A method for diagnosing root furcation lesion multi-structure segmentation based on CBCT

Abstract

The application discloses a medical image processing and computer-aided diagnosis technical field, and relates to a root furcation lesion multi-structure segmentation and diagnosis method and system based on CBCT. The method comprises the following steps: acquiring CBCT three-dimensional body data of a target multi-root and performing preprocessing to obtain standardized preprocessed body data. The application carries out double-path parallel automatic analysis and processing on the root furcation area by performing voxel-level three-dimensional segmentation and ring-cut section key point marking on the standardized preprocessed body data respectively, and verifies the two-path analysis results with each other. The complementary analysis of the complex anatomical structure is carried out from the three-dimensional voxel space and the two-dimensional sequence section two dimensions. Through geometric consistency verification and fusion correction based on trustworthiness, the problem that a single analysis method is easily disturbed by image noise or local morphological change under complex anatomical conditions is avoided, so that automatic quantitative measurement and standardized grading diagnosis of the root furcation lesion are realized.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and computer-aided diagnosis technology, specifically a multi-structure segmentation diagnosis method for root bifurcation lesions based on CBCT. Background Technology

[0002] Furcation involvement (FI) is a common periodontal lesion occurring in the interroot region of multi-rooted teeth and is a key factor influencing the prognosis and treatment strategy selection for multi-rooted teeth. Accurate diagnosis of FI relies on three-dimensional perception of interroot and alveolar bone defects, currently clinically primarily depending on probing (Nabers probe) and two-dimensional imaging. Probing is significantly affected by probe entry angle, calculus interference, and root surface morphology; horizontal and vertical defects are easily misdiagnosed. Literature reports that clinical probing accuracy for FI is only 27%, with an underestimation rate as high as 75% for grades II / II-III, and perforation defects are frequently missed. Two-dimensional imaging suffers from projection overlap and lack of buccal-lingual information, failing to reflect the true spatial structure of the furcation region. Although three-dimensional cone-beam computed tomography (CBCT) can provide three-dimensional structure, the anatomical morphology of the furcation region is extremely complex, with small interroot bone volume and extremely thin bone plates; manual interpretation of images is not only time-consuming but also highly subjective. Clinicians exhibit significant human bias when determining the most severe level of a lesion, making it difficult to perform standardized three-dimensional quantitative measurements (such as horizontal defects, vertical depth, root trunk height, etc.) and to automatically output standardized diagnostic conclusions.

[0003] The current diagnosis of root bifurcation lesions lacks precise quantitative standards, and doctors often need to rely on experience to make judgments, especially in complex cases. There are significant diagnostic differences and a high rate of misjudgment. Traditional methods are difficult to fully assess the extent and severity of lesions, and it is easy to overlook early lesions or underestimate disease progression, thus affecting the formulation of treatment plans and prognostic assessments. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-structure segmentation diagnostic method for root bifurcation lesions based on CBCT, so as to solve the problems mentioned above, such as reliance on manual judgment, lack of quantification, and high misjudgment rate.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a multi-structure segmentation diagnostic method for root bifurcation lesions based on CBCT, comprising the following steps: S1. Acquire CBCT three-dimensional volumetric data of the target multi-root tooth and preprocess it to obtain standardized preprocessed volumetric data; S2. Construct a diagnostic model. Input the preprocessed volume data into the diagnostic model for parallel and collaborative dual-path analysis to obtain a multi-structure segmentation body and a set of key points. Map the set of key points to the same three-dimensional coordinate system as the multi-structure segmentation body and perform consistency verification to obtain the final key point positions and corresponding lesion segmentation boundaries. S3. Based on the final key point location and the corresponding lesion segmentation boundary, the maximum horizontal bone defect distance, vertical defect depth and bone defect volume are automatically calculated, and the root bifurcation lesion diagnosis type is automatically obtained according to the preset diagnostic rules. S4. The diagnostic model automatically generates a structured report containing quantitative parameters and diagnostic conclusions based on the diagnostic type of root bifurcation lesion.

[0006] As a further aspect of the present invention: step S1 includes; The CBCT three-dimensional volume data includes voxels, voxel spatial coordinates, voxel spacing, voxel grayscale data, and target tooth number; The volume data matrix is ​​resampled isotropically and processed into isotropic voxel data with uniform resolution based on three-dimensional trilinear interpolation, so that the voxels have consistent resolution in three-dimensional space. The voxel grayscale data is normalized using the Z-Score method to obtain standard voxel grayscale data. The isotropic voxel data is denoised in three dimensions by performing three-dimensional nonlocal mean denoising, and the standard voxel grayscale data is denoised in three dimensions by performing three-dimensional unsharpened masking to obtain structure-enhanced standard voxel grayscale data. The processed CBCT 3D volumetric data forms standardized preprocessed volumetric data.

[0007] As a further aspect of the present invention: step S2 includes; Standardized preprocessed body data is input into the diagnostic model for parallel and collaborative dual-path analysis, the dual-path analysis including; S21. Extract the local volume containing the target multi-rooted tooth from the preprocessed volume data, perform pixel-level classification on the local volume to obtain a multi-structure segmentation body containing tooth roots, interroot bone, buccal and palatal bone plates and root bifurcation lesion candidate areas, and generate lesion closure boundaries based on the root bifurcation lesion candidate areas in the multi-structure segmentation body. S22. Determine the geometric center of the root bifurcation region and the direction of the long axis of the target multi-rooted tooth based on the multi-structure segmentation body, and establish a local coordinate system for the root bifurcation. Generate circumferential cutting surfaces with equal angular intervals around the long axis of the tooth in a plane perpendicular to the long axis of the tooth. Automatically locate and mark a set of key points on the circumferential cutting surfaces, including the root bifurcation point, the lowest point of the root bifurcation lesion, and the intersection of the lesion bone resorption area and the root surface. S23. Map the set of key points to the same root bifurcation local coordinate system as the multi-structure segment, calculate the spatial distance between the key points of the circumferential cutting surface and the lesion closure boundary, and the axial difference distance. If the spatial distance exceeds the first preset threshold and the axial difference distance is less than the second preset threshold, perform weighted fusion correction on the position of the corresponding key point to generate the final key point position and the corresponding lesion segmentation boundary verified by consistency.

[0008] As a further aspect of the present invention: step S21 includes; Let V be the preprocessed volume data, and let V be the local volume extracted from the preprocessed volume data. f The number of categories of multi-structured segments is denoted as C; Construct a 3D semantic segmentation network model, and integrate V f The input to the 3D semantic segmentation network model produces a C-class nonnormalized response r for each voxel space coordinate X. C (X), the nonnormalized response r C (X) The structural probability q is obtained through normalized mapping. C (X), the training of the 3D semantic segmentation network model adopts a combined loss function, which is: ; in These are the weighting coefficients; Weighted for; ; in, For truth labels, For category weights, For structural probability, It is a constant; Weighted cross-entropy for; ; Boundary loss term for; ; in, Number of categories The truth boundary distance field; The initial segment is obtained by discretizing the structural probability qC(X) based on the argmax function. ; , The initial segment Based on 3D connectivity filtering, the main connected components adjacent to the target tooth's space of interest are retained and discrete noise is removed to obtain the boundary of the root foreground region of the multi-root adhesion region. Distance transformation is then performed on the multi-root adhesion region. ; Where R represents the boundary of the root foreground region. The boundary is defined by y, where y represents the area between the tooth root and the surrounding tissues. Coordinates of any voxel point on the surface; D(x) is segmented into different root substructures by watershed or morphological segmentation, resulting in a multi-structure segmentation body M containing multiple categories of root, interroot bone, buccal and palatal bone plates, and candidate areas for root bifurcation lesions. seg Among them, the root bifurcation lesion candidate region of the multi-structure segment is given in the form of a closed voxel set and the lesion closure boundary is generated; The root structure, interroot bone structure, and bone defect area were identified based on the multi-structure segmentation.

[0009] As a further aspect of the present invention: step S22 includes; The geometric center of all voxel coordinates in the root bifurcation region segment is calculated based on the multi-structure segment and denoted as C. f The direction of the major axis of the target multi-rooted tooth is obtained by fitting the data, and is denoted as the unit direction vector. a Establishing a three-dimensional space based on C f Centered on, a The root bifurcation local coordinate system is the reference direction along the axial axis.

[0010] As a further aspect of the present invention: a 360-degree circumferential incision is performed on the candidate area of ​​the root bifurcation lesion in a plane perpendicular to the tooth long axis, generating circumferential incision surfaces with equal angular intervals, and several key points, including the root bifurcation point, the lowest point of the root bifurcation lesion, and the intersection of the lesion bone resorption area and the root surface, are automatically located and marked on each circumferential incision surface. A key point detection model is constructed. The key point coordinates are input into the key point detection model, and a two-dimensional Gaussian heat map is predicted and output. The peak position is extracted from the Gaussian heat map as the predicted coordinates of the key points, and several key point coordinates are obtained. Several key point coordinates form a key point set. The key point detection model is trained using the adjacent cross-section consistency constraint loss. The angular interval is between 0.5 degrees and 5 degrees.

[0011] As a further aspect of the present invention: step S23 includes; Calculate the axial difference distance between the lowest key point along the tooth long axis in the key point set and the lowest point of the lesion closure boundary along the tooth long axis. If the spatial distance exceeds a preset threshold and the axial difference distance is less than the preset axial tolerance, the position of the corresponding key point is weighted and fused based on the trust level to generate the final key point position and the corresponding lesion segmentation boundary after consistency verification.

[0012] As a further aspect of the present invention: step S3 includes; The longitudinal analysis interval is determined based on the tooth root structure, interroot bone structure, bone defect area, root bifurcation entry point and the lowest point of root bifurcation lesion. The longitudinal analysis interval is selected along the tooth long axis, and the horizontal bone defect amount and the maximum horizontal bone defect distance of the selected level are calculated. The vertical defect depth is obtained by calculating the distance along the long axis of the tooth between the root bifurcation point and the lowest point of the root bifurcation lesion. The horizontal classification of the root bifurcation lesion is determined based on whether the maximum horizontal bone defect distance forms a buccal-palatal penetration; the vertical subclass of the root bifurcation lesion is determined based on the vertical defect depth; and the bone defect volume is calculated based on the voxel number of the root bifurcation lesion candidate region of the multi-structure segment. The horizontal classification and vertical subclass of root bifurcation lesions are combined, and the complete diagnostic type of root bifurcation lesion is automatically obtained according to the preset diagnostic rules.

[0013] As a further aspect of the present invention: step S4 includes; The quantification parameters include the target tooth number; The diagnostic conclusion includes the determination of whether root bifurcation lesions exist; If a bifurcation lesion is diagnosed, the corresponding horizontal bone defect amount, vertical bone defect depth, diagnostic type of bifurcation lesion, and bone defect volume will be output simultaneously. The target tooth number, the judgment result, and the corresponding horizontal bone defect amount, vertical bone defect depth, root bifurcation lesion diagnosis type, and bone defect volume are used to form a structured report.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention performs dual-path parallel automatic analysis of the root bifurcation region by performing voxel-level three-dimensional segmentation and key point marking on the circumferential section of the preprocessed standardized preprocessed volume data. The results of the two analyses are mutually verified, and the complex anatomical structure is analyzed complementaryally from two dimensions: three-dimensional voxel space and two-dimensional sequence section. Through geometric consistency verification and trust-based fusion correction, the problem of single analysis methods being easily interfered with by image noise or local morphological changes under complex anatomical conditions is avoided. This enables automatic quantitative measurement and standardized hierarchical diagnosis of root bifurcation lesions. Based on the three-dimensional spatial structure, the horizontal, vertical and volumetric dimensions of root bifurcation lesions are objectively quantitatively measured, and the most severe defect layer is automatically selected for analysis within the actual lesion range. The diagnostic conclusion is directly derived from the image data itself, reducing the subjective differences caused by reliance on experience judgment and manual examination. This effectively improves the efficiency and standardization of imaging diagnosis of root bifurcation lesions and reduces the workload of clinicians in complex image interpretation and repetitive measurements.

[0015] 2. The present invention automates and objectifies the process, realizing fully automated processing from CBCT three-dimensional volume data input to structured diagnostic report output, greatly reducing human intervention and subjective bias, improving the standardization and repeatability of diagnosis. Based on three-dimensional spatial information, it can automatically calculate key parameters reflecting the severity of lesions and adaptively select the most severe defect level within the actual lesion range, making the quantitative results more in line with clinical reality. It can also automatically output standardized root bifurcation lesion diagnosis types, which facilitates clinical decision-making and medical record management. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow structure of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention; Figure 3 A schematic diagram of the input CBCT three-dimensional volume data for the present invention is provided. Figure 4 This invention provides a schematic diagram of the multi-structure segmentation of the system; Figure 5 A schematic diagram of key point markings for the system provided in this invention; Figure 6 This diagram illustrates the mutual verification and result fusion of the system provided for this invention.

[0017] The diagram shows: 1. Image input and preprocessing module; 2. Multi-structure automatic segmentation and labeling module; 3. Quantitative measurement and automatic diagnosis module; 4. Report generation module. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example: Please see Figure 1 This embodiment provides a multi-structure segmentation diagnostic method for root bifurcation lesions based on CBCT, including the following steps: S1. Acquire CBCT three-dimensional volumetric data of the target multi-root tooth and preprocess it to obtain standardized three-dimensional volumetric data; This process acquires CBCT 3D volumetric data of target multi-root teeth in the patient's oral cavity. CBCT 3D volumetric data can originate from an image archiving and communication system or be directly imported from the image acquisition device, supporting a file format conforming to medical imaging standards (DICOM). During import, the CBCT 3D volumetric data is read, including voxels, voxel spatial coordinates, voxel spacing, voxel grayscale data, and spatial parameter data. Spatial parameter data includes scan range and coordinate orientation data. The volumetric data matrix refers to the 3D array structure of voxels in a 3D medical image (such as CBCT), where each voxel contains voxel spacing and voxel grayscale data. Spatial parameter information includes scan range and coordinate orientation data. The CBCT 3D volumetric data is then loaded into the system as raw 3D volumetric data for subsequent preprocessing steps.

[0020] The volume data matrix is ​​resampled isotropically and processed into isotropic voxel data of uniform resolution using 3D trilinear interpolation, ensuring consistent voxel resolution in 3D space. For anisotropic voxels generated by CBCT equipment from different manufacturers, resampling is performed, and 3D trilinear interpolation is used to convert the volume data into isotropic voxels of uniform resolution. This step ensures that the millimeter-level horizontal and vertical displacement measurements performed in subsequent steps have consistent physical meaning in all directions of 3D space.

[0021] The voxel grayscale data was normalized using the Z-Score method to obtain standard voxel grayscale data. The Z-Score method was then used to standardize the grayscale, eliminating grayscale interference from equipment exposure parameters on the identification of periodontal thin bone plates. After this processing step, the 3D CBCT images achieved a unified standard in spatial scale and grayscale representation, providing stable input for subsequent analysis.

[0022] Isotropic voxel data was processed using 3D non-local means denoising to obtain 3D denoised isotropic voxel data. Standard voxel grayscale data was processed using a 3D unsharpened mask to obtain structurally enhanced standard voxel grayscale data. After geometric and grayscale correction, 3D denoising and structural enhancement were performed on the processed CBCT 3D volumetric data. To address the common problems of speckle noise and insufficient local contrast in CBCT images, 3D non-local means denoising (NLM-3D) was used to spatially enhance the isotropic voxel data, reducing the impact of background noise on the recognition of subtle anatomical structures. A 3D unsharpened mask was used to process the standard voxel grayscale data, enhancing the display of root boundaries and thin bone plate structures, improving their clarity in 3D images. This method ensures anatomical realism and improves the recognition of root bifurcation structures, providing a high-quality image foundation for subsequent localization and segmentation.

[0023] The 3D denoised isotropic voxel data, the structure-enhanced standard voxel grayscale data, and the spatial parameter data form the preprocessed volume data.

[0024] S2. Construct a diagnostic model. Input the preprocessed volume data into the diagnostic model for parallel and collaborative dual-path analysis to obtain a multi-structure segmentation body and a set of key points. Map the set of key points to the same three-dimensional coordinate system as the multi-structure segmentation body and perform consistency verification to obtain the final key point positions and corresponding lesion segmentation boundaries. S3. Based on the final key point location and the corresponding lesion segmentation boundary, the maximum horizontal bone defect distance, vertical defect depth and bone defect volume are automatically calculated, and automatic grading diagnosis is performed according to the preset diagnostic rules and preset classification to obtain automatic grading data. S4. The diagnostic model automatically generates a structured report containing quantitative parameters and diagnostic conclusions based on the diagnostic type of root bifurcation lesion.

[0025] Specifically, the present invention automates and objectifies the process, realizing fully automated processing from CBCT three-dimensional volume data input to structured diagnostic report output, greatly reducing human intervention and subjective bias, and improving the standardization and repeatability of diagnosis.

[0026] This invention performs dual-path parallel automatic analysis of root bifurcation regions by performing voxel-level three-dimensional segmentation and key point marking on the circumferential section of preprocessed standardized volumetric data. The results of the two analyses are mutually verified, and the complex anatomical structure is analyzed complementaryally from two dimensions: three-dimensional voxel space and two-dimensional sequence section. Through geometric consistency verification and trust-based fusion correction, the problem of single analysis methods being susceptible to interference from image noise or local morphological changes under complex anatomical conditions is avoided. This enables automatic quantitative measurement and standardized grading diagnosis of root bifurcation lesions. Based on the three-dimensional spatial structure, the horizontal, vertical, and volumetric dimensions of root bifurcation lesions are objectively quantitatively measured, and the most severe defect layer is automatically selected for analysis within the actual lesion range. The diagnostic conclusions are directly derived from the image data itself, reducing the subjective differences caused by reliance on experience judgment and manual examination. This effectively improves the efficiency and standardization of imaging diagnosis of root bifurcation lesions and reduces the workload of clinicians in interpreting complex images and performing repetitive measurements.

[0027] Based on three-dimensional spatial information, this invention can automatically calculate key parameters reflecting the severity of lesions (horizontal defects, vertical depth, and defect volume), and adaptively select the most severe defect level within the actual lesion range, making the quantitative results more in line with clinical reality. It can also automatically output standardized root bifurcation lesion diagnostic types (such as FI, IIB, IIIC, etc.), which facilitates clinical decision-making and medical record management.

[0028] In this embodiment, step S2 includes: Standardized preprocessed body data is input into the diagnostic model for parallel and collaborative dual-path analysis, which includes: S21. Extract the local volume containing the target multi-rooted tooth from the preprocessed volume data, perform pixel-level classification on the local volume, and obtain a multi-structure segmentation body containing tooth roots, interroot bones, buccal and palatal bone plates, and root bifurcation lesion candidate areas. Generate lesion closure boundaries based on the root bifurcation lesion candidate areas in the multi-structure segmentation body. S22. Determine the geometric center of the root bifurcation region and the direction of the long axis of the target multi-rooted tooth based on the multi-structure segmentation body, and establish a local coordinate system for the root bifurcation. Generate circumferential cutting surfaces with equal angular intervals around the long axis of the tooth in a plane perpendicular to the long axis of the tooth. Automatically locate and mark a set of key points on the circumferential cutting surfaces, including the root bifurcation point, the lowest point of the root bifurcation lesion, and the intersection of the lesion bone resorption area and the root surface. S23. Map the set of key points to the same root bifurcation local coordinate system as the multi-structure segment body, calculate the spatial distance between the key points of the circumferential cutting surface and the lesion closure boundary, as well as the axial difference distance. If the spatial distance exceeds the first preset threshold and the axial difference distance is less than the second preset threshold, perform weighted fusion correction on the position of the corresponding key point to generate the final key point position and the corresponding lesion segmentation boundary after consistency verification.

[0029] In this embodiment, step S21 includes: Let V be the preprocessed volume data, and let V be the local volume extracted from the preprocessed volume data. f The number of categories of multi-structured segments is denoted as C; Construct a 3D semantic segmentation network model, and integrate V f The input to the 3D semantic segmentation network model produces a C-class nonnormalized response r for each voxel space coordinate X. C (X), the nonnormalized response r C (X) The structural probability q is obtained through normalized mapping. C (X), the training of the 3D semantic segmentation network model adopts a combined loss function, which is: ; in These are the weighting coefficients; Weighted for; ; in, For truth labels, For category weights, For structural probability, It is a constant; used to enhance the learning intensity of small-volume structures such as interosseous bone, bone plates, and lesion candidate areas.

[0030] Weighted cross-entropy for; To further improve the geometric accuracy of the thin bone plate and the defect edge, a boundary loss term is introduced. It uses the distance transformation of the truth boundary as weight to strengthen the constraint on voxel error near the prediction boundary; Boundary loss term for; ; in, Number of categories The truth boundary distance field; used to make the bone plate and the edge of the lesion more closely fit the real anatomical contour in three-dimensional space. The structural probability q C (X) The initial segment is obtained by discretization based on the argmax function. ; , The initial segment Based on 3D connectivity filtering, the main connected components adjacent to the target tooth's space of interest are retained and discrete noise is removed to obtain the boundary of the root foreground region of the multi-root adhesion region. Distance transformation is then performed on the multi-root adhesion region. ; Where R represents the boundary of the root foreground region. The boundary is defined by y, where y represents the area between the tooth root and the surrounding tissues. Coordinates of any voxel point on the surface; D(x) is segmented into different root substructures by watershed or morphological segmentation, resulting in a multi-structure segmentation body M containing multiple categories of root, interroot bone, buccal and palatal bone plates, and candidate areas for root bifurcation lesions. seg This ensures that subsequent measurements of the root bifurcation inlet and interroot bone are not affected by root adhesion. Among them, the root bifurcation lesion candidate area of ​​the multi-structure segment is given in the form of a closed voxel set and the lesion closure boundary is generated. The post-processed results are output as a multi-structure segmentation body, Mseg. The candidate regions for root bifurcation lesions are presented as closed voxel sets, maintaining topological adjacency with interroot bone, lamina, and root trunk structures in the same coordinate system. This allows subsequent calculations of horizontal and vertical defects and defect volumes to directly extract spatial boundaries and perform standardized quantization based on Mseg. Simultaneously, this 3D segmentation output serves as a spatial reference for a second "360° circumferential section marking system," enabling cross-validation and consistency checks between the two systems.

[0031] The root structure, interroot bone structure, and bone defect area were identified based on the multi-structure segmentation.

[0032] In this embodiment, step S22 includes: The geometric center of the coordinates of all voxels in the root bifurcation region segment is calculated based on the multi-structure segment and denoted as C. f The direction of the major axis of the target multi-rooted tooth is obtained by fitting the data, and is denoted as the unit direction vector. a Establishing a three-dimensional space based on C f Centered on, a The root bifurcation local coordinate system is the reference direction along the axial axis.

[0033] Specifically, after completing the multi-structure segmentation in voxel space, the system determines the spatial reference frame for root bifurcation analysis based on the 3D segmentation results of the root bifurcation region. The system first calculates the geometric center of the coordinates of all voxels in the segmented volume of the root bifurcation region, denoted as... C f ，The geometric center serves as the rotation center for subsequent circumcision operations. Simultaneously, the system fits the overall morphology of the target multi-rooted tooth to obtain the tooth's long axis direction. This long axis can be obtained by fitting the principal direction of root voxel distribution or the principal axis from the crown to the apex, and is denoted as a unit direction vector. a Through the above steps, the system establishes a three-dimensional structure based on... C f Centered on, with a The root bifurcation local coordinate system, which serves as an axial reference, provides a unified benchmark for the subsequent generation of the circumferential cutting surface.

[0034] In this embodiment, a 360-degree circumferential incision is performed on the candidate area of ​​the root bifurcation lesion around the long axis of the tooth in a plane perpendicular to the long axis of the tooth, generating circumferential incision surfaces with equal angular intervals. On each circumferential incision surface, several key points, including the root bifurcation point, the lowest point of the root bifurcation lesion, and the intersection of the lesion bone resorption area and the root surface, are automatically located and marked. A key point detection model is constructed. The key point coordinates are input into the key point detection model, and a two-dimensional Gaussian heat map is predicted and output. The peak position is extracted from the Gaussian heat map as the predicted coordinates of the key points, resulting in several key point coordinates. Several key point coordinates form a key point set. The key point detection model is trained using the adjacent cross-section consistency constraint loss. The angular intervals range from 0.5 degrees to 5 degrees.

[0035] In this embodiment, step S23 includes: Calculate the axial difference distance between the lowest key point along the tooth long axis in the key point set and the lowest point of the lesion closure boundary along the tooth long axis. If the spatial distance exceeds a preset threshold and the axial difference distance is less than the preset axial tolerance, the position of the corresponding key point is weighted and fused based on the trust level to generate the final key point position and the corresponding lesion segmentation boundary after consistency verification.

[0036] Specifically, in one embodiment, preferably, the angle interval is set to 2 degrees, thereby generating a total of 180 circumferential cutting surfaces. After obtaining the geometric center of the root bifurcation and the direction of the tooth long axis, a 360-degree circumferential cutting operation is performed on the root bifurcation region around the tooth long axis in a plane perpendicular to the tooth long axis. The complete 360-degree angle range is divided into an equally spaced angle sequence, constructing a circumferential cutting operation through the geometric center C. f And the normal vector follows Rotational cut The 3D segment is then resampled on this cross-section to obtain the corresponding 2D cross-sectional image. This results in a sequence of two-dimensional cross-sections that completely cover the root bifurcation region. This sequence reflects the spatial distribution characteristics of root bifurcation structures and lesion areas in different orientations.

[0037] In each circumferential cut surface Above, three categories of key points closely related to root bifurcation diagnosis are defined and labeled. The first category of key points is the root bifurcation point. The first key point is defined as the coronal junction of the root bifurcation on this incisal plane, used to reflect the spatial location of the root bifurcation entrance in different orientations. The second type of key point is the lowest point of the root bifurcation lesion. This is defined as the point along the long axis of the tooth where the furcation lesion reaches its deepest point within the cut surface, used to characterize the extreme value of bone resorption in the vertical direction. The third type of key point is the intersection of the furcation lesion bone resorption area and the tooth root. This is defined as the location on the cut surface where the bone defect area contacts or intersects with the root surface, used to describe the boundary of horizontal bone support loss. These three types of points together constitute the geometric feature representation of the bifurcation lesion in that orientation on each cut surface.

[0038] The sequence of cut surfaces obtained after completing a 360° circumferential cut Then, each cross-sectional image is associated with its corresponding angle parameters. The cross-sectional geometric information and the target tooth number are registered as a training sample entry, and a sample index table is established. 180 cross-sections of the same tooth are regarded as the same "sequence sample" for subsequent consistency constraints and batch training; at the same time, it is allowed to input a single cross-section as an independent sample into the model to improve the training sample size and orientation coverage.

[0039] Generate three types of keypoint ground truth labels for each circumferential cut surface, namely the root bifurcation point. The lowest point of root bifurcation lesions along the long axis of the tooth and the intersection of the bone resorption area of ​​the root bifurcation lesion and the root surface. In a preferred implementation, the initial positions of the aforementioned key points are not manually marked one by one on the cut surface, but are based on the multi-structure segmentation body output by Sb1. Automatic generation through rule-based methods ensures consistency of annotation logic across different sections, tooth positions, and cases, and significantly reduces the workload required for manual annotation.

[0040] Specifically, multi-structure segmentation In the corresponding circumferential cutting surface Resampling is performed on the data to obtain the corresponding segmentation plane. Based on the topological relationship between the root category boundary and the root bifurcation anatomical region, the root bifurcation point is automatically determined in the segmented section. Based on the geometric extreme position of the root bifurcation lesion area along the long axis of the tooth, the lowest point of the lesion is automatically determined. Based on the intersection of the lesion boundary and the root surface boundary on the cut plane, the intersection point between the bone resorption area and the root surface is automatically determined. Using the above method, a complete set of candidate key points is automatically generated for each cross-section.

[0041] After automatic derivation, a manual review and confirmation process is introduced for each generated keypoint. This manual review does not involve re-labeling the keypoints; instead, it involves checking each candidate keypoint location point by point. Reviewers confirm whether the location matches the actual anatomical position, making minor adjustments or corrections only if necessary. The keypoint locations confirmed or corrected through manual review are considered the final ground truth labels and are written into the training sample index for subsequent training and iterative updates of the automatic labeling model.

[0042] Supervised training of keypoints can be performed using one of two target representations. In one implementation, the keypoint coordinates are... Convert to a two-dimensional Gaussian heatmap ,in These are pixel coordinates, at this location To indicate the categories of key points, a heatmap is defined as follows: in These are scale parameters for different keypoint categories, used to control the extent of the supervised region. In another implementation, the supervised model output is directly expressed as a coordinate regression model. This reduces output dimensionality and improves inference efficiency. The two expressions mentioned above can be used independently or in combination to form a two-stage training objective of "heatmap guidance + coordinate refinement".

[0043] The cross-sectional image is input into the keypoint labeling network. The network outputs the prediction results for the corresponding key points. Under the heatmap-supervised method, the network outputs... And using pixel-level error as the training objective, the loss function can be defined as:

[0044] In coordinate regression, the network output is... And using coordinate error as the training objective, the loss function can be defined as:

[0045] Among them, the use of Distance is used to enhance robustness to outliers. To ensure the model possesses both localization stability and spatial continuity, an adjacent section consistency constraint loss is introduced in the optimal implementation:

[0046] The total loss is defined as:

[0047] in The weighting coefficients can be set according to the training stability, so that the model can accurately locate key points in a single cross-section and maintain smooth continuity in a 360° sequence.

[0048] Model training, parameter updates, and stopping conditions The model is trained using a mini-batch iterative approach, and its parameters are updated based on backpropagation. During training, the data is divided into training and validation sets, and the keypoint localization error is monitored in real time on the validation set. When the average localization error on the validation set no longer decreases within several consecutive iterations or reaches a preset threshold, training is stopped, and the optimal model parameters are saved, resulting in the trained, automatically labeled model. .

[0049] During the inference phase, each cross-sectional image is input into the trained model. When the model output is a heatmap, the peak position of the heatmap is used as the predicted coordinates of the key points.

[0050] The prediction results from 180 cross-sections were combined into three sets of key points. , , It is used for subsequent consistency verification with voxel segmentation and as a geometric reference for horizontal / vertical quantization measurements.

[0051] After predicting key points for all circumferential cut surfaces, spatial consistency analysis was performed on the marking results across the 180 cut surfaces. This consistency analysis, based on the ordered nature of the cut surfaces in the angular dimension, used a sequential geometric constraint algorithm to constrain the spatial variation of key points of the same type in adjacent cut surfaces, ensuring the continuity and stability of key points in the 360-degree direction.

[0052] In the implementation, the coordinates of key points on each cross-section are uniformly mapped to the same three-dimensional coordinate system, and a key point sequence is established according to the corresponding angular order of the cross-sections. For any type of key point, its spatial displacement vector between adjacent cross-sections is calculated, and the displacement is evaluated based on a preset geometric continuity rule. When the displacement amplitude or direction change of a key point in an adjacent cross-section exceeds a reasonable range, the point is determined to not meet the sequence consistency requirements.

[0053] For key points that do not conform to consistency constraints, a sequence smoothing algorithm based on neighborhood cross-sections is used to correct their positions. This correction can be achieved by weighted averaging, interpolation reconstruction, or curve fitting of the positions of corresponding key points in multiple adjacent cross-sections, thereby maintaining the continuity of the key point's change trend in the angular sequence. The aforementioned sequence smoothing algorithm can be applied to the radial component and the component along the tooth long axis, resulting in a smooth and reasonable distribution of the root bifurcation entry point, the lowest point of the lesion, and the bone resorption boundary point in three-dimensional space.

[0054] After the above consistency constraints and corrections, the key point prediction results on 180 cross-sections are integrated into a continuous three-dimensional point set description around the root bifurcation region. This point set serves as a second set of analysis results output, independent of voxel segmentation, and is used for cross-validation with the root bifurcation structure and lesion region obtained based on voxel space segmentation. This provides a stable and precise geometric reference for subsequent horizontal quantization, vertical quantization, and volume calculation of root bifurcation lesions.

[0055] After completing the automatic multi-structure segmentation based on voxel space and the automatic keypoint marking based on the 360° circumferential section, the spatial coordinates of the two sets of output results are first unified. The root bifurcation points, the lowest points of root bifurcation lesions, and bone resorption boundary points output by the circumferential section marking are mapped back to the three-dimensional coordinate system of the voxel segmentation results, i.e., the root bifurcation local coordinate system, so that the analysis results of the two sets are comparable under the same spatial reference. Through this step, the section-level keypoints and voxel-level segmentation boundaries can be directly correlated and analyzed in three-dimensional space.

[0056] After completing the spatial mapping, the spatial relationship between the keypoints output by the cross-section markers and the lesion boundaries obtained by voxel segmentation is verified. For any keypoint obtained by cross-section markers, the boundary point with the closest spatial distance is searched in the root bifurcation lesion region boundary obtained by voxel segmentation, and the minimum spatial distance between the two is calculated. This distance is used to measure the local consistency between the two sets at this location, and its calculation method is as follows:

[0057] in, This indicates the coordinates of the key points output by the section marker. This represents the set of boundary points of the root bifurcation lesion region obtained from voxel segmentation. When the distance mentioned above is less than a preset spatial tolerance threshold, the two sets are determined to have spatial consistency at this key point location. The consistency results of each cross section within a 360° range are statistically analyzed to evaluate the degree of matching between the two sets in overall spatial positioning.

[0058] To verify the consistency of the two sets of data in the vertical localization of furcation lesions, further verification was performed based on the lowest point of the lesion along the long axis of the tooth. The deepest position of the furcation lesion along the long axis of the tooth was extracted from both the voxel segmentation results and the circumferential section marking results, and the axial difference between the two was calculated. This axial error reflects the consistency of the two sets in the vertical direction, and its calculation method is as follows:

[0059] in, This represents the minimum coordinate value of the lesion region along the long axis of the tooth in voxel segmentation. This represents the extreme value of the set of lowest points of the lesion in the section marking along the long axis of the tooth. When this error is less than the preset axial tolerance, the two sets are considered to be consistent in the vertical positioning of the lesion.

[0060] When two sets of data exhibit positional discrepancies in local areas but maintain overall consistency, the keypoint locations are fused and corrected, rather than directly discarding either output. The confidence levels of the keypoint locations for both the section markers and voxel segmentation are obtained, and the keypoint locations are then weighted and fused based on these confidence levels. The fusion result can be expressed as:

[0061] in, This indicates the location of key points in the cross-section marker output. This represents the location of the boundary point with the smallest spatial distance to the keypoint in the voxel segmentation. and These represent the confidence levels of the two sets of data at this location. Through the above fusion method, while preserving the sensitivity of the cross-section marker direction, the constraints of voxel segmentation on the integrity of the 3D structure are introduced to reasonably correct the key point locations.

[0062] After the aforementioned consistency verification and fusion correction, the output is a set of mutually verified root bifurcation key points and their corresponding lesion segmentation boundaries. The output simultaneously possesses the completeness of the voxel-level three-dimensional structure and the precision of the section-level key point localization, serving as a unified input basis for subsequent horizontal and vertical measurements of root bifurcation lesions and the calculation of defect volume. By introducing the aforementioned mutual verification mechanism of voxel segmentation and circumferential section marking, this invention significantly improves the reliability and consistency of the automatic analysis results of root bifurcation lesions in clinical applications.

[0063] In this embodiment, step S3 includes: The longitudinal analysis interval is determined based on the tooth root structure, interroot bone structure, bone defect area, root bifurcation entry point and the lowest point of root bifurcation lesion. The longitudinal analysis interval is selected along the tooth long axis, and the horizontal bone defect amount and the maximum horizontal bone defect distance of the selected level are calculated. Calculate the distance along the long axis of the tooth between the root bifurcation point and the lowest point of the root bifurcation lesion to obtain the vertical defect depth; The horizontal classification of the root bifurcation lesion is determined based on whether the maximum horizontal bone defect distance forms a buccal-palatal connection. The vertical subclass of the root bifurcation lesion is determined based on the vertical defect depth. The bone defect volume is calculated based on the voxel number of the root bifurcation lesion candidate region of the multi-structure segment. The horizontal classification and vertical subclass of root bifurcation lesions are combined, and the complete diagnostic type of root bifurcation lesion is automatically obtained according to the preset diagnostic rules.

[0064] Specifically, after obtaining multi-structure segmentation data of the root bifurcation region and completing the mutual verification of voxel segmentation and circumferential section marking, this unit first quantitatively measures the degree of bone loss in the horizontal direction of root bifurcation lesions. Based on the root structures, interroot bone structures, and bone defect areas identified in the voxel segmentation results, and combined with the verified root bifurcation points... Lowest point of root bifurcation lesion In order to determine the effective analysis range of root bifurcation lesions in three-dimensional space.

[0065] Specifically, the root fork entry point Lowest point of root bifurcation lesion The spatial interval between these points, along the long axis of the tooth, is defined as the longitudinal analysis interval for furcation lesions. Within this analysis interval, the furcation region is continuously segmented along the long axis of the tooth to generate multiple candidate horizontal measurement planes. Each candidate plane is substantially perpendicular to the long axis of the tooth and is used to reflect the lateral bone support of the furcation region at that height.

[0066] For each candidate horizontal measurement level, the two-dimensional projections of the bone defect area and the remaining interroot bone structure are extracted within that level, and the minimum horizontal distance between the bone defect area and the remaining interroot bone structure within that level is calculated as the corresponding horizontal bone defect amount. The horizontal bone defect amount is defined as the minimum Euclidean distance between the set of boundary points of the bone defect and the set of boundary points of the remaining interroot bone, and its calculation method is as follows:

[0067] Wherein, Bd represents the set of boundary points of the bone defect region within the candidate measurement plane, and Bb represents the set of boundary points of the remaining interroot bone structure within the same plane.

[0068] The horizontal bone defect amounts obtained from the candidate layers throughout the entire longitudinal analysis interval are compared, and the layer with the largest horizontal bone defect amount is automatically selected as the final reference layer for the horizontal quantitative measurement of root bifurcation lesions. This adaptive selection mechanism automatically locates the measurement position reflecting the most severe degree of root bifurcation lesions in three-dimensional space, thus avoiding the bias caused by manually setting a fixed measurement height.

[0069] Since the voxel segmentation results have distinguished and marked each root of a multi-rooted tooth (including MB, DB, palatal, or lingual roots), the aforementioned horizontal bone defect amount can be calculated separately for different root combinations. Based on clinically established rules, the measurement result corresponding to the most unfavorable direction is selected as the final quantitative value for the horizontal direction of the bifurcation lesion. The measurement results are output in millimeters and serve as the core quantitative basis for subsequent horizontal grading of bifurcation lesions (FI I–III classes). After completing the horizontal quantization, this unit further measures the vertical bone defect depth of the furcation lesion. First, the position of the furcation roof is identified in the 3D segmentation results and used as the starting reference point for vertical measurement. Subsequently, the deepest defect location along the long axis of the tooth is searched in the bone defect voxel set, which corresponds to the lowest point of the furcation lesion.

[0070] To eliminate the influence of single-section or local noise, a path search or extreme value search is performed on the bone defect area in three-dimensional space to determine the minimum coordinate value of the defect area along the long axis of the tooth. The vertical defect depth of a bifurcation lesion is defined as the distance between the top of the bifurcation and the lowest point of the lesion along the long axis of the tooth, and its calculation method is as follows:

[0071] in, This represents the coordinates of the top of the root fork in three-dimensional space. The three-dimensional coordinates of the lowest point of the lesion. This represents a unit vector along the long axis of the tooth. The vertical defect depth is output in millimeters and serves as a quantitative basis for the vertical subclass grading of root bifurcation lesions. Horizontal defect measurement of root bifurcation lesion With vertical defect depth Then, the root bifurcation lesions are automatically graded and determined according to the preset periodontal classification criteria.

[0072] In the horizontal direction, root bifurcation lesions are graded based on the amount of horizontal defect and whether a penetrating defect has formed. The pre-defined diagnostic rules are as follows:

[0073] Among them, the default will The setting is 3mm to fit the classic Hamp classification, but dynamic adjustments are allowed based on clinical guideline updates. Furthermore, after completing the voxel segmentation of the bifurcation lesion, the system performs connectivity analysis on the defective voxels in three-dimensional space. If there is no continuous alveolar bone voxel barrier in the buccal (or mesiodistal) direction, and the defective voxel forms a continuous through path between at least two opposite bifurcation entrances, then the bifurcation lesion is determined to be a through-and-through lesion, and a Class FI III diagnosis is output.

[0074] In the vertical direction, root bifurcation lesions are subclassified based on the depth of the vertical defect, and the determination rule can be expressed as follows:

[0075] By combining the horizontal grading results with the vertical subclass results, a complete diagnostic type for root bifurcation lesions, such as IIA, IIB, IIIC, etc., is generated, so that the degree of lesion can be objectively quantified in both orthogonal directions.

[0076] Furthermore, this unit automatically calculates the bone defect volume of root bifurcation lesions based on voxel segmentation results. The total volume of the lesion region is obtained by multiplying the number of voxels in the bone defect area by the volume of each voxel. The calculation method is as follows:

[0077] in, Indicates the number of voxels in the bone defect. This represents the physical volume corresponding to a single voxel. The volume index is used to reflect the overall scale of bone resorption in the root bifurcation region and can be used for quantitative comparison of lesion changes during follow-up.

[0078] Through the aforementioned quantitative measurement and automated diagnosis process, this invention achieves the automated transformation of root bifurcation lesions from three-dimensional imaging structure to objective numerical values, and then to standardized grading diagnosis, providing a reliable basis for clinical decision-making and efficacy evaluation. In this embodiment, step S4 includes: The quantification parameters include the target tooth number; The diagnostic conclusion includes the determination of whether root bifurcation lesions exist; If a bifurcation lesion is diagnosed, the corresponding horizontal bone defect amount, vertical bone defect depth, diagnostic type of bifurcation lesion, and bone defect volume will be output simultaneously. The target tooth number, the judgment result, and the corresponding horizontal bone defect amount, vertical bone defect depth, root bifurcation lesion diagnosis type, and bone defect volume are used to form a structured report.

[0079] Specifically, after completing the horizontal and vertical bone defect measurements, bone defect volume calculations, and lesion grading for furcation lesions, this unit performs a unified structured organization and parameterized output of the above analysis results. Based on tooth position as the basic unit, the quantitative indicators related to furcation lesions are categorized and integrated to form a standardized quantitative result set for furcation lesions.

[0080] The structured report includes: the target tooth number, the determination of whether a furcation lesion exists, and, if a furcation lesion is confirmed, further outputs the corresponding horizontal bone defect amount, vertical bone defect depth, diagnostic type of the furcation lesion, and bone defect volume. The presence of a furcation lesion is represented in binary form to distinguish between teeth without and with furcation lesions. When the determination is that a furcation lesion exists, the report automatically outputs the quantitative measurement results and diagnostic conclusions related to that tooth location. All parameters are expressed in uniform physical units, allowing the spatial extent and severity of the furcation lesion to be fully described numerically.

[0081] To support clinical applications and information integration, this unit outputs diagnostic results in structured data format. This output format can include, but is not limited to, common data formats such as PDF, XML, or JSON, thereby enabling integration with hospital-side image archives, image information, or hospital information. Through standardized field mapping, the quantitative indicators and diagnostic conclusions of root bifurcation lesions can be automatically written into electronic medical records or image reports, reducing manual data entry workload.

[0082] In a second aspect, the present invention provides a multi-structure segmentation and diagnosis system for root bifurcation lesions based on CBCT, the system comprising: Image input and preprocessing module 1 is used to acquire CBCT three-dimensional volume data of the target multi-root tooth and perform preprocessing to obtain standardized preprocessed volume data; The multi-structure automatic segmentation and labeling module 2 constructs a diagnostic model. The preprocessed volume data is input into the diagnostic model for parallel and collaborative dual-path analysis to obtain the multi-structure segment and key point set. The key point set is mapped to the same three-dimensional coordinate system as the multi-structure segment and consistency verification is performed to obtain the final key point position and the corresponding lesion segmentation boundary. The quantitative measurement and automatic diagnosis module 3 automatically calculates the maximum horizontal bone defect distance, vertical defect depth and bone defect volume based on the final key point position and the corresponding lesion segmentation boundary, and automatically obtains the root bifurcation lesion diagnosis type according to the preset diagnosis rules. The report generation module 4 automatically generates a structured report containing quantitative parameters and diagnostic conclusions based on the diagnostic type of root bifurcation lesion.

[0083] Input and preprocessing of CBCT 3D volumetric data (CBCT images); In this embodiment, the system acquires CBCT three-dimensional volumetric data containing the target multi-rooted teeth (e.g., Figure 3 As shown in the figure, the CBCT 3D volumetric data is in DICOM format and contains complete voxel grayscale information and pixel spatial resolution parameters. After receiving the image, the system performs preprocessing operations, including voxel size unification, grayscale normalization, and noise suppression, to ensure the stability and consistency of subsequent analysis processes.

[0084] Automatic segmentation and labeling of multiple structures; After image input and preprocessing, the system performs multi-structure segmentation and key anatomical point marking on the target multi-rooted tooth and its root bifurcation region to obtain basic data for quantitative analysis.

[0085] Automatic segmentation of multi-rooted teeth and root bifurcation regions based on voxel space; The preprocessed, standardized 3D volume data is denoted as V. The system first extracts the local volume Vi from the volume data Vi based on the coarse localization results of the target multi-rooted tooth. f The local volume V f Using the center of the target tooth and its long axis as references, the process extends axially, radially, and mesiodistally, covering the complete root structure from the cervical region to the apex. It further extends into the interradicular region and the buccal / palatal (lingual) bone plates, ensuring that the root trunk, interradicular bone, and potential lesion areas at the bifurcation are all contained within the same contiguous voxel space. This provides a unified spatial basis for subsequent multi-structure segmentation and lesion boundary closure (e.g., ...). Figure 4 ).

[0086] Based on this, the system will localize the volume V f Input to the voxel-level multi-structure segmentation module, for V f The voxels within the structure are analyzed on a voxel-by-voxel basis to obtain multi-structure voxel segmentation results. The segmentation results include at least the following structural categories: root structure of multi-rooted teeth, root trunk, interroot bone, buccal / palatal (lingual) lateral bone plate, and candidate areas for root bifurcation lesions. Each structure is represented in different voxel sets and maintains spatial consistency within the same three-dimensional coordinate system.

[0087] To address the anatomical specificity of furcation lesions, the system imposes spatial constraints on candidate areas of furcation lesions in the segmentation results. This ensures that the candidate lesion areas only appear within the anatomically relevant neighborhood of the furcation and maintain close spatial proximity to the root, root shaft, and interroot bone structures, thereby avoiding false detections far from the target tooth region (e.g., Figure 4 (As shown).

[0088] After the above processing, the system outputs the final multi-structure segmentation Mseg. Among them, the candidate regions of root bifurcation lesions are given in the form of closed voxel sets, and maintain clear topological adjacency relationships with the interradical bones, buccal / palatal (lingual) lateral plates and root trunk structures in a unified coordinate system. This allows subsequent horizontal defect measurement, vertical defect measurement and defect volume calculation to directly extract stable and continuous spatial boundaries based on Mseg and perform standardized quantization.

[0089] Automatic marking of root bifurcation key points based on 360° circumferential cutting surface; After completing the multi-structure segmentation based on voxel space, the system further performs key point marking analysis on the root bifurcation region based on 360° circumferential cutting planes. In this embodiment, the system first uses the geometric center of the root bifurcation region as the circumferential rotation center, and uses the tooth long axis direction of the target multi-root tooth as the axial reference, generating a sequence of circumferential cutting planes around this axis in a plane perpendicular to the tooth long axis.

[0090] In practical applications, the system can divide the complete 360° angular range into multiple equally spaced sections to achieve continuous coverage of the root bifurcation region in different orientations. To facilitate illustration and avoid image overlap due to an excessive number of sections, this embodiment only selects four representative circumferential sections for demonstration, such as... Figure 6 As shown in the figure, the cut surfaces correspond to different angular orientations, and their angular positions are clearly marked in the figure to illustrate the distribution relationship of the circumferential cut surfaces.

[0091] On each circumferential section, the system marks key anatomical points closely related to the analysis of furcation lesions. These key points include: cementoenamel junction (CEJ), apex (APEX), furcation entry point (F), lowest point of the furcation lesion (F-BD), and the boundary points where the bone resorption area of ​​the furcation lesion meets the root surface (F-AC1, F-AC2). These key points are defined using a uniform marking rule in each section and distinguished by different colors or symbols to maintain clear consistency across different sections.

[0092] like Figure 5 As shown, although the root bifurcation structure and lesion morphology exhibit different cross-sectional features in the images at different angles of the circumferential section, key points of the same type always correspond to the same anatomical semantic location. For example, the root bifurcation entry point F is always located at the coronal junction of the root bifurcation in each section; the lowest point of the root bifurcation lesion F-BD is located at the deepest point of the lesion area along the long axis of the tooth; while F-AC1 and F-AC2 respectively reflect the boundary position of the lesion area in contact with the root surface in that section, and are used to describe the extent of horizontal bone support loss.

[0093] Through the above method, the system, within a 360° circumferential framework, transforms the geometric features of root bifurcation lesions in different orientations into a set of key points with clear anatomical significance. Although Figure 5 The example only shows four circumferential cut surfaces and their corresponding key point marking results. However, in actual operation, the system will perform the same key point marking process on the complete circumferential cut surface sequence to obtain a continuous, multi-directional key point description around the root bifurcation region, providing a stable geometric basis for subsequent sequence consistency correction, quantitative measurement and diagnostic analysis.

[0094] Mutual verification between the voxel segmentation system and the ring section marking system; After completing the automatic segmentation of multiple structures based on voxel space and the automatic marking of root bifurcation key points based on 360° circumferential section, the system performs mutual verification and consistency correction on the two sets of analysis results to improve the reliability and stability of the spatial localization results of root bifurcation lesions.

[0095] First, the system performs a unified spatial coordinate mapping on the outputs of the two systems. Specifically, the system maps the root bifurcation entry point (F), the lowest point of the root bifurcation lesion (F-BD), and the intersection points (F-AC1, F-AC2) of the lesion bone resorption area and the root surface from the circumferential section marker system back to the three-dimensional coordinate system where the voxel segmentation results are located. This mapping process is based on the geometric relationship of the circumferential section in three-dimensional space and uses the tooth long axis as the global reference axis, thus making the section-level key point results and the voxel-level lesion segmentation boundary comparable under the same spatial reference system.

[0096] After completing the spatial mapping, the system verifies the consistency of the local spatial relationship between the keypoints and the voxel segmentation results. For any mapped keypoint, the system searches for the boundary point with the closest spatial distance in the root bifurcation lesion region boundary obtained by the voxel segmentation system and calculates the minimum spatial distance between them. The minimum distance is used to measure the local consistency between the circumferential section marking result and the voxel segmentation result at that location. When the distance is less than a preset spatial tolerance threshold, the system determines that the two systems have spatial consistency at the keypoint location; when the distance is close to or exceeds the tolerance threshold, it determines that there is a potential inconsistency at that location.

[0097] To visually reflect the degree of consistency at different spatial locations, the system maps the distance relationship between key points and lesion boundaries into a consistency heatmap, and uses a unified color coding method to represent different levels of matching results. High consistency areas are represented by cool colors, areas close to the tolerance are represented by intermediate colors, and areas exceeding the tolerance are limited to a small number of warm-colored areas, such as... Figure 6 As shown.

[0098] In addition to local spatial consistency, the system further verifies the consistency of vertical positioning of root bifurcation lesions along the long axis of the tooth. The system extracts the deepest position of the lesion region along the long axis of the tooth from the voxel segmentation results and extracts the axial extreme value position composed of multiple F-BD points from the circumferential section marking results. By comparing the coordinate differences between the two systems along the long axis of the tooth, the consistency of the two systems in vertical lesion positioning is evaluated. When the axial difference is less than a preset axial tolerance threshold, the system determines that the positioning results of the two systems are consistent in the vertical direction.

[0099] When there is a certain positional deviation in a local area but the overall consistency still meets the preset requirements, the system does not directly discard the output results of either system. Instead, it performs result fusion and position correction based on confidence level. The system obtains the confidence level information of the key point positions from the circumferential section marking system and the voxel segmentation system, respectively, and performs weighted fusion of the candidate key point positions of the two systems based on the confidence level to generate the corrected final key point positions. Through this fusion method, the system retains the sensitivity of the circumferential section marking system to directional changes while introducing the constraints of the voxel segmentation system in terms of 3D structural integrity, thereby obtaining more stable and reliable root bifurcation key point results.

[0100] After the above mutual verification and consistency correction processes, the system outputs a set of root bifurcation key points and their corresponding lesion segmentation boundaries after fusion correction, which serve as the unified input basis for subsequent horizontal measurement, vertical measurement and defect volume calculation of root bifurcation lesions.

[0101] Root bifurcation quantification measurement and automatic diagnosis; After obtaining the set of key points for root bifurcation and their corresponding lesion segmentation boundaries that have undergone mutual verification and consistency correction, the system performs standardized quantitative measurements on root bifurcation lesions and generates automatic diagnosis and classification results based on preset classification standards. The quantitative measurements include at least indicators such as root bifurcation lesion volume, vertical defect depth, and horizontal bone support loss, and all of these indicators are calculated in a unified three-dimensional coordinate system.

[0102] First, the system determines the set of closed voxels in the root bifurcation lesion region based on the voxel segmentation results, and calculates the lesion volume by combining the voxel physical size parameters of the volume data. Specifically, the system counts the number of lesion voxels and multiplies them by the volume of each voxel to obtain the overall volume of the root bifurcation lesion, which reflects the overall scale of the lesion in three-dimensional space.

[0103] Secondly, the system calculates the vertical defect depth based on the key point results output by the circumferential incision marking system and after consistency correction. The system uses the furcation entry point (F) as the reference position for vertical measurement and the lowest point of the furcation lesion (F-BD) as the extreme position of the lesion along the long axis of the tooth, calculating the distance between the two along the long axis. The system statistically analyzes the measurement results in all directions within the 360° circumferential incision range, selecting the maximum value as the maximum vertical defect depth for this case, to characterize the most severe degree of vertical bone resorption in the furcation region.

[0104] Next, the system calculates the degree of horizontal bone support loss. Using consistency-corrected bone resorption boundary points (F-AC1, F-AC2), the system determines the boundary position where the lesion intersects with the root surface in that orientation, and calculates the horizontal distance between these boundary points to characterize the extent of the horizontal defect in that orientation. The system statistically analyzes the horizontal defect distances in each orientation within a 360° range, taking the maximum value as the maximum horizontal bone resorption distance for that case. Simultaneously, the system determines whether the lesion forms a morphological feature connecting the buccal and palatal (lingual) sides based on the distribution relationship of boundary points in different orientations, and uses this as one of the auxiliary criteria for classification.

[0105] After completing the quantitative measurements of volume, vertical defects, and horizontal defects, the system automatically classifies the measurement results according to the preset classification criteria for root bifurcation lesions. The classification criteria comprehensively consider at least: (1) the maximum horizontal bone resorption distance; (2) the maximum vertical defect depth; and (3) the presence of morphological features such as buccal-palatal (lingual) lateral penetration. The system matches the measurement results with the classification threshold / judgment rules and outputs the corresponding classification labels.

[0106] In this embodiment, the system calculated the root bifurcation lesion volume to be 30.4 mm³; the maximum distance from the root bifurcation entry point to the lowest point of the lesion (F–F-BD) was 4.014 mm; the maximum horizontal bone resorption distance was 9.945 mm, and the lesion formed a penetrating connection in the buccal and palatal (lingual) directions. Based on the above measurement results and classification rules, the system automatically classified this example as FI III-B and wrote this classification result into the structured report as part of the automatic diagnostic output.

[0107] Root bifurcation lesion report output; In this embodiment, the system outputs the quantitative measurement results, automatic diagnostic conclusions, and classification results of root bifurcation lesions in JSON structured data format. The report includes both numerical measurement indicators and categorical diagnostic information. In the aforementioned JSON report, the system outputs the classification result of root bifurcation lesions as an independent field, clearly indicating the classification system used and the corresponding classification label, thus ensuring the traceability and scalability of the classification results. Through this structured output method, the system can synchronously transmit the classification information and quantitative measurement results of root bifurcation lesions to imaging reporting systems, electronic medical record systems, or other clinical information management platforms.

[0108] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-structure segmentation diagnostic method for root bifurcation lesions based on CBCT, characterized in that, Includes the following steps: S1. Acquire CBCT three-dimensional volumetric data of the target multi-root tooth and preprocess it to obtain standardized preprocessed volumetric data; S2. Construct a diagnostic model. Input the preprocessed volume data into the diagnostic model for parallel and collaborative dual-path analysis to obtain a multi-structure segmentation body and a set of key points. Map the set of key points to the same three-dimensional coordinate system as the multi-structure segmentation body and perform consistency verification to obtain the final key point positions and corresponding lesion segmentation boundaries. S3. Based on the final key point location and the corresponding lesion segmentation boundary, the maximum horizontal bone defect distance, vertical defect depth and bone defect volume are automatically calculated, and the root bifurcation lesion diagnosis type is automatically obtained according to the preset diagnostic rules. S4. The diagnostic model automatically generates a structured report containing quantitative parameters and diagnostic conclusions based on the diagnostic type of root bifurcation lesion.

2. The CBCT-based multi-structure segmentation diagnostic method for root bifurcation lesions according to claim 1, characterized in that, Step S1 includes: The CBCT three-dimensional volume data includes voxels, voxel spatial coordinates, voxel spacing, voxel grayscale data, and target tooth number; The volume data matrix is ​​resampled isotropically and processed into isotropic voxel data with uniform resolution based on three-dimensional trilinear interpolation, so that the voxels have consistent resolution in three-dimensional space. The voxel grayscale data is normalized using the Z-Score method to obtain standard voxel grayscale data. The isotropic voxel data is denoised in three dimensions by performing three-dimensional nonlocal mean denoising, and the standard voxel grayscale data is denoised in three dimensions by performing three-dimensional unsharpened masking to obtain structure-enhanced standard voxel grayscale data. The processed CBCT 3D volumetric data forms standardized preprocessed volumetric data.

3. The CBCT-based multi-structure segmentation diagnostic method for root bifurcation lesions according to claim 2, characterized in that, Step S2 includes: Standardized preprocessed body data is input into the diagnostic model for parallel and collaborative dual-path analysis, the dual-path analysis including; S21. Extract the local volume containing the target multi-rooted tooth from the preprocessed volume data, perform pixel-level classification on the local volume to obtain a multi-structure segmentation body containing tooth roots, interroot bone, buccal and palatal bone plates and root bifurcation lesion candidate areas, and generate lesion closure boundaries based on the root bifurcation lesion candidate areas in the multi-structure segmentation body. S22. Determine the geometric center of the root bifurcation region and the direction of the long axis of the target multi-rooted tooth based on the multi-structure segmentation body, and establish a local coordinate system for the root bifurcation. Generate circumferential cutting surfaces with equal angular intervals around the long axis of the tooth in a plane perpendicular to the long axis of the tooth. Automatically locate and mark a set of key points on the circumferential cutting surfaces, including the root bifurcation point, the lowest point of the root bifurcation lesion, and the intersection of the lesion bone resorption area and the root surface. S23. Map the set of key points to the same root bifurcation local coordinate system as the multi-structure segment, calculate the spatial distance between the key points of the circumferential cutting surface and the lesion closure boundary, and the axial difference distance. If the spatial distance exceeds the first preset threshold and the axial difference distance is less than the second preset threshold, perform weighted fusion correction on the position of the corresponding key point to generate the final key point position and the corresponding lesion segmentation boundary verified by consistency.

4. The CBCT-based multi-structure segmentation diagnostic method for root bifurcation lesions according to claim 3, characterized in that, Step S21 includes: Let V be the preprocessed volume data, and let V be the local volume extracted from the preprocessed volume data. f The number of categories of multi-structured segments is denoted as C; Construct a 3D semantic segmentation network model, and integrate V f The input to the 3D semantic segmentation network model produces a C-class nonnormalized response r for each voxel space coordinate X. C (X), the nonnormalized response r C (X) The structural probability q is obtained through normalized mapping. C (X), the training of the 3D semantic segmentation network model adopts a combined loss function, which is: ； in These are the weighting coefficients; Weighted for; ； in, For truth labels, For category weights, For structural probability, It is a constant; Weighted cross-entropy for; ； Boundary loss term for; ； in, Number of categories The truth boundary distance field; The initial segment is obtained by discretizing the structural probability qC(X) based on the argmax function. ; ， The initial segment Based on 3D connectivity filtering, the main connected components adjacent to the target tooth's space of interest are retained and discrete noise is removed to obtain the boundary of the root foreground region of the multi-root adhesion region. Distance transformation is then performed on the multi-root adhesion region. ； Where R is the boundary of the foreground region of the tooth root. The boundary is defined by y, where y represents the area between the tooth root and the surrounding tissues. Coordinates of any voxel point on the surface; D(x) is segmented into different root substructures by watershed or morphological segmentation, resulting in a multi-structure segmentation body M containing multiple categories of root, interroot bone, buccal and palatal bone plates, and candidate areas for root bifurcation lesions. seg Among them, the root bifurcation lesion candidate region of the multi-structure segment is given in the form of a closed voxel set and the lesion closure boundary is generated; The root structure, interroot bone structure, and bone defect area were identified based on the multi-structure segmentation.

5. The CBCT-based multi-structure segmentation diagnostic method for root bifurcation lesions according to claim 4, characterized in that, Step S22 includes: The geometric center of all voxel coordinates in the root bifurcation region segment is calculated based on the multi-structure segment and denoted as C. f The direction of the major axis of the target multi-rooted tooth is obtained by fitting the data, and is denoted as the unit direction vector. a Establishing a three-dimensional space based on C f Centered on, a The root bifurcation local coordinate system is the reference direction along the axial axis.

6. The CBCT-based multi-structure segmentation diagnostic method for root bifurcation lesions according to claim 5, characterized in that, A 360-degree circumferential incision is performed around the tooth long axis in a plane perpendicular to the tooth long axis to generate circumferential incision surfaces with equal angular intervals. On each circumferential incision surface, several key points, including the root bifurcation point, the lowest point of the root bifurcation lesion, and the intersection of the lesion bone resorption area and the root surface, are automatically located and marked. A key point detection model is constructed. The key point coordinates are input into the key point detection model, and a two-dimensional Gaussian heat map is predicted and output. The peak position is extracted from the Gaussian heat map as the predicted coordinates of the key points, and several key point coordinates are obtained. Several key point coordinates form a key point set. The key point detection model is trained using the adjacent cross-section consistency constraint loss. The angular interval is between 0.5 degrees and 5 degrees.

7. The CBCT-based multi-structure segmentation diagnostic method for root bifurcation lesions according to claim 6, characterized in that, Step S23 includes: Calculate the axial difference distance between the lowest key point along the tooth long axis in the key point set and the lowest point of the lesion closure boundary along the tooth long axis. If the spatial distance exceeds a preset threshold and the axial difference distance is less than the preset axial tolerance, the position of the corresponding key point is weighted and fused based on the trust level to generate the final key point position and the corresponding lesion segmentation boundary after consistency verification.

8. The CBCT-based multi-structure segmentation diagnostic method for root bifurcation lesions according to claim 7, characterized in that, Step S3 includes: The longitudinal analysis interval is determined based on the tooth root structure, interroot bone structure, bone defect area, root bifurcation entry point and the lowest point of root bifurcation lesion. The longitudinal analysis interval is selected along the tooth long axis, and the horizontal bone defect amount and the maximum horizontal bone defect distance of the selected level are calculated. The vertical defect depth is obtained by calculating the distance along the long axis of the tooth between the root bifurcation point and the lowest point of the root bifurcation lesion. The horizontal classification of the root bifurcation lesion is determined based on whether the maximum horizontal bone defect distance forms a buccal-palatal penetration; the vertical subclass of the root bifurcation lesion is determined based on the vertical defect depth; and the bone defect volume is calculated based on the voxel number of the root bifurcation lesion candidate region of the multi-structure segment. The horizontal classification and vertical subclass of root bifurcation lesions are combined, and the complete diagnostic type of root bifurcation lesion is automatically obtained according to the preset diagnostic rules.

9. The CBCT-based multi-structure segmentation diagnostic method for root bifurcation lesions according to claim 8, characterized in that, Step S4 includes: The quantification parameters include the target tooth number; The diagnostic conclusion includes the determination of whether root bifurcation lesions exist; If a bifurcation lesion is diagnosed, the corresponding horizontal bone defect amount, vertical bone defect depth, diagnostic type of bifurcation lesion, and bone defect volume will be output simultaneously. The target tooth number, the judgment result, and the corresponding horizontal bone defect amount, vertical bone defect depth, root bifurcation lesion diagnosis type, and bone defect volume are used to form a structured report.

Images