An AI-based automatic and detailed oral CBCT anatomy method

Abstract

A method for automatic fine anatomy of oral CBCT based on artificial intelligence mainly includes the following steps: S1 Obtain an oral CBCT fine anatomy annotation dataset; S2 Perform a first-stage automatic rough anatomy of oral CBCT based on the first anatomy network; S3 Perform a second-stage automatic fine anatomy of oral CBCT based on the second anatomy network; S4 Obtain the final oral CBCT fine anatomy mask through two-stage mask fusion; By constructing an automatic anatomy network for fine oral structures based on deep learning, the present invention can efficiently, automatically, and batch anatomize micro-millimeter-level (pixel-level), variously shaped, and anatomically structured with unclear boundaries in oral CBCT images, solve imaging structure anatomy problems such as loss of underlying detailed texture information and redundant computational volume, and effectively reduce the time and effort consumed by dentists during the fine anatomy of oral CBCT structures.

Description

Technical Field

[0001] The present invention belongs to the field of oral imaging, and particularly relates to an automatic fine anatomy method for oral CBCT based on artificial intelligence. Background Art

[0002] With the development of current precision medicine and personalized diagnosis and treatment, the understanding of oral physiological and pathological structures has reached the fine level of millimeters; precise diagnosis focuses on such fine anatomical parts and minute pathological regions; precise treatment is carried out at the organ subunit level (such as root shield technology, minimally invasive tooth extraction, robot-assisted surgery, etc.). For example, the root shield technology does not completely extract the tooth, but needs to retain the labial root fragment, and then carry out implant surgery on this basis, which indicates that the analysis of the patient's anatomical structure needs to be deepened to the fine organ subunit level.

[0003] In the process of disease diagnosis, treatment, and prognosis analysis in the field of oral medicine, the recognition and anatomy of fine structures in CBCT images are key steps in each link. The sizes of fine structures are often in the micron or millimeter range [1] , for example, the average thickness of dental enamel is 2 - 2.5 mm; the thickness of the labial bone wall in the upper anterior teeth is often less than 1 mm; the periodontal ligament space is only 0.15 - 0.38 mm wide. The manual recognition and anatomy of fine structures still consume a large amount of human and time costs, and relying on manual quantitative or qualitative analysis of fine structure anatomy may cause relatively large analysis errors.

[0004] In recent years, with the development of the field of artificial intelligence in medical imaging, research has been carried out to achieve the overall automatic anatomy of oral organs such as teeth and alveolar bone, and good results have been obtained. However, there are the following difficulties in the automatic anatomy of small structures at the organ subunit level of oral structures based on artificial intelligence methods:

[0005] First, due to the hardware conditions of radiation equipment and imaging principles, the most commonly used CBCT images in oral medicine have low resolution, and the fine structures in them only show pixel-level or sub-pixel-level imaging features. For large anatomical structures (such as complete gross organs like teeth and alveolar bone), since the organ itself is large in size and occupies a large number of voxels as a whole, the uneven anatomical edges caused by a few misrecognized pixels are sufficient to distinguish the overall shape of the organ, and it will not cause too many problems in clinical use. However, for fine structures (such as dental enamel and dental pulp at the organ subunit level), since the object itself occupies not many voxels, small errors in the anatomical edges may cause misjudgment of the overall shape, seriously affecting clinical applications. Therefore, it is necessary to update the framework of the automatic anatomy model to obtain better pixel-level fine anatomy results.

[0006] Second, the various delicate structures in the oral cavity do not have a unified morphology, their imaging features are variable, and there is no obvious boundary with adjacent tissues. There are obvious tissue spaces between gross organs (such as teeth) (manifested as differences in pixel density between imaging anatomies), and the overall consistency of the organs is relatively strong. In contrast, the delicate structures and the surrounding tissues are often integrated, with similar densities and indistinct boundaries, which poses great challenges to unsupervised algorithms based on gray-scale thresholds and watersheds. Additionally, a single rough dissection often only separates the areas with the largest differences and cannot distinguish smaller delicate subunits within them. Therefore, it is necessary to adopt a more robust algorithm (such as supervised learning) to solve this problem.

[0007] Third, when applying a conventional tissue dissection analysis network for a single rough dissection of oral cavity structures, the deep convolution of the network results in a large amount of loss of detailed texture information. Taking the Unet network as an example, it consists of two components: an encoder and a decoder. Previous studies have shown that for large-area dissections, increasing the depth of the encoder layer can obtain better dissection analysis accuracy. However, for the dissection of delicate structures, multiple convolutions in the encoder cause compression of the feature maps, losing fine-grained information at the bottom layer; at the same time, maintaining the spatial connection information in the low-level feature maps requires a large amount of unavoidable computational effort. Therefore, the dissection efficiency and effectiveness for delicate structures are poor and low. Based on this, cascading dissections using two networks with relatively shallow layers are expected to solve the deficiencies of a single network.

[0008] [1] Gao Yan (chief editor). "Oral Histopathology (8th Edition)". Beijing: People's Medical Publishing House, 2020. Summary of the Invention

[0009] The present invention aims to solve at least one of the technical problems existing in the prior art. For this purpose, the present invention proposes an artificial intelligence-based automatic fine dissection method for oral CBCT.

[0010] The technical solution adopted by an embodiment of the present invention to solve its technical problems is:

[0011] An artificial intelligence-based automatic fine dissection method for oral CBCT, comprising the following steps:

[0012] S1 Obtain an oral CBCT fine dissection annotation dataset

[0013] S101 Obtain an oral CBCT section and perform size normalization processing:

[0014] Obtain a first two-dimensional oral CBCT section from the original oral CBCT image; perform normalization processing on the first two-dimensional oral CBCT section to obtain a second two-dimensional oral CBCT section.

[0015] S102 Obtain the oral fine anatomical annotation file through manual annotation and quality control:

[0016] Import the second oral CBCT two-dimensional section obtained in step S101 into the image annotation software, and manually annotate the oral fine anatomical structures by experienced doctors to obtain the first oral fine anatomical annotation file containing only annotation information; convert the first oral fine anatomical annotation file into a second oral fine anatomical annotation file containing annotation information and original picture information; traverse the second oral fine anatomical annotation file in the computer processing software to exclude abnormally annotated or missed annotated pixel points.

[0017] S103 Construct an oral fine anatomical annotation dataset:

[0018] After the quality control in S102, the remaining annotation files are recorded as having excellent quality; archive the second oral fine anatomical annotation files with excellent quality into the specified database path to form an oral CBCT fine anatomical annotation dataset.

[0019] S2 Perform the first-stage automatic rough anatomy of oral CBCT based on the first anatomical network:

[0020] S201 Construct the first anatomical network based on the deep learning model:

[0021] The first anatomical network performs a downsampling process on the input image, extracts feature maps and classifies each pixel, and at the same time assigns pixel-level class label classification and retains the spatial information of the original input image through the upsampling process and the connection module, and finally outputs the pixel-level mask corresponding to the anatomical category of the oral CBCT.

[0022] S202 Train and validate the first anatomical network based on the deep learning model:

[0023] Input the oral CBCT fine anatomical annotation dataset obtained in step S103 into the first anatomical network based on the deep learning model constructed in S201; the first anatomical network learns the data features on the oral CBCT fine anatomical annotation training set and saves the model parameters; adopt various training strategies to train, validate and optimize the first anatomical network.

[0024] S203 Obtain the first-stage rough anatomical pixel-level mask of oral CBCT

[0025] Run the trained and validated first anatomical network to process the oral CBCT image structure to obtain the first-stage rough anatomical pixel-level mask of oral CBCT;

[0026] S3 Perform the second-stage automatic fine anatomy of oral CBCT based on the second anatomical network

[0027] S301 Construct the two-stage automatic anatomical dataset of oral CBCT:

[0028] Adopt computer vision algorithms to automatically locate, extract, and crop the fine-structure regions corresponding to the rough anatomical pixel-level masks of the first-stage oral CBCT. Perform size normalization on the images of the extracted regions to construct the two-stage automatic anatomical dataset of oral CBCT.

[0029] S302 Construct the second anatomical network based on the deep learning model:

[0030] The second anatomical network has a shallower network depth than the first anatomical network, that is, fewer convolutional layers; the second anatomical network further performs pixel-level classification on the fine anatomical structure regions of the oral cavity, and at the same time adopts a skip connection structure to retain the spatial information of the original input image, and finally outputs a mask corresponding to the category on the basis of pixel-level classification to confirm the fine boundary of the fine anatomical structure.

[0031] S303 Train and validate the second anatomical network based on the deep learning model:

[0032] Adopt multiple strategies to train and validate the second anatomical network based on the deep learning model

[0033] S304 Obtain the two-stage oral CBCT fine anatomical pixel-level mask

[0034] Run the trained and validated second anatomical network to process the two-stage automatic anatomical dataset of oral CBCT to obtain the two-stage oral CBCT fine anatomical pixel-level mask;

[0035] S4 Fuse the two-stage anatomical masks to obtain the final oral CBCT fine anatomical mask

[0036] S401 Fuse the two-stage masks to obtain the final oral CBCT fine anatomical mask:

[0037] Fuse the rough anatomical pixel-level mask of the first-stage oral CBCT and the fine anatomical pixel-level mask of the second-stage oral CBCT to obtain a fused image.

[0038] S402 Post-process the fused image to obtain the final oral CBCT fine anatomical mask:

[0039] Adopt computer graphics analysis algorithms to post-process the fused image, identify abnormal fusion regions, eliminate abnormal anatomical regions, and obtain the final oral CBCT fine anatomical mask.

[0040] S403 Qualitative and quantitative analysis based on the final oral CBCT fine anatomical mask:

[0041] Perform qualitative and quantitative analysis on the final oral CBCT fine anatomical mask.

[0042] Preferably, the oral fine anatomy is a further detailed anatomical analysis of the general oral anatomical organs, including enamel, dentin, pulp cavity, root canal system, labial bone wall, palatal bone wall, basal bone, cortical bone, cancellous bone, mandibular nerve canal, maxillary sinus.

[0043] Preferably, constructing the first anatomical network and the second anatomical network based on the deep learning model includes the following steps:

[0044] After the first anatomical network and the second anatomical network perform downsampling on the image, they extract feature maps and perform pixel-level classification. At the same time, a connection module and an upsampling process are used to assign pixel-level class labels for classification and retain the spatial information of the original input image. Finally, a mask of fine anatomy corresponding to the category is output based on pixel-level classification; the deep learning-based model includes FCN, U-net, nnUnet, Unet++, Deeplab V3, SegNet, SegFormer; the first and second anatomical networks extract feature maps through downsampling and use the upsampling method to restore the size of the original image; the downsampling process uses methods including convolution, pooling, dilated convolution, and transformer; the downsampling process and the upsampling process are connected through multiple connection modules, and the detailed texture information of the joint shallow feature map and the global texture information in the deep feature map are transmitted to the rough and fine anatomy masks of the oral CBCT in the first stage and the second stage.

[0045] Preferably, the connection module includes: lateral skip connection, dense skip connection, skip connection based on pooling index.

[0046] Preferably, the method for training the first and second anatomical networks based on the deep learning model includes K-fold cross-validation, model integration, and model pruning.

[0047] Preferably, the method for validating the first and second anatomical networks based on the deep learning model includes Dice coefficient, average pixel accuracy, pixel accuracy for each category, mean intersection over union, intersection over union for each category index, Hausdorff distance.

[0048] Preferably, the abnormal fusion region includes the anatomical discontinuous region in the complete tissue, the annotation of other regions appears inside the complete region, and the annotation of the fine anatomical tissue type is incorrect.

[0049] Preferably, the method for eliminating the abnormal annotation region includes erosion, dilation, opening operation and closing operation, edge detection algorithm.

[0050] The beneficial effects of the present invention:

[0051] (1) By constructing an oral fine anatomy network based on deep learning, the present invention can efficiently, automatically, and batch conduct anatomical analysis on the micro-millimeter level (pixel level), variously shaped, and indistinctly edged oral fine structures in CBCT images, effectively reducing the time and effort consumed by dentists during CBCT analysis and oral fine anatomy, and promoting the development of precise oral diagnosis and treatment.

[0052] (2) The two-stage anatomical strategy adopted by the invention improves the anatomical accuracy while reducing the model complexity and computing power cost in the second stage. At the same time, the relatively shallow second-stage anatomical network will not overly compress the fine structure features, maximizing the preservation of underlying fine-grained feature information. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above or additional aspects and advantages of the present invention will become apparent and easy to understand from the description of the embodiments in conjunction with the following drawings, wherein:

[0054] Figure 1 is a flowchart of an automatic fine anatomy method for oral CBCT based on artificial intelligence;

[0055] Figure 2 is a flowchart of obtaining a fine anatomy annotation dataset for the anterior teeth mid-sagittal plane of oral CBCT;

[0056] Figure 3 is a flowchart of the first and second stage automatic anatomy of the anterior teeth based on a cascaded deep learning anatomy network;

[0057] Figure 4 is a flowchart of the method for obtaining the final anterior teeth CBCT fine anatomy mask and its qualitative and quantitative analysis;

[0058] Figure 5 is an example of a fine anatomy annotation dataset for the anterior teeth mid-sagittal plane of oral CBCT and a two-stage anatomy mask. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] The detailed description of the embodiments is very important for fully disclosing, understanding, and reproducing the invention, and for supporting and interpreting the claims. The description of the embodiments should be consistent with the technical solutions adopted to solve the technical problems. For simple technical solutions, only one embodiment may be given. For technical solutions with a wider protection scope, more than one embodiment should be given to support the claimed scope.

[0060] The embodiments of the present invention provide an embodiment of a method for fine anatomy of the anterior teeth of oral CBCT based on artificial intelligence, including the following steps:

[0061] S1 Obtain a fine anatomy annotation dataset for the anterior teeth mid-sagittal plane of oral CBCT:

[0062] S101 Acquisition and size normalization of the mid-sagittal plane of the upper anterior teeth in oral CBCT:

[0063] Uniformly acquire oral CBCT images taken by the NewTom VG oral CBCT imaging system. Specifically, the imaging protocol or imaging parameters used are all the unified protocols preset by the machine. Further, batch export the oral CBCT images through the NNT medical imaging management software, name them according to the standard rules, and save them in the DICOM format at the specified path.

[0064] Further, import the obtained oral CBCT images into the implant simulation software coDianostiX, take screenshots through the standardized process for obtaining the mid-sagittal plane of the upper anterior teeth, name them according to the standard rules, and save them in the TIFF format at the specified path to obtain the first mid-sagittal plane of the upper anterior teeth. Specifically, the standardized process for obtaining the mid-sagittal plane of the upper anterior teeth includes: First, traverse from top to bottom according to the slice order. At the slice level where the enamel of the upper anterior teeth first appears, connect the centers of each tooth on this level to obtain the dental arch curve. Based on the slope of the dental arch curve and the position information of the tooth centers, reconstruct the mid-sagittal plane of the upper anterior teeth in the implant simulation software coDianostiX to obtain the first mid-sagittal plane of the upper anterior teeth.

[0065] Further, apply the python programming language to batch import the obtained first mid-sagittal plane of the upper anterior teeth, uniformly normalize the size to 416*416 pixels, name them according to the standard rules, and save them in the JPEG format at the specified path to obtain the second mid-sagittal plane of the upper anterior teeth.

[0066] S102 Manual annotation of the fine oral anatomical structures of the upper anterior teeth:

[0067] Import the second mid-sagittal plane of the upper anterior teeth obtained in step S101 into the image annotation software labelme. Further, a clinician with background knowledge in oral imaging annotates the oral anatomy in the second mid-sagittal plane of the upper anterior teeth. Specifically, the fine anatomy of the upper anterior teeth includes: enamel, dentin, pulp and root canal system, alveolar bone, and labial bone wall. Specifically, outline the edges of the fine anatomy of the upper anterior teeth in the image annotation software labelme to form a closed-loop area, export and save the annotation position information of each oral fine anatomy, name them according to the standard rules, and save them in the JSON format at the specified path to obtain the first fine anatomy annotation file of the upper anterior teeth.

[0068] S103 Quality control of the fine anatomy annotation file of the upper anterior teeth:

[0069] Run the json_to_dataset.py project file using the Python programming language to batch-convert the first upper anterior tooth fine anatomical annotation file containing only annotation information obtained in step S102 into a second upper anterior tooth fine anatomical annotation file in PNG format containing annotation information and original image information.

[0070] Furthermore, use the Python programming language to expand the second upper anterior tooth fine anatomical annotation file in matrix form to check whether there are missing annotated blank pixel points between adjacent fine anatomical categories and to check whether the edges of the annotated areas meet clinical requirements. Further, if there are missing annotated pixel points, re-execute step S102 to supplement the annotation.

[0071] S104 Construct a CBCT upper anterior tooth fine anatomical annotation dataset:

[0072] After quality control in S103, the remaining annotation files are recorded as having excellent quality; use the Python programming language to batch-move the error-free second upper anterior tooth fine anatomical annotation files after quality control in step S103 to a specified folder to obtain a CBCT upper anterior tooth fine anatomical annotation dataset.

[0073] S2. Perform a first-stage automatic rough anatomy of the upper anterior tooth CBCT based on the first anatomical network:

[0074] S201 Construct a first anatomical network based on the nnUnet model:

[0075] Construct an nnUnet model for performing first-stage automatic anatomy in the PyTorch framework. Specifically, the nnUnet model is an adaptive deep learning network based on the U-net network framework. Specifically, the nnUnet model can automatically adjust the parameters involved in the preprocessing, network framework, training, and post-processing parts according to the content of the database to be processed. Specifically, the U-shaped network framework consists of a downsampling encoder composed of convolutional layers and pooling layers, an upsampling decoder, and skip connections between the encoder and decoder. Further, the nnUnet model divides the parameters that may affect the model performance in the entire process of preprocessing, model training, and post-processing into three groups, namely, "unchanged fixed parameter group", "rule-based parameter group", and "empirical parameter group". Specifically, the three parameter groups constitute the adaptive module of the nnUnet model.

[0076] S202 Train and validate the first anatomical network based on the nnUnet model:

[0077] The first anatomical network based on the nnUnet model is trained and verified using a 5-fold cross-validation method. Specifically, the CBCT upper anterior tooth fine anatomical annotation dataset constructed in step 104 is evenly divided into 5 parts. Further, in the first training, 4 of these parts are used as the training set, and the remaining 1 part is used as the validation set. Further, in the second training, 1 part of the training set in the first cycle is used as the validation set, and the remaining 4 parts are used as the training set. Further, this rule is followed to train 5 times in sequence, and the model training parameters are saved.

[0078] Further, after inputting the CBCT upper anterior tooth fine anatomical annotation dataset obtained in step 104 into the first anatomical network based on the nnUnet model constructed in step S201, the first anatomical network automatically analyzes the data characteristics of the CBCT upper anterior tooth fine anatomical annotation dataset. Further, the first anatomical network based on the nnUnet model automatically adjusts the three parameter groups in the adaptive module described in step S201. Further, based on the adaptive module, the model is automatically optimized to achieve a proper fit on the training set and the validation set.

[0079] Further, the model anatomical training results are evaluated on the validation set. Specifically, Dice coefficient, mean pixel accuracy (mPA), class pixel accuracy (cPA), mean intersection over union (mIoU), and class intersection over union (cIoU) metrics are used. Specifically, for the evaluation of the fine anatomical effects of various categories of upper anterior teeth, including enamel, dentin, pulp and root canal system, labial bone wall, and alveolar bone.

[0080] S203 Obtain the rough anatomical pixel-level mask of the upper anterior tooth CBCT in the first stage:

[0081] According to the first anatomical network based on the nnUnet model constructed and trained in step S202, the optimal rough anatomical result of the upper anterior tooth CBCT image structure of the model is output as a pixel-level mask. Specifically, the predicted anatomical result of the first anatomical network based on the nnUnet model is named according to the standard rule and saved in the specified path to obtain the rough anatomical pixel-level mask of the upper anterior tooth CBCT in the first stage.

[0082] S3 Perform the second-stage automatic fine anatomy of the upper anterior tooth CBCT based on the second anatomical network:

[0083] S301 Automatically extract the fine regions in the rough anatomical pixel-level mask of the upper anterior tooth CBCT in the first stage to construct the second-stage automatic anatomy dataset of the upper anterior tooth CBCT:

[0084] According to the first-stage upper anterior tooth CBCT rough anatomical pixel-level mask obtained in step S203, obtain the coordinates of all fine anatomical annotation pixel points in the mid-sagittal plane of the upper anterior teeth of the oral CBCT obtained in step S101. Further, using the Python computer language, automatically crop the area of the upper anterior teeth' fine anatomy in the mid-sagittal plane of the oral CBCT obtained in step S101. Specifically, the size of the cropped area is 64*64 pixels. Specifically, the fine anatomical structures of the upper anterior teeth that need to be automatically cropped include the labial bone wall, alveolar bone, enamel, dentin, dental pulp, and root canal system.

[0085] Further, using the dilation computer image processing algorithm, obtain the peripheral area of the first-stage upper anterior tooth CBCT rough anatomical pixel-level mask. Specifically, the ratio of the obtained peripheral area of the first-stage upper anterior tooth rough anatomy to the first-stage rough anatomical area is 2:1. Further, integrate the first-stage upper anterior tooth rough anatomy and its peripheral area to obtain the second-stage automatic anatomical dataset of the upper anterior tooth CBCT.

[0086] S302 Construct the second anatomical network based on the Unet++ model:

[0087] Construct the Unet++ model for second-stage automatic anatomy in the pytorch framework. Specifically, Unet++ includes a downsampling encoder, an upsampling decoder, and a skip connection module. Specifically, the backbone network of the Unet++ encoder uses ResNet50, and the number of downsampling times is 4 times. Specifically, the number of upsampling times of Unet++ is 4 times. Specifically, use long skip connections to directly connect the corresponding layers of the encoder and the decoder, which helps to restore the information loss caused by downsampling. Specifically, use short skip connections to connect the upsampling paths of the first to third layers of the encoder respectively, enabling backpropagation to be achieved. The loss function used for each layer of the second anatomical network is the sum of the Dice loss function and the BCE loss function. Further, the overall loss function of the second anatomical network is the average of the loss functions of the feature maps obtained from different layers. Further, use the Adam optimizer to minimize the loss function.

[0088] S303 Train and validate the second anatomical network based on the Unet++ model:

[0089] Use the 5-fold cross-validation method to train and validate the second anatomical network based on the Unet++ model. Specifically, evenly divide the second-stage automatic anatomical dataset of the upper anterior tooth CBCT constructed in step S301 into 5 parts. Further, in the first training, use 4 of them as the training set, and the remaining 1 as the test set. Further, in the second training, use 1 part of the training set in the first cycle as the test set, and the remaining 4 as the training set. Further, train 5 times in this rule in sequence, and save the model training parameters.

[0090] Specifically for each cycle, the upper anterior tooth CBCT two-stage automatic anatomical dataset constructed in step S301 is input into the second anatomical network based on the Unet++ model. Further, in the 4-layer downsampling encoder with Resnet50 as the backbone for image output, feature maps are obtained. Further, the feature maps are restored to the original image size through 4 times of upsampling decoders. Specifically, the combination of long skip connections and short skip connections is adopted, which increases the feature information from different downsampling layers and helps to improve the fine anatomical accuracy. Further, after multiple iterative trainings, it reaches a proper fit on the training set and the validation set.

[0091] Further, the training results of the model are evaluated on the validation set. Specifically, the Dice coefficient, mean pixel accuracy (mPA), class pixel accuracy (cPA), mean intersection over union (mIoU), and class intersection over union (cIoU) metrics are adopted. Specifically, for the fine anatomical structures of the upper anterior teeth, including enamel, dentin, dental pulp and root canal system, labial bone wall, and alveolar bone, the above evaluation results are output.

[0092] S304 Obtain the pixel-level mask for the fine anatomy of the upper anterior tooth CBCT in the second stage:

[0093] According to the second anatomical network based on the Unet++ model constructed and trained in step S303, the optimal fine anatomical effect output of the model is output as a pixel-level mask. Specifically, the fine anatomical result of the second anatomical network based on the Unet++ model is output, named according to the standard rules, and saved in the specified path to obtain the pixel-level mask for the fine anatomy of the upper anterior tooth CBCT in the second stage.

[0094] S4 Obtain the final pixel-level mask for the fine anatomy of the upper anterior tooth CBCT by fusing the two-stage masks:

[0095] S401 Fuse the anatomical masks of the upper anterior tooth CBCT in the two stages:

[0096] Fuse the two-stage upper anterior tooth CBCT fine anatomical pixel-level mask obtained in step S304 with the one-stage upper anterior tooth CBCT rough anatomical pixel-level mask obtained in step S203. Specifically, the two-stage upper anterior tooth fine anatomy includes subtle structures such as the labial bone wall, dental pulp and root canal system, enamel, and dentin, and record the position information of the mask after its automatic anatomy. Further, add the position information to the corresponding categories (such as the labial bone wall, palatal bone wall, dental pulp and root canal system, dentin, enamel) in the one-stage upper anterior tooth CBCT rough anatomical mask.

[0097] S402 Post-process the fused image to obtain the final upper anterior tooth CBCT fine anatomical mask:

[0098] Post-process the fused image obtained in step S401 using a computer image analysis algorithm. Specifically, judge whether there are abnormal sites based on prior anatomical knowledge. Specifically, the abnormal sites include: discontinuous annotation regions in the complete tissue, and annotations of other regions appearing inside the complete region. Further, use the dilation algorithm to eliminate the abnormal anatomical results. Further, obtain the final upper anterior tooth CBCT fine anatomical mask and save it in the specified path according to the standard naming rule.

[0099] S5 Qualitative and quantitative analysis based on the final upper anterior tooth CBCT fine anatomical mask:

[0100] Based on the final upper anterior tooth CBCT fine anatomical mask obtained in step S402, perform qualitative and quantitative analysis on the upper anterior tooth CBCT fine anatomy. Specifically, the upper anterior tooth CBCT fine anatomy includes: dental pulp, enamel, and labial bone wall. Specifically, the qualitative analysis includes: whether there is a complete pulp cavity and root canal system, and whether the root canal system has calcified. Specifically, the quantitative analysis includes: the thickness of the upper anterior tooth enamel and the thickness of the labial bone wall.

[0101] Although the present invention has been described in detail with reference to the above embodiments, it is obvious to those skilled in the art through the present disclosure that various changes or modifications can be made to the present invention without departing from the principle and spirit scope of the present invention defined by the claims. Therefore, the detailed description of the embodiments of the present disclosure is only used to explain, rather than to limit the present invention, and the scope of protection is defined by the content of the claims.

Claims

1. An automatic and precise oral CBCT anatomical method based on artificial intelligence, characterized in that Including the following steps: S1 Obtain the oral CBCT fine anatomical annotation dataset Based on the original oral CBCT image, obtain the first two-dimensional section of oral CBCT; process the first two-dimensional section of oral CBCT to obtain the second two-dimensional section of oral CBCT, and import the second two-dimensional section of oral CBCT into image annotation software for processing to obtain the first oral fine anatomical annotation file; convert the first oral fine anatomical annotation file into a second oral fine anatomical annotation file containing annotation information and original picture information; archive the second oral fine anatomical annotation file to form the oral CBCT fine anatomical annotation dataset; S2 Perform the first-stage automatic rough anatomy of oral CBCT based on the first anatomical network First, construct the first anatomical network based on the deep learning model, and input the oral CBCT fine anatomical annotation dataset into the first anatomical network constructed based on the deep learning model; train and validate the first anatomical network; The first anatomical network based on the deep learning model after training and validation processes the oral imaging structure to obtain the first-stage rough anatomical pixel-level mask of oral CBCT; S3 Perform the second-stage automatic fine anatomy of oral CBCT based on the second anatomical network First, extract the fine regions in the first-stage rough anatomical pixel-level mask of oral CBCT to construct the second-stage automatic anatomical dataset of oral CBCT; secondly, construct the second anatomical network based on the deep learning model; train and validate the second anatomical network based on the deep learning model; the second anatomical network based on the deep learning model after training and validation processes the oral imaging structure to obtain the second-stage fine anatomical pixel-level mask of oral CBCT; S4 Fuse the two-stage masks to obtain the final fine anatomical mask of oral CBCT Fuse the first-stage rough anatomical pixel-level mask of oral CBCT and the second-stage fine anatomical pixel-level mask of oral CBCT to obtain a fused image, identify the abnormal fusion regions, eliminate the abnormal annotation regions, obtain the final fine anatomical mask of oral CBCT, and perform qualitative and quantitative analysis on the oral fine anatomical mask; Constructing the first anatomical network based on the deep learning model includes the following steps: The first anatomical network performs pixel-level classification on the image, while retaining the spatial information of the original input image, and finally outputs a mask corresponding to each category on the basis of pixel-level classification to indicate their respective anatomical categories; the deep learning model includes FCN, U-net, nnUnet, Unet++, Deeplab V3, SegNet, SegFormer; the first anatomical network extracts feature maps through downsampling and uses the upsampling method to restore the original image size; the downsampling process uses methods including convolution, pooling, atrous convolution, and transformer; the downsampling path and the upsampling path are connected through multiple connection modules, and the detailed texture information of the shallow feature maps and the global texture information in the deep feature maps are jointly transmitted to the final feature image to generate the first-stage rough anatomical pixel-level mask of oral CBCT; Constructing the second anatomical network based on the deep learning model includes the following steps: The second anatomical network performs pixel-level fine classification on the rough anatomical pixel-level mask of the one-stage oral CBCT and its surrounding areas, and finally outputs a mask corresponding to the anatomical category based on the pixel-level classification; the deep learning-based model includes FCN, U-net, nnUnet, Unet++, Deeplab V3, SegNet, and SegFormer; the second anatomical network has a shallower network depth than the first anatomical network; the second anatomical network extracts feature maps through downsampling and restores the original image size using the upsampling method; the downsampling path and the upsampling path are connected through multiple connection modules; the second anatomical network generates a two-stage oral CBCT fine anatomical pixel-level mask.

2. The automatic fine anatomy method of oral CBCT based on artificial intelligence according to claim 1, wherein The fine anatomical structures include enamel, dentin, pulp cavity, root canal system, labial bone wall, palatal bone wall, basal bone, cortical bone, cancellous bone, mandibular nerve canal, and maxillary sinus.

3. The automatic fine anatomy method of oral CBCT based on artificial intelligence according to claim 1, characterized in that The connection modules include: lateral skip connections, dense skip connections, and skip connections based on pooling indices.

4. The automatic fine anatomy method of oral CBCT based on artificial intelligence according to claim 1, characterized in that, The methods for training the first and second anatomical networks based on the deep learning model include K-fold cross-validation, model ensemble, and model pruning.

5. The automatic fine anatomy method of oral CBCT based on artificial intelligence according to claim 1, characterized in that, The methods for validating the first and second anatomical networks based on the deep learning model include Dice coefficient, average pixel accuracy, pixel accuracy for each category, average intersection over union, intersection over union indices for each category, and Hausdorff distance.

6. The automatic fine anatomy method of oral CBCT based on artificial intelligence according to claim 1, characterized in that, The abnormal fusion regions include discontinuous regions in the complete tissue, the appearance of annotations of other regions inside the complete region, and incorrect annotations of fine anatomical tissue types.

7. An automatic fine anatomy method of oral CBCT based on artificial intelligence according to claim 1, characterized in that, The methods for eliminating abnormal annotation regions include erosion, dilation, opening and closing operations, and edge detection algorithms.

Images