Periodontal disease image-guided accurate diagnosis targeted calibration method and system

Abstract

The invention discloses a periodontal disease image-guided accurate diagnosis targeted calibration method and system, relates to the technical field of medical image processing, and provides the following scheme: obtaining three-dimensional bone density distribution data of a jaw bone area of a patient, extracting a topological coordinate set of a bone defect area and a gray value of each point in the topological coordinate set, and calculating the three-dimensional bone density distribution data of the jaw bone area; acquiring a scattered signal intensity sequence of the patient along the surface of the tooth root and spatial pose data of the patient, and constructing an affine transformation matrix of the spatial pose data and a three-dimensional space coordinate system where the topological coordinate set is located based on a preset bimodal physical mark point in the mouth of the patient; and aligning the scattering signal intensity sequence to a topological coordinate set by using the affine transformation matrix. According to the method, the scattering signal sequence and the three-dimensional bone density distribution data are subjected to spatial registration, the included angle between the gradient vector and the bone microstructure dominant vector is analyzed, the limitation that traditional CBCT depends on the static bone density form is broken through, and spatial relevance quantization of pathological redirection of a periodontal lesion region and the bone microstructure dominant direction is achieved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically, this application relates to a method and system for targeted calibration of image-guided precision diagnosis of periodontal diseases. Background Technology

[0002] In the modern periodontal disease diagnosis and treatment system, image-guided intelligent recommendation systems have become a core support for clinical decision-making. Cone-beam computed tomography (CBCT) is widely used clinically to construct three-dimensional models of the jawbone, supporting the development of multiple generations of image-assisted recommendation systems. A typical workflow is as follows: After acquiring CBCT data of the patient's maxillofacial region, a bone density distribution map is generated through grayscale thresholding and three-dimensional reconstruction; then, an edge detection algorithm is used to identify the boundaries of bone defects, and a pathological severity score is calculated based on three-dimensional morphological parameters (such as defect depth and volume); finally, this score is matched with a preset rule base to output a recommendation list including the coordinates of the targeted intervention area, the degree of intervention, and priority ranking. With the upgrading of precision medicine needs, the underlying image analysis model still samples static morphological representations, failing to capture the temporal biological process of bone resorption and the spatially guided evolution of microstructures, leading to bottlenecks in the accuracy and severity of the targeted area.

[0003] In the analysis of periodontal disease images, existing technologies have the following limitations: traditional imaging methods (such as CBCT) can only statically present the morphology of bone defects, making it difficult to quantify the dynamic activity of bone resorption and the main direction of bone microstructure. This results in the inability to effectively distinguish between active lesions and old injuries. Furthermore, the definition of the target area depends on the doctor's empirical judgment of the morphology of bone defects, which affects the definition of the subsequent intervention area and the setting of the intervention degree. Therefore, an image-guided precision diagnosis and targeting calibration method and system for periodontal diseases is proposed to solve this problem. Summary of the Invention

[0004] To address the aforementioned technical issues, this technical solution provides a method and system for image-guided precision diagnosis and targeted calibration of periodontal diseases. The solution resolves the problems mentioned in the background section.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, this application provides a method for image-guided precision diagnosis and targeted calibration of periodontal diseases, the method comprising:

[0007] Obtain three-dimensional bone density distribution data of the patient's jaw region, and extract the topological coordinate set of the bone defect area and the gray value of each point therein;

[0008] Acquire the intensity sequence of scattered signals along the root surface of the patient and its spatial pose data;

[0009] Based on the pre-defined bimodal physical markers in the patient's mouth, an affine transformation matrix is ​​constructed between the spatial pose data and the three-dimensional spatial coordinate system containing the topological coordinate set. This affine transformation matrix is ​​then used to align the scattered signal intensity sequence to the topological coordinate set.

[0010] Calculate the Hessian matrix for the gray values ​​of each point in the topological coordinate set, take the ratio of its largest eigenvalue to the remaining eigenvalues ​​as the bone resorption activity, and take the eigenvector corresponding to the largest eigenvalue as the dominant vector of bone microstructure.

[0011] Calculate the gradient vector of the scattering signal intensity sequence after affine transformation in the three-dimensional spatial coordinate system, and calculate the spatial angle between the gradient vector and the dominant vector of bone microstructure;

[0012] Bone resorption activity, gradient vector magnitude, and spatial angle are integrated into a three-dimensional feature vector. After standardization, the vector is input into a pre-trained deep autoencoder and outputs a two-dimensional latent space vector. The Euclidean distance between the vector and the preset healthy tissue cluster center is used as the activity index, and different levels of target areas are divided according to the activity index.

[0013] Secondly, this application provides an image-guided precision diagnosis and targeted calibration system for periodontal diseases, used to implement the image-guided precision diagnosis and targeted calibration method for periodontal diseases described in any of the above claims, including:

[0014] The dual-modal data acquisition module is used to acquire three-dimensional bone density distribution data of the patient's jaw region, extract the topological coordinate set of the bone defect area and the gray value of each point therein, and acquire the scattering signal intensity sequence and spatial pose data of the patient along the tooth root surface.

[0015] The data spatial alignment module is used to construct an affine transformation matrix between the spatial pose data and the three-dimensional spatial coordinate system where the topological coordinate set is located, based on the pre-set bimodal physical markers in the patient's mouth, and to use the affine transformation matrix to align the scattered signal intensity sequence to the topological coordinate set.

[0016] The gray value matrix construction module is used to calculate the Hessian matrix of the gray values ​​of each point in the topological coordinate set, take the ratio of its largest eigenvalue to the remaining eigenvalues ​​as the bone resorption activity, and take the eigenvector corresponding to the largest eigenvalue as the dominant vector of bone microstructure.

[0017] The spatial parameter calculation module is used to calculate the gradient vector of the scattering signal intensity sequence after affine transformation in the three-dimensional spatial coordinate system, and to calculate the spatial angle between the gradient vector and the dominant vector of bone microstructure.

[0018] The target region segmentation module integrates bone resorption activity, gradient vector magnitude, and spatial angle into a three-dimensional feature vector. After standardization, the vector is input into a pre-trained deep autoencoder, which outputs a two-dimensional latent space vector. The Euclidean distance between the vector and the preset healthy tissue cluster center is used as the activity index, and different levels of target regions are segmented according to the activity index.

[0019] Thirdly, this application provides a computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described image-guided precision diagnosis and targeted calibration method for periodontal diseases.

[0020] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described image-guided precision diagnosis and targeted calibration method for periodontal diseases.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] This application achieves the spatial correlation quantification between the pathological changes in periodontal lesions and the main direction of bone microstructure by spatially registering scattering signal sequences with three-dimensional bone density distribution data and analyzing the angle between the gradient vector and the dominant vector of bone microstructure.

[0023] This application uses the Hessian matrix eigenvalue analysis of the bone defect area to simultaneously quantify the bone resorption activity and the dominant vector of bone microstructure, breaking through the limitations of traditional CBCT static characterization of bone defects, realizing pixel-level dynamic quantification of bone resorption and metabolic activity, and providing a biomechanical basis for targeted therapy.

[0024] This application overcomes the limitations of relying on subjective experience and lacking multi-parameter fusion standards in artificial target area division by mapping three-dimensional feature vectors to two-dimensional latent space and generating a bone resorption activity index based on its Euclidean distance with the healthy tissue benchmark cluster. It establishes a target differentiation and classification mechanism driven by pathological similarity. Attached Figure Description

[0025] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0026] Figure 1 This is a flowchart of the image-guided precision diagnosis and targeted calibration method for periodontal diseases proposed in this invention;

[0027] Figure 2 This is a structural block diagram of the image-guided precision diagnosis and targeted calibration system for periodontal diseases proposed in this invention. Detailed Implementation

[0028] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0029] In existing technologies, periodontal disease imaging analysis models rely on static morphological representation for analysis, which makes it difficult to quantify the dynamic activity of bone resorption process and the main direction of bone microstructure. In particular, the lack of dynamic activity of bone resorption makes it impossible to distinguish between active lesions and old damage, and the lack of main direction of bone microstructure makes it difficult to achieve automatic and accurate division of targeted intervention areas, thus affecting the objectivity of pathological risk level classification and the pertinence of intervention measures.

[0030] For example, in the treatment of middle-aged patients with periodontitis, after the maxillofacial CBCT data is collected, a bone density distribution map is generated by gray-scale threshold segmentation, and an edge detection algorithm is applied to identify the boundary of bone defects. However, the temporal dynamic characteristics of bone resorption are not captured, making it impossible to determine the metabolic activity status of the defect area. At the same time, the spatial guidance evolution information of bone microstructure is missing, making it difficult to quantify the changes in the main direction of bone trabeculae. In this scenario, there is a spatial deviation between the coordinates of the target area and the actual pathological state, and the setting of the intervention degree depends on clinical experience judgment, which increases the uncertainty of clinical decision-making.

[0031] Reference Figure 1 As shown, this application proposes an image-guided precision diagnosis and targeted calibration method for periodontal diseases, including:

[0032] Obtain three-dimensional bone density distribution data of the patient's jaw region, and extract the topological coordinate set of the bone defect area and the gray value of each point therein;

[0033] Acquire the intensity sequence of scattered signals along the root surface of the patient and its spatial pose data;

[0034] It should be noted that three-dimensional bone density distribution data of the patient's jaw region were acquired by cone-beam computed tomography (CBCT); and the gingival soft tissue was scanned in real time by a handheld OCT probe to obtain the axial scattered signal intensity sequence and its spatial pose data along the root surface.

[0035] Based on the pre-defined bimodal physical markers in the patient's mouth, an affine transformation matrix is ​​constructed between the spatial pose data and the three-dimensional spatial coordinate system containing the topological coordinate set. This affine transformation matrix is ​​then used to align the scattered signal intensity sequence to the topological coordinate set.

[0036] It should be noted that by fixing dual-modal physical markers on the intraoral stent, CBCT and OCT data acquisition are performed after the patient wears the intraoral stent, and a scanning channel is reserved on the intraoral stent; the dual-modal physical markers can be set as metal-reflective composite markers, where the CBCT imaging layer is titanium alloy microspheres and the OCT reflective layer is a gold-plated film on the surface of titanium alloy microspheres.

[0037] Calculate the Hessian matrix for the gray values ​​of each point in the topological coordinate set, take the ratio of its largest eigenvalue to the remaining eigenvalues ​​as the bone resorption activity, and take the eigenvector corresponding to the largest eigenvalue as the dominant vector of bone microstructure.

[0038] Calculate the gradient vector of the scattering signal intensity sequence after affine transformation in the three-dimensional spatial coordinate system, and calculate the spatial angle between the gradient vector and the dominant vector of bone microstructure;

[0039] Bone resorption activity, gradient vector magnitude, and spatial angle are integrated into a three-dimensional feature vector. After standardization, the vector is input into a pre-trained deep autoencoder and outputs a two-dimensional latent space vector. The Euclidean distance between the vector and the preset healthy tissue cluster center is used as the activity index, and different levels of target areas are divided according to the activity index.

[0040] Explanation of each dimension of the three-dimensional feature vector:

[0041] Dimension 1: Bone resorption activity A_bone characterizes the anisotropic strength of the trabecular bone structure, and its pathological significance is the mechanical result of osteoclasts' directional erosion of the bone matrix; when A_bone < 0.3, the bone microstructure is isotropic (healthy bone), clinically manifested as uniform distribution of trabecular bone; when A_bone is within [0.4, 0.6], the bone microstructure undergoes moderate directional resorption, clinically manifested as the starting point of active bone destruction; when A_bone > 0.7, the bone microstructure undergoes severe directional voiding, clinically manifested as old defects;

[0042] Dimension 2: The gradient vector ‖G‖ magnitude represents the rate of collagen fiber degradation. The principle of optical scattering reflects the rate of change of tissue refractive index. The larger the value, the more severe the collagen fiber degradation, which leads to increased inflammatory exudation. The pathological significance is the degree of acute inflammation. If ‖G‖>3.0, it represents active inflammation. If ‖G‖>2.5, it represents chronic inflammation or healthy bone.

[0043] Dimension 3: The spatial angle φ represents the consistency between the direction of inflammation and the direction of bone destruction. The smaller the value, the higher the efficiency of the directional spread of inflammation along the bone trabeculae. When φ < 30°, the direction of inflammation and the direction of bone destruction are in the same direction. For example, when A_bone > 0.4, ||G|| > 3.0 and < 30°, it is highly likely that there is a pathological state of active periodontitis.

[0044] It should be noted that relying solely on bone resorption activity and gradient vectors can lead to misdiagnosis of acute gingivitis with misaligned orientation as active periodontitis, or old bone defects with consistent orientation but no acute inflammation as active infection, by ignoring the spatial relationship between the direction of inflammation and bone microstructure. Combining only spatial angles and gradient vectors can easily confuse old defects with active lesions; combining only spatial angles and bone resorption activity can easily confuse osteoporosis with pathological bone resorption. Therefore, integrating these three factors into a three-dimensional feature vector, and modulating the weights of inflammation and bone destruction through directional constraints, decouples mixed pathological states and achieves pathological classification.

[0045] Through the above technical solutions, based on precise spatial alignment of bimodal data and nonlinear feature extraction of depth encoders, bone resorption activity and dominant vectors of bone microstructure can be quantitatively characterized, distinguishing the biological differences between active lesions and old injuries. Specifically, the bone resorption activity index directly reflects the activity level of microscopic bone microstructure changes, and spatial angle analysis reveals the matching relationship between the direction of inflammation and infection and the direction of bone destruction. Combined with the automatic grading mechanism of the activity index, the division of the target area no longer relies on the doctor's empirical judgment of the morphology of bone defects, realizing the definition of intervention areas based on biological processes. This overcomes the technical limitations of traditional imaging methods, which can only statically present the morphology of bone defects and cannot capture the dynamic evolution process.

[0046] In an optional embodiment, the topological coordinate set and corresponding grayscale values ​​of the bone defect region are extracted, specifically including:

[0047] Calculate the mean and standard deviation of bone density of all voxels in the three-dimensional bone density distribution data, and subtract the standard deviation by a preset multiple from the mean to obtain the initial screening threshold.

[0048] It should be noted that, based on the normal distribution characteristics of jawbone bone mineral density in healthy individuals, using the mean minus 2.5 times the standard deviation as the initial screening threshold can cover 99.4% of the normal bone mineral density range; using the mean minus 1.5 times the standard deviation as the initial screening threshold can cover 93.2% of the normal bone mineral density range. Therefore, the range of the initial screening threshold can be 1.5-2.5, with a typical value of 1.5.

[0049] All voxels with bone density below the initial screening threshold are recorded as initial screening defect voxels. The spatially continuous initial screening defect voxels are identified and aggregated by a three-dimensional connected component analysis algorithm to form discrete candidate bone defect regions.

[0050] Specifically, the 3D connected component analysis algorithm uses the 26-neighborhood connection rule (i.e., the connection of voxels in the X / Y / Z axes and all diagonal directions) and employs a breadth-first search algorithm to traverse all spatially continuous voxel sets, merging the centroid spacing among them. Connected regions of ploidyne size generate bone defect regions with continuous anatomical structures;

[0051] A spherical buffer zone is constructed by extending a predetermined distance outward from the geometric center of the candidate bone defect region as the origin;

[0052] It should be noted that the maximum physiological diameter range of the Haver system of the jawbone is 1.2-1.8 mm. Therefore, the preset distance can be 1.5 mm to ensure that the spherical buffer zone covers most of the active bone resorption area at the edge of the bone defect.

[0053] The average bone density of voxels within the spherical shell buffer zone is statistically analyzed, and its product with a preset ratio is used as the local judgment threshold for the candidate bone defect region.

[0054] It should be noted that a spherical buffer zone is constructed with the geometric center of the candidate region to dynamically obtain the local mean bone density and generate a local judgment threshold. This local judgment threshold accurately matches the normal bone density level around the candidate bone defect area, thereby distinguishing between active lesions (bone density much lower than the local mean) and old defects (bone density close to the local mean). Moreover, the bone density of the lesion is usually lower than 85% of the surrounding normal bone density, so the preset ratio can be 85%.

[0055] Voxels with bone density below the corresponding local judgment threshold in the candidate bone defect region are removed, and three-dimensional morphological closing and opening operations are performed to generate the bone defect region;

[0056] Specifically, the closing operation of the three-dimensional morphology is to use a spherical structural element with a radius of 1.5 times the voxel size to fill the micropores in the bone defect area, and the opening operation is to use a spherical structural element with a radius of 1.0 times the voxel size to smooth the jagged boundaries of the bone defect area and remove isolated protrusion artifacts.

[0057] Extract the spatial coordinates and bone density of all voxels in the bone defect area, map the bone density into gray values ​​and associate them with the corresponding spatial coordinates to generate a topological coordinate set;

[0058] It should be noted that bone mineral density (BMD) is essentially the original voxel CT value, not the physically measured BMD value. The formula for mapping BMD to grayscale values ​​is:

[0059] ;

[0060] In the formula, Grayscale value Bone density, The maximum bone density in the three-dimensional bone density distribution data. The minimum bone density in the three-dimensional bone density distribution data. The maximum value for a 12-bit grayscale image, which can be 4095;

[0061] Through the above technical solution, this application achieves accurate identification of bone defect areas and generation of spatial topological configurations by combining global initial screening and local adaptive judgment collaborative processing mechanism with three-dimensional morphological closing and opening operations. This overcomes the false positive boundary errors caused by the heterogeneity of jawbone density (mandibular density > maxillary density) and the complexity of microstructure in traditional fixed threshold methods, and provides a high-fidelity anatomical basis for bone resorption activity and target area division.

[0062] In an optional embodiment, the affine transformation matrix is ​​used to align the scattered signal intensity sequence to a topological coordinate set, specifically including:

[0063] S1. Through an affine transformation matrix, the spatial pose data of the scattered signal intensity sequence is mapped to the three-dimensional spatial coordinate system where the topological coordinate set is located, and then set into a scattered signal coordinate set;

[0064] Specifically, the affine transformation matrix is ​​constructed as follows: Coordinate data of the pre-defined dual-modal physical markers within the patient's mouth are simultaneously acquired in both the three-dimensional spatial coordinate system and the coordinate system of the scattering signal acquisition device; the coordinate data of the dual-modal physical markers in the two coordinate systems are combined to form matching point pairs, each pair containing the three-dimensional coordinates of the same marker in both systems; a rotation matrix is ​​calculated based on the matching point pairs, and the optimal rotation parameters in the least-squares sense are obtained using singular value decomposition; a translation vector is calculated based on the matching point pairs, and the spatial displacement is determined by the mean of the coordinate residuals after rotation; a scaling factor between the two coordinate systems is calculated, and a scale uniformity parameter is determined based on the ratio of the distance between the dual-modal physical markers; the rotation matrix, translation vector, and scaling factor are combined to generate the affine transformation matrix.

[0065] S2. The spatial intersection of the scattered signal coordinate set and the topological coordinate set is taken as the overlapping region. A grid point set is generated in the overlapping region according to a preset resolution, and each grid point is associated with a virtual scattered signal intensity value and a virtual gray value. The virtual scattered signal intensity value is generated by radial basis function interpolation of the scattered signal point set, and the virtual gray value is generated by trilinear interpolation of the topological coordinate set.

[0066] The virtual scattered signal intensity value and virtual gray value of each grid point are associated as data pairs, and the Pearson correlation coefficient is calculated based on all data pairs.

[0067] If the absolute value of the Pearson correlation coefficient is lower than the preset correlation threshold, the Euler angles of the affine transformation matrix are iteratively adjusted with a fixed angle step size. S1 and S2 are re-executed and the Pearson correlation coefficient is calculated again until the absolute value of the Pearson correlation coefficient is not lower than the preset correlation threshold. For example, the fixed angle step size can be 5°, and the maximum number of iterations is set to 6 to prevent deadlock. If deadlock occurs, manual handling is prompted.

[0068] For example, let all data pairs be ,in, This represents the intensity value of the virtual scattered signal. This is a virtual grayscale value. For numbering;

[0069] The formula for calculating the Pearson correlation coefficient is as follows:

[0070] ;

[0071] In the formula, The Pearson correlation coefficient is used. for The mean, for The mean;

[0072] It should be noted that in the sample testing of no fewer than 200 different patients, the samples included three-dimensional bone mineral density distribution data, scattered signal intensity sequences and their spatial pose data. When the spatial registration error between the scattered signal coordinate set and the corresponding topological coordinate set is ≤210μm, and when At this time, the spatial registration error exceeds 520μm; therefore, the preset correlation threshold can be set to 0.7.

[0073] Through the above scheme, this application effectively solves the accuracy problem of spatial alignment of dual-modal data, ensures reliable fusion of scattering signal intensity sequence and three-dimensional bone density distribution data, improves the accuracy of calculating the spatial angle between the principal direction vector of bone microstructure and the gradient vector of scattering signal, and enhances the reliability of target area division.

[0074] In an optional embodiment, calculating the Hessian matrix for the gray values ​​of each point in the topological coordinate set specifically includes:

[0075] Traverse each topological point in the topological coordinate set and extract cubic data volumes within a preset neighborhood from the three-dimensional bone density distribution data with these points as the center. For example, the preset neighborhood can be a 3×3×3 voxel. The purpose of using a small-sized cubic structure is to focus on the real microstructural features of the bone defect area and avoid the influence of global noise interference on local analysis.

[0076] A three-dimensional Gaussian kernel is used to perform multi-scale filtering on the cubic data volume based on discrete convolution, and the maximum normalized gradient norm is selected as the optimal filtering scale. For example, the multi-scale filtering range can be 0.5-2.0 voxels. The purpose is to dynamically adapt to the resolution differences of bone microstructure through scale changes, suppress noise while retaining key edge information.

[0077] Under the optimal filtering scale, the second-order partial derivative of the gray value at the topological point with respect to its spatial coordinates is calculated, and the result is used to construct a three-dimensional symmetric matrix as the Hessian matrix.

[0078] The three real eigenvalues ​​of the Hessian matrix are calculated using the cyclic Jacobian iteration method and then sorted and output in descending order of absolute value.

[0079] Specifically, this scheme traverses each topological point in the topological coordinate set and extracts cubic data volumes, limiting the analysis scope to the microstructural units of the bone defect region. Then, a three-dimensional Gaussian kernel is used for multi-scale filtering to smooth noise at different scales and select the optimal filtering scale based on the normalized gradient norm, so that noise suppression and edge preservation achieve a dynamic balance. Under the determined optimal filtering scale, the second-order partial derivatives of the gray values ​​are calculated to construct the Hessian matrix, ensuring that the curvature feature capture matches the actual resolution of the current three-dimensional bone density distribution data. Finally, the eigenvalues ​​are efficiently solved by the cyclic Jacobi iteration method.

[0080] Through the above technical solutions, this application can effectively suppress image noise interference, adaptively select the optimal filtering scale, ensure the stability of Hessian matrix eigenvalue extraction, thereby accurately quantifying bone resorption activity and determining the dominant vector of bone microstructure, and improving the accuracy of target area division.

[0081] In an optional embodiment, the pre-trained deep autoencoder includes a first fully connected layer, a second fully connected layer, and a third fully connected layer: wherein,

[0082] The first fully connected layer receives a standardized three-dimensional feature vector. The 128 neurons in this layer perform independent weight multiplication and addition operations on each dimension of the three-dimensional feature vector and apply a linear rectification function with a preset leakage coefficient to output a 128-dimensional activation feature vector.

[0083] It should be noted that the purpose of the first fully connected layer is to preserve the independent information differences of the original dimensions such as bone resorption activity, gradient vector magnitude, and spatial angle, and to avoid distortion caused by feature coupling; for example, the linear rectified function with a preset leakage coefficient can be a LeakyReLU with a leakage coefficient of 0.25.

[0084] The second fully connected layer receives the 128-dimensional activation feature vector. The 64 neurons in this layer perform independent weight multiplication and addition operations on each dimension of the 128-dimensional activation feature vector and apply a hyperbolic tangent function to output a 64-dimensional activation feature vector.

[0085] It should be noted that the purpose of the second fully connected layer is to provide stable gradient propagation characteristics and reduce numerical oscillations during training; the hyperbolic tangent function is the tanh function.

[0086] The third fully connected layer receives the 64-dimensional activation feature vector. The two neurons in this layer perform independent weight multiplication and addition operations on each dimension of the 64-dimensional activation feature vector to generate a 2-dimensional weighted sum vector, which is then output as a 2-dimensional latent space vector.

[0087] It should be noted that the two-dimensional latent space vector is The coordinate axes are defined as follows: horizontal axis The vertical axis represents the activity of bone resorption (negative values ​​represent healthy bone, positive values ​​represent diseased bone). The topological complexity of the characterization of the spread of inflammatory infection is as follows (negative values ​​represent local lesions, and positive values ​​represent lesions spreading to multiple quadrants).

[0088] Specifically, the pre-training process of the deep autoencoder is as follows: Using samples from no fewer than 200 different patients for testing, the bone resorption activity, gradient vector magnitude, and spatial angle of each sample are extracted to construct a standardized three-dimensional feature vector set; the deep autoencoder network structure is initialized, including configuring a first fully connected layer with 128 nodes (activation function: LeakyReLU with 0.25), a second fully connected layer with 64 nodes (activation function: tanh function), and a third fully connected layer with 2 nodes (no activation function); phased gradient optimization is performed: the first phase uses a learning rate... The training consists of 50 rounds, with the second phase focusing on the learning rate. Fine-tuning is performed for 20 rounds, with the training set, validation set, and test set split in an 8:1:1 ratio in each round. The validation set loss is monitored in real time during training. When the loss decreases by less than 1% for 5 consecutive rounds, a Dropout layer with a ratio of 0.2 is enabled to suppress overfitting. The weight matrix and bias vector corresponding to the optimal validation loss are saved as parameters of the deep autoencoder.

[0089] To further explain, the two-dimensional latent space vectors of healthy bone samples in all sample tests are extracted, and their mean is calculated as the preset healthy tissue cluster center;

[0090] The formula for calculating the activity index is:

[0091]

[0092] In the formula, The activity index, To pre-determine healthy tissue cluster centers;

[0093] Through the above scheme, this application effectively solves the problems of insufficient representation and unstable calculation of three-dimensional feature vectors, and ensures that the activity index can accurately quantify the topological complexity of bone resorption activity and the spread of inflammatory infection, thereby providing a reliable basis for the accurate division of the target area;

[0094] In an optional embodiment, the target region is divided into different levels based on the activity index, specifically including:

[0095] Iterate through the activity index of each topological point in the topological coordinate set, compare it with the preset risk level, and assign a risk level label based on the comparison result;

[0096] It should be noted that the sample test is conducted using no fewer than 200 different patients, and the statistical characteristics of the two-dimensional latent spatial distance distribution between the healthy sample group and the diseased sample are extracted. The preset risk level is determined by combining the pathological activity inflection point verified by histology. For example, the preset risk level can be set as low risk, medium risk and high risk, with activity index ranges of ≤0.65, 0.65-1.30 and greater than 1.30, respectively.

[0097] Perform three-dimensional connected component analysis on topological points with the same level of labels to generate several discrete risk regions;

[0098] Calculate the projected area of ​​each risk zone and remove risk zones whose projected area is less than a preset projection threshold.

[0099] It should be noted that the projected area specifically refers to the two-dimensional projected size of the risk area on a specific anatomical plane (such as the dental arch section). It can be calculated using voxel projection counting or convex hull algorithms. The purpose is to filter out small noise areas based on clinical characteristics and improve the reliability of the target area. For example, the preset projection threshold can be taken as... ;

[0100] For each retained risk area, calculate the average activity index of its internal topological points, and merge risk areas of the same level whose spatial distance is less than a preset spacing threshold and whose difference in the average activity index is less than a preset difference threshold.

[0101] It should be noted that the preset distance threshold and the preset difference threshold can be understood as dual constraint parameters characterizing spatial proximity and similarity of activity levels. The purpose is to achieve reasonable integration of adjacent lesion areas through the joint determination of distance and activity difference. For example, the preset distance threshold can be 0.8 mm and the preset difference threshold can be 0.15 mm.

[0102] The risk areas that remain discrete after merging are designated as target areas, and their risk levels are associated with them.

[0103] Through the above technical solutions, this application solves the problem of risk area fragmentation caused by directly generating the target area based on the activity index, improves the spatial continuity and clinical operability of the target area, eliminates noise interference points through the projection area filtering mechanism, and enables the generated target area to accurately match the anatomical features of continuous bone defects on the root surface through the dual constraint merging mechanism based on spatial distance and activity level similarity.

[0104] In an optional embodiment, after calculating the mean activity index of the internal topological points for each retained risk region, a dynamic adjustment mechanism is also included:

[0105] Calculate the standard deviation of the activity index for each retained risk zone. If the standard deviation is greater than the preset uniformity threshold, remove the topological points in the risk zone whose absolute value of the deviation between the activity index and the mean is greater than the preset deviation threshold.

[0106] It should be noted that the preset uniformity threshold is an empirical threshold determined through statistical analysis based on sample testing, and its purpose is to provide a quantitative benchmark for judging whether the uniformity of the risk area meets the standard; the preset deviation threshold is a tolerance range set according to the biological characteristics of periodontal disease, and its purpose is to define the acceptable boundary of the activity index fluctuation.

[0107] For example, the preset uniformity threshold can be 0.12, and the preset deviation threshold can be 0.25;

[0108] Virtually expand the current risk zone boundary outward at equal intervals to generate an expansion zone, and identify topological points within the expansion zone that have not been assigned to any risk zone and have the same risk level label;

[0109] Among the identified topological points, select points whose absolute deviation from the mean of the activity index of the current risk zone does not exceed a preset deviation threshold, and include them in the current risk zone.

[0110] Update the standard deviation of the activity index of the risk area. If it is still greater than the preset uniformity threshold, repeat the removal and inclusion operation until the standard deviation of the activity index meets the preset uniformity threshold or reaches the maximum number of iterations. For example, the maximum number of iterations can be 5. Usually, more than 5 iterations will expand to the dissection boundary.

[0111] Specifically, this embodiment achieves dynamic improvement of the uniformity within the risk zone through an iterative optimization mechanism: First, the standard deviation of the activity index of the risk zone is calculated to quantify the degree of internal dispersion. When this value exceeds a preset uniformity threshold, it indicates that there is significant non-uniformity in the region. At this time, removing topological points with excessive deviation can effectively eliminate noise interference or blurred boundary points, making the remaining points more focused on the core lesion area. Next, the boundary of the risk zone is virtually expanded to generate an expansion zone. The spatial continuity of periodontal lesions is used to identify unassigned topological points with the same risk level label within the expansion zone. The points are screened based on the deviation between the activity index and the current risk zone mean, and only points with deviations not exceeding a preset deviation threshold are included to ensure that the newly added points are consistent with the overall activity index of the region. Finally, the standard deviation of the activity index of the adjusted risk zone is updated and the above removal and inclusion operations are repeated. The internal structure of the region is gradually optimized through the iterative process until the activity distribution meets the uniformity requirements or reaches the iteration limit, thereby forming a target area with clear boundaries and high internal consistency.

[0112] See Figure 2 As shown, this solution proposes an image-guided precision diagnosis and targeted calibration system for periodontal diseases, used to implement the aforementioned image-guided precision diagnosis and targeted calibration method for periodontal diseases, including:

[0113] The dual-modal data acquisition module is used to acquire three-dimensional bone density distribution data of the patient's jaw region, extract the topological coordinate set of the bone defect area and the gray value of each point therein, and acquire the scattering signal intensity sequence and spatial pose data of the patient along the tooth root surface.

[0114] The data spatial alignment module is used to construct an affine transformation matrix between the spatial pose data and the three-dimensional spatial coordinate system where the topological coordinate set is located, based on the pre-set bimodal physical markers in the patient's mouth, and to use the affine transformation matrix to align the scattered signal intensity sequence to the topological coordinate set.

[0115] The gray value matrix construction module is used to calculate the Hessian matrix of the gray values ​​of each point in the topological coordinate set, take the ratio of its largest eigenvalue to the remaining eigenvalues ​​as the bone resorption activity, and take the eigenvector corresponding to the largest eigenvalue as the dominant vector of bone microstructure.

[0116] The spatial parameter calculation module is used to calculate the gradient vector of the scattering signal intensity sequence after affine transformation in the three-dimensional spatial coordinate system, and to calculate the spatial angle between the gradient vector and the dominant vector of bone microstructure.

[0117] The target region segmentation module integrates bone resorption activity, gradient vector magnitude, and spatial angle into a three-dimensional feature vector. After standardization, the vector is input into a pre-trained deep autoencoder, which outputs a two-dimensional latent space vector. The Euclidean distance between the vector and the preset healthy tissue cluster center is used as the activity index, and different levels of target regions are segmented according to the activity index.

[0118] In another embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0119] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described above.

[0120] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps described above.

[0121] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A targeted calibration method for image-guided precision diagnosis of periodontal diseases, characterized in that, The method includes: Obtain three-dimensional bone density distribution data of the patient's jaw region, and extract the topological coordinate set of the bone defect area and the gray value of each point therein; Acquire the intensity sequence of scattered signals along the root surface of the patient and its spatial pose data; Based on the pre-defined bimodal physical markers in the patient's mouth, an affine transformation matrix is ​​constructed between the spatial pose data and the three-dimensional spatial coordinate system containing the topological coordinate set. This affine transformation matrix is ​​then used to align the scattered signal intensity sequence to the topological coordinate set. Calculate the Hessian matrix for the gray values ​​of each point in the topological coordinate set, take the ratio of its largest eigenvalue to the remaining eigenvalues ​​as the bone resorption activity, and take the eigenvector corresponding to the largest eigenvalue as the dominant vector of bone microstructure. Calculate the gradient vector of the scattering signal intensity sequence after affine transformation in the three-dimensional spatial coordinate system, and calculate the spatial angle between the gradient vector and the dominant vector of bone microstructure; Bone resorption activity, gradient vector magnitude, and spatial angle are integrated into a three-dimensional feature vector. After standardization, the vector is input into a pre-trained deep autoencoder and outputs a two-dimensional latent space vector. The Euclidean distance between the vector and the preset healthy tissue cluster center is used as the activity index, and different levels of target areas are divided according to the activity index.

2. The method according to claim 1, characterized in that, Extracting the topological coordinate set and corresponding grayscale values ​​of the bone defect region, specifically including: Calculate the mean and standard deviation of bone density of all voxels in the three-dimensional bone density distribution data, and subtract the standard deviation by a preset multiple from the mean to obtain the initial screening threshold. All voxels with bone density below the initial screening threshold are recorded as initial screening defect voxels. The spatially continuous initial screening defect voxels are identified and aggregated by a three-dimensional connected component analysis algorithm to form discrete candidate bone defect regions. A spherical buffer zone is constructed by extending a predetermined distance outward from the geometric center of the candidate bone defect region as the origin; The average bone density of voxels within the spherical shell buffer zone is statistically analyzed, and its product with a preset ratio is used as the local judgment threshold for the candidate bone defect region. Voxels with bone density below the corresponding local judgment threshold in the candidate bone defect region are removed, and three-dimensional morphological closing and opening operations are performed to generate the bone defect region; Extract the spatial coordinates and bone density of all voxels in the bone defect area, map the bone density into grayscale values ​​and associate them with the corresponding spatial coordinates to generate a topological coordinate set.

3. The method according to claim 1, characterized in that, The affine transformation matrix is ​​used to align the scattered signal intensity sequence to a topological coordinate set, specifically including: S1. Through an affine transformation matrix, the spatial pose data of the scattered signal intensity sequence is mapped to the three-dimensional spatial coordinate system where the topological coordinate set is located, and then set into a scattered signal coordinate set; S2. The spatial intersection of the scattered signal coordinate set and the topological coordinate set is taken as the overlapping region. A grid point set is generated in the overlapping region according to a preset resolution, and each grid point is associated with a virtual scattered signal intensity value and a virtual gray value. The virtual scattered signal intensity value is generated by radial basis function interpolation of the scattered signal point set, and the virtual gray value is generated by trilinear interpolation of the topological coordinate set. The virtual scattered signal intensity value and virtual gray value of each grid point are associated as data pairs, and the Pearson correlation coefficient is calculated based on all data pairs. If the absolute value of the Pearson correlation coefficient is lower than the preset correlation threshold, the Euler angles of the affine transformation matrix are iteratively adjusted with a fixed angle step size, and S1 and S2 are re-executed and the Pearson correlation coefficient is calculated until the absolute value of the Pearson correlation coefficient is not lower than the preset correlation threshold.

4. The method according to claim 1, characterized in that, Calculating the Hessian matrix for the gray values ​​of each point in the topological coordinate set specifically includes: Traverse each topological point in the topological coordinate set and extract cubic data volumes within a preset neighborhood from the three-dimensional bone density distribution data, using each point as the center. A three-dimensional Gaussian kernel is used to perform multi-scale filtering on the cubic data volume based on discrete convolution, and the one with the largest normalized gradient norm is selected as the optimal filtering scale. Under the optimal filtering scale, the second-order partial derivative of the gray value at the topological point with respect to its spatial coordinates is calculated, and the result is used to construct a three-dimensional symmetric matrix as the Hessian matrix. The three real eigenvalues ​​of the Hessian matrix are calculated using the cyclic Jacobian iteration method, and then sorted and output in descending order of absolute value.

5. The method according to claim 1, characterized in that, The pre-trained deep autoencoder includes a first fully connected layer, a second fully connected layer, and a third fully connected layer: wherein, The first fully connected layer receives a standardized three-dimensional feature vector. The 128 neurons in this layer perform independent weight multiplication and addition operations on each dimension of the three-dimensional feature vector and apply a linear rectification function with a preset leakage coefficient to output a 128-dimensional activation feature vector. The second fully connected layer receives the 128-dimensional activation feature vector. The 64 neurons in this layer perform independent weight multiplication and addition operations on each dimension of the 128-dimensional activation feature vector and apply a hyperbolic tangent function to output a 64-dimensional activation feature vector. The third fully connected layer receives the 64-dimensional activation feature vector. The two neurons in this layer perform independent weight multiplication and addition operations on each dimension of the 64-dimensional activation feature vector to generate a 2-dimensional weighted sum vector, which is then output as a 2-dimensional latent space vector.

6. The method according to claim 1, characterized in that, Target regions are classified into different levels based on their activity index, specifically including: Iterate through the activity index of each topological point in the topological coordinate set, compare it with the preset risk level, and assign a risk level label based on the comparison result; Perform three-dimensional connected component analysis on topological points with the same level of labels to generate several discrete risk regions; Calculate the projected area of ​​each risk zone and remove risk zones whose projected area is less than a preset projection threshold. For each retained risk area, calculate the average activity index of its internal topological points, and merge risk areas of the same level whose spatial distance is less than a preset spacing threshold and whose difference in the average activity index is less than a preset difference threshold. The risk areas that remain discrete after merging are designated as target areas, and their risk levels are associated with them.

7. The method according to claim 6, characterized in that, After calculating the mean activity index of the internal topological points of each retained risk zone, a dynamic adjustment mechanism is also included: Calculate the standard deviation of the activity index for each retained risk zone. If the standard deviation is greater than the preset uniformity threshold, remove the topological points in the risk zone whose absolute value of the deviation between the activity index and the mean is greater than the preset deviation threshold. Virtually expand the current risk zone boundary outward at equal intervals to generate an expansion zone, and identify topological points within the expansion zone that have not been assigned to any risk zone and have the same risk level label; Among the identified topological points, select points whose absolute deviation from the mean of the activity index of the current risk zone does not exceed a preset deviation threshold, and include them in the current risk zone. Update the standard deviation of the activity index of the risk zone. If it is still greater than the preset uniformity threshold, repeat the removal and inclusion operations until the standard deviation of the activity index meets the preset uniformity threshold or the maximum number of iterations is reached.

8. A targeted calibration system for image-guided precision diagnosis of periodontal diseases, characterized in that: The method for image-guided precise diagnosis and targeted calibration of periodontal disease as described in any one of claims 1-7 includes: The dual-modal data acquisition module is used to acquire three-dimensional bone density distribution data of the patient's jaw region, extract the topological coordinate set of the bone defect area and the gray value of each point therein, and acquire the scattering signal intensity sequence and spatial pose data of the patient along the tooth root surface. The data spatial alignment module is used to construct an affine transformation matrix between the spatial pose data and the three-dimensional spatial coordinate system where the topological coordinate set is located, based on the pre-set bimodal physical markers in the patient's mouth, and to use the affine transformation matrix to align the scattered signal intensity sequence to the topological coordinate set. The gray value matrix construction module is used to calculate the Hessian matrix of the gray values ​​of each point in the topological coordinate set, take the ratio of its largest eigenvalue to the remaining eigenvalues ​​as the bone resorption activity, and take the eigenvector corresponding to the largest eigenvalue as the dominant vector of bone microstructure. The spatial parameter calculation module is used to calculate the gradient vector of the scattering signal intensity sequence after affine transformation in the three-dimensional spatial coordinate system, and to calculate the spatial angle between the gradient vector and the dominant vector of bone microstructure. The target region segmentation module integrates bone resorption activity, gradient vector magnitude, and spatial angle into a three-dimensional feature vector. After standardization, the vector is input into a pre-trained deep autoencoder, which outputs a two-dimensional latent space vector. The Euclidean distance between the vector and the preset healthy tissue cluster center is used as the activity index, and different levels of target regions are segmented according to the activity index.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images