An automatic segmentation analysis method and system for an oral image root canal system

Abstract

The application discloses an automatic segmentation analysis method and system of an oral image root canal system, the method comprising marking and preprocessing oral CBCT image data to obtain a marked data set and an unmarked data set; training a pseudo-label semi-supervised segmentation model using the marked data set and the unmarked data set to realize three-dimensional segmentation of a root canal structure in the oral CBCT image and obtain a root canal three-dimensional segmentation result, which greatly reduces the dependence of the model on high-quality marked data and realizes high-precision three-dimensional segmentation of the root canal system; further, three-dimensional connected domain analysis is performed on the root canal three-dimensional segmentation result, a target root canal region is extracted, quantitative analysis of a root canal bending shape of the target root canal region is performed, the bending angle and bending position of the root canal are automatically calculated, the efficiency of root canal shape interpretation is greatly improved, measurement errors caused by artificial subjective factors and angle deviation are avoided, the measurement result is more accurate, and the beneficial effect of reducing the risk of treatment complications is realized.

Description

Technical Field

[0001] This application relates to the field of oral medical image processing technology, specifically to an automatic segmentation and analysis method and system for an oral CBCT image root canal system. Background Technology

[0002] Root canal treatment is the core method for treating pulpitis and periapical periodontitis. However, the anatomical morphology of the root canal system is extremely complex, often exhibiting anatomical variations such as multiple root canals, accessory root canals, and root canal branching. In particular, the presence of curved root canals presents significant clinical challenges to root canal treatment. During the treatment of curved root canals, complications such as instrument separation and root canal perforation are highly likely to occur, directly affecting the thoroughness of root canal debridement and leading to poor treatment outcomes. Therefore, accurate interpretation of the root canal system morphology before root canal treatment is crucial.

[0003] Cone-beam computed tomography (CBCT), a revolutionary imaging technology in oral medicine, rapidly acquires high-resolution three-dimensional images of the oral and maxillofacial region through rotating cone-beam X-ray scanning. It provides three-dimensional morphological information of root canals, compensating to some extent for the shortcomings of traditional two-dimensional imaging and becoming an important tool for root canal morphology assessment. However, in clinical application, the interpretation of CBCT images still faces several challenges: firstly, junior dentists lack sufficient clinical experience, making CBCT image interpretation time-consuming and difficult to guarantee accuracy; secondly, CBCT image quality is susceptible to interference from section selection, reconstruction parameters, and artifacts, and manual measurements often differ from the actual root canal anatomy due to angular deviations, limiting its direct application in precision root canal treatment.

[0004] Therefore, for those skilled in the art, developing a method and system that can achieve high-precision segmentation of the root canal system and automatically complete quantitative analysis of root canal curvature morphology, and solving the technical problems of segmentation relying on a large number of annotations, low efficiency and poor accuracy of morphology interpretation in the existing technology, is of great practical significance for improving the precision of root canal treatment and reducing the risk of treatment complications. Summary of the Invention

[0005] This invention addresses the technical problems of existing technologies where the segmentation of the root canal system in oral CBCT imaging relies on a large amount of high-quality labeled data, and the morphological interpretation efficiency and accuracy are low. It provides an automatic segmentation and analysis method and system for the root canal system in oral CBCT imaging.

[0006] To achieve the above-mentioned objectives, the present invention provides an automatic segmentation and analysis method for an oral imaging root canal system, comprising the following steps: Step S1: Label and preprocess the oral CBCT image data to obtain labeled datasets and unlabeled datasets; Step S2: Train the pseudo-label semi-supervised segmentation model using the labeled dataset and the unlabeled dataset to achieve three-dimensional segmentation of root canal structures in oral CBCT images and obtain the three-dimensional segmentation results of root canals. Step S3: Perform three-dimensional connected component analysis on the three-dimensional segmentation results of the root canal, extract the target root canal region, perform quantitative analysis on the root canal curvature of the target root canal region, and automatically calculate the curvature angle and curvature position of the root canal.

[0007] Furthermore, the preprocessing includes performing layer alignment and grayscale normalization on the CBCT image data.

[0008] Further, in step S2, the process of training the pseudo-label semi-supervised segmentation model using the labeled dataset and the unlabeled dataset includes: S21. Supervised training of the initial nn-UNet model using the labeled dataset to obtain a baseline model with preliminary segmentation capabilities; S22. Use the baseline model to predict the unlabeled dataset and generate a pseudo-labeled dataset; S23. Merge the labeled dataset and the pseudo-labeled dataset into an expanded training set, and retrain the pseudo-labeled semi-supervised segmentation model under the nn-UNet framework to obtain the trained pseudo-labeled semi-supervised segmentation model.

[0009] Further, in step S3, three-dimensional connected component analysis is performed on the three-dimensional segmentation results of the root canal to extract the target root canal region, and quantitative analysis of the root canal curvature morphology of the target root canal region is performed, including: Step S31: Count the number of voxels in each connected region in the three-dimensional segmentation result of the root canal, remove the connected regions in the noise region, sort the remaining candidate connected regions according to their spatial position, and select the target root canal region.

[0010] Step S32: Identify branches in the selected target root canal region and locate the key endpoints of each branch; Step S33: Construct the root canal centerline based on the key endpoints and perform a smooth fit on the centerline; Step S34: Calculate the bending angle and bending position of the root canal based on the smoothed centerline.

[0011] Further, in step S32, the branch identification of the selected target root canal region and the location of the key endpoints of each branch include: The segmentation mask of the target root canal region is scanned layer by layer along the z-axis, and two-dimensional connected components are marked in each layer to identify the connected components in the current layer. The inheritance and updating of branch numbers are achieved by considering the overlap between the connected components in the current layer and the connected components in the previous layer. Specifically, if the connected components in the current layer overlap with those in the previous layer, the branch number is inherited; if the connected components in the current layer do not overlap with those in the previous layer, a new branch number is assigned; if the number of connected components in the current layer is reduced relative to the previous layer, the downward scanning is terminated, and branch identification is completed. After branch identification is completed, two key endpoints are further determined for each branch. The key endpoints include an upper reference point and a lower reference point. The lower reference point is the center point of the layer where the branch first appears, and the upper reference point is the center point where the branch extends to the bottom layer.

[0012] Further, in step S33, constructing the root canal centerline based on the key endpoints and smoothly fitting the centerline includes: Extract all foreground voxels of the target root canal region and group them according to the z-axis direction. Calculate the average position of the branch point set in each layer, select the real voxel point closest to the average position as the center point of that layer, and connect the center points of each layer in order according to the z-axis direction to obtain the initial discrete center line. The initial discrete centerline is smoothly reconstructed using the B-spline curve fitting method, and the beginning and end portions of the smooth curve are trimmed to obtain the smoothly fitted centerline.

[0013] Further, in step S34, the bending angle and bending position of the root canal are calculated based on the smoothly fitted centerline, specifically including: The method for calculating the bending angle of the root canal is as follows: Let any point on the centerline after smooth fitting be... The upper reference point and lower reference point of the key endpoint are C and A, respectively. Construct two vectors for the vertices , as follows:

[0014] but , The angle formed by two corresponding vectors The bending angle of the root canal:

[0015] in, The centerline after smooth fitting is characterized at point The local bending state relative to the line connecting the upper and lower reference points; Calculate the angle point by point along the centerline after smooth fitting, and select the point corresponding to the smallest included angle as the corresponding point of the bending position.

[0016] Another aspect of the present invention provides an automatic segmentation and analysis system for oral imaging root canal systems, comprising: The data processing module is used to collect, annotate, and preprocess oral CBCT image data to obtain an annotated dataset and an unannotated dataset. The pseudo-label semi-supervised segmentation module is used to train the pseudo-label semi-supervised segmentation model using the labeled dataset and the unlabeled dataset, so as to achieve three-dimensional segmentation of the root canal system in oral CBCT images and obtain the root canal three-dimensional segmentation results. The morphological quantitative analysis module performs quantitative analysis of the root canal bending morphology based on the three-dimensional segmentation results of the root canal, and automatically calculates the bending angle and bending position of the root canal.

[0017] Furthermore, the pseudo-label semi-supervised segmentation model includes: The initial training unit is used to train the initial nnUNet model using the labeled dataset to obtain the baseline model; A pseudo-label generation unit is used to predict and generate a pseudo-label dataset from the unlabeled dataset using the baseline model. The joint retraining unit is used to fuse the labeled dataset and the pseudo-labeled dataset for retraining to obtain the root canal 3D segmentation result.

[0018] Furthermore, the morphological quantitative analysis module includes: The region extraction unit is used to extract the target root canal region based on the three-dimensional segmentation results of the root canal; The branch identification and endpoint localization unit is used to complete the identification of root canal branches and the localization of key endpoints. A centerline construction and smoothing unit is used to construct the root canal centerline based on the located key endpoints and to smooth the root canal centerline. The bending parameter calculation unit is used to calculate the root canal bending angle and bending position.

[0019] Furthermore, the system also includes a results output module, which visualizes the three-dimensional segmentation results, bending angles, and bending positions of the root canals and generates analysis reports containing various quantitative parameters, making it convenient for doctors to view intuitively and use clinically. It also includes a model storage and retrieval module, which stores the trained root canal segmentation model and supports rapid segmentation and analysis of new oral CBCT image data, improving the convenience of clinical application.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention provides an automatic segmentation and analysis method for root canal systems in oral imaging. The method includes: annotating and preprocessing oral CBCT image data to obtain an annotated dataset and an unannotated dataset; training a pseudo-label semi-supervised segmentation model using the annotated and unannotated datasets to achieve three-dimensional segmentation of root canal structures in oral CBCT images, obtaining three-dimensional segmentation results of root canals. This significantly reduces the model's dependence on high-quality annotated data and achieves high-precision three-dimensional segmentation of the root canal system; furthermore, performing three-dimensional connected component analysis on the root canal three-dimensional segmentation results to extract target root canal regions, and performing quantitative analysis of the root canal curvature morphology of the target root canal regions, automatically calculating the curvature angle and curvature position of the root canals, which greatly improves the efficiency of root canal morphology interpretation, while avoiding measurement errors caused by human subjective factors and angle deviations, resulting in more accurate measurement results and achieving the beneficial effect of reducing the risk of treatment complications.

[0021] The automatic segmentation and analysis system of this invention has a clear structure and complete functions, realizing full automation from CBCT image data processing, model training, root canal segmentation to quantitative morphological analysis. The measurement and analysis results of this system are highly consistent with the judgment of senior endodontic specialists and are significantly better than the manual interpretation results of junior doctors. It can effectively assist clinicians in improving their three-dimensional understanding of the root canal system morphology, providing accurate morphological parameters for preoperative root canal treatment, helping doctors to develop personalized treatment plans, reducing the risk of treatment complications such as instrument separation and root canal perforation, and improving the efficacy and accuracy of root canal treatment. Attached Figure Description

[0022] Figure 1 This is a flowchart of an automatic segmentation and analysis method for an oral imaging root canal system according to an embodiment of this application; Figure 2 This is a flowchart illustrating the training of a pseudo-label semi-supervised segmentation model using labeled and unlabeled datasets in an embodiment of this application. Figure 3 This is a flowchart illustrating the three-dimensional connected component analysis of the root canal three-dimensional segmentation results, the extraction of the target root canal region, and the quantitative analysis of the root canal curvature morphology in the target root canal region in this embodiment of the application. Figure 4 This is a schematic diagram of the structure of an automatic segmentation and analysis system for an oral imaging root canal system according to an embodiment of this application; Figure 5 This is a schematic diagram of the pseudo-label semi-supervised segmentation model in the embodiments of this application; Figure 6 This is a schematic diagram of the morphological quantitative analysis module in an embodiment of this application. Detailed Implementation

[0023] The following is in conjunction with the appendix Figure 1-6 The technical solutions of this application are further described in a complete and detailed manner to enable those skilled in the art to fully implement this application. It should be understood that the following embodiments are only for explaining this application and are not intended to limit the scope of protection of this application. All other implementation methods obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0024] The technical solutions of the various embodiments of this application can be combined with each other, but must be based on the ability of a person skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0025] Example 1 Please see Figures 1-3 This embodiment provides an automatic segmentation and analysis method for root canal systems based on oral CBCT images. The overall process is as follows: Figure 1 As shown, the specific steps include: Step S1: Annotation and preprocessing of oral CBCT image data. This step is used to construct a standardized dataset that meets the requirements of model training and provides a data foundation for subsequent semi-supervised training.

[0026] Specifically, raw cone-beam computed tomography (CBCT) images in DICOM format were acquired for oral clinical use, including but not limited to root canal images of different tooth positions such as anterior teeth, premolars, and molars. Two specialists with more than 5 years of clinical experience in endodontics manually annotated the root canal regions layer by layer to form an annotated dataset; the unannotated CBCT images were used as an unannotated dataset for pseudo-label semi-supervised learning.

[0027] Perform preprocessing operations uniformly on both labeled and unlabeled datasets, specifically including: (1) Organize the dataset format: First, convert and organize the original DICOM data into a NIfTI (.nii or .nii.gz) format dataset. Only when the data organization structure and file format meet the requirements can the subsequent automatic analysis, experimental planning and preprocessing processes proceed smoothly.

[0028] (2) Extract dataset information: Statistical analysis of key attributes of the entire dataset, including image size, voxel spacing and gray intensity distribution, to provide a basis for the automatic generation of subsequent preprocessing strategies.

[0029] (3) Preprocessing planning: Based on the overall characteristics of the dataset obtained in the previous step, a preprocessing scheme adapted to the current task is automatically generated, including the target image size, target voxel spacing and intensity processing method, etc., and a unified preprocessing operation is performed on all data accordingly to ensure the consistency of the input data in spatial scale and intensity distribution.

[0030] (4) Gray-level normalization: First, the mean, standard deviation, and 0.5% and 99.5% quantiles of gray levels are calculated within the foreground voxel range. Then, the gray-level values ​​are truncated to the corresponding quantile intervals, followed by standardization. This process helps to reduce abnormal gray-level fluctuations caused by factors such as metal artifacts and soft tissue noise, and unifies the gray-level distribution of images under different scanning devices and different scanning parameters, thereby improving the stability and generalization ability of model training. After the above preprocessing, the spatially aligned and gray-level normalized labeled dataset and unlabeled dataset are output as inputs for model training.

[0031] Step S2: Training and inference of the root canal 3D segmentation model based on pseudo-label semi-supervised learning. This step is based on the nn-UNet 3D medical image segmentation framework to construct a pseudo-label semi-supervised segmentation model. This achieves high-precision 3D segmentation of the root canal system while significantly reducing the amount of manual annotation. The training process combines... Figure 2 The explanation is as follows: S21. Supervised training of the initial nn-UNet model is performed using the labeled dataset to obtain a baseline model with preliminary segmentation capabilities.

[0032] Specifically, the initial nn-UNet model is trained under supervision using labeled datasets as supervised samples. The nn-UNet model adopts a three-dimensional encoder-decoder structure. The encoder layer extracts shallow and deep features of the root canal through continuous three-dimensional convolution and downsampling. The decoder layer restores the feature map size and fuses detailed information through deconvolution and skip connections, outputting a root canal segmentation probability map with the same size as the input image.

[0033] The training parameters were set as follows: learning rate 1e-4, batch size 2, AdamW optimizer, and a combination of Dice loss and cross-entropy loss. The training iterations were 200 epochs. After training, a baseline model with preliminary root canal segmentation capabilities was obtained, which can generate preliminary segmentation results from unlabeled CBCT images.

[0034] S22. Use the baseline model to predict the unlabeled dataset and generate a pseudo-labeled dataset; Using the trained baseline model for forward inference on the unlabeled dataset, the image to be segmented is first processed according to the preprocessing scheme corresponding to the training stage. Then, a sliding window prediction is used to perform forward inference on the entire 3D volume data in blocks, and the prediction results of each window are fused. The network output is usually voxel-level logits / probability maps for each category. Subsequently, argmax is used to discriminate and assign a final category label to each voxel, instead of using a fixed 0.5 threshold for general segmentation. After obtaining the prediction results, the predicted probability map or segmentation results in the preprocessing space are resampled back to the original image space and original shape, thereby generating a 3D segmentation mask consistent with the manually labeled format. In addition, according to the post-processing scheme automatically determined in the validation stage, the inference results can also undergo optional post-processing, such as retaining only the largest connected components, to further improve the stability and anatomical rationality of the segmentation results.

[0035] S23. Merge the labeled dataset and the pseudo-labeled dataset into an expanded training set, and retrain the pseudo-labeled semi-supervised segmentation model under the nn-UNet framework to obtain the trained pseudo-labeled semi-supervised segmentation model. The original labeled dataset and the pseudo-labeled dataset are merged at a preset ratio to form an expanded training set; under the nn-UNet framework, the model is jointly retrained using the expanded training set, and the training parameters are kept consistent with S21.

[0036] The retrained model performs forward inference on an unlabeled dataset. First, it processes the image to be segmented according to the preprocessing scheme from the training phase. Then, it uses sliding window prediction to perform forward inference on blocks of the entire 3D volume data, fusing the prediction results from each window. The network output is typically voxel-level logits / probability maps for each category. Subsequently, argmax is used to assign a final category label to each voxel, instead of using a fixed 0.5 threshold for general segmentation. After obtaining the prediction results, the predicted probability map or segmentation results in the preprocessing space are resampled back to the original image space and original shape, thereby generating a 3D segmentation mask consistent with the manually labeled format. Furthermore, based on the post-processing scheme automatically determined during the validation phase, the inference results can undergo optional post-processing, such as retaining only the largest connected components, to further improve the stability and anatomical rationality of the segmentation results.

[0037] After training, a final pseudo-label semi-supervised segmentation model is obtained. This model can directly input oral CBCT images and output three-dimensional segmentation results of root canals, achieving fully automatic and accurate differentiation of root canals from background areas such as dentin, periodontal ligament, and bone tissue. During training, the model simultaneously learns the precise boundary features of labeled data and the potential structural patterns of unlabeled data, reducing pseudo-label noise interference and improving the robustness of segmentation for complex root canal regions (such as branches and curved segments).

[0038] Step S3: Perform three-dimensional connected component analysis on the three-dimensional segmentation results of the root canals to extract the target root canal region. Quantitatively analyze the root canal curvature morphology of the target root canal region, automatically calculating the curvature angle and location of the root canals. This step involves post-processing and anatomical quantification of the three-dimensional segmentation results, and fully automatically extracting the root canal curvature angle and location. The specific process is detailed below. Figure 3 The explanation is as follows: Step S31: Count the number of voxels in each connected region of the root canal 3D segmentation results, remove connected regions from noisy areas, sort the remaining candidate connected regions according to their spatial relationships, and select the target root canal region. Specifically, perform 3D connected region analysis on the root canal 3D segmentation results: (1) Count the number of voxels in all connected components and remove noisy connected components with less than 1000 voxels; (2) Sort the remaining candidate connected domains according to the spatial location of the tooth and the size of the connected domain, and select the connected domain with the largest volume that matches the clinical target tooth position as the target root canal region.

[0039] The above steps eliminate segmentation noise, ensuring that subsequent morphological analysis is performed only on the actual root canal structure.

[0040] Step S32: Identify branches in the selected target root canal region and locate the key endpoints of each branch; The three-dimensional segmentation mask of the target root canal region is scanned layer by layer along the z-axis (the direction of the tooth's long axis). Two-dimensional connected component marking is performed in each layer, and the branch numbering is inherited and updated through the overlapping relationship of connected components between layers.

[0041] Specifically, if the connected components of the current layer overlap with those of the previous layer, the branch number of the previous layer is inherited; if the connected components of the current layer do not overlap with those of the previous layer, it is determined to be a new branch and a new branch number is assigned; if the number of connected components in the current layer is less than that in the previous layer, it is determined to be a branch termination and the downward scanning stops.

[0042] After branch identification is completed, locate the key endpoints for each root canal branch: Lower reference point C: The geometric center of the layer where the branch first appears; Upper reference point A: The geometric center point where the branch extends to the bottom layer.

[0043] The aforementioned endpoints serve as anatomical reference points for subsequent centerline construction and bending calculations.

[0044] Step S33: Construct the root canal centerline based on the key endpoints, and perform a smooth fit on the centerline, specifically including: (1) Extract all foreground voxels within the target root canal region and group them according to the z-axis layer number; (2) Calculate the average coordinates of all voxels in the current branch within each layer, and select the real foreground voxel closest to the average position as the center point of the layer; (3) Connect the center points of each layer in ascending order of z-axis to obtain the initial discrete center line; (4) The initial discrete centerline is smoothly fitted by a cubic B-spline curve to reconstruct a continuous and smooth three-dimensional centerline. The non-anatomical effective segments at the beginning and end are trimmed to obtain the final smooth fitted centerline. The smoothing process eliminates the jitter of discrete points and ensures the continuity and accuracy of the bending calculation.

[0045] Step S34: Calculate the bending angle and bending position of the root canal based on the smoothed centerline.

[0046] Based on the smoothly fitted centerline, for each sampling point on the centerline Perform the following calculations: (1) with Construct two vectors for the vertices , as follows:

[0047] (2) Calculate using the vector dot product formula and The included angle The included angle is The local curvature angle of the point;

[0048] (3) Calculate the bending angle of all sampling points along the center line, and select the sampling point corresponding to the minimum angle as the root canal bending position.

[0049] This calculation method directly reflects the true curvature of the root canal relative to its two ends, is not affected by the perspective of the cross section, and completely avoids the angular deviation and subjective error of manual measurement.

[0050] Through the aforementioned steps S1-S3, this method achieves full automation from CBCT image input to root canal 3D segmentation, bending angle, and position output. Specifically, by annotating and preprocessing oral CBCT image data, labeled and unlabeled datasets are obtained. A pseudo-label semi-supervised segmentation model is trained using these datasets to achieve 3D segmentation of root canal structures in oral CBCT images, resulting in 3D segmentation results. This significantly reduces the model's dependence on high-quality labeled data, achieving high-precision 3D segmentation of the root canal system. Furthermore, 3D connected component analysis is performed on the root canal 3D segmentation results to extract target root canal regions. Quantitative analysis of the root canal bending morphology is then performed on these target root canal regions, automatically calculating the bending angle and position of the root canals. This significantly improves the efficiency of root canal morphology interpretation while avoiding measurement errors caused by subjective human factors and angle deviations, resulting in more accurate measurement results and reducing the risk of treatment complications.

[0051] Example 2 This embodiment provides a system for performing the method described in Embodiment 1, with the following structure: Figure 4 As shown, it includes a data processing module, a pseudo-label semi-supervised segmentation module, a morphological quantitative analysis module, a result output module, and a model storage and retrieval module. The modules communicate with each other and coordinate their functions to form a complete intelligent analysis system for oral root canal images.

[0052] The data processing module is used to collect, manually annotate, align layers, and normalize grayscale of raw oral CBCT images, and output standardized annotated and unannotated datasets to provide a unified input for model training and eliminate the impact of data differences on segmentation accuracy.

[0053] A pseudo-label semi-supervised segmentation model is used to train the pseudo-label semi-supervised segmentation model using the labeled dataset and the unlabeled dataset to achieve three-dimensional segmentation of the root canal system in oral CBCT images and obtain the three-dimensional segmentation results of the root canals.

[0054] Combination Figure 5 The diagram shown illustrates the structure of the pseudo-label semi-supervised segmentation model, which includes: Initial training unit: Configure the nn-UNet 3D network structure, perform supervised training on the labeled dataset, and output the baseline model; Pseudo-label generation unit: Loads the baseline model and infers from the unlabeled dataset to generate a pseudo-labeled dataset; Joint retraining unit: By fusing real labeled datasets and pseudo labeled datasets and applying dynamic weight constraints, retraining is performed to obtain high-precision 3D segmentation results of root canals.

[0055] This model combines semi-supervised learning with 3D medical image segmentation. Even with a 70% reduction in labeled data, it still maintains a segmentation accuracy of Dice coefficient ≥ 0.98, which is beneficial for subsequent accurate quantitative analysis.

[0056] Combination Figure 6 The diagram shown illustrates the structure of the morphological quantitative analysis module, which includes: The region extraction unit performs 3D connected component analysis, noise reduction, and sorting to output the target root canal region. Specifically, it counts the number of voxels in all connected components in the 3D segmentation results of the root canal, removes noisy connected components with fewer than 1000 voxels, retains candidate regions related to the real root canal structure, sorts the remaining candidate connected components according to the tooth position (e.g., anterior teeth, premolars, molars, etc.) and the volume of the connected components, and finally selects the connected component with the largest volume that matches the clinical target tooth position as the target root canal region. This ensures that subsequent morphological analysis is performed only on the real root canal structure, laying the foundation for accurately extracting the root canal curvature angle and position.

[0057] Branch identification and endpoint localization unit: used to scan layer by layer along the z-axis, identify branches through the inter-layer connected domain relationship, and locate the lower reference point C and the upper reference point A; Specifically, the 3D segmentation mask of the target root canal region is scanned layer by layer along the z-axis (the long axis of the tooth). Two-dimensional connected component marking is performed within each layer, and branch numbers are inherited and updated based on the overlap of connected components between layers. If the connected components of the current layer overlap with those of the previous layer, the branch number of the previous layer is inherited; if the connected components of the current layer do not overlap with those of the previous layer, it is determined to be a new branch, and a new branch number is assigned; if the number of connected components in the current layer is less than that in the previous layer, it is determined to be a branch termination, and the downward scanning stops. After branch identification is completed, the key endpoints of each root canal branch are located: Lower reference point C: the geometric center point of the layer where the branch first appears; upper reference point A: the geometric center point where the branch extends to the lowest layer. These endpoints serve as anatomical reference points for subsequent centerline construction and bending calculations.

[0058] Centerline Construction and Smoothing Unit: The initial centerline is generated by calculating center points layer by layer and connecting them. A smooth centerline is obtained by B-spline fitting. Specifically, all foreground voxels within the target root canal region are extracted and grouped by z-axis layer number. Within each layer, the average coordinate of all voxels in the current branch is calculated, and the nearest real foreground voxel to this average position is selected as the center point of that layer. The center points of each layer are connected in ascending order along the z-axis to obtain the initial discrete centerline. A cubic B-spline curve is used to smooth and fit the initial discrete centerline, reconstructing a continuous and smooth three-dimensional centerline. Non-anatomical effective segments at both ends are trimmed to obtain the final smoothed centerline. The smoothing process eliminates discrete point jitter, ensuring the continuity and accuracy of the bending calculation.

[0059] Bending parameter calculation unit: Calculates the bending angle point-by-point based on the vector angle method, determining the minimum angle and corresponding bending position. Specifically, based on the smoothly fitted centerline, it calculates the bending angle at each sampling point on the centerline. Perform the following calculations: (1) with Construct two vectors for the vertices , as follows:

[0060] (3) Calculate using the vector dot product formula and The included angle The included angle is The local curvature angle of the point;

[0061] (3) Calculate the bending angle of all sampling points along the center line, and select the sampling point corresponding to the minimum angle as the root canal bending position.

[0062] This calculation method directly reflects the true curvature of the root canal relative to its two ends, is not affected by the perspective of the cross section, and completely avoids the angular deviation and subjective error of manual measurement.

[0063] The above units are executed sequentially, converting the segmentation results into anatomical quantitative parameters that can be directly used in clinical practice.

[0064] In addition, the segmentation analysis system of this application also includes a result output module and a model storage and retrieval module; The result output module is used to perform three-dimensional visualization rendering of the root canal three-dimensional segmentation mask, smooth center line, bending angle, and three-dimensional coordinates of bending position, and automatically generate a standardized analysis report containing tooth position, segmentation confidence, bending angle value, and bending coordinates. It supports export and printing, making it convenient for doctors to quickly formulate treatment plans.

[0065] The model storage and retrieval module is used to store the trained model weight file, optimization parameters and preprocessing rules. It supports one-click loading of the model and rapid inference for newly entered CBCT images. The analysis time for a single image is ≤10 seconds, which meets the needs of efficient clinical diagnosis.

[0066] The technical solution of this application, through the synergistic formation of various technical features, can effectively solve a series of technical problems in existing technologies, such as segmentation relying on a large amount of annotation, low efficiency and poor accuracy of morphological interpretation. It has significant practical implications for improving the precision of root canal treatment and reducing the risk of treatment complications. Specifically: By combining preprocessing with a pseudo-label semi-supervised segmentation model, layer alignment and grayscale normalization provide stable and consistent data input for pseudo-label training, significantly improving the convergence speed of the baseline model and reducing pseudo-label noise. Combined with a high-completeness segmentation mask output by the pseudo-label semi-supervised model, a reliable foundation is provided for 3D connected component screening and branch recognition, avoiding branch misjudgment and centerline offset caused by segmentation breaks and omissions. Furthermore, by combining centerline fitting and bending calculation, smooth fitting eliminates discrete centerline jitter, making vector angle calculation continuous and stable, ensuring low measurement errors in bending angle and position, far superior to the deviation of manual measurement.

[0067] In summary, this application breaks through the technical bottlenecks of traditional root canal image analysis, which relies on a large amount of manual annotation, physician experience, and strong measurement subjectivity. It achieves low annotation cost, fully automatic, high precision, and repeatable root canal segmentation and quantitative analysis. The measurement results are highly consistent with the judgment of senior specialists, which can effectively reduce the risk of complications such as instrument separation and root canal perforation during root canal treatment, and significantly improve the accuracy and success rate of root canal treatment.

[0068] The foregoing description illustrates and describes preferred embodiments of this application. However, as previously understood, this application is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the conception herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of this application should be within the protection scope of the appended claims.

Claims

1. An automatic segmentation and analysis method for an oral imaging root canal system, characterized in that, Includes the following steps: Step S1: Label and preprocess the oral CBCT image data to obtain labeled datasets and unlabeled datasets; Step S2: Train the pseudo-label semi-supervised segmentation model using the labeled dataset and the unlabeled dataset to achieve three-dimensional segmentation of root canal structures in oral CBCT images and obtain the three-dimensional segmentation results of root canals. Step S3: Perform three-dimensional connected component analysis on the three-dimensional segmentation results of the root canal, extract the target root canal region, perform quantitative analysis on the root canal curvature of the target root canal region, and automatically calculate the curvature angle and curvature position of the root canal.

2. The automatic segmentation and analysis method for an oral imaging root canal system according to claim 1, characterized in that, The preprocessing includes performing layer alignment and grayscale normalization on the CBCT image data.

3. The automatic segmentation and analysis method for an oral imaging root canal system according to claim 1, characterized in that, Step S2, the process of training the pseudo-label semi-supervised segmentation model using the labeled dataset and the unlabeled dataset, includes: S21. Supervised training of the initial nn-UNet model using the labeled dataset to obtain a baseline model with preliminary segmentation capabilities; S22. Use the baseline model to predict the unlabeled dataset and generate a pseudo-labeled dataset; S23. Merge the labeled dataset and the pseudo-labeled dataset into an expanded training set, and retrain the pseudo-labeled semi-supervised segmentation model under the nn-UNet framework to obtain the trained pseudo-labeled semi-supervised segmentation model.

4. The automatic segmentation and analysis method for a dental CBCT imaging root canal system according to claim 1, characterized in that, In step S3, three-dimensional connected component analysis is performed on the three-dimensional segmentation results of the root canal to extract the target root canal region. Quantitative analysis of the root canal curvature morphology within the target root canal region is then performed, including: Step S31: Count the number of voxels in each connected region in the three-dimensional segmentation result of the root canal, remove the connected regions in the noise region, sort the remaining candidate connected regions according to their spatial position, and select the target root canal region. Step S32: Identify branches in the selected target root canal region and locate the key endpoints of each branch; Step S33: Construct the root canal centerline based on the key endpoints and perform a smooth fit on the centerline; Step S34: Calculate the bending angle and bending position of the root canal based on the smoothed centerline.

5. The automatic segmentation and analysis method for an oral imaging root canal system according to claim 4, characterized in that, In step S32, the branch identification of the selected target root canal region and the location of the key endpoints of each branch include: The segmentation mask of the target root canal region is scanned layer by layer along the z-axis, and two-dimensional connected components are marked in each layer to identify the connected components in the current layer. The inheritance and updating of branch numbers are achieved by considering the overlap between the connected components in the current layer and the connected components in the previous layer. Specifically, if the connected components in the current layer overlap with those in the previous layer, the branch number is inherited; if the connected components in the current layer do not overlap with those in the previous layer, a new branch number is assigned; if the number of connected components in the current layer is reduced relative to the previous layer, the downward scanning is terminated, and branch identification is completed. After branch identification is completed, two key endpoints are further determined for each branch. The key endpoints include an upper reference point and a lower reference point. The lower reference point is the center point of the layer where the branch first appears, and the upper reference point is the center point where the branch extends to the bottom layer.

6. The automatic segmentation and analysis method for an oral imaging root canal system according to claim 4, characterized in that, In step S33, the root canal centerline is constructed based on the key endpoints, and the centerline is smoothly fitted, including: Extract all foreground voxels of the target root canal region and group them according to the z-axis direction. Calculate the average position of the branch point set in each layer, select the real voxel point closest to the average position as the center point of that layer, and connect the center points of each layer in order according to the z-axis direction to obtain the initial discrete center line. The initial discrete centerline is smoothly reconstructed using the B-spline curve fitting method, and the beginning and end portions of the smooth curve are trimmed to obtain the smoothly fitted centerline.

7. The automatic segmentation and analysis method for an oral imaging root canal system according to any one of claims 4, characterized in that, In step S34, the bending angle and bending position of the root canal are calculated based on the smoothly fitted centerline, specifically including: The method for calculating the bending angle of the root canal is as follows: Let any point on the centerline after smooth fitting be... The upper reference point and lower reference point of the key endpoint are C and A, respectively. Construct two vectors for the vertices , as follows: but , The angle formed by two corresponding vectors The bending angle of the root canal: in, The centerline after smooth fitting is characterized at point The local bending state relative to the line connecting the upper and lower reference points; Calculate the angle point by point along the centerline after smooth fitting, and select the point corresponding to the smallest included angle as the corresponding point of the bending position.

8. An automatic segmentation and analysis system for an oral imaging root canal system, characterized in that, include: The data processing module is used to collect, annotate, and preprocess oral CBCT image data to obtain an annotated dataset and an unannotated dataset. The pseudo-label semi-supervised segmentation module is used to train the pseudo-label semi-supervised segmentation model using the labeled dataset and the unlabeled dataset, so as to achieve three-dimensional segmentation of the root canal system in oral CBCT images and obtain the root canal three-dimensional segmentation results. The morphological quantitative analysis module performs quantitative analysis of the root canal bending morphology based on the three-dimensional segmentation results of the root canal, and automatically calculates the bending angle and bending position of the root canal.

9. The automatic segmentation and analysis system for oral CBCT images according to claim 8, characterized in that, The pseudo-label semi-supervised segmentation model includes: The initial training unit is used to train the initial nnUNet model using the labeled dataset to obtain the baseline model; A pseudo-label generation unit is used to predict and generate a pseudo-label dataset from the unlabeled dataset using the baseline model. The joint retraining unit is used to fuse the labeled dataset and the pseudo-labeled dataset for retraining to obtain the root canal 3D segmentation result.

10. The automatic segmentation and analysis system for oral imaging root canal systems according to claim 8, characterized in that, The morphological quantitative analysis module includes: The region extraction unit is used to extract the target root canal region based on the three-dimensional segmentation results of the root canal; The branch identification and endpoint localization unit is used to complete the identification of root canal branches and the localization of key endpoints. A centerline construction and smoothing unit is used to construct the root canal centerline based on the located key endpoints and to smooth the root canal centerline. The bending parameter calculation unit is used to calculate the root canal bending angle and bending position.

Images