Oral cavity image analysis method based on Micro-CT and transfer learning

Abstract

The invention discloses an oral cavity image analysis method based on Micro-CT and transfer learning, and relates to the field of artificial intelligence assisted medical image diagnosis, the method comprises the following steps: using Micro-CT and CBCT scanning technologies to obtain isolated tooth three-dimensional data; taking Micro-CT three-dimensional data as input, taking a morphological classification result obtained through the Micro-CT data as a label, and training the classification model to obtain a weight parameter of a first classification model; taking the CBCT three-dimensional data as input, taking the Micro-CT classification result as a label, training the classification model in the 3D network, and obtaining a trained second classification model; the initial weight of the second classification model is a weight parameter of the first classification model; and according to the trained second classification model, predicting the morphological characteristics of the target tooth. According to the method, the accuracy of oral image analysis is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence-assisted medical imaging diagnosis, and in particular to an oral imaging analysis method based on Micro-CT and transfer learning. Background Art

[0002] Endodontic periapical disease is one of the most common diseases in dentistry, and root canal therapy is currently the most effective and commonly used method for treating this type of disease. The clinician's understanding of the root canal morphology plays an important role in the success of root canal therapy. The anatomical structure of the root canal is complex and changeable. Understanding the number, morphology, direction and branching of the root canal can help doctors better remove infections, effectively fill the root canal, and reduce the occurrence of complications during treatment.

[0003] Oral cone beam computed tomography (CBCT) can provide accurate three-dimensional structural information, but the radiation dose is slightly higher and the cost is more expensive. In addition, its imaging quality may be limited by relatively low resolution and artifacts formed by tooth hard tissue and high-radio-blocking materials. In addition, the root canal system has a complex structure and large variability. Therefore, when using CBCT to identify root canal structure images, it is time-consuming and difficult to rely on the analysis of clinicians, resulting in unavoidable missed diagnosis and misdiagnosis. Micro-CT (Micro Computed Tomography, Micro-CT) is the gold standard for morphological research, but due to the small scanning field and long scanning time, it is currently only used for in vitro experimental research. The information provided in Micro-CT images is currently mostly transmitted in the form of conclusive knowledge, and it is difficult to apply this information to personalized individuals. Micro-CT and CBCT each have their own advantages and disadvantages, and it is very difficult to directly connect Micro-CT three-dimensional data with clinical CBCT. Summary of the invention

[0004] The purpose of this application is to provide an oral image analysis method based on Micro-CT and transfer learning, and a method that combines Micro-CT three-dimensional data and CBCT to assist in root canal morphology prediction based on deep learning and transfer learning ideas, so as to improve clinicians' understanding of root canal morphology, prompt key links in clinical operations, and improve the success rate of root canal treatment.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides an oral image analysis method based on Micro-CT and transfer learning, comprising:

[0007] Obtain ex vivo teeth from patients undergoing tooth extraction.

[0008] Based on the Micro-CT scanning technology, the ex vivo tooth is scanned to obtain Micro-CT three-dimensional data of the ex vivo tooth.

[0009] Based on the CBCT scanning technology, the ex-vivo tooth is scanned to obtain CBCT three-dimensional data of the ex-vivo tooth.

[0010] Taking the Micro-CT three-dimensional data of ex vivo teeth as input and the morphological classification results obtained through the Micro-CT three-dimensional data as labels, the classification model is trained in the 3D classification network to obtain the weight parameters of the trained first classification model.

[0011] The CBCT three-dimensional data of the ex vivo tooth is used as input, and the morphological classification results obtained through the Micro-CT three-dimensional data are used as labels, and the classification model is trained in the 3D network to obtain a trained second classification model; the initial weight of the second classification model is the weight parameter of the trained first classification model.

[0012] The morphological features of the target tooth are predicted based on the trained second classification model.

[0013] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0014] The present application proposes an oral image analysis method based on Micro-CT and transfer learning. First, the ex vivo teeth of tooth extraction patients are collected as research samples. Subsequently, the ex vivo teeth are scanned in all directions using high-precision Micro-CT scanning technology to obtain their detailed Micro-CT three-dimensional data. At the same time, the same ex vivo tooth is scanned using CBCT scanning technology to obtain its CBCT three-dimensional data as a basis for comparative analysis. In the data processing and model training stage, the Micro-CT three-dimensional data of the ex vivo tooth is first used as input and input into the 3D classification network. Through the training process, an optimized classification model, namely the first classification model, is obtained, and its weight parameters are extracted as the initial values ​​for subsequent training. Next, the transfer learning strategy is adopted, and the CBCT three-dimensional data of the ex vivo tooth is used as a new input, and the morphological classification results obtained through the Micro-CT three-dimensional data are used as labels. Under the 3D classification network architecture, the weight parameters of the first classification model previously extracted are used as the initial weights to further train the classification model to obtain the second classification model. This step makes full use of the high-precision characteristics of Micro-CT data, and improves the accuracy of CBCT data in morphological analysis through transfer learning. Finally, the trained second classification model is used to predict the morphology of the target tooth. This method not only improves the accuracy of oral image analysis, but also provides strong support for clinical diagnosis and treatment. This application combines the high-precision imaging of Micro-CT with the powerful generalization ability of transfer learning to improve the accuracy of oral image analysis and enable it to better understand the complex features in medical images. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 This is a flowchart of an oral image analysis method based on Micro-CT and transfer learning provided in one embodiment of the present application.

[0017] Figure 2 This is a framework diagram of an oral image analysis method based on Micro-CT and transfer learning provided in one embodiment of the present application.

[0018] Figure 3 A schematic diagram of a 2D data classification model provided in an embodiment of the present application.

[0019] Figure 4A schematic diagram of a 3D data classification model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0021] Deep learning is a cutting-edge research direction in the field of artificial intelligence (AI). By constructing an artificial neural network to combine underlying features to form more abstract high-level features, AI can analyze and diagnose through medical images and provide assistance to clinicians. In recent years, deep learning research based on oral CBCT has mainly focused on two aspects. One is to improve the imaging quality of CBCT through deep learning; the other is to automatically segment, extract features and classify the region of interest (ROI) through deep learning, which is used to analyze and diagnose the physiological or pathological structures presented by oral CBCT, or to plan treatment plans. This type of research is mainly used for large-scale ROI, while deep learning research based on CBCT root canal structure is relatively rare. Existing methods that use deep learning to assist root canal morphology prediction may be limited by the limitations of CBCT data itself. Combining Micro-CT three-dimensional data information with it can provide a richer information perspective for the model. For CBCT data, it is possible to consider directly performing transfer learning on the model trained on the high-precision Micro-CT three-dimensional dataset; thereby improving the accuracy of CBCT prediction of root canal morphology, assisting clinicians in formulating personalized diagnosis and treatment plans for patients, suggesting key links in clinical operations, and improving the success rate of root canal treatment.

[0022] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0023] Embodiment 1

[0024] like Figure 1 As shown, this embodiment provides an oral image analysis method based on Micro-CT and transfer learning, including:

[0025] Step 101: Obtain an ex vivo tooth from a patient undergoing tooth extraction.

[0026] Step 102: Scan the ex vivo tooth based on Micro-CT scanning technology to obtain Micro-CT three-dimensional data of the ex vivo tooth.

[0027] Step 103: Scan the ex-vivo tooth based on CBCT scanning technology to obtain CBCT three-dimensional data of the ex-vivo tooth.

[0028] Step 104: Using the Micro-CT three-dimensional data of the ex vivo tooth as input and the morphological classification results obtained through the Micro-CT three-dimensional data as labels, the classification model is trained in the 3D classification network to obtain weight parameters of the trained first classification model.

[0029] Step 105: Using the CBCT three-dimensional data of the ex vivo tooth as input and the morphological classification results obtained through the Micro-CT three-dimensional data as labels, the classification model is trained in the 3D network to obtain a trained second classification model; the initial weight of the second classification model is the weight parameter of the trained first classification model.

[0030] Step 106: predicting the morphological features of the target tooth based on the trained second classification model.

[0031] Among them, Figure 2 As shown, in this embodiment, when step 101 is executed, the specific steps may be as follows:

[0032] Extracted teeth due to severe periodontitis, orthodontic reduction teeth, impacted teeth and other reasons unrelated to the study were collected from the Oral Medicine Center according to different tooth positions. The extracted teeth were teeth with fully developed roots, no external root resorption or root fracture, no pulp cavity involvement due to tooth hard tissue lesions, and no endodontic treatment.

[0033] In this example, the root canal morphology in the fused root of the mandibular second molar is taken as an example. The teeth collected from patients who needed to be extracted due to severe periodontitis, orthodontic reduction teeth, impacted teeth, etc., which were not related to this experiment, must meet the following inclusion criteria: ① fused roots of mandibular second molars; ② If there is caries, the pulp cavity is not involved; ③ No endodontic treatment has been performed; ④ No root resorption, root caries and root fracture; ⑤ Closed apical foramen. In this experiment, a total of 243 ex vivo fused roots of mandibular second molars were included in this study.

[0034] When executing steps 102-103, the specific steps may be as follows:

[0035] The ex vivo tooth is scanned based on Micro-CT scanning technology, and a geometric transformation method is used to obtain Micro-CT three-dimensional data of the ex vivo tooth after data augmentation.

[0036] The ex-vivo tooth is scanned based on CBCT scanning technology, and the geometric transformation method is used to obtain the CBCT three-dimensional data of the ex-vivo tooth after data augmentation.

[0037] That is, for CBCT and Micro-CT three-dimensional data, the root canal part is selected as a rectangular block in the 3D Slicer software, a mask is drawn, and geometric transformations in traditional image augmentation methods, such as rotation and flipping, are used to augment the three-dimensional data, and all data are adjusted to a size of 110*40*40.

[0038] When executing steps 104-105, the specific steps may be as follows:

[0039] Taking the Micro-CT three-dimensional data of ex vivo teeth as input and the morphological classification results obtained from the Micro-CT three-dimensional data as labels, the classification model was trained in the 3D ResNet18 network pre-trained with the MedicalNet dataset to obtain the weight parameters of the trained first classification model.

[0040] The CBCT three-dimensional data of the ex vivo tooth is used as input, the morphological classification results obtained through the Micro-CT three-dimensional data are used as labels, the weight parameters of the first classification model are used as initial weight parameters, and the classification model is trained in the 3D ResNet18 network to obtain a trained second classification model.

[0041] Specifically, this embodiment uses two different 3D imaging technologies (Micro-CT and CBCT) and deep learning models (specifically on the 3D ResNet18 architecture) to train two classification models to distinguish the root canal morphology of teeth. The following is a detailed explanation:

[0042] Among them, train the first classification model:

[0043] Input: Micro-CT (micro-computed tomography) 3D data of ex vivo teeth. Micro-CT technology provides high-resolution 3D images, which are very suitable for accurate analysis of the internal structure of teeth, such as root canal morphology.

[0044] Tags: Results of morphological classification of teeth in Micro-CT 3D data. These classifications may be based on features such as the number, shape, curvature of the root canals, etc.

[0045] Model: Use a 3D ResNet18 network pre-trained on the MedicalNet dataset. Pre-training means that the network has been preliminarily learned on a large amount of medical image data and can capture common features in medical images. MedicalNet is a model for 3D medical imaging pre-training.

[0046] Training: In the 3D ResNet18 network, Micro-CT three-dimensional data and its corresponding root canal morphology classification results (as labels) are used to train the classification model. The purpose of the training is to obtain the weight parameters of the first classification model.

[0047] Output: The weight parameters of the first classification model after training. These parameters represent the mapping relationship from Micro-CT images to root canal morphology classification learned by the model.

[0048] Among them, when training the second classification model:

[0049] Input: CBCT (cone beam computed tomography) 3D data of ex vivo teeth. CBCT is another commonly used medical imaging technology used in dentistry and other medical fields, but the resolution is usually lower than that of Micro-CT.

[0050] Label: Micro-CT classification results of ex vivo teeth. Here, Micro-CT data is used as the "gold standard" or "true value" to guide the classification of CBCT images.

[0051] Model: 3D ResNet18 network is also used. The initial weight of this model is the weight parameter of the first classification model.

[0052] Training: In the 3D ResNet18 network, the classification model was trained using the CBCT 3D data and its corresponding Micro-CT 3D data classification results (as labels). The purpose of this model is to learn how to predict root canal morphology classification similar to high-resolution Micro-CT images from lower-resolution CBCT images. The Adam optimizer was used during training, with the initial learning rate set to 0.0001 and the regularization parameter of the loss function to 0.001. The classification model was trained for 50 cycles with a batch size of 8. The performance of the model was evaluated using accuracy, precision, recall, F1 score, receiving operating characteristic curves (ROC) and area under the curve (AUC).

[0053] Output: The weight parameters of the trained second classification model. These parameters represent the mapping relationship learned by the model from CBCT images to the classification of root canal morphology similar to Micro-CT images.

[0054] like Figure 4 As shown, this embodiment combines the two imaging technologies of Micro-CT and CBCT to train a classification model capable of identifying tooth root canal morphology based on transfer learning.

[0055] After obtaining the trained first classification model and second classification model, it also includes:

[0056] Based on the data validation set, the trained first classification model and the trained second classification model are validated.

[0057] Wherein, when executing step 106, the predicted results of the morphological characteristics of the target tooth may include root morphological variation, dentin thickness in the danger zone, number of root canals, root canal morphology, root canal length, root canal curvature, root canal calcification and apical foramen diameter.

[0058] In addition, based on the idea of ​​the oral image analysis method based on Micro-CT and transfer learning provided in this application, this application also proposes a method such as Figure 3 The oral image analysis method based on the two-dimensional classification model is shown.

[0059] Among them, Figure 3 As shown, in this embodiment, the root canal morphology in the fused root of the mandibular second molar is taken as an example. The teeth collected from patients who need to be extracted due to severe periodontitis, orthodontic reduction teeth, impacted teeth, etc., which are not related to this experiment, must meet the following inclusion criteria: ① fused roots of mandibular second molars; ② If there is caries, the pulp cavity is not involved; ③ No endodontic treatment has been performed; ④ No root resorption, root caries and root fracture; ⑤ The apical foramen is closed. In this experiment, a total of 243 ex vivo fused roots of mandibular second molars were included in this study.

[0060] In this embodiment, the original teeth on the ex vivo mandibular bone are extracted, the ex vivo fused root type mandibular second molar is placed in the alveolar socket, the ex vivo tooth is fixed with red wax, and multi-angle X-rays are taken to simulate the clinical scene, at several angles including median, mesial rotation, distal rotation, etc., where the rotation angle is 15°.

[0061] Then, the ex vivo tooth is scanned using Micro-CT scanning technology to obtain Micro-CT three-dimensional data of the ex vivo tooth.

[0062] Specifically, after Micro-CT scanning of all ex vivo teeth, the Micro-CT three-dimensional images were visualized using MeVisLab software, and the root canal morphology of all teeth was classified according to the classification method proposed by Gao (2006), specifically as “fused type”, “symmetrical type”, and “asymmetrical type”.

[0063] Then, based on the Micro-CT three-dimensional data, a virtual X-ray slide of the ex vivo tooth is generated through three-dimensional reconstruction technology.

[0064] Specifically, in MeVisLab software, the DRR component is used to generate virtual X-ray projections of ex vivo teeth based on Micro-CT three-dimensional data. After determining the median projection of the tooth, the corresponding mesial and distal projections are obtained by rotating the tooth 15 degrees toward the mesial and distal directions, respectively.

[0065] The multi-angle projection method is used to perform data augmentation on virtual X-ray slides and real X-ray films, and a set of virtual X-ray slides and a set of real X-ray films of ex-vivo teeth are obtained.

[0066] Then, the virtual X-ray film set obtained by Micro-CT was used as input, and the multi-angle projection method was used for data augmentation. The classification model was trained in the ResNet18 network pre-trained with the ImageNet dataset, and the weight parameters with the best training effect were saved; the real X-ray film set simulating clinical scenes was used as input, and the multi-angle projection method was used for data augmentation. The classification model was trained in the ResNet18 network pre-trained with the ImageNet dataset; the real X-ray films simulating clinical scenes were used as input, and the multi-angle projection method was used for data augmentation. The weight parameters trained on the virtual X-ray film set were used as the initial weights for transfer learning, and the classification model was trained in the ResNet18 network to obtain specific root canal morphology categories.

[0067] Among them, Figure 3 As shown, the virtual X-ray projection set is used as input, the classification result obtained based on the Micro-CT data is used as the label of the ex vivo tooth, and the classification model is trained in the ResNet18 network pre-trained with the ImageNet dataset to obtain the first classification model after training, which is as follows:

[0068] The classification model was trained in the ResNet18 network using transfer learning and pre-trained model weights from the ImageNet dataset (ImageNet is a large visualization database for visual object recognition software research). The virtual X-ray film set was used as input. Data augmentation was performed on the images obtained by multi-angle projection method for a small number of categories (asymmetric and fusion types). All X-rays were adjusted to 224*224 size in the image preprocessing stage. The Adam optimizer was used during training, the initial learning rate was set to 0.0002, the learning rate decay method was cosine annealing, and the regularization parameter of the loss function was 0.001. The classification model was trained for a total of 20 cycles with a batch size of 16. The performance of the model was evaluated using accuracy, precision, recall, F1 score, receiving operating characteristic curves (ROC) and area under the curve (AUC), and the weight parameters with the best training effect were saved.

[0069] Methods: The classification model was trained in a ResNet18 network using transfer learning and pre-trained model weights from the ImageNet dataset. Real radiographs simulating clinical scenarios were used as input. Data augmentation was performed on images obtained by multi-angle projection for a small number of categories (asymmetric and fused). All radiographs were resized to 224*224 in the image preprocessing stage. The Adam optimizer was used during training, with an initial learning rate of 0.0002, a cosine annealing method for learning rate decay, and a regularization parameter of 0.001 for the loss function. The classification model was trained for 30 epochs with a batch size of 16. The performance of the model was evaluated using accuracy, precision, recall, F1 score, receiving operating characteristic curves (ROC), and area under the curve (AUC).

[0070] Using transfer learning and model weights trained on virtual X-rays, the classification model was trained in a ResNet18 network using real X-rays that simulated clinical scenarios as input. Data augmentation was performed on images obtained by multi-angle projection for a small number of categories (asymmetric and fused), and all X-rays were resized to 224*224 in the image preprocessing stage. The Adam optimizer was used during training, with an initial learning rate set to 0.0003, a cosine annealing method for learning rate decay, and a regularization parameter of 0.001 for the loss function. The classification model was trained for 20 epochs with a batch size of 16. The performance of the model was evaluated using accuracy, precision, recall, F1 score, receiving operating characteristic curves (ROC), and area under the curve (AUC).

[0071] In this embodiment, taking the root canal morphology recognition of the isolated mandibular second molar fused root as an example, the following experiments were performed, as follows:

[0072] The in vitro dental X-ray images were obtained by micro-computed tomography (Micro-CT) and combined with convolutional neural network (CNN) and transfer learning strategy to improve the accuracy of identifying the root canal morphology of the fused roots of the mandibular second molars from simulated clinical real X-rays.

[0073] In this embodiment, 243 fused root mandibular second molars were included, Micro-CT three-dimensional data were obtained and two-dimensional X-ray projection images were generated, and real X-rays were taken on the ex vivo mandible to simulate the clinical scenario. According to Micro-CT, all root canals were divided into fused type, symmetrical type and asymmetrical type, and multi-angle projection images were used to expand the smaller number of categories. Four CNNs, ResNet18, ResNet50, VGG19 and EfficientNet-b5, were selected for root canal morphology recognition. Four CNNs were trained according to the ImageNet pre-training weights + X-ray clinical simulation films, ImageNet pre-training weights + X-ray projection images, and X-ray projection image weights + X-ray clinical simulation films, and the classification results of CNNs were compared with those of residents.

[0074] Finally, the overall accuracy of CNN in the test set of ImageNet pre-trained weights + X-ray clinical simulation film group was 65.96%, 63.83%, 61.70% and 65.96%; ImageNet pre-trained weights + X-ray projection image group was 79.25%, 75.47%, 73.58% and 77.36%; X-ray projection image weights + X-ray clinical simulation film group was 72.34%, 68.09%, 65.96% and 72.34%. The classification results of X-ray clinical simulation films by residents were 63.83%, 57.45%, 63.83% and 59.57%. The recognition ability of CNN using ImageNet pre-trained weights on X-ray clinical simulation films was comparable to that of residents (P>0.05), and its recognition ability was significantly improved after combining with X-ray projection image weights (P<0.05).

[0075] Therefore, the weight parameters obtained by training on the X-ray projection images generated by Micro-CT are helpful to improve the morphological recognition accuracy in the training process of X-ray clinical simulation films. The strategy of combining different modal images by using X-ray projection images as a bridge between three-dimensional Micro-CT data and clinical two-dimensional X-ray data is feasible.

[0076] In summary, this application has the following technical effects:

[0077] This application proposes an oral image analysis method based on Micro-CT and transfer learning, which uses CBCT clinical image data to predict root canal morphology and assist clinical diagnosis and treatment. In three-dimensional CBCT data, this application proposes to directly transfer learning the weight parameters trained in Micro-CT three-dimensional data, and integrates more accurate Micro-CT three-dimensional data information into the training process of CBCT data, thereby improving the accuracy of morphological prediction of CBCT data.

[0078] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. An oral image analysis method based on Micro-CT and transfer learning, characterized in that: The oral image analysis method based on Micro-CT and transfer learning includes: Obtain ex vivo teeth from patients undergoing tooth extraction; Scanning the ex vivo tooth based on Micro-CT scanning technology to obtain Micro-CT three-dimensional data of the ex vivo tooth; Scanning the ex-vivo tooth based on CBCT scanning technology to obtain CBCT three-dimensional data of the ex-vivo tooth; Using Micro-CT three-dimensional data of ex vivo teeth as input and the morphological classification results obtained through the Micro-CT three-dimensional data as labels, the classification model is trained in the 3D classification network to obtain the weight parameters of the first classification model after training; Taking the CBCT three-dimensional data of the ex vivo tooth as input and the morphological classification results obtained through the Micro-CT three-dimensional data as labels, the classification model is trained in the 3D network to obtain a trained second classification model; the initial weight of the second classification model is the weight parameter of the trained first classification model; The morphological features of the target tooth are predicted based on the trained second classification model.

2. The method for oral image analysis based on Micro-CT and transfer learning according to claim 1, characterized in that: Based on the Micro-CT scanning technology, the ex vivo tooth is scanned to obtain Micro-CT three-dimensional data of the ex vivo tooth, which specifically includes: The ex vivo tooth is scanned based on Micro-CT scanning technology, and a geometric transformation method is used to obtain Micro-CT three-dimensional data of the ex vivo tooth after data augmentation.

3. The method for oral image analysis based on Micro-CT and transfer learning according to claim 1, characterized in that: Based on the CBCT scanning technology, the ex-vivo tooth is scanned to obtain CBCT three-dimensional data of the ex-vivo tooth, specifically including: The ex-vivo tooth is scanned based on CBCT scanning technology, and the geometric transformation method is used to obtain the CBCT three-dimensional data of the ex-vivo tooth after data augmentation.

4. The method for oral image analysis based on Micro-CT and transfer learning according to claim 1, characterized in that: The reasons for extracting the ex vivo teeth include severe periodontitis, orthodontic reduction teeth and impacted teeth.

5. The method for oral image analysis based on Micro-CT and transfer learning according to claim 1, characterized in that: Also includes: Based on the data validation set, the trained first classification model and the trained second classification model are validated.

6. The method for oral image analysis based on Micro-CT and transfer learning according to claim 1, characterized in that: The prediction results of the morphological characteristics of the target tooth include root morphological variation, dentin thickness in the risk zone, root canal number, root canal morphology, root canal length, root canal curvature, root canal calcification and apical foramen diameter.

7. The method for oral image analysis based on Micro-CT and transfer learning according to claim 1, characterized in that: The Micro-CT three-dimensional data of the ex vivo tooth is used as input, and the morphological classification results obtained through the Micro-CT three-dimensional data are used as labels. The classification model is trained in the 3D classification network to obtain the weight parameters of the first classification model after training, which specifically include: Taking the Micro-CT three-dimensional data of ex vivo teeth as input and the morphological classification results obtained from the Micro-CT three-dimensional data as labels, the classification model was trained in the 3D ResNet18 network pre-trained with the MedicalNet dataset to obtain the weight parameters of the trained first classification model.

8. The method for oral image analysis based on Micro-CT and transfer learning according to claim 1, characterized in that: Taking the CBCT three-dimensional data of the ex vivo tooth as input and the morphological classification results obtained through the Micro-CT three-dimensional data as labels, the classification model is trained in the 3D network to obtain a trained second classification model, which specifically includes: The CBCT three-dimensional data of the ex vivo tooth is used as input, the morphological classification results obtained through the Micro-CT three-dimensional data are used as labels, and the weight parameters of the obtained first classification model are used as initial weight parameters. The classification model is trained in the 3D ResNet18 network to obtain a trained second classification model.

Images