Oral cavity knowledge graph construction method and device, equipment and storage medium

Abstract

The invention relates to the technical field of image data processing, in particular to an oral cavity knowledge graph construction method and device, equipment and a storage medium. Preprocessing the oral CT original sectional image data based on a preset adaptive threshold method and a preset rigid registration algorithm to obtain standard image data; performing feature extraction on the standard image data based on a preset deep learning model and a preset principal component analysis method to obtain a structured feature vector; constructing the structured feature vector, the structured data and the unstructured data according to a preset entity linking method, a preset inference rule and a preset graph neural network model to obtain an optimized knowledge graph; an optimized knowledge graph is constructed through CT image preprocessing, deep learning feature extraction and multi-source data fusion, an image-to-structured knowledge conversion link is realized, and a comprehensive, reliable and convenient knowledge data support is provided for precise diagnosis and treatment of oral diseases.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, specifically to a method, apparatus, device, and storage medium for constructing an oral cavity knowledge graph. Background Technology

[0002] In the diagnosis and treatment of oral diseases, doctors need to comprehensively consider oral anatomy, imaging features, and clinical knowledge to judge the condition; this is the core demand scenario. With the popularization of oral CT imaging technology, clinicians routinely use CT tomographic images to observe the morphology of teeth, jawbones, and other tissues, as well as lesions such as caries and periodontitis. However, the value of imaging data is limited to visual observation, and its deeper value has not been explored. At the same time, although knowledge graphs have become an important tool for integrating oral clinical data and medical literature, they are seriously disconnected from CT imaging, a core diagnostic basis, and have significant compatibility issues: the graph dimension is single, it does not transform CT images into structured knowledge, and the images are only stored as independent files, unable to participate in knowledge reasoning; moreover, knowledge updates rely on manually entered text data, resulting in poor timeliness and difficulty in adapting to new imaging diagnostic standards. In summary, existing technologies cannot achieve the full-process transformation of oral CT images from raw data to structured knowledge, nor can they construct multi-source fusion oral knowledge graphs. There is an urgent need for a construction method that combines the accuracy of image processing with the completeness of knowledge fusion. Summary of the Invention

[0003] To address the shortcomings of the existing technology, this invention proposes a method for constructing an oral knowledge graph.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for constructing an oral cavity knowledge graph includes: acquiring raw oral CT tomographic image data, and preprocessing the raw oral CT tomographic image data based on a preset adaptive threshold method and a preset rigid registration algorithm to obtain standard image data; extracting features from the standard image data based on a preset deep learning model and a preset principal component analysis method to obtain structured feature vectors; acquiring structured data and unstructured data, and constructing the structured feature vectors, structured data, and unstructured data according to a preset entity linking method, preset inference rules, and a preset graph neural network model to obtain an optimized knowledge graph.

[0005] Furthermore, the preprocessing of the original oral CT tomographic image data based on the preset adaptive threshold method and the preset rigid registration algorithm to obtain standard image data includes: performing noise processing on the original oral CT tomographic image data according to the preset PythonSimpleITK library to obtain denoised image data; performing feature extraction on the denoised image data based on the adaptive threshold method to obtain dynamic features of image grayscale distribution; performing artifact removal on the denoised image data according to the preset interpolation algorithm and the adaptive threshold to obtain artifact-removed image data; and performing alignment processing on the artifact-removed image data based on the rigid registration algorithm to obtain standard image data.

[0006] Furthermore, the step of extracting features from standard image data based on a preset deep learning model and a preset principal component analysis method to obtain structured feature vectors includes: performing three-dimensional reconstruction of the standard image data according to a preset MarchingCubes algorithm and preset model parameters to obtain a three-dimensional oral cavity mesh model; analyzing the three-dimensional oral cavity mesh model based on a deep learning model to obtain accurate segmentation results; extracting features from the accurate segmentation results based on a preset ResNet50 model to obtain anatomical structure features and lesion features; and performing dimensionality reduction processing on the anatomical structure features and lesion features according to the principal component analysis method to obtain structured feature vectors.

[0007] Furthermore, the step of constructing structured feature vectors, structured data, and unstructured data according to a preset entity linking method, preset inference rules, and preset graph neural network model to obtain an optimized knowledge graph includes: identifying relationships between structured and unstructured data based on a preset BiLSTM model to obtain entity relationship features; constructing entity relationship features and structured feature vectors according to the entity linking method to obtain a knowledge graph; and optimizing the knowledge graph according to inference rules and graph neural network model to obtain an optimized knowledge graph.

[0008] Furthermore, the method of identifying relationships between structured and unstructured data based on a preset BiLSTM model to obtain entity relationship features includes: cleaning structured and unstructured data according to preset word segmentation, preset regular expressions, and a preset terminology database to obtain cleaned medical data; extracting entities from the cleaned medical data based on a preset BERT-BiLSTM-CRF model to obtain entity data; and identifying relationships between entity data based on the BiLSTM model to obtain entity relationship features.

[0009] Furthermore, the step of constructing the knowledge graph by means of entity relationship features and structured feature vectors according to the entity linking method includes: binding structured feature vectors and entity data according to entity relationship features to obtain associated features; and constructing the associated features with a preset multi-level ontology according to the entity linking method to obtain the knowledge graph.

[0010] Furthermore, the optimization of the knowledge graph based on reasoning rules and graph neural network models to obtain an optimized knowledge graph includes: reasoning about the knowledge graph based on reasoning rules and graph neural network models to obtain implicit relationships; The knowledge graph is optimized based on implicit relationships to obtain an optimized knowledge graph.

[0011] Furthermore, an oral knowledge graph construction device includes: a preprocessing module for acquiring raw oral CT tomographic image data and preprocessing the raw oral CT tomographic image data based on a preset adaptive thresholding method and a preset rigid registration algorithm to obtain standard image data; a feature extraction module for extracting features from the standard image data based on a preset deep learning model and a preset principal component analysis method to obtain structured feature vectors; a graph construction module for acquiring structured and unstructured data and constructing the structured feature vectors, structured data, and unstructured data according to a preset entity linking method, preset inference rules, and a preset graph neural network model to obtain an optimized knowledge graph; and a storage module for storing the structured feature vectors and the optimized knowledge graph in a preset graph database.

[0012] Furthermore, an oral knowledge graph construction device includes: a memory and at least one processor, wherein the memory stores instructions; at least one processor invokes the instructions in the memory to cause the oral knowledge graph construction device to perform the various steps of the oral knowledge graph construction method as described above.

[0013] Furthermore, a computer-readable storage medium stores instructions that, when executed by a processor, implement the steps of a method for constructing an oral knowledge graph as described above.

[0014] The beneficial effects of the oral cavity knowledge graph construction method of the present invention are as follows: By acquiring raw tomographic images of the oral cavity using an adaptive thresholding method and a rigid registration algorithm, precise preprocessing is performed to effectively filter noise, remove metal artifacts, and align multiple image sequences, outputting standardized data with a unified format and complete structure, laying a solid data foundation for knowledge mining. Then, a deep learning model captures multi-dimensional image features, and PCA dimensionality reduction generates standardized structured feature vectors, filling the gap in the integration of image data and textual knowledge. Subsequently, multi-source structured and unstructured data are integrated, entity links are used to eliminate ambiguity, and inference rules and graph neural networks are used to complete the implicit relationship between "image-disease-treatment," constructing a multi-dimensional optimized knowledge graph to solve the adaptability problems of traditional graphs with single dimensions and fragmented associations. Finally, the data is stored in a graph database, supporting precise retrieval and efficient access to diagnostic and treatment information. This entire process streamlines the transformation of oral CT images from raw data to structured knowledge, allowing image data to truly participate in knowledge reasoning, meeting the core needs of doctors for comprehensive multi-dimensional information diagnosis and treatment, improving diagnostic accuracy and clinical decision-making efficiency, reducing reliance on doctors' experience, and providing comprehensive, reliable, and convenient knowledge data support for the precise diagnosis and treatment of oral diseases. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a first flowchart of a method for constructing an oral knowledge graph according to an embodiment of the present invention; Figure 2 This is a second flowchart of a method for constructing an oral knowledge graph according to an embodiment of the present invention; Figure 3 This is a third flowchart of a method for constructing an oral knowledge graph according to an embodiment of the present invention; Figure 4 This is a fourth flowchart of a method for constructing an oral knowledge graph according to an embodiment of the present invention; Figure 5 The fifth flowchart of a method for constructing an oral knowledge graph provided in an embodiment of the present invention; Figure 6 The sixth flowchart of a method for constructing an oral knowledge graph provided in an embodiment of the present invention; Figure 7 The seventh flowchart of a method for constructing an oral knowledge graph provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an oral knowledge graph construction device provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an oral knowledge graph construction device provided in an embodiment of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0017] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 An embodiment of a method for constructing an oral knowledge graph according to the present invention includes: 101. Acquire raw tomographic images of the oral cavity CT, and preprocess the raw tomographic images of the oral cavity CT based on a preset adaptive threshold method and a preset rigid registration algorithm to obtain standard image data. In this embodiment, multiple original oral CT tomographic images (DICOM format, slice thickness 0.5-1mm, HU value range -1000~1000) are used. DICOM is a standard medical image format that includes patient information, examination content, and image sequences (one image per slice). An adaptive thresholding method and a rigid registration algorithm are used to perform precise preprocessing of the original oral CT tomographic image data. This effectively filters out noise and removes metal artifacts from the original images, solves the problem of spatial misalignment in multiple image sequences, and outputs standardized image data with a unified format and complete structure. This provides high-quality data support for subsequent 3D reconstruction, feature extraction, and knowledge graph construction, avoids subsequent analysis biases caused by defects in the original images, and ensures the accuracy and reliability of knowledge mining related to oral disease diagnosis. 102. Based on a pre-defined deep learning model and a pre-defined principal component analysis method, feature extraction is performed on standard image data to obtain structured feature vectors; In this embodiment, the deep learning model can accurately capture the multidimensional core features of oral anatomical structures and lesions, while principal component analysis completes feature dimensionality reduction while retaining key information, generating highly recognizable structured feature vectors. This solves the problems of strong subjectivity and fragmented information in traditional image feature extraction, provides standardized data support for subsequent entity association and knowledge graph construction, fills the gap in the integration of image data and text knowledge, and helps improve the accuracy of oral disease diagnosis and the efficiency of knowledge application. 103. Obtain structured and unstructured data, and construct structured feature vectors, structured data and unstructured data according to preset entity linking methods, preset inference rules and preset graph neural network models to obtain an optimized knowledge graph; In this embodiment, entity links eliminate data ambiguity, and inference rules and graph neural networks complete implicit relationships such as "image-disease-treatment," solving the problems of single-dimensionality and fragmented associations in traditional knowledge graphs. The optimized knowledge graph achieves deep fusion of multi-source knowledge, supports rapid matching of diagnostic and treatment information from image features, improves diagnostic accuracy and efficiency, reduces reliance on doctors' experience, and provides comprehensive knowledge support for oral clinical practice. Optimizing the knowledge graph also includes: collecting new oral CT images (100 cases / time) and clinical feedback data every 3 months, repeating the preprocessing, feature extraction, and knowledge graph construction steps, and dynamically updating the entities, relationships, and inference rules of the knowledge graph. 104. Store the structured feature vectors and optimized knowledge graphs in a pre-defined graph database; In this embodiment, the graph database is the Neo4j graph database, which supports Cypher query statements (such as "MATCH(n:disease{name:'dental caries'})-[r:treatment plan]->(m)RETURNm"). The Neo4j graph database is used to store structured feature vectors and optimized knowledge graphs. By storing the structured feature vectors and optimized knowledge graphs in the Neo4j graph database and leveraging its Cypher query support, efficient data retrieval is achieved. This solves the problem of scattered storage of image structured data and multi-source knowledge, enabling accurate retrieval of related data such as "disease-image features-treatment plan". Doctors can quickly obtain target information through simple query statements, improving the efficiency of clinical decision-making. At the same time, it ensures the standardization and relevance of data storage, providing convenient and reliable knowledge data support for the diagnosis and treatment of oral diseases. In this embodiment, raw tomographic images of the oral cavity are acquired and precisely preprocessed using an adaptive thresholding method and a rigid registration algorithm. This effectively filters noise, removes metal artifacts, and aligns multiple image sequences, outputting standardized data with a unified format and complete structure, laying a solid data foundation for knowledge mining. Then, a deep learning model captures multi-dimensional image features, and PCA dimensionality reduction generates standardized structured feature vectors, filling the gap in the fusion of image data and textual knowledge. Subsequently, multi-source structured and unstructured data are integrated, entity links are used to eliminate ambiguity, and inference rules and graph neural networks are used to complete the implicit relationship between "image-disease-treatment," constructing a multi-dimensional optimized knowledge graph. This solves the adaptation problems of traditional graphs, which are often single-dimensional and have fragmented associations. Finally, the data is stored in a graph database, supporting precise retrieval and efficient access to diagnostic and treatment information. This entire process streamlines the transformation of oral CT images from raw data to structured knowledge, allowing image data to truly participate in knowledge reasoning. This meets the core needs of doctors for comprehensive multi-dimensional information diagnosis and treatment, improves diagnostic accuracy and clinical decision-making efficiency, reduces reliance on doctors' experience, and provides comprehensive, reliable, and convenient knowledge data support for the precise diagnosis and treatment of oral diseases.

[0019] Please see Figure 2 In a second embodiment of the oral cavity knowledge graph construction method of the present invention, step 101 specifically includes: 201. Based on the pre-set PythonSimpleITK library, the median filtering algorithm (with a filtering window size of 3×3) is used to process the noise in the original oral CT tomographic image data to obtain denoised image data. In this embodiment, by calling the median filtering interface of the PythonSimpleITK library and setting a 3×3 window size, Gaussian noise (such as soft tissue edge blurring and microparticle noise) in the raw tomographic image data of oral CT is specifically filtered to avoid noise interfering with subsequent anatomical structure segmentation (such as tooth and jawbone boundary recognition) and lesion feature extraction (such as density judgment of early caries foci). The core function of this library is to simplify the preprocessing, analysis and postprocessing development of medical images (such as CT, MRI and ultrasound). The library natively supports DICOM format and can directly read raw tomographic image data of oral CT. It has built-in common algorithms such as median filtering, Gaussian filtering and threshold segmentation. There is no need to implement complex algorithms from scratch. Noise processing can be completed by calling the interface and setting parameters (such as 3×3 window), improving the image signal-to-noise ratio and providing a high-quality data foundation for subsequent artifact elimination, 3D reconstruction and feature extraction. 202. Based on the adaptive thresholding method, feature extraction is performed on the denoised image data to obtain the dynamic features of image grayscale distribution; In this embodiment, the dynamic characteristics of image grayscale distribution (including the distribution pattern of HU value of normal bone tissue 200-400 and HU value of metal prosthesis >1000) are used to generate an adaptive threshold; 203. Generate an adaptive threshold based on the dynamic characteristics of image grayscale distribution; In this embodiment, an adaptive thresholding method is used to dynamically set a threshold based on the dynamic characteristics of image grayscale distribution (HU value of normal bone tissue is 200-400, and HU value of metal restoration is >1000) to remove strip artifacts caused by metal restorations. First, the image grayscale distribution characteristics (including the HU value distribution pattern of normal bone tissue and metal restoration) of the denoised image data are analyzed by the adaptive thresholding method. Then, an appropriate adaptive threshold is generated based on the image grayscale distribution characteristics to achieve accurate removal of metal artifacts. 204. Perform artifact removal on the denoised image data according to the preset interpolation algorithm and adaptive threshold to obtain artifact-removed image data; In this embodiment, the adaptive threshold is a preliminary localization tool for artifact removal. The adaptive threshold is used to locate the artifacts first, and then the artifacts are eliminated by subsequent repair algorithms. The repair logic is as follows: using the grayscale features of the normal tissue surrounding the artifact region, the blank pixels in the artifact region are filled by interpolation algorithms (such as linear interpolation and Gaussian interpolation). For example, if the HU value of a pixel in the artifact region is 800 (belonging to the artifact range of 500-1100 HU), and the left side is normal bone tissue (350 HU) and the right side is normal soft tissue (50 HU), the interpolation algorithm will calculate a reasonable grayscale value for the artifact pixel (such as 150-200 HU, which conforms to the grayscale transition law of the surrounding tissue) based on the grayscale values ​​of the normal pixels on both sides, and replace the original artifact pixel value. 205. Alignment processing of artifact-removed image data based on rigid registration algorithm to obtain standard image data; In this embodiment, a rigid registration algorithm is employed (based on mutual information optimization; in oral CT image registration, mutual information is specifically defined as: quantifying the correlation of grayscale distribution between the "source image (e.g., coronal CT)" and the "target image (e.g., axial CT)". When the anatomical structures of the two images are well aligned, the grayscale values ​​of the corresponding regions show a strong correlation, and the mutual information reaches its maximum value; when the images are misaligned, the grayscale correlation is weak, and the mutual information is small). The rigid registration algorithm is a mature existing technology. The "rigid registration algorithm based on mutual information optimization" is a classic existing algorithm in the field of medical image registration. Its core framework (using mutual information as a similarity measure and solving rigid transformations to achieve image alignment) has been gradually formed and widely used since 1995. In the field of medical imaging, rigid registration based on mutual information (such as the axial / coronal alignment scheme of oral CT in this paper) has become a standard method. It can be directly called through open source libraries (Nibabel, ITK, SimpleITK) to adapt to the alignment needs of multimodal images such as CT and MRI. The rigid registration algorithm in the Nibabel library is used to align multi-sequence CT images (such as axial and coronal images) to ensure the consistency of the spatial position of anatomical structures. The rigid registration algorithm based on mutual information optimization is a classic existing algorithm in the field of medical image registration. Its core framework (using mutual information as a similarity measure and solving rigid transformations to achieve image alignment) has been gradually formed and widely used since 1995. Mutual information (MI) is an index in information theory that measures the amount of information shared between two random variables. Its core logic is: when two images are well aligned, the gray values ​​of corresponding regions have a strong correlation (such as the maxillary first molar region in axial CT and the same region in coronal CT, where the gray distribution is highly matched), and the mutual information reaches its maximum value; when the images are not aligned, the gray correlation of corresponding regions is weak, and the mutual information is small. The calculation process for mutual information in dental CT registration is as follows: ① Select the target image (e.g., axial CT, as the alignment reference) and the source image (e.g., coronal CT, to be transformed and aligned) from the artifact-removed image data. ② Statistically analyze the grayscale probability distribution of the two images: Marginal probability (Grayscale values ​​in the source image) (probability of occurrence) grayscale values ​​in the target image (probability of occurrence); joint probability (Source image grayscale) grayscale corresponding to the target image (probability of simultaneous occurrence) Calculate mutual information using the formula. : Definition of core symbols in the formula (in the context of oral CT scans) In practice, the calculation of mutual information for dental CT image registration must conform to the "rigid registration" process in this scheme: 1. Data Preprocessing: Discretize the HU values ​​of coronal (A) and axial (B) CT images (e.g., divide [-1000HU~400HU] into 1400 intervals with 1HU intervals) to ensure... and The range of values ​​is consistent; 2. Probability and Statistics: statistics Traverse coronal images and calculate the voxel percentage for each HU value interval; statistics Similarly, traverse the axis images and calculate the voxel percentage for each HU value interval; statistics : Statistically analyze the "coronal position" according to a one-to-one correspondence of spatial coordinates. And axis position The percentage of voxel logarithms; 3. Formula calculation: Substitute into the formula and sum to obtain the mutual information value under the current registration state; 4. Iterative optimization: Adjust the position of the coronal images through rigid transformation (rotation and translation), and repeat steps 2-3 until the mutual information reaches its maximum value (at which point the image alignment is optimal). In this embodiment, the median filtering interface of the PythonSimpleITK library is used to specifically filter Gaussian noise, improve the signal-to-noise ratio of the image, and avoid interference with anatomical structure segmentation and early lesion feature extraction. The library natively supports the DICOM format, eliminating the need for complex algorithm development and balancing efficiency and practicality. An adaptive thresholding method dynamically adapts to the image grayscale distribution, accurately locating metal artifact regions. Interpolation algorithms are then used to fill in artifact pixels, maintaining a natural grayscale transition in surrounding tissues and addressing the inaccuracy of fixed-threshold artifact removal. A rigid registration algorithm based on mutual information optimization is employed to achieve spatial alignment of multi-sequence CT images, ensuring the consistency of anatomical structure positions and avoiding distortion of subsequent models caused by tomographic misalignment. The entire preprocessing workflow is progressive, efficiently resolving various quality issues of the original images and outputting standardized image data with uniform format, complete structure, and meeting accuracy standards. This lays a solid data foundation for subsequent stages such as 3D reconstruction, semantic segmentation, and feature extraction, directly ensuring the accuracy and reliability of the oral cavity knowledge graph construction.

[0020] Please see Figure 3 In a third embodiment of the oral cavity knowledge graph construction method of the present invention, step 102 specifically includes: 301. Perform three-dimensional reconstruction on standard image data according to the preset MarchingCubes algorithm and preset model parameters to obtain a three-dimensional mesh model of the oral cavity; In this embodiment, the MarchingCubes algorithm (implemented via the VTK library) is used. The pre-packaged MarchingCubes algorithm module is retrieved from the VTK library. Model parameters include a reconstruction threshold range of -1000 to 400 HU (covering teeth, jawbone, and soft tissue), preset according to clinical standards for HU values ​​(CT values) of different oral tissues; and a mesh resolution of 0.3mm × 0.3mm × 0.3mm, preset based on a reasonable balance between device accuracy and efficiency. These model parameters are not hard-coded fixed values, but rather configurable optimal initial values, conforming to the MarchingCubes algorithm. The calling logic of the S-algorithm is also adapted to the practical needs of oral CT 3D reconstruction. The construction of the oral 3D mesh model is based on preprocessed (standard image data) as the original data, with the MarchingCubes algorithm as the core transformation tool, and reconstruction threshold and mesh resolution as key control parameters, to finally construct the oral 3D mesh model. The core input of the MarchingCubes algorithm is continuous CT volume data (rather than single tomographic images), so it is necessary to first obtain a uniform DICOM tomographic image sequence (slice thickness 0) through preprocessing (denoising, artifact removal, image registration).5-1mm, HU value range covering the entire oral cavity; spatially aligned volumetric data (a three-dimensional voxel matrix after fusion of multiple CT images, ensuring no misalignment of anatomical structures); without this standard image data, the algorithm will suffer from high input data noise, artifact interference, and tomographic misalignment, leading to distorted reconstructed models (such as blurred tooth edges, jawbone position displacement), or even failure to generate a complete model; the MarchingCubes algorithm transforms two-dimensional CT tomographic sequences (volume data) into three-dimensional surface models, acting as a "data translator," specifically: traversing CT volumes For each voxel (3D pixel) in the data, its HU value is read; based on a preset reconstruction threshold, it is determined whether the voxel belongs to the target tissue to be reconstructed (such as air, soft tissue, or bone tissue with HU values ​​between -1000 and 400 HU); for the target voxel region, continuous triangular patches are generated using isosurface extraction technology (the core of the algorithm), and these patches are stitched together to form a 3D surface structure; the final output is a 3D mesh model (STL format) containing the outline of the anatomical structure. Without this algorithm, CT tomographic images alone cannot achieve the surface transformation from two-dimensional to three-dimensional, and can only remain at the tomographic observation stage, unable to... A rotatable and sectionable visualization model is obtained; two parameters directly determine the content and detail quality of the 3D mesh model, and are key instructions in the algorithm conversion process: the reconstruction threshold (-1000~400HU) determines which tissues will be included in the reconstruction (equivalent to a content filter). If the threshold range is too small (e.g., only 200~400HU), soft tissue and oral air areas will be lost, and the model cannot fully represent the spatial relationship of teeth, gums, and jawbone; if the threshold is too large (e.g., >1000HU), metal restorations and artifacts will be included, resulting in messy and distorted model edges; mesh resolution ( The resolution (0.3mm × 0.3mm × 0.3mm) determines the model's detail accuracy and data volume (equivalent to a precision adjuster). If the resolution is too low (e.g., 0.5mm), crucial details such as tooth fissures and subtle alveolar bone resorption will be lost, affecting subsequent lesion feature extraction (e.g., early caries foci identification). If the resolution is too high (e.g., 0.1mm), the model's data volume will increase more than tenfold, which does not meet the needs of efficient clinical diagnosis and treatment. Converting two-dimensional tomographic images into three-dimensional visualization models allows doctors to intuitively observe the spatial relationship between lesions and normal structures (e.g., the boundary between caries and dentin), reducing the rate of missed diagnoses. 302. Analyze the 3D mesh model of the oral cavity based on a deep learning model to obtain accurate segmentation results; In this embodiment, the U-Net++ deep learning model (PyTorch framework) is used to analyze the oral cavity 3D mesh model and output accurate segmentation results. The essence of the analysis here is semantic segmentation based on deep learning. That is, the model learns the 3D features of the oral cavity anatomy and automatically classifies and labels different tissues and lesions in the oral cavity 3D mesh model, rather than simply observing the surface or measuring the shape. The segmentation objects include teeth (including tooth position classification), jawbone (maxilla and mandible), periodontal ligament, alveolar bone, and lesion areas (dental caries, periodontal pockets, and jaw cysts). The U-Net++ deep learning model (PyTorch framework) is designed for the target task of oral cavity anatomy segmentation. The U-Net++ deep learning model has undergone dedicated training (also called fine-tuning or de novo training) beforehand, using labeled data of the target scene (500 oral CT scans) and training parameters adapted to the task (100 rounds, 0.001 learning rate), directly achieving the performance index of the task (segmentation accuracy ≥92%). 303. Based on the preset ResNet50 model, feature extraction is performed on the accurate segmentation results to obtain anatomical structure features and lesion features; In this embodiment, the ResNet50 model (pre-trained weights based on ImageNet) is used to extract two types of features: anatomical structural features (including three-dimensional morphological parameters: three-dimensional tooth length, jawbone volume; spatial location parameters: three-dimensional lesion coordinates); and lesion features (including three-dimensional volume in mm³, three-dimensional edge smoothness, location (e.g., caries located on the occlusal surface), and density distribution (HU value standard deviation)). The accurate segmentation result is oral CT three-dimensional voxel data fused with U-Net++ segmentation masks. This data is not newly introduced; the segmentation mask is structured intermediate data generated by the deep learning model analysis and is a key carrier connecting the three-dimensional visualization model and the entity features of the knowledge graph. 304. Dimensionality reduction of anatomical and lesion features is performed using principal component analysis to obtain structured feature vectors; In this embodiment, feature standardization is performed by reducing the dimensionality of the 2048-dimensional feature vector to 512 dimensions using principal component analysis (PCA) to generate structured feature vectors (for subsequent knowledge association). In this embodiment, a full-link image structuring process—"3D reconstruction - precise segmentation - feature extraction - dimensionality reduction optimization"—effectively addresses the challenge of insufficient deep value mining in oral CT images, providing high-quality feature support for knowledge graph construction. The MarchingCubes algorithm, encapsulated in the VTK library, is used with configurable reconstruction thresholds and grid resolution to transform preprocessed standard image data into a complete, high-precision 3D oral mesh model, intuitively presenting tissue spatial relationships and reducing the risk of missed diagnoses. A U-Net++ model, finely tuned with labeled data, achieves precise semantic segmentation of multiple objects such as teeth and lesion areas. The ResNet50 model extracts multi-dimensional core features of anatomical structures and lesions, followed by PCA dimensionality reduction of high-dimensional features, preserving key information while reducing data complexity. The entire process achieves efficient conversion from 2D CT tomographic images to structured feature vectors, overcoming the limitation of traditional images only being visually observable, and providing standardized, highly recognizable feature data for subsequent entity association and knowledge graph construction, thus helping to improve the accuracy of oral disease diagnosis and the effectiveness of knowledge fusion.

[0021] Please see Figure 4 In the fourth embodiment of the oral cavity knowledge graph construction method of the present invention, step 103 specifically includes: 401. Based on a pre-defined BiLSTM model, perform relationship identification on structured and unstructured data to obtain the relationship features between entities; In this embodiment, structured data includes electronic medical records (patient age, gender, diagnosis results, treatment records, sourced from the hospital's HIS system), ICD-10 oral disease codes (such as K02.9 dental caries), and oral drug databases (drug names, indications, usage and dosage, sourced from the "Handbook of Oral Clinical Drugs"), etc.; unstructured data includes oral medicine literature (1000 articles published in the "Chinese Journal of Stomatology" in the past 5 years), clinical treatment guidelines (National Health Commission's "Guidelines for the Diagnosis and Treatment of Oral Diseases (2022 Edition)"), authoritative textbooks ("Internal Stomatology", "Oral and Maxillofacial Surgery"), etc. 402. Construct the relationship features and structured feature vectors between entities based on the entity linking method to obtain a knowledge graph; In this embodiment, the entity linking method is used to associate the relationship features between entities with structured feature vectors, eliminate synonym conflicts, construct an oral knowledge graph with image dimensions, fill the gap of existing graphs that only contain text, realize the full-link association of image-disease-treatment, help improve diagnostic accuracy, reduce reliance on doctors' experience, and provide more reliable knowledge support for clinical practice. 403. Optimize the knowledge graph based on reasoning rules and graph neural network models to obtain an optimized knowledge graph; In this embodiment, the inference rules include but are not limited to: ① If the imaging feature is "alveolar bone resorption depth ≥ 3 mm", then the inferred entity relationship is "this feature - suggesting disease - chronic periodontitis"; ② If the disease entity is "dental caries" and the imaging feature is "density HU value 200 - 300", then the inferred entity relationship is "this disease - recommended treatment - root canal treatment"; Combining the inference rules with the graph neural network model to optimize the knowledge graph, accurately complement the implicit relationships such as "imaging feature - disease - treatment", solve the problem of broken relationships in traditional graphs, and the optimized graph improves the full-link association, which can support quickly matching diagnosis and treatment information from imaging features, improve the diagnostic efficiency and accuracy, and effectively reduce the clinical dependence on doctors' experience; In this embodiment, first, based on the BiLSTM model, the entity relationships between structured data and unstructured data are mined to solidify the data foundation of the graph; then, through the entity linking method, the relationship features are associated with the imaging structured feature vectors to eliminate synonymous conflicts, construct a knowledge graph containing the imaging dimension, fill the blank of existing graphs with only text, and achieve the full-link association of "imaging - disease - treatment"; finally, combining the inference rules with the graph neural network model to complement the implicit relationships and solve the problem of broken relationships in traditional graphs; the optimized graph can support quickly matching diagnosis and treatment information from imaging features, improve the diagnostic accuracy and efficiency, effectively reduce the clinical dependence on doctors' experience, and provide reliable knowledge guarantee for the accurate diagnosis and treatment of oral diseases.

[0022] Please refer to Figure 5 , the fifth embodiment of a method for constructing an oral knowledge graph in the embodiment of the present invention, step 401 specifically includes: 501. Perform data cleaning on the structured data and unstructured data according to the preset word segmentation, preset regular expression, and preset term library to obtain cleaned medical data; In this embodiment, the word segmentation is jieba word segmentation. Stop words (such as "de", "he") are removed using jieba word segmentation, and typos are corrected through regular expressions (such as correcting "periodontal strictness" to "periodontitis"); the structured data includes electronic medical records of hospital HIS systems, ICD-10 oral disease codes, and data from "Oral Clinical Drug Handbook"; the unstructured data includes papers from "Chinese Journal of Stomatology" in the past 5 years, "Oral Disease Diagnosis and Treatment Guidelines (2022 Edition)" of the National Health Commission, and authoritative textbooks such as "Oral Medicine"; 502. Perform entity extraction on the cleaned medical data based on the preset BERT - BiLSTM - CRF model to obtain entity data; In this embodiment, the BERT-BiLSTM-CRF model, based on pre-trained weights bert-base-chinese, was fine-tuned for downstream tasks on 2000 annotated texts in the field of oral medicine. Testing showed that the model achieved an entity recognition accuracy of 92.5% and an F1 score of 91.8% on the test set, significantly improving entity recognition performance compared to the BiLSTM-CRF model without pre-trained weights. Based on the BERT-BiLSTM-CRF model, entity extraction was performed on the cleaned medical data, extracting entity types including anatomical structures (e.g., maxillary first molar, alveolar bone); diseases (e.g., chronic periodontitis, dental caries); symptoms (e.g., gingival bleeding, loose teeth); treatment plans (e.g., root canal treatment, periodontal scaling); and medications (e.g., metronidazole, amoxicillin). 503. Relationship identification of entity data based on BiLSTM model to obtain relationship features between entities; In this embodiment, a BiLSTM model based on a multi-head attention mechanism is used to identify relationships between entity data and identify relationship features between entities (such as maxillary first molar - lesion location - caries, chronic periodontitis - symptoms - gingival bleeding). In this embodiment, a precise data cleaning and efficient entity relationship recognition process provides high-quality core data for the construction of an oral health knowledge graph, effectively solving the problems of messy oral medical data and ambiguous entity relationships. Jieba word segmentation to remove stop words combined with regular expressions corrects errors, improving data cleanliness. The BERT-BiLSTM-CRF model, after specific fine-tuning, can comprehensively extract multiple types of oral entities. Combined with a multi-head attention BiLSTM model, it can accurately capture the relationships between entities. This process provides accurate entity and relationship data for the knowledge graph, avoiding ambiguity and errors in subsequent graph construction, laying the foundation for achieving a full-link association between "images-diseases-treatment," and helping to improve diagnostic accuracy.

[0023] Please see Figure 6 In the sixth embodiment of the oral cavity knowledge graph construction method of the present invention, step 402 specifically includes: 601. Bind the structured feature vector and entity data according to the relationship characteristics between entities to obtain the associated features; In this embodiment, the structured feature vector is bound to the corresponding entity (e.g., dental caries - image feature - density HU value 250±50, volume 8±2mm³); the result of the feature association step (the binding relationship between the structured feature vector and the entity) is the core data component for the construction of the oral knowledge graph, and its participation runs through the entire process of graph construction: from supplementing the image attributes of entities and establishing entity relationships in the image dimension, to supporting the consistency verification of knowledge fusion, and finally storing it as queryable data of the graph; 602. Construct a knowledge graph by combining associated features with a pre-defined multi-level ontology using entity linking methods; In this embodiment, the entity linking method is a cosine similarity algorithm (similarity threshold 0.7) used to match image feature entities with text entities (such as low-density areas on the occlusal surface and caries foci), eliminating synonym conflicts. Based on this association result (association features), the constructed knowledge graph can realize the full-link association of image-structure-disease-treatment, filling the gap of existing oral knowledge graphs that only contain text and lack image dimensions, achieving the overall technical effect of improving diagnostic accuracy and reducing reliance on experience. The ontology design of the knowledge graph includes a top-level ontology (oral medical field), and a second-level ontology containing 6 categories: anatomical structure (third-level sub-ontologies: teeth, jawbone, periodontal tissue), disease type (third-level sub-ontologies: dental caries, periodontitis, jawbone cysts), image features (third-level sub-ontologies: morphological features, density features, texture features), treatment plan (third-level sub-ontologies: drug treatment, surgical treatment), drugs (third-level sub-ontologies: antibiotics, analgesics), and patient information (third-level sub-ontologies: basic information, medical history). Each sub-ontology contains attributes (such as tooth attributes: tooth position, shape, health status). In this embodiment, a knowledge graph is constructed through entity binding and entity linking, effectively solving the problems of existing oral health knowledge graphs being single-dimensional and disconnected from images and text. Structured feature vectors extracted from CT images are precisely bound to entities, supplementing them with quantified image attributes to form core associated features. Furthermore, a cosine similarity algorithm is used to achieve precise matching between image and text entities, eliminating synonym conflicts. This, combined with multi-level ontology construction, fills the gap in existing graphs that only contain text, constructing a full-link association of "image features – anatomical structure – disease – treatment plan," enabling the graph to support the need to directly query diagnostic and treatment information from image features. Ultimately, this improves the accuracy of oral disease diagnosis, reduces reliance on doctors' subjective experience, and provides more comprehensive and reliable knowledge support for clinical diagnosis and treatment.

[0024] Please see Figure 7 In the seventh embodiment of the oral cavity knowledge graph construction method of the present invention, step 403 specifically includes: 701. Based on reasoning rules and graph neural network models, reason about the knowledge graph to obtain implicit relationships; In this embodiment, reasoning based on reasoning rules (e.g., alveolar bone resorption ≥3mm indicates chronic periodontitis) and graph neural network reasoning (GCN model) are combined to complete implicit relationships (e.g., dental caries - imaging feature - low density - dental caries - treatment plan - root canal treatment). The reasoning rules include, but are not limited to: ① If the imaging feature is "alveolar bone resorption depth ≥3mm", then the reasoning entity relationship is "this feature - suggests disease - chronic periodontitis"; ② If the disease entity is "dental caries" and the imaging feature is "density HU value 200-300", then the reasoning entity relationship is "this disease - recommended treatment - root canal treatment". Knowledge reasoning is essentially the process and core components of constructing an oral knowledge graph. Reasoning is performed on elements (entities, relationships, attributes) during the construction of a knowledge graph (specifically after knowledge fusion and before graph storage). This reasoning fills in missing implicit relationships within the graph, ultimately forming a more complete and interconnected knowledge graph. The objects and logic of this reasoning are deeply intertwined with the graph construction process. Specifically, this can be broken down into three aspects: the direct objects of reasoning, the impact of the reasoning results on the graph, and the incomplete knowledge components generated in earlier stages. Knowledge reasoning is not a process that waits until the graph is fully built before proceeding; rather, it intervenes at key stages of graph construction, targeting the incomplete knowledge components generated in earlier steps. These components form the future graph. The core components specifically include: extracted entity-relation pairs (textual basic relations) derived from the knowledge extraction step, such as tooth decay - typical symptoms - gingival bleeding, root canal treatment - medication - metronidazole. These are explicit textual relations in the graph, but lack the association between image features and treatment plans, and between image features and diseases (e.g., there is no direct relationship between tooth decay - image feature - low density and tooth decay - treatment plan - root canal treatment). Inference will, based on these explicit relations, combine rules or GCN models to complete the implicit links (e.g., from low density - tooth decay, tooth decay - root canal treatment, derive the indirect relationship of low density - root canal treatment). The associated image feature - entity binding results come from features. The associated steps, such as dental caries - imaging features - density HU value 250±50, chronic periodontitis - imaging features - alveolar bone resorption 3mm; these are the imaging dimension knowledge of the atlas, but only reflect the binding between entities and imaging features, lacking the full-link association of imaging features - disease - treatment (e.g., only alveolar bone resorption 3mm - binding - chronic periodontitis, without the reasoning relationship of alveolar bone resorption 3mm - prompt - chronic periodontitis, chronic periodontitis - need - periodontal cleaning). The reasoning will be based on these binding results, using rules (e.g., alveolar bone resorption ≥3mm is chronic periodontitis) to complete the diagnostic relationship of imaging features - disease, and then using the GCN model to complete the recommendation relationship of disease - treatment plan;The prototype of the entity-relationship network after knowledge fusion originates from the knowledge fusion steps. For example, after matching the low-density area of ​​the occlusal surface (image entity) and the caries focus (text entity) through entity links, a preliminary network is formed (e.g., low-density area of ​​occlusal surface - synonym association - caries focus - symptoms - gingival bleeding). This network still has gaps (e.g., there is no direct or indirect relationship between the low-density area of ​​the occlusal surface and root canal treatment). Inference will be based on this prototype of the network and use the GCN model to learn the potential association patterns between entities (e.g., the low-density area of ​​the occlusal surface is often associated with the caries focus, and the caries focus is often associated with root canal treatment), thereby completing the implicit relationship of low-density area of ​​occlusal surface - indirect recommendation - root canal treatment. 702. Optimize the knowledge graph based on implicit relationships to obtain an optimized knowledge graph; In this embodiment, a dual-reasoning mechanism combining rule-based reasoning and GCN model reasoning effectively addresses the issues of disconnection between existing oral knowledge graphs and CT images, as well as broken relationship links. This improves the completeness and clinical adaptability of the knowledge graph. The reasoning precisely intervenes in the semi-finished knowledge during the graph construction process. Based on explicit textual relationships and image-entity binding results, preset rules are used to complete the diagnostic association between "image features and diseases," while the GCN model learns latent patterns to complete recommendation relationships such as "disease-treatment," thus perfecting the entire link of "image features-anatomical structure-disease-treatment plan." The optimized knowledge graph fills the gap in the integration of image and textual knowledge, supporting accurate queries that directly link diagnostic and treatment information from CT image features. This assists doctors in quickly establishing diagnostic logic, reducing reliance on experience, and providing efficient knowledge support for the precise diagnosis and treatment of oral diseases.

[0025] Atlas optimization also includes a dynamic update mechanism: every 3 months, 100 new oral CT images and clinical feedback data are collected, and the preprocessing is repeated to the atlas construction step to update entities and relationships; based on the diagnostic biases reported by doctors, parameters such as the loss function weights and inference rule thresholds of the U-Net++ model are adjusted to ensure the timeliness and accuracy of atlas knowledge.

[0026] The above describes a method for constructing an oral knowledge graph according to an embodiment of the present invention. The following describes an apparatus for constructing an oral knowledge graph according to an embodiment of the present invention. Please refer to [link / reference]. Figure 8 One embodiment of the oral cavity knowledge graph construction device of the present invention includes: Preprocessing module 1 is used to acquire raw tomographic images of the oral CT and preprocess the raw tomographic images of the oral CT based on a preset adaptive threshold method and a preset rigid registration algorithm to obtain standard image data. Feature extraction module 2 is used to extract features from standard image data based on a preset deep learning model and a preset principal component analysis method to obtain structured feature vectors. The graph construction module 3 is used to acquire structured and unstructured data, and construct structured feature vectors, structured data and unstructured data according to preset entity linking methods, preset inference rules and preset graph neural network models to obtain an optimized knowledge graph. Storage module 4 is used to store the structured feature vectors and optimized knowledge graphs into a preset graph database; The update module is used to collect 100 new oral CT images and clinical feedback data every 3 months, repeat the preprocessing to the atlas construction step, and update entities and relationships; based on the diagnostic biases reported by doctors, the loss function weights, inference rule thresholds and other parameters of the U-Net++ model are adjusted. In this embodiment, raw tomographic images of the oral cavity are acquired and precisely preprocessed using an adaptive thresholding method and a rigid registration algorithm. This effectively filters noise, removes metal artifacts, and aligns multiple image sequences, outputting standardized data with a unified format and complete structure, laying a solid data foundation for knowledge mining. Then, a deep learning model captures multi-dimensional image features, and PCA dimensionality reduction generates standardized structured feature vectors, filling the gap in the fusion of image data and textual knowledge. Subsequently, multi-source structured and unstructured data are integrated, entity links are used to eliminate ambiguity, and inference rules and graph neural networks are used to complete the implicit relationship between "image-disease-treatment," constructing a multi-dimensional optimized knowledge graph. This solves the adaptation problems of traditional graphs, which are often single-dimensional and have fragmented associations. Finally, the data is stored in a graph database, supporting precise retrieval and efficient access to diagnostic and treatment information. This entire process streamlines the transformation of oral CT images from raw data to structured knowledge, allowing image data to truly participate in knowledge reasoning. This meets the core needs of doctors for comprehensive multi-dimensional information diagnosis and treatment, improves diagnostic accuracy and clinical decision-making efficiency, reduces reliance on doctors' experience, and provides comprehensive, reliable, and convenient knowledge data support for the precise diagnosis and treatment of oral diseases.

[0027] Figure 9This is a schematic diagram of the structure of an oral knowledge graph construction device 900 provided in an embodiment of the present invention. This oral knowledge graph construction device 900 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) for storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the oral knowledge graph construction device 900. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute the series of instruction operations in the storage media 930 on the oral knowledge graph construction device 900 to implement the steps of the oral knowledge graph construction method provided in the above-described method embodiments.

[0028] An oral knowledge graph construction device 900 may further include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, MacOSX, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 9 The illustrated structure of an oral knowledge graph construction device does not constitute a limitation on an oral knowledge graph construction device 900, and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0029] A computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of a method for constructing an oral knowledge graph as described above.

[0030] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention. The actual content is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A method for constructing an oral cavity knowledge graph, characterized in that, include: The raw tomographic images of the oral CT are acquired, and preprocessed based on a preset adaptive threshold method and a preset rigid registration algorithm to obtain standard image data. Based on a pre-defined deep learning model and a pre-defined principal component analysis method, feature extraction is performed on standard image data to obtain structured feature vectors. The system acquires structured and unstructured data, and constructs structured feature vectors, structured data, and unstructured data based on preset entity linking methods, preset inference rules, and preset graph neural network models to obtain an optimized knowledge graph.

2. The method for constructing an oral knowledge graph as described in claim 1, characterized in that, The method based on a preset adaptive threshold and a preset rigid registration algorithm preprocesses the raw tomographic images of the oral CT to obtain standard image data, including: The original oral CT tomographic image data is processed for noise reduction using the pre-defined PythonSimpleITK library to obtain denoised image data. Feature extraction is performed on denoised image data based on an adaptive thresholding method to obtain dynamic features of image grayscale distribution; An adaptive threshold is generated based on the dynamic characteristics of the image grayscale distribution. Artifact removal is performed on the denoised image data according to the preset interpolation algorithm and adaptive threshold to obtain artifact-removed image data. The image data with artifact removal is aligned using a rigid registration algorithm to obtain standard image data.

3. The method for constructing an oral knowledge graph as described in claim 1, characterized in that, The method of extracting features from standard image data based on a preset deep learning model and a preset principal component analysis method to obtain structured feature vectors includes: The standard image data is reconstructed in three dimensions according to the preset MarchingCubes algorithm and preset model parameters to obtain a three-dimensional mesh model of the oral cavity. A deep learning model is used to analyze a 3D mesh model of the oral cavity to obtain accurate segmentation results; Based on the pre-set ResNet50 model, feature extraction is performed on the accurate segmentation results to obtain anatomical structure features and lesion features; Principal component analysis is used to reduce the dimensionality of anatomical and lesion features to obtain structured feature vectors.

4. The method for constructing an oral knowledge graph as described in claim 1, characterized in that, The process of constructing an optimized knowledge graph from structured feature vectors, structured data, and unstructured data based on a preset entity linking method, preset inference rules, and a preset graph neural network model includes: Based on a pre-defined BiLSTM model, relationships are identified between structured and unstructured data to obtain the relationship features between entities. A knowledge graph is obtained by constructing the relationship features and structured feature vectors between entities based on the entity linking method. The knowledge graph is optimized based on reasoning rules and graph neural network models to obtain an optimized knowledge graph.

5. The method for constructing an oral knowledge graph as described in claim 4, characterized in that, The pre-defined BiLSTM model is used to identify relationships between structured and unstructured data to obtain inter-entity relationship features, including: Data cleaning is performed on structured and unstructured data based on preset word segmentation, preset regular expressions, and preset terminology database to obtain cleaned medical data. Entity extraction is performed on the cleaning medical data based on the preset BERT-BiLSTM-CRF model to obtain entity data; The BiLSTM model is used to identify relationships between entity data in order to obtain the relationship features between entities.

6. The method for constructing an oral knowledge graph as described in claim 5, characterized in that, The process of constructing a knowledge graph by analyzing entity relationship features and structured feature vectors based on entity linking methods includes: The structured feature vectors and entity data are bound together based on the relationship characteristics between entities to obtain associated features; The knowledge graph is constructed by linking related features with a pre-defined multi-level ontology using the entity linking method.

7. The method for constructing an oral knowledge graph as described in claim 4, characterized in that, The optimization of the knowledge graph based on inference rules and graph neural network models to obtain an optimized knowledge graph includes: The knowledge graph is reasoned based on reasoning rules and graph neural network models to obtain implicit relationships; the knowledge graph is then optimized based on these implicit relationships to obtain an optimized knowledge graph.

8. An oral cavity knowledge graph construction device, characterized in that, include: The preprocessing module is used to acquire raw tomographic images of the oral CT and preprocess them based on a preset adaptive threshold method and a preset rigid registration algorithm to obtain standard image data. The feature extraction module is used to extract features from standard image data based on a preset deep learning model and a preset principal component analysis method to obtain structured feature vectors. The graph construction module is used to acquire structured and unstructured data, and construct structured feature vectors, structured data and unstructured data according to preset entity linking methods, preset inference rules and preset graph neural network models to obtain an optimized knowledge graph. The storage module is used to store structured feature vectors and optimized knowledge graphs into a pre-defined graph database.

9. An oral knowledge graph construction device, characterized in that, include: A memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the oral knowledge graph construction device to perform the steps of the oral knowledge graph construction method as claimed in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the oral knowledge graph construction method as described in any one of claims 1-7.

Images