Embedded tooth root bone adhesion feature identification method and system based on deep learning

Abstract

The invention discloses an embedded tooth root bone adhesion feature recognition method and system based on deep learning, and relates to the field of medical image recognition, the position of an embedded tooth is detected based on a normalized image, image cutting of the embedded tooth and the surrounding area thereof is carried out along the long axis of a tooth body, the cutting area is a three-dimensional rectangular frame, and the three-dimensional rectangular frame is used as a matrix; classifying the cut images according to FDI tooth positions; performing hierarchical slicing on the classified images along the long axis of the tooth, taking the images after hierarchical slicing as input of an embedded tooth root bone adhesion feature recognition model, extracting depth features of the images in a feature extraction network ResNet50, removing a final full-connection layer of the ResNet50 model, retaining a feature map output by a convolutional layer, and identifying the embedded tooth root bone adhesion feature in the feature extraction network ResNet50; and inputting a feature map output by the convolutional layer into the region suggestion network and the classification and regression head to obtain a feature recognition result, including the probability that the embedded tooth root bone adhesion feature prediction candidate region belongs to the target category and the accurate position of the embedded tooth root bone adhesion feature candidate region bounding box.

Description

Technical Field

[0001] The present invention belongs to the technical field of CBCT examination, and particularly relates to a method and system for identifying the characteristics of impacted tooth root-bone adhesion based on deep learning. Background Art

[0002] Human teeth are normally located on the alveolar bone and are kept stable by cementum, periodontal ligament, and alveolar bone, with a certain degree of physiological mobility. During the process of human evolution, the degeneration of the jawbone and the tooth volume is inconsistent, or due to premature loss of deciduous teeth, retention of deciduous teeth, abnormal position of permanent tooth germs, dysplasia of alveolar bone, etc., the permanent teeth that are still impacted in the jawbone after the tooth eruption period will become impacted teeth. Some patients also have a family genetic tendency, which may be related to gene mutations or genetic defects. At present, the examination and diagnosis of impacted teeth mainly rely on X-ray equipment imaging to clarify the position, size, number, and relationship with surrounding tissues of the impacted teeth. And the treatment methods for impacted teeth mainly include observation for those without discomfort and no interference to adjacent teeth and surrounding tissues, eruption promotion surgery when the position is relatively shallow and there is a potential for eruption, orthodontic traction treatment when the position is deep but there is a value for retention, and extraction when there is no value for retention or there is a serious impact on adjacent teeth; the above treatment plans are all based on the correct diagnosis of impacted teeth. If tooth adhesion exists but is not detected, it will lead to an extension of the unnecessary treatment cycle and delay the correct treatment time. Therefore, how to correctly and quickly judge whether there is adhesion of impacted teeth has become an urgent problem to be solved in orthodontic clinics.

[0003] Cone-beam computed tomography, i.e., CBCT imaging examination, uses cone-beam X-rays to scan an object, obtains cross-sectional image data from multiple angles, and then uses a computer reconstruction algorithm to generate a three-dimensional model. It has the advantages of high resolution, low radiation dose, fast scanning speed, and three-dimensional imaging, and is widely used in oral clinical auxiliary examinations. At present, with the rapid development of artificial intelligence, many scholars have combined deep learning with oral medicine and achieved many remarkable results. For example, Cui et al. developed an artificial intelligence system based on deep learning, which has good clinical stability and accuracy and can be used for fully automatic segmentation of teeth and alveolar bone from CBCT images, especially for tooth segmentation. Su et al. proposed an automatic segmentation of the periodontal ligament in cone-beam computed tomography based on an instance segmentation network, realizing an end-to-end fully automatic segmentation process. Compared with existing segmentation methods, this method can obtain higher segmentation accuracy and segmentation efficiency. It shows that deep learning has great development prospects in oral medicine. However, existing solutions such as the Chinese patent with the publication number CN118781134A, a method for fully automatic three-dimensional tooth segmentation based on deep learning, uses a multi-layer deep learning model to solve the problem of three-dimensional instance segmentation of a single tooth, and pays more attention to how to use deep learning to completely extract tooth roots, while ignoring the specific relationship between the root and the bone. There is still room for development in judging whether the tooth roots are adhered.

[0004] At present, there are problems in the orthodontic treatment of impacted teeth based on traditional methods, such as long traction time and uncertain curative effect of whether the impacted teeth can be pulled out in place. The traditional method for doctors to judge whether there is root-bone adhesion using CBCT images has problems such as high dependence on doctors' experience, long time consumption, and easy missed diagnosis. Summary of the Invention

[0005] In order to solve the problems existing in the prior art, the present invention provides a method for identifying the characteristics of root-bone adhesion of impacted teeth based on deep learning, which can assist in identifying the possible adhesion between the tooth root and the alveolar bone and obtain the classification result of the tooth root and alveolar bone.

[0006] In order to achieve the above object, in the first aspect, the present invention provides a method for identifying the characteristics of root-bone adhesion of impacted teeth based on deep learning, including the following steps:

[0007] Detect the position of the impacted tooth based on the normalized image, perform image cropping on the impacted tooth and its surrounding area along the long axis of the tooth, the cropping area is a three-dimensional rectangular frame, and classify the cropped image according to the FDI tooth position;

[0008] Perform layer-by-layer slicing on the classified images along the long axis of the tooth, use the sliced images as the input of the root-bone adhesion feature recognition model of the impacted tooth, extract the deep features of the images in the feature extraction network ResNet50, remove the last fully connected layer of the ResNet50 model, retain the feature map output by the convolutional layer, and input the feature map output by the convolutional layer into the region proposal network and the classification and regression head to obtain the feature recognition result, including the probability that the candidate region of the root-bone adhesion feature of the impacted tooth belongs to the target category and the exact position of the bounding box of the candidate region of the root-bone adhesion feature of the impacted tooth.

[0009] Further, obtaining the normalized image includes:

[0010] Obtain the CBCT data of the human maxillofacial region,

[0011] Normalize the pixel values of the obtained original CBCT image so that the pixel values are within [0,1].

[0012] Further, the root-bone adhesion feature recognition model of the impacted tooth adopts an improved Faster R-CNN model, and the specific training includes the following steps:

[0013] Detect the position of the impacted tooth based on the normalized image, perform image cropping on the impacted tooth and its surrounding area along the long axis of the tooth, the cropping area is a three-dimensional rectangular frame, and classify the cropped image according to the FDI tooth position;

[0014] The classified images are sliced layer by layer along the long axis of the tooth. After all the sliced layers are obtained, a total dataset is formed. The total dataset is divided into a training set, a validation set, and a test set. Then, in the order from the root tip to the incisal edge, each layer of images is manually annotated. Two experts respectively annotate each normalized layer of images with two types: normal root-bone relationship and adhesion root-bone relationship. If there are inconsistent annotations, the two experts discuss until the final result is consistent. Finally, an experienced expert reviews the annotation results to obtain the training set of target teeth at different tooth positions;

[0015] Use the pre-trained feature extraction network ResNet50 model as the feature extractor of Faster R-CNN. The feature extraction network ResNet50 is implemented by loading the pre-trained weights. Based on the feature maps extracted by ResNet50, a region proposal network is constructed, and candidate regions are generated based on the region proposal network;

[0016] Perform non-maximum suppression on the generated candidate regions to remove the candidate regions with too high overlap. The classification and regression heads convert the feature maps of the candidate regions after non-maximum suppression into fixed-size feature vectors through the RoIPooling operation, and connect two fully connected layers. One fully connected layer is used to predict the probability that the candidate region of the impacted tooth root-bone adhesion feature belongs to the target class, and the other fully connected layer is used to predict the exact position of the bounding box of the candidate region of the impacted tooth root-bone adhesion feature.

[0017] Furthermore, the manual annotation includes the tooth edge and the alveolar bone edge of each layer of images; before the manual annotation, the size of each picture is normalized.

[0018] Furthermore, the region proposal network consists of a convolutional layer and is divided into two parts: one part is used to predict the score of the target at each position, that is, whether there is adhesion can be judged according to the score, and the other part is used to predict the regression offset of the adhesion bounding box.

[0019] Furthermore, the loss of the impacted tooth root-bone adhesion feature recognition model mainly includes the loss of the region proposal network and the loss of Fast R-CNN. Both parts of the loss include classification loss and regression loss; the total loss function is:

[0020]

[0021] Among them, The classification loss function, N cls is the number of anchors p selected during the training of RPN i is the probability that the anchor is predicted as the adhesion root-bone relationship, GT label: Among them is the logarithmic loss of two categories and is a cross-entropy loss function for classic binary classification In the Fast RCNN classification loss function, the function formula is the same as that of the RPN classification loss function, and N cls becomes the number of ROIs selected by the RPN network is the regression loss function takes the value of λ = 1, and N reg = the total number of positive and negative samples t i ={t x , t y , t w , t h} is a vector representing the anchor, and predicts the offset of the bounding box during the RPN training stage and the Fast-R-CNN stage is the same dimension vector as t i representing the anchor, which is the actual offset relative to the GT during the RPN training stage and the Fast-R-CNN stage

[0022] Furthermore, it also includes introducing the openCV library to the model, extracting information such as the coordinates, class labels, and probabilities of the adhered bounding boxes in the output results, and then using OpenCV to draw the bounding boxes and add class labels on the image

[0023] In a second aspect, the present invention provides a deep learning-based impacted tooth root and bone adhesion feature recognition system, including:

[0024] Data acquisition module: Detect the position of the impacted tooth based on the normalized image, crop the image of the impacted tooth and its surrounding area along the long axis of the tooth, the cropped area is a three-dimensional rectangular box, and classify the cropped image according to the FDI tooth position

[0025] Recognition module: Perform layered slicing of the classified images along the long axis of the tooth respectively, use the sliced images as the input of the impacted tooth root and bone adhesion feature recognition model, extract the deep features of the images in the feature extraction network ResNet50, remove the last fully connected layer of the ResNet50 model, retain the feature map output by the convolutional layer, and input the feature map output by the convolutional layer into the region proposal network and the classification and regression head to obtain the feature recognition results, including the probability that the impacted tooth root and bone adhesion feature prediction candidate region belongs to the target category and the precise position of the bounding box of the impacted tooth root and bone adhesion feature candidate region

[0026] In a third aspect, the present invention can also provide a computer device, including a processor and a memory. The memory is used to store computer-executable programs. The processor reads the computer-executable programs from the memory and executes them. When the processor executes the computer-executable programs, it can implement the method for identifying the features of impacted tooth root and alveolar bone adhesion based on deep learning according to the present invention.

[0027] Meanwhile, a computer-readable storage medium is provided. A computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, it can implement the method for identifying the features of impacted tooth root and alveolar bone adhesion based on deep learning according to the present invention.

[0028] Compared with the prior art, the present invention has at least the following beneficial effects: The present invention aims to empower the field of medical image detection and diagnosis through deep learning methods, accelerate the process of image recognition and auxiliary diagnosis, and effectively avoid the missed diagnosis caused by the adhesion of the tooth tissue of impacted teeth and the alveolar bone tissue, as well as the misdiagnosis caused by the lack of relevant clinical or imaging examination experience and the corresponding problem of unnecessary extended treatment cycle.

[0029] Furthermore, based on manual annotation, the mutual relationship between the tooth root and the alveolar bone is accurately annotated and classified. First, FDI tooth position classification is performed, and then the images of a single tooth and its surrounding area are sliced along the long axis of the tooth. The root-bone relationship is annotated for each layer of the image to obtain a training set. The model for identifying the features of impacted tooth root and alveolar bone adhesion trained with the training set has a good detection rate for whether the impacted tooth is adhered, thus assisting doctors in making rapid clinical decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is the structural and working flow chart of a CBCT examination and diagnosis system for impacted tooth root and alveolar bone adhesion based on deep learning;

[0031] Figure 2 is a schematic structural diagram of an impacted tooth and its surrounding area;

[0032] Figure 3 is a schematic diagram of the root slice annotation of this system. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be further described in detail below with reference to the drawings and specific embodiments. It should be noted that, without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0034] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described within the scope hereof. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0035] To diagnose whether there is adhesion between the tooth root and the alveolar bone, a comprehensive judgment should be made in combination with the tooth root and the alveolar bone. In the traditional tooth root and alveolar bone classification model, more attention is paid to the complete structure of the tooth root, while the possible adhesion between the tooth root and the alveolar bone is ignored, resulting in its inability to be directly applied to the diagnosis of whether there is adhesion between the tooth root and the alveolar bone. In this regard, when training the recognition model for the characteristics of impacted tooth root and bone adhesion in this application, the mutual relationship between the tooth root and the alveolar bone is accurately labeled and classified based on manual annotation. First, the FDI tooth position classification is carried out, and then the images of a single tooth and its surrounding area are sliced along the long axis of the tooth body. The root-bone relationship of each layer of the image is labeled to obtain the training set. The recognition model for the characteristics of impacted tooth root and bone adhesion trained with the training set has a good detection rate for whether the impacted tooth is adhered, thus assisting doctors in making rapid clinical decisions.

[0036] Please refer to Figure 1 , Figure 1 The present invention provides a CBCT diagnosis system for impacted tooth root and bone adhesion based on deep learning. The system includes:

[0037] A data acquisition module for acquiring CBCT data of the teeth in the human maxillofacial region;

[0038] An artificial annotation module for obtaining images of a single tooth and its surrounding alveolar bone based on the CBCT images of the human maxillofacial region, and performing FDI classification annotation of the tooth position; slicing the images of a single tooth and its surrounding area along the long axis of the tooth body, and three experts perform two kinds of annotations, namely normal root-bone relationship and adhesion type root-bone relationship, for each layer of the image to obtain the training set;

[0039] An inspection and diagnosis module for training the recognition model for the characteristics of impacted tooth root and bone adhesion based on the above-mentioned training set to obtain the recognition model for the characteristics of impacted tooth root and bone adhesion, and using the recognition model for the characteristics of impacted tooth root and bone adhesion to classify and identify the teeth in the test set to obtain the inspection and diagnosis results.

[0040] Among them, the CBCT data of the human maxillofacial region is directly obtained from the CBCT examination equipment.

[0041] Please refer to Figure 2 , Figure 2 which is a schematic diagram of the structure of the impacted tooth and its surrounding area, Figure 2 in which ① is the alveolar bone, ② is the gap between the alveolar bone and the dental tissue, ③ is the enamel, ④ is the dentin, ⑤ is the pulp, and ⑥ is the cementum.

[0042] The artificial annotation for image processing is carried out based on the CBCT image data. In some embodiments, the artificial image processing specifically includes:

[0043] Image normalization processing: Normalize the pixel values of the acquired original CBCT images so that the pixel values fall within the set interval [0, 1], which helps the model better learn features and avoid training difficulties caused by excessive differences in pixel values between different layers of images.

[0044] Based on the normalized images, detect the positions of impacted teeth, and perform image cropping on the impacted teeth and their surrounding areas along the long axis of the tooth. The cropping area is a three-dimensional rectangular box, including all dental tissues and the image of the 1-mm area outside the dental tissues. Classify the cropped image according to the FDI tooth position.

[0045] Perform layered slicing along the long axis of the tooth for the classified images. The thickness of each slice is 1 mm. After all the layered slicing, obtain the total dataset. The total dataset is divided into a 70% training set, a 15% validation set, and a 15% test set. And in the order from the root tip to the incisal edge, manually annotate each layer of images. The manual annotation includes the dental tissue edge and the alveolar bone edge of each layer of images. Before manual annotation, use software such as PhotoScape to normalize the size of each picture. Two experts respectively perform two types of annotations, namely normal root-bone relationship and adherent root-bone relationship, on the normalized images of each layer. If there are inconsistent annotation situations, the two experts discuss until the final results are consistent. Finally, an experienced expert reviews the annotation results to obtain the training set of the target teeth in different tooth positions.

[0046] Please refer to Figure 3 , Figure 3 This is a schematic diagram for root slice annotation of this system, including the normal root-bone relationship (left) and the adherent root-bone relationship (right).

[0047] In some embodiments, the network model adopts an improved Faster R-CNN model.

[0048] Among them, first load the pre-trained model: The pre-trained feature extraction network ResNet50 model can be used as the feature extractor of Faster R-CNN. The feature extraction network ResNet50 can be realized by loading the pre-trained weights. ResNet50 consists of multiple convolutional layers and residual blocks, and its structure can effectively extract the deep features of images. Then remove the last fully connected layer of the ResNet50 model and retain the feature map output by its convolutional layer. The feature map output by the convolutional layer will be used as the input of the subsequent Region Proposal Network (RPN) and the classification and regression head. The feature map output by the convolutional layer contains rich semantic information and spatial information of the image.

[0049] Region Proposal Network (RPN): Build the RPN on the feature maps extracted by ResNet50. The RPN consists of a convolutional layer to generate candidate regions. The region proposal network is divided into two parts: one part is used to predict the score of the target at each position, that is, it can judge whether there is adhesion according to the score, and the other part is used to predict the regression offset of the adhesion bounding box.

[0050] Perform non-maximum suppression (NMS) on the candidate regions generated by the RPN to remove candidate regions with too high overlap, and retain a set number of high-quality candidate regions, such as 2000, to reduce the subsequent calculation amount and improve the detection accuracy.

[0051] Classification and regression head: For each candidate region after NMS processing, convert its feature map into a feature vector of a fixed size (7×7) through the RoIPooling operation. RoIPooling will perform a pooling operation according to the position and size of the candidate region on the feature map, converting the feature maps of candidate regions of different sizes into a unified size for subsequent fully connected layer processing.

[0052] Connect two fully connected layers. One fully connected layer is used to predict the probability that the candidate region of the impacted tooth root and alveolar bone adhesion feature belongs to the target category (adhesive root and alveolar bone relationship), and the number of output channels is 1. The other fully connected layer is used to predict the exact position of the bounding box of the candidate region of the impacted tooth root and alveolar bone adhesion feature, and the number of output channels is 4.

[0053] Among them, for the bounding box of the candidate region of the impacted tooth root and alveolar bone adhesion feature, a regression loss function is used, which is mainly used to adjust the position of the prediction box so that the prediction box can more accurately locate the region bounding box of the tooth body and alveolar bone adhesion in a single-layer image.

[0054] The loss of Faster-R-CNN is mainly divided into the loss of the RPN and the loss of Fast R-CNN. The calculation formula is as follows, and both parts of the loss include classification loss and regression loss.

[0055] Total loss function:

[0056] The classification loss function for the RPN is N cls is the number of anchors p selected during the training of the RPN i is the probability that the anchor is predicted to be an adhesive root and alveolar bone relationship, GT label: Among them is the logarithmic loss of two categories (adhesive type and non-adhesive type), which is a classic binary cross-entropy loss function In the Fast RCNN classification loss function, the function formula is the same as that of the RPN classification loss function, but N clsBecomes the number of ROIs selected by the RPN network.

[0057] The regression loss function is To make the weights of the classification loss function and the regression loss function basically the same in order to obtain a better model representation, The value of λ is 1, N reg = the total number of positive and negative samples, t i = {t x , t y , t w , t h} is a vector representing the anchor, and predicts the offset of the bounding box during the RPN training stage and the Fast-R-CNN stage. is the same dimension vector as t i representing the anchor, which is the actual offset relative to the GT during the RPN training stage and the Fast-R-CNN stage. Among them, in some embodiments, σ = 3 during RPN training, and σ = 1 during the subsequent Fast-R-CNN processing stage.

[0058] Use the Adam optimization algorithm to update the model weights and finally make the loss converge.

[0059] Finally, introduce the openCV library to visualize the output results of the model: extract information such as the coordinates of the adhesive bounding box, class label, and probability in the output results, and then use OpenCV to draw the bounding box and add the class label on the image:

[0060] By setting the epoch, the complete dataset is passed through the neural network multiple times. According to the characteristics of the training dataset and the monitored learning curve, finally set the epoch to 40 to make the model go from underfitting to optimal without overfitting.

[0061] Embodiment 2, the present invention provides a deep learning-based impacted tooth root bone adhesion feature recognition system, including:

[0062] Data acquisition module: Detect the position of the impacted tooth based on the normalized image, crop the image of the impacted tooth and its surrounding area along the long axis of the tooth, the cropping area is a three-dimensional rectangular box, and classify the cropped image according to the FDI tooth position;

[0063] Recognition module: For the classified images, perform layered slicing along the long axis of the tooth body respectively. Take the images after layered slicing as the input of the impacted tooth root-bone adhesion feature recognition model. Extract the deep features of the images in the feature extraction network ResNet50. Remove the last fully connected layer of the ResNet50 model, and retain the feature map output by the convolutional layer. Input the feature map output by the convolutional layer into the region proposal network and the classification and regression head to obtain the feature recognition results, including the probability that the impacted tooth root-bone adhesion feature prediction candidate region belongs to the target category and the exact position of the bounding box of the impacted tooth root-bone adhesion feature candidate region.

[0064] It also includes a data acquisition module for acquiring CBCT data of the teeth in the human maxillofacial region;

[0065] Manual annotation module: Based on the CBCT images of the human maxillofacial region, obtain the region of interest and perform FDI classification annotation for the tooth positions; perform layered slicing of the images of a single tooth and its surrounding area along the long axis of the tooth body. Three experts perform two types of annotations, normal root-bone relationship and adhesion root-bone relationship, for each layer of images to obtain the training set.

[0066] Model training module: Based on the normalized images, detect the position of the impacted tooth, and perform image cropping of the impacted tooth and its surrounding area along the long axis of the tooth. The cropping area is a three-dimensional rectangular box. Classify the cropped image according to the FDI tooth position;

[0067] For the classified images, perform layered slicing along the long axis of the tooth body respectively. After all the layered slicing, obtain the total data set. The total data set is divided into a training set, a validation set, and a test set. And in the order from the root tip to the incisal edge, perform manual annotation for each layer of images. Two experts perform two types of annotations, normal root-bone relationship and adhesion root-bone relationship, for each layer of the normalized images respectively. If there is a situation of inconsistent annotations, the two experts discuss until the final result is consistent. Finally, an experienced expert reviews the annotation results to obtain the training set of the target teeth in different tooth positions;

[0068] Use the pre-trained feature extraction network ResNet50 model as the feature extractor of Faster R-CNN. Implement the feature extraction network ResNet50 by loading the pre-trained weights. Build a region proposal network based on the feature map extracted by ResNet50, and generate candidate regions based on the region proposal network;

[0069] Perform non-maximum suppression on the generated candidate regions to remove the candidate regions with too high overlap. For the candidate regions after non-maximum suppression, the classification and regression head converts the feature map of the candidate regions into a feature vector of a fixed size through the RoIPooling operation, and connects two fully connected layers. One fully connected layer is used to predict the probability that the candidate region of the impacted root bone adhesion feature belongs to the target class, and the other fully connected layer is used to predict the exact position of the bounding box of the candidate region of the impacted root bone adhesion feature.

[0070] On the other hand, the present invention provides a computer-readable storage medium storing a computer program, which when executed by a processor, can implement the method for identifying impacted root bone adhesion features based on deep learning according to the present invention.

[0071] The present invention can also provide a computer device, including a processor and a memory. The memory is used to store computer-executable programs. The processor reads and executes the computer-executable programs from the memory, and when the processor executes the computer-executable programs, it can implement the method for identifying impacted root bone adhesion features based on deep learning according to the present invention.

[0072] The computer device can be a laptop computer, a desktop computer or a workstation.

[0073] The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC) or a field-programmable gate array (FPGA).

[0074] For the memory of the present invention, it can be an internal storage unit of a laptop computer, a desktop computer or a workstation, such as a memory or a hard disk; it can also use an external storage unit, such as a mobile hard disk or a flash card.

[0075] Computer-readable storage media may include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media may include: read-only memory (ROM), random access memory (RAM), solid state drives (SSD), or optical discs, etc. Among them, random access memory may include resistive random access memory (ReRAM) and dynamic random access memory (DRAM).

[0076] The present invention is described by way of examples. For those skilled in the art, various changes or equivalent replacements can be made to these features and examples without departing from the spirit and scope of the present invention. Additionally, under the teachings of the present invention, these features and examples can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the protection scope of the present invention.

Claims

1. A method for identifying the features of impacted tooth root bone adhesion based on deep learning, characterized in that: The following steps are involved: The position of the impacted tooth is detected based on the normalized image, and the image of the impacted tooth and its surrounding area is cropped along the long axis of the tooth. The cropped area is a three-dimensional rectangular frame, and the cropped image is classified according to the FDI tooth position. The classified images were sliced ​​along the long axis of the tooth and used as the input of the buried root bone adhesion feature recognition model. The deep features of the image were extracted in the feature extraction network ResNet50, and the last fully connected layer of the ResNet50 model was removed. The feature map output by the convolutional layer was retained, and the feature map output by the convolutional layer was input into the region proposal network and the classification and regression head to obtain the feature recognition results, including the probability that the candidate area for the buried root bone adhesion feature prediction belongs to the target category and the precise position of the bounding box of the candidate area for the buried root bone adhesion feature.

2. The method for identifying the features of impacted tooth root bone adhesion based on deep learning according to claim 1, characterized in that: Obtaining the normalized image includes: Acquire human maxillofacial CBCT data, The pixel values ​​of the original CBCT images are normalized so that the pixel values ​​are within [0, 1].

3. The method for identifying the features of impacted tooth root bone adhesion based on deep learning according to claim 1, characterized in that: The impacted tooth root bone adhesion feature recognition model uses the improved Faster R-CNN model. The specific training includes the following steps: Based on the normalized image, the position of the impacted tooth is detected, and the image of the impacted tooth and its surrounding area is cropped along the long axis of the tooth. The cropped area is a three-dimensional rectangular frame, and the cropped image is classified according to the FDI tooth position; The classified images were sliced ​​along the long axis of the tooth, and the total data set was obtained after all the slices were sliced. The total data set was divided into training set, validation set and test set, and each layer of the image was manually annotated in the order from the root apex to the incisal end. Two experts annotated each layer of the normalized image for normal root-bone relationship and adhesion root-bone relationship. If there was any inconsistency in the annotation, the two experts discussed until the final result was consistent. Finally, an experienced expert reviewed the annotation results to obtain the training set of target teeth in different tooth positions. Use the pre-trained feature extraction network ResNet50 model as the feature extractor of Faster R-CNN, implement the feature extraction network ResNet50 by loading the pre-trained weights, build a region proposal network based on the feature map extracted by ResNet50, and generate candidate regions based on the region proposal network; The generated candidate regions are processed with non-maximum suppression to remove candidate regions with too high overlap. The classification and regression heads convert the feature maps of the candidate regions processed with non-maximum suppression into feature vectors of fixed sizes through RoIPooling operations. Two fully connected layers are connected. One fully connected layer is used to predict the probability that the candidate region of the buried root bone adhesion feature belongs to the target category, and the other fully connected layer is used to predict the precise position of the bounding box of the candidate region of the buried root bone adhesion feature.

4. The method for identifying the features of impacted tooth root bone adhesion based on deep learning according to claim 3, characterized in that: Manual annotation includes the tooth margin and alveolar bone margin of each layer of image; The size of each image is normalized before manual annotation.

5. The method for identifying the features of impacted tooth root bone adhesion based on deep learning according to claim 3, characterized in that: The region proposal network consists of a convolutional layer and is divided into two parts: one part is used to predict the score of the target at each position, that is, the score can be used to determine whether there is adhesion, and the other part is used to predict the regression offset of the adhesion bounding box.

6. The method for identifying the features of impacted tooth root bone adhesion based on deep learning according to claim 3, characterized in that: The loss of the impacted tooth root bone adhesion feature recognition model mainly includes the loss of the region proposal network and the loss of Fast R-CNN. Both losses include classification loss and regression loss. The total loss function is: in, The classification loss function, N cls The number of anchors p selected during RPN training i The probability of anchor prediction being an adhesion-type root-bone relationship, GT label: in It is the logarithmic loss of two categories, which is a classic cross entropy loss function for two categories. In the Fast RCNN classification loss function, the function formula is the same as the RPN classification loss function, N cls becomes the number of ROIs selected by the RPN network, is the regression loss function, The value of is λ=1,N reg = the total number of positive and negative samples, t i ={t x ,t y ,t w ,t h } is a vector representing the anchor, which predicts the offset of the bounding box during the RPN training phase and the Fast-R-CNN phase. is with t i A vector of the same dimension represents the actual offset of the anchor relative to the GT in the RPN training stage and the Fast-R-CNN stage.

7. The method for identifying the features of impacted tooth root bone adhesion based on deep learning according to claim 1, characterized in that: It also includes the introduction of the openCV library to the model, extracting information such as the adhesion bounding box coordinates, category labels, and probabilities in the output results, and then using OpenCV to draw bounding boxes and add category labels on the image.

8. A deep learning-based feature recognition system for impacted tooth root bone adhesion, characterized in that: include: Data acquisition module: Detect the position of the impacted tooth based on the normalized image, crop the image of the impacted tooth and its surrounding area along the long axis of the tooth, the cropped area is a three-dimensional rectangular frame, and the cropped image is classified according to the FDI tooth position; Recognition module: The classified images are sliced ​​along the long axis of the tooth, and the sliced ​​images are used as the input of the buried root bone adhesion feature recognition model. The deep features of the image are extracted in the feature extraction network ResNet50, the last fully connected layer of the ResNet50 model is removed, the feature map output by the convolutional layer is retained, and the feature map output by the convolutional layer is input into the region proposal network and the classification and regression head to obtain the feature recognition results, including the probability that the buried root bone adhesion feature prediction candidate area belongs to the target category and the precise position of the bounding box of the buried root bone adhesion feature candidate area.

9. A computer device, characterized in that: It includes a processor and a memory, the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes it, and when the processor executes part or all of the computer executable program, it can implement the deep learning-based method for identifying the features of impacted tooth root bone adhesions as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored in a computer-readable storage medium, and when the computer program is executed by a processor, it can implement the method for identifying the features of impacted tooth root bone adhesion based on deep learning as described in any one of claims 1 to 7.

Images