segmentation methods are difficult to integrate CBCT image information at different time points of orthodontic treatment. By performing CBCT scans before the start of orthodontic treatment and in the sixth month, combined with a dynamic MRI examination before treatment, using generative adversarial networks (GANs) to generate virtual data for the third and ninth months, and designing discriminators and using multi-task learning, the data is effectively integrated. This solves the problem of difficulty in grasping the dynamic changes of teeth and surrounding tissues during treatment due to incomplete data, provides rich information for comprehensive analysis, and improves the accuracy of overall treatment effect evaluation and subsequent treatment plan formulation.

[0016] Accurately focus on key structural features: In order to solve the problem that existing technologies are difficult to accurately focus on key structural features, the attention mechanism in deep learning is used to automatically focus on key structures such as teeth, alveolar bones, and temporomandibular joints, extract multiple features at different time points, and fuse features through weighted averaging. This allows segmentation to be performed more accurately around key parts, significantly improving the accuracy and precision of segmentation results in key parts.

[0017] In order to solve the problem that CBCT has poor ability to distinguish soft tissue details, this method introduces dynamic MRI examination and uses the attention mechanism to focus on the changes of key structures closely related to soft tissue in subsequent feature extraction and fusion. This indirectly improves the understanding of the surrounding environment of soft tissue, provides doctors with more information in the diagnosis of soft tissue lesions, and makes up for the shortcomings of CBCT in distinguishing soft tissue details.

[0018] Breaking through the limitations of static imaging: Since CBCT is a static imaging technology that cannot capture dynamic physiological processes, this method can simulate and analyze dynamic physiological processes through multi-time point data acquisition and processing, the use of dynamic MRI images, the extraction and fusion of key structure dynamic features, and the training of a three-dimensional segmentation model based on rich information. This overcomes the limitations of CBCT in the diagnosis and treatment effect evaluation of dynamic-related diseases such as pulp vitality recovery after pulp revascularization surgery and functional evaluation of temporomandibular joint disorders.

[0019] This method is efficient and convenient, and simplifies the data collection process: by using GAN to generate virtual data, the actual collection times are reduced, the data collection process is simplified, and the data collection cost and workload are reduced while ensuring data integrity. The multi-task learning joint training method does not require complex additional data processing steps and can be implemented on the basis of conventional data processing processes, further simplifying the steps.

[0020] Simplify the feature extraction and fusion process: The attention mechanism can achieve accurate focus on key features in the conventional image data processing process without complex preprocessing or additional screening steps. The weighted average feature fusion method is simple to calculate and does not require complex algorithm adjustments or additional data processing steps. It simplifies the fusion steps while ensuring effective feature fusion.

[0021] The real-time adaptive scanning resolution adjustment method automatically adjusts the resolution according to the situation, improving the pertinence of data collection for small lesion areas. The improved deep learning model is used to screen small lesion areas in low-resolution images, further optimizing the efficiency and pertinence of data collection, increasing the initiative and accuracy of detection, and avoiding the loss of small lesion information.

[0022] Enhanced data preprocessing effect: Combining local contrast enhancement processing with texture analysis algorithm improves data processing efficiency, provides richer information for subsequent segmentation, helps to distinguish between small lesion areas and normal tissues, improves segmentation accuracy, and makes up for the shortcomings of traditional preprocessing algorithms in highlighting the characteristics of small lesions.

[0023] For the tiny lesion area, we focus on the newly extracted texture features and lesion possibility features, and design an adaptive weight adjustment mechanism to fully consider the importance of tiny lesion features when fusing features, thereby improving the accuracy and foresight of the tiny lesion area segmentation.

[0024] The 3D-CNN model was added with a network layer that specifically processes the features related to microlesions, and the model was retrained to give a higher weight to the segmentation error of the microlesion area, which enhanced the model's detection and segmentation performance for microlesions. Adding the evaluation of the accuracy of the segmentation of microlesions to the evaluation indicators can more comprehensively evaluate the model effect, timely identify deficiencies and make improvements, and help doctors detect microlesions earlier and more accurately and take measures to avoid the development of lesions.

[0025] This method provides doctors with more accurate and comprehensive three-dimensional segmentation results of teeth and surrounding tissues in orthodontics, disease diagnosis and treatment. Doctors can use these results to more clearly understand the structure, pathological conditions and dynamic changes of teeth and surrounding tissues, so as to more accurately formulate orthodontic treatment plans, diagnose oral diseases (including soft tissue lesions and minor lesions), and take effective preventive or treatment measures in a timely manner, thus improving the quality and effectiveness of oral medical care. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a schematic diagram of a method and system for three-dimensional segmentation of dental CBCT images according to the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] Traditional segmentation methods are often unable to effectively integrate CBCT image information taken at different time points during orthodontic treatment, resulting in insufficient understanding of the dynamic changes of teeth and surrounding tissues during treatment, affecting the accurate evaluation of the overall treatment effect and the precise formulation of subsequent treatment plans. When segmenting teeth and surrounding tissues, existing technologies are difficult to accurately focus on the features of key structures such as teeth, alveolar bones, and temporomandibular joints, resulting in poor accuracy and precision of segmentation results in key areas.

[0029] See also Figure 1 : Example 1: A three-dimensional segmentation method for dental CBCT images based on intelligent prediction and feature focusing In the fields of orthodontics and oral disease diagnosis, accurate 3D segmentation of dental CBCT images is crucial for understanding the structure and status of teeth and surrounding tissues. However, traditional segmentation methods often fail to fully utilize image information at different time points and accurately focus on key structural features, which affects the segmentation effect.

[0030] step: Data collection: Patients were scanned with CBCT before and at the sixth month of orthodontic treatment, and a dynamic MRI examination was performed before treatment. However, in order to explore ways to reduce data collection costs and ensure prediction results, a generative adversarial network (GAN) can be used to generate virtual data at intermediate time points. The generator in the GAN learns the distribution of data before and at the sixth month, and generates data similar to the real intermediate state (such as the third and ninth months) to make up for the data loss caused by reducing the time point collection. At the same time, a special discriminator structure is designed to enable it to evaluate the rationality of the generated data based on the physiological characteristics of the teeth and surrounding tissues and the logic of changes during orthodontic treatment, ensuring the effectiveness of the generated data for segmentation model training.

[0031] It is necessary to add that: The implementation method of designing a special discriminator structure to evaluate the rationality of generated data is: Assessment based on the physiological characteristics of teeth and surrounding tissues: Analysis of tooth structure characteristics: First, for the teeth in the generated virtual data, the geometric shape characteristics of the teeth are analyzed. For example, the contour of a normal tooth should conform to a specific mathematical model, including the curvature and length ratio of the crown and root of the tooth. By learning from a large amount of real tooth data, a standard template library of tooth shapes is established. The discriminator can calculate the similarity between the tooth shape in the generated data and the standard shape in the template library. If the similarity is within a reasonable range, it is considered reasonable in terms of the physiological characteristics of the tooth shape.

[0032] Check the density characteristics of teeth. Different types of teeth (such as incisors, molars, etc.) have specific density ranges, which can be determined by analyzing the grayscale value distribution of the tooth area in the real data. The discriminator can evaluate the tooth density in the generated data to ensure that its density value falls within the reasonable density range of the corresponding type of teeth.

[0033] Analysis of physiological characteristics of surrounding tissues: For the alveolar bone, we focus on its connection with the teeth and its own bone structure. The alveolar bone should fit the tooth root reasonably, and the bone density should show a specific gradient change around the tooth root. The discriminator can evaluate its rationality by analyzing the grayscale changes at the boundary between the alveolar bone and the teeth in the generated data and the texture features inside the alveolar bone. For example, the texture of a healthy alveolar bone should have a certain directionality and continuity, which is similar to the texture features in the real data.

[0034] Analyze soft tissue features such as the periodontal ligament. The periodontal ligament appears as a thin and uniform layer of soft tissue surrounding the tooth root in the image, and its thickness and grayscale value have a certain range. The discriminator determines the rationality of the data by detecting whether these features of the periodontal ligament in the generated data are consistent with the physiological situation.

[0035] Evaluation based on the logic of change during orthodontic treatment: Treatment phase feature matching: According to the different stages of orthodontic treatment (such as the initial stage, mid-term adjustment stage, late consolidation stage, etc.), teeth and surrounding tissues will have specific change patterns. For the time points claimed by the generated data (such as the 3rd month or the 9th month), the discriminator needs to check its rationality based on the known logic of orthodontic treatment changes. For example, in the early stages of orthodontic treatment, the movement of teeth is relatively slow, mainly due to the adaptive reconstruction of periodontal tissues. The generated data should reflect the initial change characteristics such as the absorption of alveolar bone on the pressure side and the proliferation of the tension side.

[0036] Compare the consistency between the data generated at different time points. If it is the data from the third month, it should have a reasonable transition relationship with the data before the start of treatment, and compared with the data from the sixth month, the amount of tooth movement and the degree of tissue change should be consistent with the time series logic during orthodontic treatment.

[0037] Evaluation of the effect of force: In orthodontic treatment, correction force is the key factor that causes tooth movement and tissue changes. The discriminator can reversely infer the applied correction force based on the direction and degree of tooth movement in the generated data. Then, based on the principles of orthodontic mechanics, check whether the inferred correction force is reasonable, including whether the magnitude, direction and point of action of the force are consistent with the actual clinical situation. For example, if the generated data shows that the teeth have a large degree of tilt movement, then there should be reasonable evidence of the effect of correction force in the corresponding direction, and the response of the surrounding tissue (such as the pressure and tension distribution of the alveolar bone) should also match this movement.

[0038] Through the above comprehensive method, the generated data is evaluated from two dimensions: the physiological characteristics of teeth and surrounding tissues and the logic of orthodontic treatment changes, so as to ensure the effectiveness of the generated data for segmentation model training. In actual implementation, this special discriminator can be constructed using machine learning algorithms (such as support vector machines, decision trees, etc.) or discriminant network structures in deep learning, and trained with a large amount of labeled data (including real collected data and verified generated data) to enable it to accurately evaluate the rationality of generated data.

[0039] In addition, multi-task learning can be used to jointly train the time series prediction task and the tooth tissue segmentation task. While reducing the time points for data collection, the spatial relationship between different tissues in the segmentation task (such as the relative position change law of teeth and alveolar bones) and the known orthodontic treatment change patterns (common changes in different stages) are used to provide additional constraint information for the prediction model, thereby improving the prediction accuracy.

[0040] Data preprocessing: CBCT image preprocessing: The median filter algorithm is used to remove noise points on each acquired CBCT image, and grayscale normalization is performed to ensure image clarity and grayscale consistency, which is convenient for subsequent segmentation operations and feature extraction.

[0041] Dynamic MRI image preprocessing: Motion correction is performed on dynamic MRI images to eliminate the impact of slight head movements that may occur during the examination. Target structures such as articular discs and muscles are then highlighted by adjusting parameters such as contrast and brightness, making these structures easier to identify and process in the subsequent fusion with CBCT images and feature extraction.

[0042] Establish a unified coordinate system: The nasal spine of the maxilla and the condyle of the mandible were selected as fixed anatomical landmarks, and professional image registration software was used to register and calibrate the CBCT images and dynamic MRI images with these anatomical landmarks as references, so that they were in the same spatial coordinate system. This step ensured the accurate correspondence of different modal images in spatial position, laying the foundation for subsequent three-dimensional segmentation based on multimodal information.

[0043] Feature extraction and fusion: Using the attention mechanism in deep learning, we can automatically focus on key structures such as teeth, alveolar bones, and temporomandibular joints. For CBCT images, we extract core features such as displacement changes, density changes, and angle changes of these key areas at different time points (including the generated virtual intermediate time point data); for dynamic MRI images, we extract the displacement features of the articular disc and the contraction and relaxation features of the muscles.

[0044] Then, the weighted average method is used for feature fusion, and the corresponding weights are assigned to the importance of different features for the three-dimensional segmentation of tooth CBCT images and orthodontic treatment analysis to form a fused feature vector. For example, the displacement change feature of the tooth is crucial for judging the movement of the tooth during the treatment process, and a higher weight can be assigned to better reflect the dynamic changes of the tooth during the segmentation process.

[0045] 3D segmentation model construction and training: Based on the fused feature vectors, a three-dimensional convolutional neural network (3D-CNN) model is constructed for the three-dimensional segmentation of dental CBCT images. The labeled dental CBCT image dataset (including accurate labeling information of different tissues such as teeth, alveolar bones, and dental pulp cavities) is used as a training set and input into the model for training. During the training process, the model continuously adjusts its own parameters by learning the relationship between the feature vectors and the corresponding tissue labels in different samples to achieve accurate segmentation of unknown samples.

[0046] Use the validation set to verify the model during training, monitor the model's performance indicators (such as Dice similarity coefficient, Hausdorff distance, etc.), and adjust the model's hyperparameters (such as learning rate, convolution kernel size, etc.) in a timely manner according to the validation results to ensure the model's generalization ability and segmentation accuracy.

[0047] Segmentation effect evaluation and application: The CBCT image of the tooth to be segmented and the corresponding dynamic MRI image (after preprocessing and coordinate alignment) are input into the trained 3D-CNN model. The model outputs the three-dimensional segmentation results of the teeth and surrounding tissues, including the segmentation of individual teeth, the boundary demarcation between the alveolar bone and the tooth, and the positioning of the dental pulp cavity.

[0048] By comparing with the manually annotated gold standard segmentation results, the 3D segmentation results are evaluated using evaluation indicators such as the Dice similarity coefficient and Hausdorff distance. Based on the accurate segmentation results, doctors can have a clearer understanding of the structure of teeth and surrounding tissues, providing important basis for the formulation of orthodontic treatment plans and the diagnosis of oral diseases, such as accurately judging the arrangement of teeth and whether there is alveolar bone resorption.

[0049] Example 2: A three-dimensional segmentation method for dental CBCT images to enhance detection of microlesions On the basis of Example 1, in order to more effectively detect minor lesions (such as early caries, enamel microcracks, etc.) that may occur in teeth and surrounding tissues during orthodontic treatment, it is necessary to further improve the three-dimensional segmentation method of tooth CBCT images to improve the segmentation accuracy and detection capability of minor lesion areas, and provide solutions to the possible problems of identifying minor lesion areas only by improving CBCT scanning parameters and image preprocessing algorithms.

[0050] step: Data collection (based on Example 1): When performing CBCT scanning, a real-time adaptive scanning resolution adjustment method is used. The scan is initially performed at a normal resolution, and as the scan progresses, the acquired images are preliminarily analyzed. If a suspected microlesion area is found, the system automatically increases the scanning resolution of the area; for other areas, the original resolution is maintained or adjusted appropriately.

[0051] At the same time, the improved deep learning model is used to preliminarily screen areas that may have tiny lesions in low-resolution images so that these areas can be focused on in subsequent high-resolution scanning and segmentation processes.

[0052] Data preprocessing (based on Example 1): When processing high-resolution CBCT images, local contrast enhancement processing is combined with texture analysis algorithms in one step. For example, in the process of enhancing the contrast of areas prone to microlesions such as tooth surfaces, texture features of these areas are directly extracted to avoid repeated calculations and intermediate data storage, providing richer information for subsequent feature extraction and segmentation.

[0053] Feature extraction and fusion (based on Example 1): CBCT image feature extraction (added part): In addition to the conventional features extracted in Example 1, focus on the newly extracted texture features for areas where there may be minor lesions (such as the enamel surface, near the pulp cavity, etc.), and use the lesion possibility features output by the deep learning model. By analyzing the changes in these features at different time points, the development trend of minor lesions can be more accurately judged.

[0054] Feature fusion (modified part): When using the weighted average method for feature fusion, an adaptive weight adjustment mechanism is designed. The weights are automatically adjusted according to the importance of the micro-lesion features at different stages, and the newly extracted micro-lesion-related features are weighted averaged together with other conventional features to form an updated fusion feature vector. In this way, the micro-lesion features can be more fully considered in the fusion process, improving the segmentation accuracy of the micro-lesion area.

[0055] Improvement and training of 3D segmentation model (based on Example 1): Based on the updated fusion feature vector, the 3D-CNN model constructed in Example 1 is improved. A network layer specifically processing the features related to the minor lesions is added, for example, a parameter adjustment mechanism for the minor lesion features is introduced in the convolution layer or the fully connected layer, so that the model can better learn the relationship between the minor lesion features and other tissue features, and improve the segmentation capability of the minor lesion area.

[0056] The improved model was retrained using a dental CBCT image dataset containing known minor lesions. During the training process, a higher weight was given to the error in segmenting the minor lesion area, prompting the model to pay more attention to the accurate segmentation of the minor lesion area.

[0057] Segmentation effect evaluation and application (based on Example 1): The CBCT image of the tooth to be segmented and the corresponding dynamic MRI image (after preprocessing and coordinate alignment) are input into the retrained 3D-CNN model. The model outputs the three-dimensional segmentation results of the teeth and surrounding tissues. In addition to the conventional segmentation of teeth, alveolar bones, etc., the possible micro-lesion areas (such as early caries areas, enamel micro-cracks areas, etc.) are also individually labeled and segmented, so that doctors can more intuitively observe the location and range of micro-lesions.

[0058] The accuracy of segmentation of micro-lesion areas is added to the evaluation indicators, such as the recall rate and precision rate of micro-lesion areas. By comparing with the gold standard segmentation results of manual annotation, the segmentation effect of the model on micro-lesion areas is more comprehensively evaluated. Based on the segmentation results, doctors can detect micro-lesions in time and take corresponding preventive or treatment measures, such as filling early caries lesions and repairing enamel microcracks, to prevent further development of lesions.

[0059] Improvements in identification of microlesion areas: Develop a feature extraction method based on self-supervised learning. Without additional annotation, use the image's own structural information (such as the natural changes in texture and shape of teeth and surrounding tissues) to learn feature representations. For microlesion detection, a special self-supervised task can be designed, such as performing random transformations on tooth images (such as rotation, scaling, etc.), and then letting the model predict the type and parameters of the transformation. In this way, the model can learn features that are sensitive to microlesions, which may be related to local image changes caused by the lesions. In the subsequent segmentation process, these self-supervised learned features are used to enhance the recognition of microlesion areas without the need for additional deep learning models for lesion screening.

[0060] It is necessary to add that: The implementation steps of the feature extraction method based on self-supervised learning are: Data preparation: We use an unlabeled dental CBCT image dataset, which contains the normal structure of teeth and surrounding tissues as well as possible micro-lesions.

[0061] Designing self-supervised tasks: Random transformation generation: Perform random geometric transformations on each tooth image, such as rotation (within a certain angle range, such as -30° to 30°), scaling (within a certain ratio range, such as 0.8-1.2 times), translation (within a certain ratio range of image size), etc. Record the type (rotation, scaling, translation) and specific parameters (such as rotation angle, scaling ratio, translation distance) of each transformation.

[0062] Construct training data pairs: The original image and the transformed image form a training data pair. For the model, the input is the original image, and the target output is the type and parameters of the transformation.

[0063] Model training: Choose a suitable network architecture: A model structure suitable for processing image data, such as a convolutional neural network (CNN), can be used. The input layer of the model receives the dental CBCT image, and the output layer is designed to output information related to the transformation type and parameters. For example, a multi-classification output layer can be used to represent different transformation types (such as rotation categories, scaling categories, translation categories, etc.), and a regression output layer can be used to represent the corresponding parameter values.

[0064] Training process: During the training process, the original image in the data pair is input into the model, and the model attempts to predict the type and parameters of the transformation. The loss function is calculated by comparing the predicted result with the actual transformation information (i.e. the recorded transformation type and parameters). The parameters of the model are adjusted using optimization algorithms such as gradient descent so that the model can accurately predict the transformation.

[0065] Feature extraction and microlesion recognition: Feature learning: During the training process, the feature representations learned by the model contain information that is sensitive to small changes in the image. These features not only capture the structural information of normal teeth and surrounding tissues, but also have a certain perception of features related to local image changes caused by lesions.

[0066] Using learned features to identify microlesions: In the subsequent segmentation process, the dental CBCT images to be segmented are input into the trained model to extract the feature vectors output by the intermediate layer of the model. These feature vectors can be input into the segmentation model as additional information to enhance the recognition of microlesion areas. For example, these features can be fused with features extracted by other traditional methods, or these features can be directly used to design special branches in the segmentation model to improve the segmentation accuracy of microlesion areas.

[0067] The graph convolutional network (GCN) is used to combine the spatial information of the image. The CBCT image is regarded as a graph structure, in which each pixel is a node and the relationship between adjacent pixels is an edge. The local and global relationships between different areas in the image (including the area of ​​microlesions) can be learned through GCN. In the image preprocessing stage, GCN is used to enhance the image features to highlight the characteristics of the microlesion area, while avoiding the addition of complex deep learning models for lesion screening. This method utilizes the spatial structural information of the image and may be able to capture lesion characteristics better than simply improving scanning parameters and preprocessing algorithms.

[0068] Further explanation is needed: The steps in the second embodiment are improvements over the first embodiment: Real-time adaptive scanning resolution adjustment method: Based on Example 1, this method can more accurately collect data for tiny lesions. Using conventional resolution in the early stage of scanning can reduce unnecessary data volume and acquisition time, and increasing the resolution when suspected tiny lesion areas are found can obtain more detailed image information, which is helpful for the subsequent accurate identification and segmentation of tiny lesions. Compared with Example 1, it avoids the problem of loss or unclear information of tiny lesions that may be caused by using a uniform resolution throughout the scanning process, making the collected data more targeted and providing a better data foundation for the detection of tiny lesions.

[0069] Using the improved deep learning model to screen areas of minor lesions: Preliminary screening of areas where minor lesions may exist in low-resolution images further optimizes the efficiency and pertinence of data acquisition. This allows subsequent high-resolution scans to focus on these suspicious areas, reducing the amount of data processing, while also ensuring that minor lesion information is not missed in the subsequent segmentation process. Compared to the method of Example 1 that simply relies on time point acquisition and virtual data generation, this screening mechanism increases the initiative and accuracy of minor lesion detection and improves the ability of the entire three-dimensional segmentation method to cope with minor lesions.

[0070] Combining the local contrast enhancement process with the texture analysis algorithm in one step has a significant optimization effect on Example 1. When processing high-resolution CBCT images, this combination can directly extract texture features while enhancing the contrast of areas prone to micro-lesions. On the one hand, it avoids repeated calculations and intermediate data storage, improves the efficiency of data processing, and saves computing resources and time. On the other hand, it provides richer information for subsequent feature extraction and segmentation, especially for texture features related to micro-lesions. This additional information helps to more accurately distinguish between micro-lesion areas and normal tissues, making up for the shortcomings of Example 1 in capturing micro-lesion features through conventional preprocessing alone, thereby improving the segmentation accuracy of micro-lesion areas.

[0071] In addition to the conventional features in Example 1, focusing on the newly extracted texture features and the lesion possibility features output by the deep learning model for areas where tiny lesions may exist can more comprehensively describe the characteristics of tiny lesions. Analyzing the changes in these features at different time points can better grasp the development trend of tiny lesions. Compared with Example 1, this can more deeply explore information related to tiny lesions, making the model more sensitive to tiny lesions during the segmentation process, and improving the accuracy and foresight of the segmentation of tiny lesion areas.

[0072] Feature fusion modification part: The design of the adaptive weight adjustment mechanism fully considers the importance of the micro-lesion features at different stages when fusing features. Based on the weighted average method of Example 1, this mechanism can dynamically adjust the weights according to the development of micro-lesions, so that the micro-lesion related features can be more reasonably reflected in the fused feature vector. This helps to highlight the role of micro-lesion features in the entire segmentation model, avoids the problem of insufficient attention to micro-lesion features due to fixed weights during the feature fusion process in Example 1, and further improves the segmentation accuracy of micro-lesion areas.

[0073] The 3D-CNN model is improved based on the updated fusion feature vector, and a network layer is added to specifically process the features related to minor lesions, so that the model can better learn the relationship between the features of minor lesions and other tissue features. Introducing a parameter adjustment mechanism for minor lesion features in the convolutional layer or the fully connected layer gives the model a stronger ability to process minor lesion information. Compared with Example 1, this improved model structure is specifically optimized for minor lesions, and can more effectively utilize the newly extracted minor lesion features, thereby improving the ability to segment minor lesion areas. At the same time, the model is retrained using a dental CBCT image dataset containing known minor lesions, and a higher weight is given to the error in segmenting minor lesion areas, prompting the model to pay more attention to the accurate segmentation of minor lesion areas, further enhancing the model's detection and segmentation performance for minor lesions.

[0074] In terms of output results: after the tooth CBCT image to be segmented and the corresponding dynamic MRI image are input into the retrained model, in addition to the conventional segmentation results, the possible micro-lesion areas are also individually labeled and segmented, which provides doctors with more intuitive and detailed information. Doctors can observe the location and range of micro-lesions more clearly. Compared with the first embodiment, micro-lesions can be discovered earlier and more accurately, providing strong support for timely prevention or treatment measures, and avoiding further development of lesions and causing greater impact on patients' oral health.

[0075] Evaluation indicators: Add the evaluation of the accuracy of segmentation of small lesion areas to the evaluation indicators, such as the recall rate and precision rate of small lesion areas, and compare with the manually annotated gold standard segmentation results to more comprehensively evaluate the segmentation effect of the model on small lesion areas. This makes the evaluation of the model more scientific and accurate, and can timely discover the deficiencies of the model in small lesion segmentation and make improvements. Compared with the use of only conventional evaluation indicators in Example 1, it better guarantees the effectiveness and reliability of the model in detecting small lesions, thereby improving the performance of the entire three-dimensional segmentation method in detecting small lesions.

[0076] Feature extraction method based on self-supervised learning: This method uses the image's own structural information to learn feature representation without additional annotation. For micro-lesion detection, by designing special self-supervised tasks, such as randomly transforming tooth images and letting the model predict the type and parameters of the transformation, the model can learn features that are sensitive to micro-lesions. These features are used in the subsequent segmentation process to enhance the recognition of micro-lesion areas, without the need for additional deep learning models for lesion screening. This is an important supplement to Example 1. It further mines the information of the image itself, provides a new way to identify micro-lesions, improves the accuracy and efficiency of identifying micro-lesion areas, and reduces dependence on labeled data.

[0077] Use graph convolutional networks (GCN) combined with the spatial information of the image: treat the CBCT image as a graph structure, and use GCN to learn the local and global relationships between different areas in the image (including the area of ​​microlesions). Use GCN to enhance the image features in the image preprocessing stage to highlight the features of the microlesion area. This method utilizes the spatial structural information of the image and can more effectively capture the lesion features than the simple improvement of scanning parameters and preprocessing algorithms in Example 1. It provides richer spatial information for the identification and segmentation of microlesions, further optimizes the processing of microlesion areas, and improves the detection ability of the entire three-dimensional segmentation method for microlesions.

[0078] In the implementation process of the first and second embodiments, a dental CBCT image three-dimensional segmentation system is applied: Data acquisition module: used to perform CBCT scans on patients before and at the sixth month of orthodontic treatment, and to perform a dynamic MRI examination before the treatment begins; this module also has the function of using the generator of the generative adversarial network to learn the distribution law of the data before and at the sixth month to generate virtual data for the third and ninth months. At the same time, a discriminator based on the physiological characteristics of teeth and surrounding tissues and the logic of orthodontic treatment changes is set to evaluate the rationality of virtual data. In addition, multi-task learning can be used to jointly train the time series prediction task and the tooth tissue segmentation task, and the spatial relationship between tissues in the segmentation task and the orthodontic treatment change pattern are used to provide constraint information for the prediction model; Image preprocessing module: Its function is to perform median filtering algorithm to remove noise points and grayscale normalization on the acquired CBCT images, and to perform motion correction, contrast adjustment and brightness adjustment on the dynamic MRI images to highlight the target structure, which includes but is not limited to the articular disc and muscle; Image registration module: used to select the nasal spine of the maxilla and the condyle of the mandible as anatomical landmarks, and use image registration software to register and calibrate the CBCT image and dynamic MRI image; Feature extraction and fusion module: It can use the attention mechanism in deep learning to extract the displacement change, density change, angle change characteristics of key structures such as teeth, alveolar bones, temporomandibular joints in CBCT images at different time points (including virtual data time points), as well as the displacement characteristics of articular discs and muscle contraction and relaxation characteristics in dynamic MRI images. It can also perform weighted average fusion into feature vectors based on the importance of features to segmentation and orthodontic treatment analysis. 3D segmentation model construction and training module: A 3D convolutional neural network model is constructed based on the fused feature vector generated by the feature extraction and fusion module. The dental CBCT image dataset with labeled information of different tissues is used as a training set to input the model for training. The validation set is used to monitor the model performance indicators to adjust the model hyperparameters. Segmentation effect evaluation module: It is used to input the tooth CBCT image to be segmented and the corresponding dynamic MRI image into the model trained by the three-dimensional segmentation model construction and training module, output the three-dimensional segmentation results of the teeth and surrounding tissues, and compare them with the manually labeled gold standard segmentation results. The three-dimensional segmentation results are evaluated by using evaluation indicators such as Dice similarity coefficient and Hausdorff distance. In terms of output results, in addition to conventional segmentation of teeth, alveolar bones, etc., this module can also separately annotate and segment possible micro-lesion areas (such as early caries areas, enamel micro-cracks areas, etc.); in terms of evaluation indicators, in addition to conventional evaluation indicators, the accuracy of segmentation of micro-lesion areas can also be evaluated, such as by calculating the recall rate, precision rate and other indicators of micro-lesion areas to more comprehensively evaluate the segmentation effect of the model on micro-lesion areas.

[0079] Also includes: Targeted processing module for micro-lesions: In the data acquisition module, when performing CBCT scanning, a real-time adaptive scanning resolution adjustment method is used. The scanning is performed at a conventional resolution at the beginning. As the scanning progresses, the acquired images are preliminarily analyzed. If a suspected micro-lesion area is found, the system automatically increases the scanning resolution of the area; for other areas, the original resolution is maintained or adjusted appropriately; at the same time, a deep learning model is used to preliminarily screen areas where micro-lesions may exist in low-resolution images, so as to focus on these areas in subsequent high-resolution scanning and segmentation processes; in the data preprocessing module, when processing high-resolution CBCT images, the local contrast enhancement processing and the texture analysis algorithm are combined in one step. In the process of enhancing the contrast of areas prone to micro-lesions such as tooth surfaces, the texture features of these areas are directly extracted; in the feature In the extraction and fusion module, for areas where there may be minor lesions (such as the enamel surface, near the pulp cavity, etc.), focus on the newly extracted texture features, and use the lesion possibility features output by the deep learning model. When using the weighted average method for feature fusion, design an adaptive weight adjustment mechanism to automatically adjust the weights according to the importance of the minor lesion features at different stages, and weight the newly extracted minor lesion-related features together with other conventional features to form an updated fusion feature vector; in the three-dimensional segmentation model construction and training module, add a network layer for processing minor lesion-related features, including introducing a parameter adjustment mechanism for minor lesion features in the convolutional layer or the fully connected layer, and retrain the model using a dental CBCT image dataset containing known minor lesions. During the training process, a high level of weight is assigned to the error in segmenting the minor lesion area; Image spatial information utilization module: Use graph convolutional networks to combine the spatial information of images and regard the CBCT image as a graph structure, in which each pixel is a node and the relationship between adjacent pixels is an edge; use GCN to learn the local and global relationships between different areas in the image; in the image preprocessing stage, use GCN to enhance the features of the image and highlight the features of the tiny lesion area.

[0080] Summarize: This method uses a generative adversarial network (GAN) to generate virtual data. Based on the CBCT scans before the start of orthodontic treatment and the 6th month and a dynamic MRI examination before treatment, it can generate data in intermediate states such as the 3rd and 9th months, effectively making up for the data loss caused by reducing the time point collection. While reducing the actual number of collection times, relatively complete treatment process data can still be obtained, providing a richer source of information for subsequent comprehensive analysis, and solving the problem that traditional methods are difficult to grasp the dynamic changes of teeth and surrounding tissues due to incomplete data.

[0081] There is no need to collect data at a large number of different time points. This step simplifies the data collection process by generating virtual data with the help of GAN, which reduces the cost and workload of data collection while ensuring data integrity.

[0082] By utilizing multi-task learning joint training, the time series prediction task is combined with the tooth tissue segmentation task. The spatial relationship between different tissues in the segmentation task and the known change patterns of orthodontic treatment are used to provide additional constraint information for the prediction model, so as to more accurately grasp the changes of teeth and surrounding tissues during the treatment process, thereby improving the prediction accuracy of the overall treatment effect evaluation and the formulation of subsequent treatment plans, and overcoming the problem of inaccurate evaluation caused by the lack of an effective mechanism to associate data at different time points in traditional methods.

[0083] This joint training method does not require complex additional data processing steps and can be implemented based on the conventional data processing process, which simplifies the steps and improves the prediction effect.

[0084] By using the attention mechanism in deep learning, we can automatically focus on key structures such as teeth, alveolar bones, and temporomandibular joints. For CBCT images and dynamic MRI images, we can accurately extract various core features of these key areas at different time points (including virtual intermediate time point data), such as displacement, density, angle change, and other features as well as articular disc displacement, muscle contraction and relaxation, etc. This effectively solves the problem that traditional methods are difficult to accurately capture key structural features, so that segmentation can be performed more accurately around key parts.

[0085] The application of this attention mechanism does not require complex image preprocessing or additional feature screening steps. It can achieve precise focusing on key features in the conventional image data processing process, simplifying the feature extraction process and improving the extraction effect.

[0086] By adopting the weighted average method for feature fusion and assigning corresponding weights to the importance of three-dimensional segmentation of dental CBCT images and orthodontic treatment analysis according to different features, a fused feature vector is formed. The importance of each feature can be fully considered, and the accuracy of segmentation is further improved. The accuracy and precision of the segmentation results in key parts are significantly improved, which makes up for the shortcomings of traditional methods that lead to poor segmentation of key parts due to unreasonable fusion of features.

[0087] This weighted average feature fusion method is simple to calculate and does not require complex algorithm adjustments or additional data processing steps. While ensuring effective feature fusion, it simplifies the fusion steps and ensures improved segmentation effects.

[0088] The real-time adaptive scanning resolution adjustment method in the second embodiment performs CBCT scanning at a conventional resolution at the initial stage, and automatically increases the resolution of a suspected microlesion area when the area is found, thereby avoiding the problem of microlesion information loss or unclearness that may be caused by a unified resolution, making the collected data more targeted for microlesion detection, providing a better data basis for subsequent analysis, and effectively solving the problem that traditional scanning with only a unified resolution is difficult to fully capture microlesion information.

[0089] This method automatically adjusts the resolution according to the situation during the scanning process, without the need for frequent manual intervention or complex resolution setting steps. While simplifying the data acquisition operation, it improves the pertinence and efficiency of data acquisition for the detection of tiny lesions.

[0090] The improved deep learning model is used to preliminarily screen areas where possible tiny lesions may exist in low-resolution images, further optimizing the efficiency and specificity of data acquisition, ensuring that subsequent high-resolution scanning and segmentation processes can focus on suspicious areas, reducing the amount of data processing and not missing information about tiny lesions. Compared with the traditional method that simply relies on time point acquisition and virtual data generation, it increases the initiative and accuracy of tiny lesion detection and improves the ability of the entire three-dimensional segmentation method to deal with tiny lesions.

[0091] This screening method based on deep learning models only requires low-resolution images to be input into the model to complete preliminary screening, without the need for additional complex image analysis steps. This simplifies the data acquisition process while enhancing the effect of detecting tiny lesions.

[0092] Combining local contrast enhancement processing with texture analysis algorithms to process high-resolution CBCT images in one step can not only enhance the contrast of areas prone to microlesions, but also directly extract texture features, avoiding repeated calculations and intermediate data storage, improving data processing efficiency, and saving computing resources and time. This effectively makes up for the shortcomings of traditional preprocessing algorithms in highlighting microlesion features, so that subsequent feature extraction and segmentation can better use this information to distinguish microlesion areas from normal tissues, thereby improving the segmentation accuracy of microlesion areas.

[0093] This combined processing method combines two steps that may have to be performed separately into one, eliminating the need for cumbersome data transfer and additional parameter adjustments between the two processing steps. While simplifying the data preprocessing process, it improves the effect of data preprocessing on tiny lesion areas.

[0094] In addition to conventional features, we focus on the newly extracted texture features and the lesion possibility features output by the deep learning model for areas where tiny lesions may exist, and analyze their changes at different time points. This can more comprehensively describe the characteristics of tiny lesions and more deeply explore information related to tiny lesions, enabling the model to have a keener perception of tiny lesions during the segmentation process, thereby improving the accuracy and foresight of segmentation of tiny lesion areas, and overcoming the problem of inaccurate tiny lesion detection caused by incomplete feature extraction in traditional methods.

[0095] This feature extraction method increases the focus on features related to tiny lesions on the basis of conventional extraction. There is no need to make major changes to the overall feature extraction process. It only requires targeted extraction of relevant features in specific areas and analysis of changes. This simplifies the feature extraction steps while improving the segmentation effect of tiny lesion areas.

[0096] When using the weighted average method for feature fusion, an adaptive weight adjustment mechanism is designed to automatically adjust the weights according to the importance of the subtle lesion features at different stages, fully consider the subtle lesion features, and make the newly extracted subtle lesion-related features more reasonably reflected in the fused feature vector, further improving the segmentation accuracy of the subtle lesion area.

[0097] This adaptive weight adjustment mechanism is optimized based on the conventional weighted average fusion. It does not require complicated weight setting rules or additional data processing steps. It simplifies the feature fusion steps while ensuring the improvement of the segmentation effect of small lesion areas.

[0098] The 3D-CNN model is improved based on the updated fusion feature vector, and a network layer is added to specifically process the features related to minor lesions. For example, a parameter adjustment mechanism for minor lesion features is introduced in the convolutional layer or the fully connected layer, so that the model can better learn the relationship between minor lesion features and other tissue features, and improve the segmentation ability of minor lesion areas.

[0099] This model improvement method makes targeted expansion based on the original 3D-CNN model, without the need to rebuild a complex new model architecture. It simplifies the model improvement steps while improving the segmentation performance of tiny lesion areas.

[0100] The improved model was retrained using a dental CBCT image dataset containing known minor lesions. During the training process, a higher weight was given to the error in segmenting the minor lesion area, prompting the model to pay more attention to the accurate segmentation of the minor lesion area, further enhancing the model's detection and segmentation performance for minor lesions.

[0101] This retraining method only requires inputting a specific data set into the improved model and adjusting the error weight. It does not require complex resetting of training parameters or additional training process adjustments. It simplifies the training steps while enhancing the model's segmentation effect on tiny lesion areas.

[0102] Improvement of segmentation effect evaluation: By adding an assessment of the accuracy of segmentation of tiny lesion areas to the evaluation indicators, such as the recall rate and precision of tiny lesion areas, and comparing them with the manually labeled gold standard segmentation results, the segmentation effect of the model on tiny lesion areas can be more comprehensively evaluated, and the deficiencies of the model in tiny lesion segmentation can be discovered and improved in time, thereby better ensuring the effectiveness and reliability of the model in detecting tiny lesions. This allows doctors to more intuitively observe the location and range of tiny lesions, discover tiny lesions earlier and more accurately, and take corresponding preventive or therapeutic measures to avoid further development of the lesions.

[0103] This method of improving the evaluation indicators adds evaluation content for small lesion areas on the basis of conventional evaluation, without the need to make major changes to the entire evaluation process. It simplifies the evaluation steps and improves the evaluation accuracy of the segmentation effect of small lesion areas.

Claims

1. A method for three-dimensional segmentation of dental CBCT images, characterized in that: The following steps are involved: Data collection and processing: CBCT scans were performed on the patients before and at the sixth month of orthodontic treatment, and a dynamic MRI examination was performed before the treatment began. The generator of the generative adversarial network was used to learn the distribution law of the data before and at the sixth month to generate virtual data for the third and ninth months. At the same time, a discriminator based on the physiological characteristics of teeth and surrounding tissues and the logic of orthodontic treatment changes was designed to evaluate the rationality of the virtual data. Multi-task learning was used to jointly train the time series prediction task and the tooth tissue segmentation task, and the spatial relationship between tissues in the segmentation task and the change pattern of orthodontic treatment were used to provide constraint information for the prediction model. Image preprocessing: Use median filtering algorithm to remove noise points and grayscale normalization on the acquired CBCT images, perform motion correction on dynamic MRI images, adjust contrast and brightness to highlight target structures, including but not limited to articular discs and muscles; Image registration: The nasal spine of the maxilla and the condyle of the mandible were selected as anatomical landmarks, and the CBCT images and dynamic MRI images were registered and calibrated using image registration software; Feature extraction and fusion: The attention mechanism in deep learning is used to extract the displacement, density, and angle change characteristics of key structures such as teeth, alveolar bones, and temporomandibular joints in CBCT images at different time points, including virtual data time points, as well as the displacement characteristics of articular discs and muscle contraction and relaxation characteristics in dynamic MRI images. The features are then weighted and averaged to form feature vectors based on their importance to segmentation and orthodontic treatment analysis. 3D segmentation model construction and training: A 3D convolutional neural network model is constructed based on the fused feature vectors. The dental CBCT image dataset with labeled information of different tissues is used as the training set to train the model. The validation set is used to monitor the model performance indicators to adjust the model hyperparameters. Segmentation effect evaluation: The CBCT image of the tooth to be segmented and the corresponding dynamic MRI image are input into the trained model to output the three-dimensional segmentation results of the teeth and surrounding tissues, which are then compared with the manually annotated gold standard segmentation results and evaluated using the Dice similarity coefficient and Hausdorff distance.

2. A method for three-dimensional segmentation of dental CBCT images according to claim 1, characterized in that: In data acquisition, specifically, when performing CBCT scanning, a real-time adaptive scanning resolution adjustment method is used. The scanning is performed at a normal resolution at the beginning. As the scanning progresses, the acquired images are preliminarily analyzed. If a suspected microlesion area is found, the system automatically increases the scanning resolution of the area; for other areas, the original resolution is maintained or adjusted appropriately. At the same time, a deep learning model is used to preliminarily screen areas that may have tiny lesions in low-resolution images, so as to focus on these areas in subsequent high-resolution scanning and segmentation processes.

3. The method for three-dimensional segmentation of dental CBCT images according to claim 1, characterized in that: In data preprocessing, specifically, when processing high-resolution CBCT images, local contrast enhancement processing and texture analysis algorithms are combined in one step; This includes directly extracting texture features of areas prone to microlesions, such as tooth surfaces, during the process of enhancing the contrast of these areas.

4. A method for three-dimensional segmentation of dental CBCT images according to claim 1, characterized in that: Focusing on areas where subtle lesions may exist, including but not limited to the enamel surface and near the pulp cavity, we focus on the newly extracted texture features and utilize the lesion possibility features output by the deep learning model.

5. The method for three-dimensional segmentation of dental CBCT images according to claim 1, characterized in that: When the weighted average method is used for feature fusion, an adaptive weight adjustment mechanism is designed; the weights are automatically adjusted according to the importance of the microlesion features at different stages, and the newly extracted microlesion-related features are weighted averaged together with other conventional features to form an updated fusion feature vector.

6. A method for three-dimensional segmentation of dental CBCT images according to claim 1, characterized in that: In the improvement and training of 3D segmentation models: adding network layers to process features related to small lesions, including introducing parameter adjustment mechanisms for small lesion features in convolutional layers or fully connected layers; The model was retrained using a dental CBCT image dataset containing known minor lesions, and during the training process, a high level of weight was given to the error in segmenting the minor lesion area.

7. A method for three-dimensional segmentation of dental CBCT images according to claim 1, characterized in that: In the segmentation effect evaluation: the tooth CBCT image to be segmented and the corresponding dynamic MRI image are input into the retrained 3D-CNN model, and the model outputs the three-dimensional segmentation results of the teeth and surrounding tissues, including the segmentation of teeth and alveolar bones, and also includes but is not limited to the early caries areas and enamel microcracks areas for separate labeling and segmentation.

8. The method for three-dimensional segmentation of dental CBCT images according to claim 1, characterized in that: Using graph convolutional networks combined with the spatial information of images, CBCT images are regarded as a graph structure, in which each pixel is a node and the relationship between adjacent pixels is an edge; GCN is used to learn the local and global relationships between different areas in the image; in the image preprocessing stage, GCN is used to enhance the features of the image to highlight the characteristics of the tiny lesion area.

9. A dental CBCT image three-dimensional segmentation system, characterized in that: include: Data acquisition module, image preprocessing module, image registration module, feature extraction and fusion module, three-dimensional segmentation model construction and training module, segmentation effect evaluation module, micro-lesion targeted processing module, and image spatial information utilization module.

Images