Age prediction method and system based on oral panoramic X-rays

Abstract

The present invention relates to the field of artificial intelligence technology, and discloses an age prediction method and system based on oral panoramic X-rays. The oral panoramic X-ray and the relatively fixed spatial coordinate information of each permanent tooth are used in an optimized YoloV8 to obtain the developmental stage data of each permanent tooth. The developmental stage data of all permanent teeth are digitized and added to obtain the sum of the developmental stages of all permanent teeth (Sum). The oral panoramic X-ray and Sum are then passed together to a classification model TANet to obtain an inference result of the individual's age. Compared to existing age prediction technologies, the use of deep learning's learning capabilities to classify and predict age avoids heavy reliance on Demirjian's judgment criteria and reduces the systematic errors caused by Demirjian's judgment criteria. Compared with the traditional model's huge demand for data volume, the TANet model is more suitable for the age classification task of oral panoramic X-rays. This problem is alleviated by using the pre-trained model vit_b16 and introducing permanent tooth development stage-assisted reasoning.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an age prediction method and system based on oral panoramic X-rays. Background Art

[0002] The development of permanent teeth (adult teeth) is a long process, usually beginning in childhood and continuing into early adolescence. It includes the formation of dental follicles, calcification of crowns, root formation and lengthening, tooth eruption, continued root growth, and gradual closure of the apical foramen. While the development time of different permanent teeth varies, each tooth takes approximately 10 years to develop. During childhood and adolescence, permanent teeth are still developing, and determining their developmental stage is crucial for formulating treatment plans in oral clinical practice. For example, when dental pulp or periapical inflammation is caused by tooth decay or trauma, different treatment plans are selected for teeth at different developmental stages, and the prognosis for the affected teeth is also different. The development levels of permanent teeth in children and adolescents vary among different races, regions, and periods. Establishing a mathematical model of the relationship between permanent tooth development and natural age at a certain time and place can reveal the characteristics of tooth development in children and adolescents in that region from an anthropological perspective; it can help infer an individual's age from a forensic perspective; and from a clinical medical perspective, it can help determine whether an individual's development is normal and whether it still has growth potential. At the same time, it can provide technical support and guarantees for health management departments to analyze the overall development trend of permanent teeth in the research population and adjust public health policies based on research results.

[0003] Manual identification of the permanent tooth development stage based on panoramic oral X-rays was performed using criteria and methods proposed by Nolla and Demirjian in 1952 and 1973, respectively. These are currently the two most commonly used methods in dentistry. Limited by the image quality of panoramic oral X-rays, the influence and interference of alveolar bone on dental imaging, and the limitations of image processing technology, the industry has long relied on the methods proposed by Nolla and Demirjian to manually analyze and determine the permanent tooth development stage in children and adolescents. However, manual analysis and determination are extremely inefficient and cannot be used for large-scale analytical research. With the advancement of artificial intelligence technology, especially the explosive development of deep learning models, research on the use of artificial intelligence technology to evaluate the development stage of permanent teeth has begun to emerge in recent years. In 2022, Santosh KC proposed a machine learning method to analyze oral panoramic X-rays to determine human age and gender. In 2024, Sivri MB compared the effects of convolutional neural networks and artificial Demirjian methods in judging the development stage of permanent teeth. In the same year, Shi Y proposed a deep learning method for fully automatic dental age estimation based on bone tomography. This method first uses the YOLOv3 detection network to identify the permanent tooth type, and then uses the SOS-Net model to evaluate the development stage of permanent teeth.

[0004] Currently, the existing age prediction schemes based on oral panoramic X-rays mainly include:

[0005] (1) Milošević D, Vodanović M, Galić I et al. proposed a deep learning-based automatic age estimation technology for panoramic dental X-ray images in their paper “Automated estimation of chronological age from panoramic dental X-ray images using deep learning” published in 2022. 4035 panoramic dental X-ray images of male and female subjects aged 19 to 90 were collected. These images were unfiltered and contained various dental lesions and restorations. Some of the data were annotated with dental status, including fillings, caries, bridges, implants, retainers, orthodontic appliances, missing teeth, or crowns to assist in inferring age. A pre-trained convolutional neural network (CNN) was used as a feature extractor, and multiple architectures were tested, such as DenseNet201, InceptionResNetV2, ResNet50, VGG16, VGG19, and Xception. An optional attention mechanism was used to enhance the model's focus on key features in the image. Compared with traditional methods, deep learning models have improved in automation and performance, especially when dealing with tooth lesions and loss. In order to determine which areas contain useful age indicators, three variants of the oral panoramic X-ray image model were trained. One model was trained on images containing only teeth, with the surrounding bone structure blocked. Another model was trained on images in which the central area containing the teeth was blocked, retaining the surrounding bone structure. The third model is an improvement on the tooth blocking variant. Not only is the area where the teeth are located blocked, but each tooth is blocked individually to ensure that no tooth roots or strange teeth appear in the image. However, this scheme does not pay enough attention to the information of tooth position. Teeth in different positions have different characteristics. Expanding the dimension of tooth position information can improve the accuracy of age inference in the vertical direction. Secondly, the deep learning model used to extract features is relatively simple, and it is difficult to capture enough feature information during the feature extraction process, thereby reducing the inference accuracy.

[0006] (2) A deep learning method based on fully automated dental age estimation proposed by Yuchao Shi, Zelin Ye et al. in "Deep learning methods for fully automated dental age estimation on orthopantomograms". A fully automated dental age estimation framework is proposed, aiming to estimate the dental development age of children aged 3 to 15 years old using panoramic dental radiographs (OPG) images. The framework consists of three main steps: tooth positioning and numbering, dental development staging, and dental age assessment based on a reference dataset. First, the first step is to complete the tooth positioning and numbering tasks through the YOLOv3 network. YOLOv3 is a detection network based on Darknet-53, which is used to extract the bounding box of each numbered permanent tooth on the digital panoramic dental radiograph. The second step is to use the designed SOS-Net (Symmetric and Ordinal Staging Network) network to perform accurate dental development staging based on the modified Demirjian method. The network uses ResNext50 as a feature extractor and designs a symmetric feature combination block (SFC Block) to adaptively combine the feature maps of paired symmetrical teeth to enhance the distinguishing information. An auxiliary regression branch and a new loss function, the adjacent stage perception loss (ASA loss), were introduced to reduce the misclassification rate of adjacent stages. Step 3: Based on the previously created reference dataset, a single-group meta-analysis was performed to calculate the weighted average of the mean age of each subject, using the standard error as a weighting factor for the random defect model. The overall mean age calculated from this analysis was determined as the subject's dental age. However, the efficiency of tooth segmentation and position encoding in this scheme is limited because the X-ray images of children's / adolescent teeth are closely squeezed or occluded by the unerupted permanent teeth and deciduous teeth, posing a considerable challenge to detection efficiency. At the same time, this squeezing or occlusion causes errors in position prediction in the second step. Target detection and developmental stage prediction are then performed in segments.

[0007] (3) ViT, a medical image classification model proposed in the paper “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale” by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. ViT is a general image classification model based on the Transformer. Traditional convolutional design relies on local receptive fields and convolution operations, which may limit the capture of global contextual information. Transformers have achieved great success in the field of natural language processing (NLP), especially their self-attention mechanism that can capture long-range dependencies. The authors hope to introduce Transformers into the field of computer vision and explore their potential. When pre-trained on large-scale datasets (such as JFT-300M), ViT surpassed the current CNN models (such as ResNet) on benchmark datasets such as ImageNet. However, as a general classification model, ViT performs better than the convolutional model on large-scale data, but worse than the convolutional model on small-scale data. In addition, the classification model has low accuracy when classifying age using panoramic dental X-ray images. This is because it is difficult to classify panoramic dental X-ray images even manually, as it is difficult to form obvious semantic features in the images for identification.

[0008] Therefore, how to solve the strong dependence of the current automatic age estimation technology based on oral panoramic X-rays on Demirjian's judgment criteria, avoid the huge demand for model performance and data volume for direct age prediction through oral panoramic X-rays, and avoid the limitations of directly estimating age through developmental stages, is a technical problem that needs to be solved urgently. Summary of the Invention

[0009] The main purpose of the present invention is to provide an age prediction method and system based on oral panoramic X-rays, aiming to solve at least one of the above technical problems.

[0010] To achieve the above object, the present invention provides an age prediction method based on oral panoramic X-rays, comprising:

[0011] Obtain an oral panoramic X-ray, and based on the fact that the spatial position of each permanent tooth in the oral panoramic X-ray is relatively fixed, obtain the spatial coordinate information of each permanent tooth in the oral panoramic X-ray;

[0012] Construct a YoloV8 classification model and a TANet classification model, and train the YoloV8 classification model and the TANet classification model respectively according to a first training data set and a second training data set constructed from an oral orthogonal panoramic X-ray film training set stored in an oral orthogonal panoramic X-ray film training library;

[0013] Input the oral panoramic X-ray and the spatial coordinate information of each permanent tooth into the trained YoloV8 classification model, identify the developmental stage data of each permanent tooth, digitize the developmental stage data of each permanent tooth, and add them together to calculate the sum of the developmental stage data of all permanent teeth in the oral panoramic X-ray;

[0014] The oral panoramic X-ray and the sum of the developmental stage values ​​of all permanent teeth were input into the trained TANet classification model to predict the age of the individual to whom the oral panoramic X-ray belonged.

[0015] Optionally, an oral anteroposterior panoramic X-ray is obtained. Based on the fact that the spatial position of each permanent tooth in the oral anteroposterior panoramic X-ray is relatively fixed, the spatial coordinate information of each permanent tooth in the oral anteroposterior panoramic X-ray is obtained, specifically including:

[0016] Obtain an oral anteroposterior panoramic X-ray, and determine the coordinates of each permanent tooth at the corresponding relatively fixed position based on the spatially relatively fixed position of each permanent tooth in the oral anteroposterior panoramic X-ray;

[0017] The coordinates of each permanent tooth are used as the spatial coordinate information in the oral panoramic X-ray.

[0018] Optionally, a YoloV8 classification model and a TANet classification model are constructed, and the YoloV8 classification model and the TANet classification model are trained respectively according to a first training data set and a second training data set constructed from an oral orthogonal panoramic X-ray film training set stored in an oral orthogonal panoramic X-ray film training library, specifically including:

[0019] Obtain several oral orthodontic panoramic training X-rays with age labels from an oral orthodontic panoramic X-ray training library, and determine the spatial position information of each permanent tooth in each oral orthodontic panoramic training X-ray;

[0020] According to Demirjian's judgment criteria, the developmental stage data of each permanent tooth was marked on each oral panoramic training X-ray, and each oral panoramic training X-ray and the spatial position information of each permanent tooth were constructed as the first training data set. The developmental stage data of each permanent tooth in each oral panoramic training X-ray were numerically added to obtain the total developmental stage value, and each oral panoramic training X-ray and the total developmental stage value were constructed as the second training data set;

[0021] Construct a YoloV8 classification model and a TANet classification model, use the first training data set to train the YoloV8 classification model, and use the second training data set to train the TANet classification model.

[0022] Optionally, the loss function used to train the TANet classification model is expressed as follows: ;

[0023] Where n is the number of samples, is the actual value of the age label of the i-th sample, is the predicted age output by the model for the i-th sample.

[0024] Optionally, the constructed YoloV8 classification model includes:

[0025] The feature extraction module consists of Backbone and Neck, the spatial attention module set between Backbone and Neck, and the classification detection head set after Neck;

[0026] The spatial attention module is configured to map the spatial coordinate information to the feature space using a fully connected layer, broadcast the converted coordinate features to the spatial dimension, fuse the coordinate features with the image features by direct addition, and finally generate a spatial attention map through 1×1 convolution and sigmoid function, and multiply the fused image features with the spatial attention map;

[0027] Among them, the expression for mapping spatial coordinate information to feature space is specifically: ;

[0028] Where, is the coordinate feature corresponding to the spatial coordinate information, B is the batch, 2 is the coordinate dimension (x, y), is the fully connected layer weight, is the bias term, is the transformed coordinate feature, and C is the number of channels corresponding to the image feature;

[0029] Among them, the transformed coordinate features are broadcasted to the representation of the spatial dimension, specifically: Where, For broadcast operations, is a matrix of all 1s, used to expand the spatial dimension H×W, is the coordinate feature broadcast to the spatial dimension. Optionally, the constructed TANet classification model includes: for extracting the feature x of the oral panoramic X-ray film outA CNN network, a pre-trained model for generating global semantic information CLS Token, and a prediction network for inferring and predicting the concatenated total developmental stage value and the global semantic information CLS Token.

[0030] Optional, used to extract features of oral panoramic X-rays out The CNN network is configured to include:

[0031] Downsampling first layer, downsampling second layer, upsampling first layer, upsampling second layer, first squeeze and excitation channel attention module connecting downsampling first layer and upsampling second layer, second squeeze and excitation channel attention module connecting downsampling second layer and upsampling first layer, and convolution module;

[0032] The first downsampling layer and the second downsampling layer are configured to include two convolutional layers and one maximum pooling layer, and the specific expression is: ; ; ; Where, , The convolution kernel is 3×3, the input is 3 channels and the output is 64 channels. The rest of the convolution blocks are similar. is the maximum pooling layer with a stride of 2, is the ReLU function;

[0033] The first upsampling layer and the second upsampling layer are configured to include a concatenation module, a convolution layer, and an upsampling module, and the specific expression is: ; ; ; ; ; Where, The convolution kernel is 2×2, which changes the number of channels from 128 to 64. The rest of the transposed convolutions are similar. is the concatenation function.

[0034] Optionally, the first squeeze and excitation channel attention module and the second squeeze and excitation channel attention module are configured to filter loss information in downsampling, and the specific expression thereof is: ; Where, The first squeeze and excitation channel attention module with 64 input and output channels, The second squeeze and excitation channel attention module has 128 input and output channels;

[0035] The expression of the convolution module is specifically: ;

[0036] Where, The features extracted by the CNN network from the oral panoramic X-ray.

[0037] Optionally, the pre-trained model used to generate the global semantic information CLS Token is configured to use the ViT model loaded with the pre-trained vit_b16 and with the classification head removed. The specific expression is: ;

[0038] In the formula, the output of the pre-trained model includes global semantic information CLS Token;

[0039] Among them, the prediction network used to infer and predict the spliced ​​total developmental stage value and the global semantic information CLS Token specifically includes: a splicing layer for splicing the total developmental stage value and the global semantic information CLS Token, a ReLU function, and a fully connected layer that outputs the inference prediction age.

[0040] In addition, in order to achieve the above-mentioned purpose, the present invention also provides an age prediction system based on oral panoramic X-rays, the system comprising:

[0041] An acquisition module is used to acquire an oral anteroposterior panoramic X-ray, and obtain the spatial coordinate information of each permanent tooth in the oral anteroposterior panoramic X-ray based on the fact that the spatial position of each permanent tooth in the oral anteroposterior panoramic X-ray is relatively fixed;

[0042] A training module is used to build a YoloV8 classification model and a TANet classification model, and train the YoloV8 classification model and the TANet classification model respectively according to a first training data set and a second training data set constructed from an oral orthogonal panoramic X-ray film training set stored in an oral orthogonal panoramic X-ray film training library;

[0043] The recognition module is used to input the spatial coordinate information of the oral panoramic X-ray and each permanent tooth into the trained YoloV8 classification model, identify the developmental stage data of each permanent tooth, digitize the developmental stage data of each permanent tooth, add them together, and calculate the sum of the developmental stage data of all permanent teeth in the oral panoramic X-ray;

[0044] The prediction module is used to input the oral panoramic X-ray and the sum of the developmental stage values ​​of all permanent teeth into the trained TANet classification model to predict the age of the individual to whom the oral panoramic X-ray belongs.

[0045] The beneficial effects of the present invention are: a method and system for age prediction based on oral panoramic X-rays are proposed. Compared with existing age prediction technologies, the present invention avoids heavy reliance on Demirjian's judgment criteria, uses the learning ability of deep learning methods to classify and predict age, reduces the systematic errors brought about by Demirjian's judgment criteria, and promotes the clinical application of automatic age estimation technology based on oral panoramic X-rays. Compared with traditional models, the TANet model is more suitable for the age classification task of oral panoramic X-rays, and traditional models have a huge demand for data volume. This problem is alleviated by using the pre-trained model vit_b16 and introducing auxiliary reasoning of permanent tooth development stages. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the process of the age prediction method based on oral panoramic X-rays of the present invention;

[0047] Figure 2 Schematic diagram of the execution process of age prediction based on oral panoramic X-rays in the present invention;

[0048] Figure 3 This is a schematic diagram of determining spatial coordinate information based on an oral panoramic X-ray in the present invention;

[0049] Figure 4 This is a schematic diagram of the principle of the automatic age estimation technology based on oral panoramic X-rays of the present invention;

[0050] Figure 5 This is a schematic diagram of the TANet network structure proposed in the present invention;

[0051] Figure 6 The figure is a schematic structural diagram of the age prediction system based on oral panoramic X-rays of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] The embodiment of the present invention provides an age prediction method based on oral panoramic X-rays, referring to Figure 1 , Figure 1 The figure is a flow chart of an embodiment of the age prediction method based on oral panoramic X-rays of the present invention.

[0054] In this embodiment, the age prediction method based on oral panoramic X-rays includes the following steps:

[0055] S100: Obtaining an oral panoramic X-ray, and obtaining spatial coordinate information of each permanent tooth in the oral panoramic X-ray based on the relatively fixed spatial position of each permanent tooth in the oral panoramic X-ray;

[0056] S200: constructing a YoloV8 classification model and a TANet classification model, and respectively training the YoloV8 classification model and the TANet classification model according to a first training data set and a second training data set constructed from an oral orthogonal panoramic X-ray film training set stored in an oral orthogonal panoramic X-ray film training library;

[0057] S300: Inputting the oral panoramic X-ray and the spatial coordinate information of each permanent tooth into the trained YoloV8 classification model, identifying the developmental stage data of each permanent tooth, digitizing the developmental stage data of each permanent tooth and adding them together, and calculating the sum of the developmental stage data of all permanent teeth in the oral panoramic X-ray;

[0058] S400: Input the oral panoramic X-ray and the sum of the developmental stage values ​​of all permanent teeth into the trained TANet classification model to predict the age of the individual to whom the oral panoramic X-ray belongs.

[0059] It should be noted that the present invention mainly addresses the problem that the current automatic age estimation technology based on oral panoramic X-rays strongly relies on Demirjian's judgment criteria. The current technology mainly uses deep learning to determine the developmental stage of permanent teeth, and then directly estimates the natural age based on the developmental stage. However, although there is a significant correlation between the developmental stage of permanent teeth and the age of the individual, there is an objective error. The actual age range corresponding to the same developmental stage can reach 3 years. When the developmental stage is the only estimation data, this error is inevitable.

[0060] Although the method of directly estimating age by developmental stage has great limitations, current deep learning methods are difficult to efficiently classify age based on oral panoramic X-rays. This is because the image differences between different ages are relatively small, and there are fewer effective features in oral panoramic X-rays, making it difficult to form effective semantic features to help the network model learn image features. Therefore, the permanent tooth development stage is still of great help in estimating age.

[0061] The present invention designs a classification model that introduces data on the developmental stages of permanent teeth for auxiliary prediction, and uses oral panoramic X-rays to predict age. This avoids the huge demands on model performance and data volume for direct age prediction using oral panoramic X-rays, and also avoids the limitations of directly estimating age through developmental stages.

[0062] The following provides a specific implementation process of the age automatic estimation technology based on oral panoramic X-rays proposed by the present invention. Figure 2shown.

[0063] First, the present invention combines the physical space characteristics. The spatial positions of various types of permanent teeth in the oral panoramic X-ray are relatively fixed. Therefore, the approximate coordinate information of each permanent tooth position is constructed, which is available in all images. Assuming that each X-ray is marked with the coordinates of each permanent tooth, such as (x1, y1), (x2, y2), ..., (xn, yn), the spatial coordinate information is also used for training and prediction, such as Figure 3 As shown, the white square is the mapping of the obtained coordinates.

[0064] Then, according to Demirjian's judgment criteria, the corresponding single permanent tooth development stage value and the specific position of each permanent tooth in the oral panoramic X-ray are annotated as the YoloV8 data set. The spatial position information and the YoloV8 data set are used as the first data set. This data set is used to train the optimized YoloV8 model for the classification task to obtain the model parameters. Then, a second data set is made. The development stage value of each permanent tooth in each image is added to obtain the total development stage value of the permanent tooth. This data set and the oral panoramic X-ray are used as the second data set to train the TANet model, and finally the trained TANet model is obtained. The training loss function is: ;

[0065] Where: n is the sample size. is the actual value of the i-th sample (here the age label). is the predicted value of the i-th sample (here it is the predicted age output by the model).

[0066] It should be noted that the reasoning process of the optimized YoloV8 model is as follows Figure 4 As shown in the figure, an oral X-ray and spatial position information are fed into the optimized YoloV8 model to obtain the permanent tooth development stage classification results. The YoloV8 model is optimized by adding spatial information. After obtaining the spatial coordinate information of each permanent tooth, the coordinates are normalized to ensure that the coordinate values ​​are within the range [0, 1]. The feature extraction part of YoloV8 consists of Backbone and Neck (such as FPN or PAN). A spatial attention module is inserted after the output of Backbone to enhance the spatial perception ability of the feature map.

[0067] The spatial attention module first maps the coordinate information to the feature space through a fully connected layer to ensure that the number of channels is the same as that of the image feature. The formula is as follows: ;

[0068] Coordinate features , where B is the batch, 2 is the coordinate dimension (x, y), is the fully connected layer weight, is the bias term, output , C is the number of channels (aligned with image features).

[0069] The coordinate features are then broadcast to the spatial dimension to ensure that the size of the coordinate information matches the image features. The formula is as follows: ;

[0070] Represents a broadcast operation (implemented by the unsqueeze and expand_as functions), Is a full 1 matrix, used to expand the spatial dimension H×W, output The coordinate features and image features are then fused by direct addition, and finally a 1×1 convolution and sigmoid function are used to generate a spatial attention map. The image features are then multiplied by the spatial attention map to obtain the result.

[0071] Then, the present invention uses the optimized YoloV8 model to obtain the developmental stage data of each permanent tooth from the oral panoramic X-ray film, digitizes the developmental stage data of all permanent teeth and adds them together to obtain the sum of the developmental stages of all permanent teeth in the image. Then, the oral panoramic X-ray film and Sum are passed together to the classification model TANet to obtain the inference result, such as Figure 4 shown.

[0072] Among them, the TANet structure is as follows Figure 5 As shown in the figure, in traditional classification models, reasoning is mainly based on semantic features composed of low-level features in the image. However, in oral X-rays, semantic features are not obvious, so the reasoning efficiency of traditional classification models is not ideal. Therefore, the present invention designs a TANet model. The first half of the model uses a CNN network to extract low-level features of the image, helping the network to obtain the desired features for reasoning. The CNN network part borrows from the UNet network, and extracts features through the downsampling process: Input image , where B is the batch size and H×W is the image size. x completes the first layer of downsampling through two convolutional layers and a maximum pooling layer, as shown in the following formula: ; ;in, , The convolution kernel is 3×3, and the input is a 3-channel convolution block with a 64-channel output (the rest of the convolution blocks are similar). is a maximum pooling layer with a stride of 2. is the ReLU function.

[0073] Next, we go through the second layer of downsampling: ; ;

[0074] After completing downsampling, the present invention gradually restores the original size during the upsampling process and facilitates subsequent classification tasks. During the downsampling and upsampling process, the high-quality downsampled image is transferred to the downsampler. Since the oral anteroposterior panoramic X-ray image itself has few low-level features, it is easy to lose information during the downsampling process. Therefore, by squeezing and stimulating channel attention, the channel of lost information in the downsampling is filtered out, thereby improving the quality of feature extraction. The jump connection formula is as follows: ; ;in, The SENet module has 64 input and output channels, and the rest are similar.

[0075] The first layer upsampling formula is as follows: ; ; ;in, It is a transposed convolution with a convolution kernel of 2×2, which changes the number of channels from 128 to 64. The rest of the transposed convolutions are similar. is the concatenation function.

[0076] The second layer upsampling formula is as follows: ; ; ; Finally, the result after feature extraction is obtained through the 3*3 convolution layer: Since the reasoning of oral panoramic X-rays depends on the morphology of each permanent tooth and the relationship between them, the advantage of Transformer for long-distance dependence is brought into play. Therefore, in the second half of the model, the features extracted by the CNN network are used to represent the features of the permanent teeth. Pass in the ViT model with the pre-trained vit_b16 loaded and the classification head removed. The formula is as follows: ;

[0077] Loading the pre-trained vit_b16 can reduce the number of iterations and training data required for model training. After obtaining the Outputs, the CLS Token is extracted, the Sum is concatenated with the CLS Token, and finally the inference result is obtained through the fully connected layer.

[0078] Compared to existing age prediction techniques, this invention avoids heavy reliance on Demirjian's criteria and instead leverages the learning capabilities of deep learning methods to classify and predict age. This reduces the systematic errors introduced by Demirjian's criteria and promotes the clinical application of automatic age estimation based on oral panoramic X-rays. Compared to traditional models, the TANet model is more suitable for age classification based on oral panoramic X-rays. However, traditional models require a significant amount of data, a problem alleviated by using the pretrained model vit_b16 and introducing the permanent tooth development stage to aid inference.

[0079] Reference Figure 6 , Figure 6 This is a block diagram of an embodiment of an age prediction system based on an oral anteroposterior panoramic X-ray film of the present invention. The age prediction system based on an oral anteroposterior panoramic X-ray film proposed in this embodiment of the present invention includes:

[0080] An acquisition module 10 is used to acquire an oral anteroposterior panoramic X-ray, and obtain the spatial coordinate information of each permanent tooth in the oral anteroposterior panoramic X-ray based on the fact that the spatial position of each permanent tooth in the oral anteroposterior panoramic X-ray is relatively fixed;

[0081] A training module 20 is used to construct a YoloV8 classification model and a TANet classification model, and train the YoloV8 classification model and the TANet classification model respectively according to a first training data set and a second training data set constructed from an oral orthogonal panoramic X-ray film training set stored in an oral orthogonal panoramic X-ray film training library;

[0082] Recognition module 30, for inputting the spatial coordinate information of the oral panoramic X-ray and each permanent tooth into the trained YoloV8 classification model, identifying the developmental stage data of each permanent tooth, digitizing the developmental stage data of each permanent tooth, and adding them together to calculate the sum of the developmental stage data of all permanent teeth in the oral panoramic X-ray;

[0083] The prediction module 40 is used to input the oral panoramic X-ray and the sum of the developmental stage values ​​of all permanent teeth into the trained TANet classification model to predict the age of the individual to whom the oral panoramic X-ray belongs.

[0084] Other embodiments or specific implementations of the age prediction system based on oral anteroposterior panoramic X-rays of the present invention can refer to the above-mentioned method embodiments and will not be repeated here.

[0085] It should be understood that, in the description of this specification, reference to terms such as "one embodiment," "another embodiment," "other embodiments," or "first to Nth embodiments" means that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any appropriate manner in any one or more embodiments or examples.

[0086] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0087] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for age prediction based on oral panoramic X-rays, characterized in that: include: Obtain an oral panoramic X-ray, and based on the fact that the spatial position of each permanent tooth in the oral panoramic X-ray is relatively fixed, obtain the spatial coordinate information of each permanent tooth in the oral panoramic X-ray; Construct a YoloV8 classification model and a TANet classification model, and train the YoloV8 classification model and the TANet classification model respectively according to a first training data set and a second training data set constructed from an oral orthogonal panoramic X-ray film training set stored in an oral orthogonal panoramic X-ray film training library; Among them, the constructed TANet classification model includes: features x for extracting oral panoramic X-rays out A CNN network, a pre-trained model for generating global semantic information CLS Token, and a prediction network for inferring and predicting the concatenated total developmental stage value and global semantic information CLS Token; Among them, the pre-trained model used to generate the global semantic information CLS Token is configured to use the ViT model loaded with the pre-trained vit_b16 and with the classification head removed. Its specific expression is: Outputs=ViT(x out ); In the formula, the output of the pre-trained model includes global semantic information CLS Token; The prediction network for inferring and predicting the concatenated total developmental stage value and the global semantic information CLS Token specifically includes: a concatenation layer for concatenating the total developmental stage value and the global semantic information CLS Token, a ReLU function, and a fully connected layer for outputting the inferred predicted age; Input the oral panoramic X-ray and the spatial coordinate information of each permanent tooth into the trained YoloV8 classification model, identify the developmental stage data of each permanent tooth, digitize the developmental stage data of each permanent tooth, and add them together to calculate the sum of the developmental stage data of all permanent teeth in the oral panoramic X-ray; The oral panoramic X-ray and the sum of the developmental stage values ​​of all permanent teeth were input into the trained TANet classification model to predict the age of the individual to whom the oral panoramic X-ray belonged.

2. The age prediction method based on oral panoramic X-rays according to claim 1, characterized in that: Obtain an oral anteroposterior panoramic X-ray. Based on the fact that the spatial position of each permanent tooth in the oral anteroposterior panoramic X-ray is relatively fixed, obtain the spatial coordinate information of each permanent tooth in the oral anteroposterior panoramic X-ray, specifically including: Obtain an oral anteroposterior panoramic X-ray, and determine the coordinates of each permanent tooth at the corresponding relatively fixed position based on the spatially relatively fixed position of each permanent tooth in the oral anteroposterior panoramic X-ray; The coordinates of each permanent tooth are used as the spatial coordinate information in the oral panoramic X-ray.

3. The age prediction method based on oral panoramic X-rays according to claim 1, characterized in that: Constructing a YoloV8 classification model and a TANet classification model, and performing training steps on the YoloV8 classification model and the TANet classification model respectively according to a first training data set and a second training data set constructed from an oral orthogonal panoramic X-ray film training set stored in an oral orthogonal panoramic X-ray film training library, specifically including: Obtain several oral orthodontic panoramic training X-rays with age labels from an oral orthodontic panoramic X-ray training library, and determine the spatial position information of each permanent tooth in each oral orthodontic panoramic training X-ray; According to Demirjian's judgment criteria, the developmental stage data of each permanent tooth was marked on each oral panoramic training X-ray, and each oral panoramic training X-ray and the spatial position information of each permanent tooth were constructed as the first training data set. The developmental stage data of each permanent tooth in each oral panoramic training X-ray were numerically added to obtain the total developmental stage value, and each oral panoramic training X-ray and the total developmental stage value were constructed as the second training data set; Construct a YoloV8 classification model and a TANet classification model, use the first training data set to train the YoloV8 classification model, and use the second training data set to train the TANet classification model.

4. The age prediction method based on oral panoramic X-rays according to claim 3, characterized in that: The loss function used to train the TANet classification model is expressed as follows: Where n is the number of samples, y i is the actual value of the age label of the i-th sample, is the predicted age output by the model for the i-th sample.

5. The age prediction method based on oral panoramic X-rays according to claim 1, characterized in that: The constructed YoloV8 classification model includes: The feature extraction module consists of Backbone and Neck, the spatial attention module set between Backbone and Neck, and the classification detection head set after Neck; The spatial attention module is configured to map the spatial coordinate information to the feature space using a fully connected layer, broadcast the converted coordinate features to the spatial dimension, fuse the coordinate features with the image features by direct addition, and finally generate a spatial attention map through 1×1 convolution and sigmoid function, and multiply the fused image features with the spatial attention map; Among them, the expression for mapping spatial coordinate information to feature space is specifically: z fc =W fc ·c+b fc ; Where c∈R B×2 is the coordinate feature corresponding to the spatial coordinate information, B is the batch, 2 is the coordinate dimension (x, y), W fc ∈R C×2 is the weight of the fully connected layer, b fc ∈R C is the bias term, z fc ∈R B×C is the transformed coordinate feature, and C is the number of channels corresponding to the image feature; Among them, the transformed coordinate features are broadcasted to the representation of the spatial dimension, specifically: Where, For broadcast operation, 1 H×W is a matrix of all 1s, used to expand the spatial dimension H×W, z expand ∈R B×C×H×W are the coordinate features broadcast to the spatial dimensions.

6. The age prediction method based on oral panoramic X-rays according to claim 1, characterized in that: Used to extract features of oral panoramic X-rays out The CNN network is configured to include: Downsampling first layer, downsampling second layer, upsampling first layer, upsampling second layer, first squeeze and excitation channel attention module connecting downsampling first layer and upsampling second layer, second squeeze and excitation channel attention module connecting downsampling second layer and upsampling first layer, and convolution module; The first downsampling layer and the second downsampling layer are configured to include two convolutional layers and one maximum pooling layer, and the specific expression is: Where x1∈R B×64×H×W , The convolution kernel is 3×3, and the input is 3 channels and the output is 64 channels. The rest of the convolution blocks are similar. MaxPool2 is the maximum pooling layer with a stride of 2, and ReLU is the ReLU function. The first upsampling layer and the second upsampling layer are configured to include a concatenation module, a convolution layer, and an upsampling module, and the specific expression is: Where, It is a transposed convolution with a convolution kernel of 2×2, which changes the number of channels from 128 to 64. The rest of the transposed convolutions are similar. Concat is a concatenation function.

7. The age prediction method based on oral panoramic X-rays according to claim 6, characterized in that: The first squeeze and excitation channel attention module and the second squeeze and excitation channel attention module are configured to filter the loss information in downsampling, and the specific expression is: Where, SENet 64 The first squeeze and excitation channel attention module with 64 input and output channels, SENet 128 The second squeeze and excitation channel attention module has 128 input and output channels; The expression of the convolution module is specifically: Where x out The features extracted by the CNN network from the oral panoramic X-ray.

8. An age prediction system based on oral panoramic X-rays, characterized in that: include: An acquisition module is used to acquire an oral anteroposterior panoramic X-ray, and obtain the spatial coordinate information of each permanent tooth in the oral anteroposterior panoramic X-ray based on the fact that the spatial position of each permanent tooth in the oral anteroposterior panoramic X-ray is relatively fixed; A training module is used to build a YoloV8 classification model and a TANet classification model, and train the YoloV8 classification model and the TANet classification model respectively according to a first training data set and a second training data set constructed from an oral orthogonal panoramic X-ray film training set stored in an oral orthogonal panoramic X-ray film training library; Among them, the constructed TANet classification model includes: features x for extracting oral panoramic X-rays out A CNN network, a pre-trained model for generating global semantic information CLS Token, and a prediction network for inferring and predicting the concatenated total developmental stage value and global semantic information CLS Token; Among them, the pre-trained model used to generate the global semantic information CLS Token is configured to use the ViT model loaded with the pre-trained vit_b16 and with the classification head removed. Its specific expression is: Outputs=ViT(x out ); In the formula, the output of the pre-trained model includes global semantic information CLS Token; The prediction network for inferring and predicting the concatenated total developmental stage value and the global semantic information CLS Token specifically includes: a concatenation layer for concatenating the total developmental stage value and the global semantic information CLS Token, a ReLU function, and a fully connected layer for outputting the inferred predicted age; The recognition module is used to input the spatial coordinate information of the oral panoramic X-ray and each permanent tooth into the trained YoloV8 classification model, identify the developmental stage data of each permanent tooth, digitize the developmental stage data of each permanent tooth, add them together, and calculate the sum of the developmental stage data of all permanent teeth in the oral panoramic X-ray; The prediction module is used to input the oral panoramic X-ray and the sum of the developmental stage values ​​of all permanent teeth into the trained TANet classification model to predict the age of the individual to whom the oral panoramic X-ray belongs.

Images