Automatic segmentation method and system of dental maxillofacial anatomical structure based on deep learning model

Abstract

The invention discloses an automatic segmentation method and system for a dental maxillofacial anatomical structure based on a deep learning model, and belongs to the technical field of medical image processing. Multi-modal image data of to-be-anatomical structures from different images are collected; determining a rough position of the to-be-dissected structure, and taking the rough position as a preliminary segmentation region; in the preliminary segmentation region, fusing the multi-modal image data; combining the fused multi-modal image data with the marked real label data to form an expert demonstration data set; constructing and training a segmentation strategy model based on the expert demonstration data set to obtain an optimized segmentation strategy model; and realizing automatic segmentation of the dental maxillofacial anatomical structure by using the optimized segmentation strategy model. According to the invention, through automatic standardization processing and fusion of multi-modal data, precise segmentation of a dental maxillofacial anatomical structure is realized, and the data precision and the accuracy of subsequent analysis are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and more specifically, to an automatic segmentation method and system for craniofacial anatomical structures based on a deep learning model. Background Art

[0002] In recent years, with the rapid development of medical imaging technology, the accurate recognition and segmentation of craniofacial anatomical structures have become increasingly important in dentistry, oral surgery and related fields. Medical images, such as CT (Computed Tomography) and MRI (Magnetic Resonance Imaging), etc., provide doctors with rich anatomical information to help them carry out effective diagnosis and treatment. However, the traditional manual segmentation method is not only time-consuming, but also easily affected by human factors, resulting in low segmentation accuracy. Therefore, automated segmentation technology has emerged to improve efficiency and accuracy.

[0003] The automatic segmentation method based on deep learning can learn the complex features of craniofacial anatomical structures from a large amount of labeled data with the help of advanced machine learning models such as convolutional neural networks (CNNs), so as to achieve efficient and accurate automatic segmentation. Nevertheless, due to the significant differences in anatomical structures among patients and diverse imaging acquisition methods, the current automatic segmentation methods often lack adaptability under different patients or different imaging types, resulting in unsatisfactory segmentation results.

[0004] Therefore, how to provide an automatic segmentation method and system for craniofacial anatomical structures based on a deep learning model that can solve the above problems is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides an automatic segmentation method and system for craniofacial anatomical structures based on a deep learning model to solve the technical problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An automatic segmentation method for craniofacial anatomical structures based on a deep learning model, comprising the following steps:

[0008] S100: Collect multi-modal image data of the to-be-anatomized structure from different imaging;

[0009] S200: Determine the rough position of the to-be-anatomized structure and use it as the preliminary segmentation area;

[0010] S300: In the preliminary segmentation area, fuse the multi-modal image data;

[0011] S400: Combine the fused multi-modal image data with the labeled true label data to form an expert demonstration data set;

[0012] S500: Build and train a segmentation strategy model based on the expert demonstration data set to obtain an optimized segmentation strategy model;

[0013] S600: Use the optimized segmentation strategy model to achieve automatic segmentation of the craniofacial anatomical structure.

[0014] Furthermore, the multi-modal image data includes: intraoral scan, panoramic radiograph, lateral cephalogram, facial photograph, and intraoral CT image, CBCT image or MRI image.

[0015] Furthermore, the S200 includes:

[0016] S210: Obtain the multi-modal image data of the anatomical structure to be dissected;

[0017] S220: Denoise and enhance the contrast of the multi-modal image data;

[0018] S230: Use Canny edge detection to extract the edges in the multi-modal image data;

[0019] S240: Use the Faster R-CNN model to detect the rough position of the craniofacial anatomical structure;

[0020] S250: Take the rough position as the preliminary segmentation area.

[0021] Furthermore, the S300 includes:

[0022] S310: Obtain the position information of the anatomical structure in the multi-modal image data in the first coordinate system;

[0023] S320: Obtain the semantic encoding data of the first coordinate system;

[0024] S330: Build a three-dimensional model of the craniofacial anatomical structure, and compare and match the built three-dimensional model of the craniofacial anatomical structure with the existing standard model of the craniofacial anatomical structure to obtain the optimal craniofacial anatomical structure;

[0025] S340: Based on the position information, semantic encoding in the first coordinate system, and the optimal craniofacial anatomical structure, obtain the fused multi-modal image data.

[0026] Furthermore, the origin of the first coordinate system is set at a preset fixed reference point in the multi-modal image data, such as the center of the image or a specific anatomical landmark point.

[0027] Furthermore, the S320 includes:

[0028] S321: Perform semantic segmentation on multi-modal image data;

[0029] S322: Perform per-pixel classification to generate a semantic segmentation map;

[0030] S323: Perform a top-down projection on the segmented semantic segmentation map. When projecting, use the centroid of the overall point cloud of the anatomical structure to replace the position of the actual object;

[0031] S324: Perform semantic encoding on the projected anatomical structure to generate semantic labels of each anatomical structure in the first coordinate system.

[0032] Further, the S330 includes:

[0033] S331: Construct a three-dimensional model of the anatomical structure. The nodes in the three-dimensional model represent the anatomical parts to be recognized, and the edges represent the spatial relationships between the parts. Use the anatomical structure at the preset time point t as the nodes, and the edges represent the anatomical constraints between the nodes. These constraints include, but are not limited to, anatomical proximity, connectivity, and shape features, etc.;

[0034] S332: Construct an optimized objective function according to the nodes and edges to obtain a three-dimensional model of the anatomical structure;

[0035] S333: Compare and match the three-dimensional model of the constructed anatomical structure with the existing standard model of the dental and maxillofacial anatomical structure to obtain the optimal dental and maxillofacial anatomical structure.

[0036] Further, the S500 includes:

[0037] S510: Use a deep neural network to initialize the reward function and the segmentation policy model based on the reward function;

[0038] S520: Estimate the state distribution probability density of the segmentation policy model using the expert demonstration dataset;

[0039] S530: Update the segmentation policy model using the DRL model based on the state distribution probability density;

[0040] S540: Determine whether the updated segmentation policy model meets the preset convergence condition. If it meets, record the weights of the current deep neural network to obtain an optimized segmentation policy model; if it does not meet, continue to estimate the state distribution probability density using the expert demonstration data and repeat the update process until an optimized segmentation policy model is obtained.

[0041] On the other hand, the present invention provides an automatic segmentation system for dental and maxillofacial anatomical structures based on a deep learning model, including:

[0042] Data collection module: Collect multi-modal image data of the anatomical structure to be dissected from different imaging modalities;

[0043] Preliminary confirmation module: Determine the approximate location of the anatomical structure to be dissected and use it as the preliminary segmentation area;

[0044] Data fusion module: Within the preliminary segmentation area, fuse the multi-modal image data;

[0045] Dataset generation module: Combine the fused multi-modal image data with the labeled true label data to form an expert demonstration dataset;

[0046] Model optimization module: Based on the expert demonstration dataset, construct and train a segmentation strategy model to obtain an optimized segmentation strategy model;

[0047] Output module: Use the optimized segmentation strategy model to achieve automatic segmentation of the dentomaxillofacial anatomical structure.

[0048] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses an automatic segmentation method and system for dentomaxillofacial anatomical structures based on a deep learning model, providing strong support for research and clinical practice in the field of stomatology. The specific beneficial effects are as follows:

[0049] (1) By fusing multi-modal image data and combining it with labeled true label data, it is possible to more accurately identify and segment complex dentomaxillofacial anatomical structures, which helps to capture information from different imaging techniques, improve the accuracy of segmentation, and reduce the limitations that may be brought by a single data source;

[0050] (2) By determining the approximate location of the anatomical structure to be dissected, it is possible to effectively narrow down the range that the model needs to process, which not only reduces the computational burden but also improves the positioning accuracy of key anatomical structures in the subsequent segmentation process;

[0051] (3) By training a segmentation strategy model based on an expert demonstration dataset, the segmentation strategy model can learn the experience and knowledge of professional doctors. When dealing with actual clinical cases, it can be closer to the actual operation of doctors, enhancing reliability and effectiveness and increasing the practical application ability of the model. Description of the Drawings

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0053] Figure 1Schematic diagram of the method flow of the present invention;

[0054] Figure 2 Schematic diagram of the system structure of the present invention. Detailed implementation manners

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0056] The object of the present invention is to provide an automatic segmentation method and system for dental and maxillofacial anatomical structures based on a deep learning model, including: collecting multi-modal image data of the anatomical structures to be dissected from different imaging modalities; determining the approximate position of the anatomical structures to be dissected and using it as the preliminary segmentation region; within the preliminary segmentation region, fusing the multi-modal image data; combining the fused multi-modal image data with the labeled true label data to form an expert demonstration data set; constructing and training a segmentation strategy model based on the expert demonstration data set to obtain an optimized segmentation strategy model; and using the optimized segmentation strategy model to achieve automatic segmentation of dental and maxillofacial anatomical structures. The present invention realizes precise segmentation of dental and maxillofacial anatomical structures through automatic standardization and fusion of multi-modal data, improving data accuracy and the accuracy of subsequent analysis.

[0057] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0058] Refer to Figure 1 , the embodiments of the present invention disclose an automatic segmentation method for dental and maxillofacial anatomical structures based on a deep learning model, including the following steps:

[0059] S100: Collect multi-modal image data of the anatomical structures to be dissected from different imaging modalities;

[0060] S200: Determine the approximate position of the anatomical structures to be dissected and use it as the preliminary segmentation region;

[0061] S300: Within the preliminary segmentation region, fuse the multi-modal image data;

[0062] S400: Combine the fused multi-modal image data with the labeled true label data to form an expert demonstration data set;

[0063] S500: Construct and train a segmentation strategy model based on the expert demonstration data set to obtain an optimized segmentation strategy model;

[0064] S600: Automatically segment the craniofacial anatomical structure by using the optimized segmentation strategy model.

[0065] In a specific embodiment, the multi-modal imaging data includes: intraoral scan, panoramic radiograph, lateral cephalogram, facial photograph, and intraoral CT image, CBCT image or MRI image.

[0066] Specifically, the labeled ground truth data can be the ground truth data labeled by doctors.

[0067] Specifically, an intraoral scan is a digital intraoral scanning technology that uses an optical scanner to capture the three-dimensional geometric information inside the oral cavity. It can capture the minute details of teeth and gums, without the need to use impression materials, providing a better patient experience. The scan results can be immediately displayed on a computer screen for instant viewing and analysis, and are commonly used for making digital dental models, designing restorations (such as crowns, bridges), orthodontic treatment, etc.

[0068] Specifically, a panoramic radiograph is a two-dimensional imaging technology that generates a panoramic image of the oral cavity and maxillofacial region by rotating the X-ray source and detector. It provides an overview image of the entire oral cavity. Compared with CBCT and traditional CT, the panoramic radiograph has a lower radiation dose, and the equipment cost and usage cost are relatively low. It is commonly used for preliminary screening, tooth alignment, wisdom tooth assessment, periodontal disease diagnosis, etc.

[0069] Specifically, a lateral cephalogram is a two-dimensional imaging technology that provides a lateral view of the skull and maxillofacial region by taking an X-ray image of the side of the head. It provides a lateral view image of the skull and maxillofacial region, with a relatively low radiation dose, and detailed cephalometric measurements can be performed. It is mainly used for the diagnosis and planning of orthodontic treatment, and for evaluating the development of the jawbone.

[0070] Specifically, a facial photograph is an image of the patient's face taken by an ordinary camera. It is radiation-free and non-invasive, and can intuitively display the patient's facial appearance and expression. The shooting and storage costs are low, and it is commonly used for recording the patient's facial features, evaluating changes before and after surgery, visual assessment of orthodontic treatment, etc.

[0071] Specifically, an intraoral CT image is a high-resolution three-dimensional imaging technology dedicated to detailed imaging of the oral cavity. It can provide high-resolution three-dimensional images of the oral cavity, mainly focusing on specific regions inside the oral cavity. The radiation dose is between that of a panoramic radiograph and CBCT, and it is commonly used for dental implants, root canal treatment, periodontal disease diagnosis, etc.

[0072] Specifically, intraoral CBCT imaging is a three-dimensional imaging technology that generates high-resolution three-dimensional images of the oral cavity and maxillofacial region by emitting a cone-shaped X-ray beam. It provides detailed three-dimensional anatomical structure information with low radiation dose, can clearly show the details of bones, teeth, and soft tissues, and is widely used in dental implants, orthodontics, periodontal diseases, maxillofacial surgery, etc.

[0073] Specifically, intraoral MRI is an imaging technology that uses magnetic fields and radio waves to generate images of the internal structures of the human body. It does not use X-rays and has no radiation hazard to the body. It is especially suitable for imaging soft tissues such as nerves, can generate images in any plane, provides rich anatomical information, and is mainly used for evaluating soft tissue lesions, etc.

[0074] Specifically, these imaging technologies each have their own advantages. By using them comprehensively, anatomical structure information of the dentomaxillofacial region can be obtained comprehensively and accurately, providing strong support for clinical diagnosis and treatment. In the automatic segmentation of the dentomaxillofacial anatomical structure, the fusion of multi-modal image data can make full use of the advantages of different imaging technologies to improve the accuracy and reliability of segmentation.

[0075] In a specific embodiment, in S200, it is assumed that in this embodiment, CBCT image data is used for preliminary positioning, including:

[0076] S210: Obtain CBCT image data of the anatomical structure to be dissected;

[0077] S220: Denoise and enhance the contrast of the CBCT image;

[0078] S230: Use Canny edge detection to extract the edges in the image;

[0079] S240: Use the Faster R-CNN model to detect the rough position of the dentomaxillofacial anatomical structure;

[0080] S250: Take the rough position as the preliminary segmentation area.

[0081] Specifically, it also includes using data such as CT images for preliminary positioning, and the present invention does not make any restrictions.

[0082] Specifically, it also includes combining multiple preliminary positioning methods and using ensemble learning (such as Bagging, Boosting, Stacking) to improve the robustness and accuracy of preliminary positioning.

[0083] Specifically, to traditionally determine the rough position of the anatomical structure to be dissected and use it as the preliminary segmentation area, a standard anatomical template (such as a standardized dentomaxillofacial anatomical structure model) is used as a reference to identify and label key anatomical landmark points in the image, such as the tips of teeth, the contours of the upper and lower jaws, etc.

[0084] In a specific embodiment, S300 includes:

[0085] S310: Obtain the position information of the anatomical structure in the CT image, CBCT image, or MRI image in the multi-modal image data in the first coordinate system;

[0086] S320: Obtain the semantic encoding data of the first coordinate system;

[0087] S330: Construct a three-dimensional model of the dentomaxillofacial anatomical structure, and compare and match the constructed three-dimensional model of the dentomaxillofacial anatomical structure with the existing standard model of the dentomaxillofacial anatomical structure to obtain the optimal dentomaxillofacial anatomical structure;

[0088] S340: Based on the position information, semantic encoding, and the optimal dentomaxillofacial anatomical structure of the first coordinate system, obtain the fused multi-modal image data.

[0089] In a specific embodiment, the origin of the first coordinate system is set at a preset fixed reference point in the CT image, CBCT image, or MRI image, such as the center of the image or a specific anatomical landmark point.

[0090] In a specific embodiment, S320 includes:

[0091] S321: Perform semantic segmentation on the multi-modal image data;

[0092] Specifically, the semantic targets for segmentation include teeth, maxilla and mandible, soft tissues, salivary glands, etc.

[0093] S322: Perform per-pixel classification to generate a semantic segmentation map;

[0094] S323: Perform a top-down projection on the segmented semantic segmentation map, and use the centroid of the overall point cloud of the anatomical structure to replace the position of the actual object during projection;

[0095] S324: Perform semantic encoding on the projected anatomical structure.

[0096] Specifically, perform semantic encoding on the projected anatomical structure to generate a semantic label for each anatomical structure in the first coordinate system. For example:

[0097] Tooth: 1

[0098] Maxilla: 2

[0099] Mandible: 3

[0100] Soft tissue: 4

[0101] Salivary gland: 5

[0102] Specifically, the coding method can use integer or string tags to represent different anatomical structures and associate these tags with the corresponding centroids of the 3D point clouds.

[0103] Specifically, the specific steps of semantic segmentation are as follows:

[0104] Input: Multi-modal imaging data (such as intraoral scans, CBCTs, panoramic radiographs, lateral cephalograms, facial photos, intraoral photos).

[0105] Model: Use U-Net or TransUNet++ for semantic segmentation.

[0106] Output: Semantic segmentation maps of each type of imaging data, containing tags of different anatomical structures.

[0107] Specifically, the specific steps of top-down projection are as follows:

[0108] Input: 3D point cloud model.

[0109] Method: Project the 3D point cloud model into the first coordinate system. When projecting, use the centroid of the overall point cloud of the anatomical structure to replace the position of the actual object.

[0110] Output: The centroid positions of each anatomical structure in the first coordinate system.

[0111] Specifically, the specific steps of semantic encoding are as follows:

[0112] Input: Projected centroid positions.

[0113] Method: Assign a unique semantic label to each anatomical structure and associate these labels with the corresponding centroid positions.

[0114] Output: Semantic encodings of each anatomical structure in the first coordinate system.

[0115] In a specific embodiment, S330 includes:

[0116] S331: Construct a 3D model of the anatomical structure. The nodes in the 3D model represent the anatomical parts to be recognized, and the edges represent the spatial relationships between the parts. Use the anatomical structure at a preset time point t as the nodes, and the edges represent the anatomical constraints between the nodes, including but not limited to anatomical proximity, connectivity, and shape features, etc.;

[0117] S332: Construct an optimized objective function based on the nodes and edges to obtain the 3D model of the anatomical structure;

[0118] S333: Compare and match the 3D model of the constructed anatomical structure with the existing standard model of the dentofacial anatomical structure to obtain the optimal dentofacial anatomical structure.

[0119] Specifically, the specific steps of 3D reconstruction are as follows:

[0120] Input: semantic segmentation map.

[0121] Method: Use 3D reconstruction technology (such as the Marching Cubes algorithm) to convert the 2D segmentation result into a 3D model.

[0122] Output: 3D point cloud model of each anatomical structure.

[0123] Specifically, through the above steps, the information of anatomical structures can be automatically extracted and encoded from the auxiliary imaging data of the dentomaxillofacial region.

[0124] In a specific embodiment, S500 includes:

[0125] S510: Initialize the reward function and the segmentation policy model based on the reward function using a deep neural network;

[0126] S520: Estimate the state distribution probability density of the segmentation policy model using the expert demonstration dataset, that is, the prediction distribution of the model on different anatomical structures;

[0127] S530: Based on the above state distribution probability density, use the deep reinforcement learning DRL model to update the segmentation policy model. Specifically, by iteratively calculating the expectation of the access count of the anatomical structure - segmentation behavior, use the maximum entropy gradient to update the weights of the deep neural network to optimize the segmentation policy;

[0128] S540: Determine whether the updated segmentation policy model meets the preset convergence condition. If it meets, record the weights of the current deep neural network to obtain the optimized segmentation policy model; if it does not meet, continue to estimate the state distribution probability density using the expert demonstration data and repeat the update process until the optimized segmentation policy model is obtained.

[0129] Specifically, the deep neural network models involved in this embodiment are all existing models. For example, U-Net is a classic network architecture for medical image segmentation, which is good at capturing local information. TransUNet++ combines the advantages of Transformer and U-Net, can effectively handle long-range dependencies, and at the same time maintain local details.

[0130] On the other hand, referring to Figure 2 , the embodiment of the present invention also discloses an automatic segmentation system for dentomaxillofacial anatomical structures based on a deep learning model, including:

[0131] Data collection module: Collect multi-modal imaging data of the anatomical structure to be analyzed from different imaging modalities;

[0132] Initial confirmation module: Determine the approximate location of the anatomical structure to be dissected and use it as the initial segmentation area;

[0133] Data fusion module: Within the initial segmentation area, fuse the multi-modal image data;

[0134] Dataset generation module: Combine the fused multi-modal image data with the labeled true label data to form an expert demonstration dataset;

[0135] Model optimization module: Build and train a segmentation strategy model based on the expert demonstration dataset to obtain an optimized segmentation strategy model;

[0136] Output module: Use the optimized segmentation strategy model to achieve automatic segmentation of the dental and maxillofacial anatomical structure.

[0137] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, please refer to the description in the method section.

[0138] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An automatic segmentation method for maxillofacial anatomical structure based on a deep learning model, characterized in that: The following steps are involved: S100: Collect multimodal image data of the anatomical structure to be imaged from different images; S200: determining a rough position of the structure to be dissected and using it as a preliminary segmentation area; S300: fusing multimodal image data within the preliminary segmented area; S400: combining the fused multimodal image data with the annotated real label data to form an expert demonstration data set; S500: constructing and training a segmentation strategy model based on the expert demonstration data set to obtain an optimized segmentation strategy model; S600: Automatic segmentation of maxillofacial anatomy using the optimized segmentation strategy model.

2. The method for automatic segmentation of dental and maxillofacial anatomical structures based on a deep learning model according to claim 1, characterized in that: The multimodal image data include: oral scans, panoramic films, cephalographic films, facial photographs, and intraoral CT images, CBCT images, or MRI images.

3. The method for automatic segmentation of dental and maxillofacial anatomical structures based on a deep learning model according to claim 1, characterized in that: The S200 includes: S210: Acquire multimodal image data of the structure to be dissected; S220: denoising and contrast enhancement of multimodal image data; S230: extracting edges in the multimodal image data using Canny edge detection; S240: Detect the rough position of the maxillofacial anatomical structure using the FasterR-CNN model; S250: Taking the rough position as a preliminary segmentation area.

4. The method for automatic segmentation of dental and maxillofacial anatomical structures based on a deep learning model according to claim 1, characterized in that: The S300 includes: S310: Acquire position information of the anatomical structure in the multimodal image data in the first coordinate system; S320: Acquire semantic coding data of the first coordinate system; S330: constructing a three-dimensional model of the dental and maxillofacial anatomical structure, and comparing and matching the constructed three-dimensional model of the dental and maxillofacial anatomical structure with an existing standard model of the dental and maxillofacial anatomical structure to obtain an optimal dental and maxillofacial anatomical structure; S340: Based on the position information of the first coordinate system, the semantic coding and the optimal maxillofacial anatomical structure, the fused multimodal image data is obtained.

5. The method for automatic segmentation of dental and maxillofacial anatomical structures based on a deep learning model according to claim 4, characterized in that: The coordinate origin of the first coordinate system is set at a fixed reference point preset in the multimodal image data.

6. The method for automatic segmentation of dental and maxillofacial anatomical structures based on a deep learning model according to claim 4, characterized in that: The S320 includes: S321: performing semantic segmentation on multimodal image data; S322: Perform pixel-by-pixel classification to generate a semantic segmentation map; S323: performing a top-down projection on the segmented semantic segmentation map, and replacing the position of the actual object with the centroid of the overall point cloud of the anatomical structure during projection; S324: Perform semantic encoding on the projected anatomical structures to generate a semantic label for each anatomical structure in the first coordinate system.

7. The method for automatic segmentation of maxillofacial anatomical structure based on a deep learning model according to claim 4, characterized in that: The S330 includes: S331: constructing a three-dimensional model of the anatomical structure, wherein nodes in the three-dimensional model represent anatomical parts to be identified, and edges represent spatial relationships between the parts. The anatomical structure at a preset time point t is used as a node, and the edges represent anatomical constraints between nodes. S332: constructing a three-dimensional model of the anatomical structure according to the objective function of the node and edge construction optimization; S333: Compare and match the constructed three-dimensional model of the anatomical structure with the existing standard model of the maxillofacial anatomical structure to obtain the optimal maxillofacial anatomical structure.

8. The method for automatic segmentation of dental and maxillofacial anatomical structures based on a deep learning model according to claim 1, characterized in that: The S500 comprises: S510: Initialize a reward function and a segmentation strategy model based on the reward function using a deep neural network; S520: estimating the state distribution probability density of the segmentation strategy model using the expert demonstration data set; S530: Based on the state distribution probability density, using the DRL model to update the segmentation strategy model; S540: Determine whether the updated segmentation strategy model meets the preset convergence conditions. If so, record the weights of the current deep neural network to obtain the optimized segmentation strategy model. If not, continue to use the expert demonstration data to estimate the state distribution probability density and repeat the updating process until the optimized segmentation strategy model is obtained.

9. An automatic segmentation system for the dental and maxillofacial anatomical structure based on a deep learning model using the automatic segmentation method for the dental and maxillofacial anatomical structure based on a deep learning model according to any one of claims 1 to 8, characterized in that: include: Data collection module: collects multimodal imaging data of the anatomical structure to be imaged from different images; Preliminary confirmation module: determines the rough location of the anatomical structure to be analyzed and uses it as the preliminary segmentation area; Data fusion module: fuses multimodal image data within the initial segmentation area; Dataset generation module: combines the fused multimodal image data with the annotated real label data to form an expert demonstration dataset; Model optimization module: builds and trains a segmentation strategy model based on the expert demonstration dataset to obtain an optimized segmentation strategy model; Output module: Use the optimized segmentation strategy model to achieve automatic segmentation of the maxillofacial anatomical structure.

Images