Tooth automatic detection and fdi number consistency correction model training and application method and related device

Abstract

The application discloses a tooth automatic detection and FDI number consistency correction model training and application method and related devices, relates to the oral medical image processing technical field, and the method comprises the following steps: determining the predicted bounding box, the FDI number prediction result and the visual feature of each tooth in a sample image by using a target detection module; taking the predicted bounding box of each tooth in the sample image as a graph node; selectively drawing a connection edge between two graph nodes based on the anatomical structure relationship of the tooth; determining the node feature of the graph node and the edge attribute of the connection edge; obtaining the FDI number correction result of each tooth in the sample image; further calculating the total loss; updating by using the total loss; obtaining the tooth automatic detection and FDI number consistency correction model; and subsequently performing tooth automatic detection and FDI number. The application can improve the precision of tooth automatic detection and FDI number.

Description

Technical Field

[0001] This application relates to the field of oral medical image processing technology, and in particular to a method and related apparatus for training and applying an automatic tooth detection and FDI number consistency correction model. Background Technology

[0002] Panoramic radiographs are a commonly used imaging technique in dental clinics, providing crucial anatomical information for assessing dental development, analyzing malocclusion, locating lesions, and conducting follow-up observations. Automatic tooth detection and FDI (Front-Index Digitization) numbering are essential foundations for the structured analysis of panoramic radiographs, impacting not only the efficiency of tooth positioning but also subsequent downstream tasks such as lesion localization, treatment planning, and longitudinal follow-up.

[0003] In recent years, most automatic tooth detection and FDI numbering methods are built on general object detection models, such as Faster R-CNN and YOLO, and typically use instance-level detection metrics such as mAP (mean Average Precision), Recall, and Precision to evaluate performance. These methods can achieve good detection performance in adult permanent dentition scenarios. However, high detection performance does not necessarily mean reliable tooth numbering results. Automatic tooth detection and FDI numbering not only require teeth to be detected but also require a stable one-to-one correspondence between each tooth and its FDI number. However, in practical applications, these methods still frequently encounter problems such as confusion between adjacent tooth positions, duplicate bounding boxes, false positives caused by artifacts, and false negatives. Furthermore, compared to adult permanent dentition, children's mixed dentition has higher structural complexity. On the one hand, deciduous and permanent teeth coexist, exhibiting eruption, replacement, retention, and crowding phenomena; on the other hand, physiological root resorption of deciduous teeth leads to local appearance instability, making object detection models relying solely on local visual features more prone to false negatives or false positives. Therefore, these methods suffer from low accuracy.

[0004] To improve accuracy, some studies introduce fixed template matching or static rule methods for post-processing after using target detection models to complete the detection. These post-processing methods are effective in relatively regular adult permanent dentition scenarios, but they are often difficult to maintain stability when faced with the highly variable dental arch arrangement in children's mixed dentition, and the problem of low accuracy still exists. Summary of the Invention

[0005] The purpose of this application is to provide a method and related apparatus for training and applying a tooth automatic detection and FDI number consistency correction model, which can improve the accuracy of tooth automatic detection and FDI numbering.

[0006] To achieve the above objectives, this application provides the following solution.

[0007] Firstly, this application provides a method for training a tooth automatic detection and FDI number consistency correction model, including: Acquire multiple sample images and the corresponding annotation data for each sample image; the sample images are panoramic tomographic images of mixed dentition in children, and the annotation data includes the true bounding box and true FDI number of each tooth in the sample image; Using sample images as input, the target detection module determines the predicted bounding box, FDI number prediction result, and visual features of each tooth in the sample image. Each predicted bounding box of a tooth in the sample image is treated as a graph node. Based on the anatomical relationship of the tooth, connection edges are selectively drawn between two graph nodes to obtain the initial multi-relationship heterogeneous graph of the sample. The types of connection edges include overlapping competition edges, dental arch adjacency edges, same-jaw neighborhood edges, and spatial proximity edges. For each graph node and each connecting edge in the initial multi-relation heterogeneous graph of the sample, the predicted bounding box corresponding to the graph node, the prediction result of the FDI number, and the visual features are used as inputs. The first encoding module is used to determine the node features of the graph node and assign node features to the graph node. Based on the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge, the geometric attributes of the connecting edge are determined. The type and geometric attributes of the connecting edge are used as inputs. The second encoding module is used to determine the edge attributes of the connecting edge and assign edge attributes to the connecting edge, thus obtaining the intermediate multi-relation heterogeneous graph of the sample. Using the sample intermediate multi-relation heterogeneity graph as input, the graph reasoning module is used to determine the sample multi-relation heterogeneity graph. Using the sample multi-relation heterogeneity graph as input, the classification module is used to determine the FDI number correction result of each tooth in the sample image. The total loss is calculated based on the true bounding box, true FDI number, predicted bounding box, and FDI number correction results for each tooth in the sample image. The target detection module, the first encoding module, the second encoding module, the graph reasoning module, and the classification module are updated using the total loss, resulting in updated target detection module, updated first encoding module, updated second encoding module, updated graph reasoning module, and updated classification module. If the iteration termination condition is met, the updated target detection module, updated first encoding module, updated second encoding module, updated graph reasoning module, and updated classification module are combined to form an automatic tooth detection and FDI number consistency correction model.

[0008] Secondly, this application provides a method for applying an automatic tooth detection and FDI number consistency correction model, including: Acquire the image to be processed; the image to be processed is a panoramic tomographic image of a child's mixed dentition. Using the image to be processed as input, the updated object detection module is used to determine the predicted bounding box, FDI number prediction result and visual features of each tooth in the image to be processed. The predicted bounding box of each tooth in the image to be processed is taken as a graph node. Based on the anatomical relationship of the tooth, connection edges are selectively drawn between two graph nodes to obtain an initial multi-relationship heterogeneous graph. For each graph node and each connecting edge in the initial multi-relation heterogeneous graph, the node features of the graph node are determined by the updated first encoding module, and node features are assigned to the graph node, based on the predicted bounding boxes of the two graph nodes connected by the connecting edge. The geometric attributes of the connecting edge are determined by the updated second encoding module, based on the type and geometric attributes of the connecting edge, and edge attributes are assigned to the connecting edge, thus obtaining the intermediate multi-relation heterogeneous graph. Using the intermediate multi-relation heterogeneous graph as input, the updated graph inference module determines the multi-relation heterogeneous graph. Using the multi-relation heterogeneous graph as input, the updated classification module determines the FDI number correction result for each tooth in the image to be processed. The updated target detection module, the updated first encoding module, the updated second encoding module, the updated graph inference module, and the updated classification module constitute the automatic tooth detection and FDI number consistency correction model. The automatic tooth detection and FDI number consistency correction model is a model trained using the above-mentioned automatic tooth detection and FDI number consistency correction model training method. Based on the FDI number correction results of each tooth in the image to be processed, the FDI number of each tooth in the image to be processed is determined.

[0009] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the above-described automatic tooth detection and FDI number consistency correction model training method or the above-described automatic tooth detection and FDI number consistency correction model application method.

[0010] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for training the automatic tooth detection and FDI number consistency correction model or the above-described method for applying the automatic tooth detection and FDI number consistency correction model.

[0011] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method and related apparatus for training and applying an automatic tooth detection and FDI number consistency correction model. Using sample images as input, a target detection module determines the predicted bounding box, FDI number prediction result, and visual features of each tooth in the sample image. Each predicted bounding box of a tooth in the sample image is treated as a graph node. Based on the anatomical structure of the tooth, selectively drawing connecting edges between two graph nodes yields an initial multi-relationship heterogeneous graph. For each graph node and each connecting edge in the initial multi-relationship heterogeneous graph, using the predicted bounding box, FDI number prediction result, and visual features corresponding to the graph node as input, a first encoding module determines the node features of the graph node and assigns node features to the graph node. Based on the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge, the geometric attributes of the connecting edge are determined. Using the type and geometric attributes of the connecting edge as input, a second encoding module determines the edge attributes of the connecting edge and assigns edge attributes to the connecting edge. The system obtains a heterogeneous graph of intermediate relationships among samples. Using this graph as input, the graph reasoning module determines the heterogeneous graph of intermediate relationships. Using this graph as input, the classification module determines the FDI number correction result for each tooth in the sample image. Based on the true bounding box, true FDI number, predicted bounding box, and FDI number correction result for each tooth in the sample image, the total loss is calculated. The total loss is used to update the target detection module, the first encoding module, the second encoding module, the graph reasoning module, and the classification module respectively until the iteration termination condition is reached, resulting in a tooth automatic detection and FDI number consistency correction model. Subsequently, the tooth automatic detection and FDI number consistency correction model is used for tooth automatic detection and FDI numbering. Thus, target detection is performed first, followed by graph reasoning, realizing the functional separation of target detection and number correction. It can correct inconsistent numbers, thereby improving the accuracy of tooth automatic detection and FDI numbering. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 An application environment diagram for a tooth automatic detection and FDI number consistency correction model training and application method provided in an embodiment of this application; Figure 2 This is a flowchart illustrating a method for training an automatic tooth detection and FDI number consistency correction model, provided in an embodiment of this application. Figure 3This is a flowchart illustrating an application method for an automatic tooth detection and FDI number consistency correction model, provided in an embodiment of this application. Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] The tooth automatic detection and FDI number consistency correction model training method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, the terminal communicates with the server via a network. The data storage system stores the data the server needs to process. The data storage system can be set up independently, integrated into the server, or placed in the cloud or on another server. The terminal can send training requests to be processed to the server. After receiving the training request, the server acquires multiple sample images and the corresponding annotation data for each sample image. Using the sample images as input, the target detection module determines the predicted bounding box, FDI number prediction result, and visual features of each tooth in the sample image. Each predicted bounding box of each tooth in the sample image is treated as a graph node. Based on the anatomical structure of the teeth, selectively drawing connecting edges between two graph nodes yields an initial multi-relationship heterogeneous graph. For each graph node and each connecting edge in the initial multi-relationship heterogeneous graph, using the predicted bounding box, FDI number prediction result, and visual features corresponding to the graph node as input, the first encoding module determines the node features of the graph node and assigns node features to the graph node. Based on the predicted bounding boxes of the two graph nodes connected by the connecting edge, the geometric attributes of the connecting edge are determined. Using the type and geometric attributes of the connecting edge as input, the second... The encoding module determines the edge attributes of the connecting edges and assigns edge attributes to the connecting edges, obtaining a sample intermediate multi-relation heterogeneous graph. Using this sample intermediate multi-relation heterogeneous graph as input, the graph inference module determines the sample multi-relation heterogeneous graph. Using this sample multi-relation heterogeneous graph as input, the classification module determines the FDI number correction result for each tooth in the sample image. Based on the true bounding box, true FDI number, predicted bounding box, and FDI number correction result for each tooth in the sample image, the total loss is calculated. The total loss is used to update the object detection module, the first encoding module, the second encoding module, the graph inference module, and the classification module, respectively, resulting in updated object detection, first encoding, second encoding, graph inference, and classification modules. If the iteration termination condition is met, the updated object detection, first encoding, second encoding, graph inference, and classification modules form a tooth automatic detection and FDI number consistency correction model. The server can feed back this training result—the tooth automatic detection and FDI number consistency correction model for the training request—to the terminal.

[0017] In addition, in some embodiments, the tooth automatic detection and FDI number consistency correction model training method can also be implemented by the server or the terminal alone. For example, the terminal can directly process the training request to be processed, or the server can obtain the training request to be processed from the data storage system and process it.

[0018] The terminals can be, but are not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle systems. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. Servers can be implemented using independent servers, server clusters composed of multiple servers, or cloud servers.

[0019] In one exemplary embodiment, such as Figure 2 As shown, a method for training an automatic tooth detection and FDI number consistency correction model is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 The following steps, S1 to S7, are used as an example to illustrate the process of using a server in the example.

[0020] Step S1: Obtain multiple sample images and the corresponding annotation data for each sample image; the sample images are panoramic tomographic images of mixed dentition in children, and the annotation data includes the true bounding box and true FDI number of each tooth in the sample image.

[0021] Step S2: Using the sample image as input, the target detection module determines the predicted bounding box, FDI number prediction result, and visual features of each tooth in the sample image.

[0022] Step S3: Take the predicted bounding box of each tooth in the sample image as a graph node, and selectively draw connecting edges between two graph nodes based on the anatomical relationship of the teeth to obtain the initial multi-relationship heterogeneous graph of the sample; the types of connecting edges include overlapping competition edges, dental arch adjacency edges, same-jaw neighborhood edges, and spatial proximity edges.

[0023] Step S4: For each graph node and each connecting edge in the initial multi-relation heterogeneous graph of the sample, the predicted bounding box corresponding to the graph node, the prediction result of the FDI number, and the visual features are used as inputs. The first encoding module is used to determine the node features of the graph node and assign node features to the graph node. Based on the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge, the geometric attributes of the connecting edge are determined. The type and geometric attributes of the connecting edge are used as inputs. The second encoding module is used to determine the edge attributes of the connecting edge and assign edge attributes to the connecting edge to obtain the intermediate multi-relation heterogeneous graph of the sample.

[0024] Step S5: Using the sample intermediate multi-relation heterogeneous graph as input, the graph reasoning module is used to determine the sample multi-relation heterogeneous graph. Using the sample multi-relation heterogeneous graph as input, the classification module is used to determine the FDI number correction result of each tooth in the sample image.

[0025] Step S6: Calculate the total loss based on the true bounding box, true FDI number, predicted bounding box, and FDI number correction results for each tooth in the sample image.

[0026] Step S7: Update the target detection module, first encoding module, second encoding module, graph reasoning module, and classification module using the total loss to obtain the updated target detection module, updated first encoding module, updated second encoding module, updated graph reasoning module, and updated classification module. If the iteration termination condition is met, the updated target detection module, updated first encoding module, updated second encoding module, updated graph reasoning module, and updated classification module are combined to form the tooth automatic detection and FDI number consistency correction model.

[0027] By implementing steps S1 to S7 above, this embodiment can improve the accuracy of automatic tooth detection and FDI numbering.

[0028] Currently, automatic tooth detection and FDI numbering methods can be broadly divided into two categories.

[0029] (1) The first type is the end-to-end detection and numbering method, which unifies tooth bounding box prediction and tooth position category prediction into the same detection framework, that is, it uses the target detection model. This type of method has a simple structure, but it often equates the numbering task with the detection task, lacks explicit modeling of the constraints on the dental arch structure, and simply regards the numbering problem as a detection problem, ignoring that FDI numbering is essentially an allocation problem with anatomical consistency constraints. Therefore, confusion is likely to occur in adjacent tooth positions, symmetrical tooth positions and local crowded areas, and there is a lack of explicit relationship correction mechanism. It only relies on local visual features, which makes it difficult to handle consistency problems such as duplicate bounding boxes, adjacent numbering conflicts, and artifact false detections, resulting in low accuracy.

[0030] (2) The second category is methods that use external rules for post-processing after target detection, such as multi-model cascading, fixed template matching, or alignment correction based on manual static rules. This type of method is effective in rule-based scenarios, but it is prone to problems such as error propagation, high system complexity, and limited generalization ability across datasets. It is particularly limited in the scenario of mixed dentition in children, where there are problems of overly strong or rigid hard rule constraints. If fixed order constraints are directly applied to mixed dentition in children, it may incorrectly punish real individual differences, eruption differences, and normal variations in mixed dentition in children, resulting in low accuracy.

[0031] To address the aforementioned issues, there is an urgent need for a new method that can both preserve the bounding boxes of weak targets and incorporate dental anatomical relationships to perform consistency correction on target detection results.

[0032] Considering that graph neural networks or relational modeling methods model structural priors through information propagation between nodes to improve prediction consistency, these methods show that explicit relational modeling is helpful for dental tasks. However, they are often designed for segmentation tasks, CBCT (Cone-Beam Computed Tomography) scenarios, or more regular spatial structures, making it difficult to directly transfer to the automatic tooth detection and FDI numbering tasks for mixed dentition in children based on curved tomographic images. In other words, while graph neural networks or relational modeling methods can explicitly introduce structural priors in medical imaging tasks and have the potential to improve prediction consistency, they mostly rely on segmentation masks or more regular spatial structures, making them unsuitable for automatic tooth detection and FDI numbering tasks for mixed dentition in children based on bounding boxes. This embodiment innovatively introduces graph reasoning into the automatic tooth detection and FDI numbering tasks to explicitly introduce structural priors, improve prediction consistency, and correct FDI numbering.

[0033] This embodiment uses artificial intelligence technology for oral medical image processing and proposes an automatic tooth detection and FDI number consistency correction model for children's mixed dentition panoramic radiographs. The automatic tooth detection and FDI numbering tasks are decomposed into two decoupled but complementary stages. The first stage (detection stage) generates tooth bounding boxes and FDI numbers with high recall as the goal. The second stage (graph reasoning stage) performs explicit relation reasoning and number consistency correction on the tooth bounding boxes under the constraints of dentition anatomical relationships, thereby improving the stability and clinical usability of the final FDI numbering results and increasing accuracy.

[0034] Specifically, this embodiment proposes a two-stage tooth numbering framework with decoupled perception and reasoning. In the first stage, a Faster R-CNN-based object detection model (or other types of object detection models) is used to generate high-recall tooth bounding boxes and FDI numbers. In the second stage, the tooth bounding boxes are constructed into a multi-relationship heterogeneous graph. A graph neural network is used to jointly model the overlapping competition relationships, arch adjacency relationships, same-jaw neighborhood relationships, and spatial proximity relationships between tooth bounding boxes. The final FDI number correction result is obtained through iterative reasoning. This achieves the separation of candidate retention and number correction functions in object detection, which can retain weak targets as much as possible and correct inconsistent numbers, thereby improving accuracy.

[0035] Based on this, this embodiment provides a training method for an automatic tooth detection and FDI number consistency correction model based on an artificial intelligence model for children with mixed dentition panoramic radiographs. In the model training stage, a two-stage processing approach is adopted: first, high-recall tooth bounding boxes and FDI numbers are generated through a target detection model; then, a multi-relationship heterogeneous graph is constructed based on the tooth bounding boxes, and inference correction is performed in combination with the anatomical relationship of the dentition, thereby improving the accuracy and consistency of the final FDI number correction result. The following is a detailed introduction to the automatic tooth detection and FDI number consistency correction model training method provided in this embodiment, including model construction method, module composition, connection relationship, training steps and training parameters.

[0036] (a) Sample Acquisition In this embodiment, multiple sample images and corresponding annotation data for each sample image are obtained. The sample images are panoramic tomographic images of mixed dentition in children, and the annotation data includes the true bounding box and true FDI number of each tooth in the sample image.

[0037] In this embodiment, data is first read. The panoramic tomographic images of children with mixed dentition and their corresponding true bounding boxes (which can be minimum adjacent rectangles) and true FDI numbers are read. The annotation data of each panoramic tomographic image of children with mixed dentition is obtained by a professional dentist who carefully and meticulously annotates the image by hand, and the minimum adjacent rectangle and FDI number of the teeth are manually annotated.

[0038] In this embodiment, the panoramic tomographic images of children with mixed dentition are used as sample images. The sample images can be denoised, grayscale normalized, and contrast enhanced. Denoising can be achieved by bilateral filtering, grayscale normalization can be achieved by max-min normalization, and contrast enhancement can be achieved by CLAHE adaptive histogram equalization.

[0039] This embodiment can also divide the sample images into training, validation, and test sets according to the patient level (whether it is pathological tooth loss) to avoid data leakage. Alternatively, the sample images can be directly divided into training, validation, and test sets, and the ratio of the training, validation, and test sets can be customized by the user.

[0040] The automatic tooth detection and FDI number consistency correction model in this embodiment adopts a two-stage cascaded structure, including a candidate detection module and a relation inference module. The candidate detection module performs tooth target detection on the input panoramic tomographic images of mixed dentition in children, outputting the predicted bounding boxes of the teeth, the predicted FDI numbers, and the visual features of the corresponding regions. The relation inference module receives the output from the candidate detection module, infers and corrects the predicted FDI numbers based on the anatomical relationships between the tooth targets, and outputs the predicted bounding boxes of the teeth and their corresponding corrected FDI numbers. The output of the former module serves as the input to the latter, forming a cascaded connection.

[0041] (II) Target Detection In this embodiment, the sample image is used as input, and the target detection module is used to determine the predicted bounding box, FDI number prediction result and visual features of each tooth in the sample image.

[0042] The object detection process is completed by the candidate detection module, which can use any existing object detection model. The object detection model is used to detect teeth in the input sample image and outputs the predicted bounding box of the teeth, the predicted FDI number, and visual features.

[0043] The candidate detection module uses Faster R-CNN as the object detection model, including a backbone feature extraction unit, a feature fusion unit, a candidate region generation unit, a candidate region classification and regression unit, and a candidate region feature extraction unit. The backbone feature extraction unit uses ResNet-50, and the feature fusion unit uses a Feature Pyramid Network (FPN) to extract multi-scale features. The backbone feature extraction unit extracts features from the sample images to obtain a shared feature map. The feature fusion unit generates multi-scale features based on the shared feature map. The candidate region generation unit generates preliminary candidate boxes based on the multi-scale features. The candidate region classification and regression unit generates the final predicted bounding boxes and FDI number prediction results based on the preliminary candidate boxes. The candidate region feature extraction unit generates the visual features corresponding to the predicted bounding boxes. The candidate detection module outputs the predicted bounding boxes, FDI number prediction results, and the visual features required for subsequent graph inference, while ensuring the integrity of the candidate targets.

[0044] (III) Graphical Reasoning In this embodiment, the predicted bounding box of each tooth in the sample image is treated as a graph node. Based on the anatomical structure of the teeth, selective connecting edges are drawn between two graph nodes to obtain an initial multi-relationship heterogeneous graph. The types of connecting edges include overlapping competing edges, dental arch adjacency edges, same-jaw neighborhood edges, and spatial proximity edges. For each graph node and each connecting edge in the initial multi-relationship heterogeneous graph, the predicted bounding box, FDI number prediction result, and visual features corresponding to the graph node are used as input. The first encoding module determines the node features of the graph node and assigns node features to the graph node. Based on the predicted bounding boxes of the two graph nodes connected by the connecting edge, the geometric attributes of the connecting edge are determined. Using the type and geometric attributes of the connecting edge as input, the second encoding module determines the edge attributes of the connecting edge and assigns edge attributes to the connecting edge, resulting in an intermediate multi-relationship heterogeneous graph. Using the intermediate multi-relationship heterogeneous graph as input, the graph inference module determines the final multi-relationship heterogeneous graph. Using the final multi-relationship heterogeneous graph as input, the classification module determines the FDI number correction result for each tooth in the sample image.

[0045] The graph inference process is completed by the relation inference module, which can adopt a recursive graph attention network structure, including a node construction unit, a node feature encoding unit, a relation edge construction unit, a graph inference unit, a recursive update unit, and a classification output unit. The node construction unit maps each predicted bounding box output by the candidate detection module to a graph node. The node feature encoding unit encodes visual features, geometric features, and prior features respectively, and then concatenates them to obtain the node features of the graph node. The relation edge construction unit establishes overlapping competition relationships, dental arch adjacency relationships, same-jaw neighborhood relationships, and spatial proximity relationships between graph nodes, draws connecting edges, and determines the edge attributes of the connecting edges, including overlapping competition edges, dental arch adjacency edges, same-jaw neighborhood edges, and spatial proximity edges. The graph inference unit performs multi-relation message passing based on node features and edge attributes to update the multi-relation heterogeneous graph in the sample. The recursive update unit determines whether the number of updates has reached the preset number after each round of updates. If not, it sends the data to the graph inference unit for updating again. The classification output unit outputs the final FDI number correction result corresponding to each graph node.

[0046] Based on the aforementioned relational reasoning module, this embodiment proposes a multi-relational heterogeneous graph construction method for children with mixed dentition. The method maps the predicted bounding boxes output by the candidate detection module to graph nodes and fuses three types of features for each graph node: visual features extracted by the candidate detection module; prior features consisting of the original score (i.e., logits in the FDI number prediction result) and predicted probability (i.e., the predicted probability vector in the FDI number prediction result, which is obtained by normalizing logits and includes the predicted probability of each category); and geometric features consisting of the center point coordinates, area, aspect ratio, and relative arch center position encoding of the predicted bounding box output by the candidate detection module. These results in the node features of the graph node. Simultaneously, various types of connecting edges are constructed, including overlapping competing edges, arch adjacency edges, same-jaw neighborhood edges, and spatial proximity edges. Geometric attributes are further added to the connecting edges, including center offset, Euclidean distance, scale ratio, and overlap. The type and geometric attributes of the connecting edges are fused to obtain the edge attributes, enabling explicit coupling of relational strength and geometric structure during subsequent graph reasoning.

[0047] The calculation method for the relative dental arch center position encoding includes: first, calculating the average value of the center point coordinates of all predicted bounding boxes to obtain the geometric center point coordinates, using the geometric center point as the approximate center of the dental arch region in the entire image; for each predicted bounding box, calculating the distance between the center point coordinates of the predicted bounding box and the geometric center point coordinates (Euclidean distance can be used), and normalizing the distance to obtain the relative dental arch center position encoding of the predicted bounding box, which is the normalized offset of the center point of the predicted bounding box relative to the geometric center point.

[0048] In the construction of the multi-relationship heterogeneous graph, each predicted bounding box output from the detection stage is mapped to a graph node, and an initial representation consisting of visual features, prior features, and geometric features is constructed for each graph node to obtain the node features of each graph node. Various types of connecting edges are constructed between graph nodes, including overlapping competition edges, dental arch adjacency edges, same-jaw neighborhood edges, and spatial proximity edges, and corresponding geometric attributes are attached to each connecting edge to characterize the anatomical structural relationship between candidate targets. The type and geometric attributes of the connecting edges are fused to obtain the edge attributes of each connecting edge.

[0049] When determining graph nodes, the predicted bounding box of each tooth in the sample image is treated as a graph node. The number of graph nodes is the same as the number of predicted bounding boxes, and this is accomplished by node building units.

[0050] When determining the connecting edges, connecting edges are selectively drawn between two graph nodes based on the anatomical relationship of the teeth, which is accomplished by relation edge building units.

[0051] At this point, based on the anatomical relationship of the teeth, selectively drawing connecting edges between the two graph nodes, specifically including: For each graph node, perform the following steps: (1) Record the graph node as the first graph node and the other graph nodes besides the first graph node as the second graph node.

[0052] (2) For each second graph node, based on the predicted bounding box corresponding to the first graph node and the predicted bounding box corresponding to the second graph node, the cross-union ratio between the first graph node and the second graph node is calculated, and the second graph node with a cross-union ratio greater than the preset cross-union ratio threshold is selected as the first target graph node, and overlapping competition edges are drawn between the first graph node and the first target graph node.

[0053] Set of overlapping competing edges It can be determined based on the intersection-over-union ratio (IoU): ; in, For graph nodes Graph Nodes The set consists of two elements, and overlapping competing edges need to be drawn between them; For graph nodes Graph Nodes The crossover ratio between them; This is the preset crossover-union ratio threshold.

[0054] The intersection-over-union ratio (IoS) can be a smaller frame IoS, in which case the set of overlapping competing edges... It can be determined based on a smaller frame intersection-union ratio: ; in, For graph nodes Graph Nodes The smaller crossover ratio between the boxes; A preset threshold for smaller cross-union ratio is set.

[0055] (3) Second graph nodes that are not connected to first graph nodes through overlapping competing edges are denoted as third graph nodes. For each third graph node, the first distance between the first graph node and the third graph node is calculated based on the predicted bounding box corresponding to the first graph node and the predicted bounding box corresponding to the third graph node. The third graph nodes are sorted in ascending order of the first distance, and the top three nodes are selected. A third graph node is used as a second target graph node, and dental arch adjacent edges are drawn between the first graph node and the second target graph node.

[0056] The first distance can be calculated in anisotropic scaling space, and the formula for calculating the first distance is: ; in, This is the first distance; This is the lateral distance penalty coefficient; The x-coordinate difference between the predicted bounding box corresponding to the node in the first image and the predicted bounding box corresponding to the node in the third image. The x-coordinate can be the x-coordinate of the center point. This is the longitudinal distance penalty coefficient; This is the difference in the ordinate of the predicted bounding box corresponding to the node in the first image and the predicted bounding box corresponding to the node in the third image. The ordinate can be the ordinate of the center point.

[0057] Set of adjacent edges of the dental arch Before selecting in anisotropic scaling space The nearest neighbor nodes (excluding self-loops) ) and overlapping pairs (i.e. ): ; in, For graph nodes Graph Nodes For the set that is formed, dental arch adjacent edges need to be drawn between the two; To exclude self-loops and overlapping pairs, the first closest front Each graph node A set of.

[0058] (4) Based on the predicted bounding box corresponding to the first image node, determine the jawbone label of the first image node. For each second image node, based on the predicted bounding box corresponding to the second image node, determine the jawbone label of the second image node. Select the second image node whose jawbone label is the same as the jawbone label of the first image node as the fourth image node. For each fourth image node, based on the predicted bounding box corresponding to the first image node and the predicted bounding box corresponding to the fourth image node, calculate the second distance between the first image node and the fourth image node. Sort the fourth image nodes in ascending order of the second distance and select the first one. The fourth graph node is used as the third target graph node, and the same-jaw neighborhood edge is drawn between the first graph node and the third target graph node.

[0059] The method for determining the jawbone label includes: calculating the median threshold, which is the average of the maximum and minimum ordinates. The maximum ordinate is the maximum value of the ordinates of all predicted bounding boxes, and the minimum ordinate is the minimum value of the ordinates of all predicted bounding boxes. For a given predicted bounding box, if the ordinate of the predicted bounding box is greater than the median threshold, then the jawbone label is maxilla; otherwise, the jawbone label is mandible.

[0060] The second distance can be the Euclidean distance.

[0061] Set of adjacent edges of the same jaw It is to select the anterior part of the same jawbone region. The nearest neighbor nodes: ; in, For graph nodes Graph Nodes For the set that is composed of two parts, it is necessary to draw the same jaw neighborhood edges between them; For the jawbone label, exclude the second closest anterior distance from the ring. Each graph node A set; Graph nodes based on ordinate partitioning Jawbone label; Graph nodes based on ordinate partitioning Jawbone label.

[0062] (5) For each second graph node, based on the predicted bounding box corresponding to the first graph node and the predicted bounding box corresponding to the second graph node, calculate the third distance between the first graph node and the second graph node, sort the second graph nodes in ascending order of the third distance, and select the first... The second graph node is used as the fourth target graph node, and spatial proximity edges are drawn between the first graph node and the fourth target graph node.

[0063] The third distance can be the Euclidean distance.

[0064] Set of spatially adjacent edges It is to select the front in Euclidean space The nearest neighbor nodes: ; in, For graph nodes Graph Nodes For a given set, spatial proximity edges need to be drawn between the two elements; To exclude the third closest front of the self-loop Each graph node A set of.

[0065] To integrate the model with panoramic radiographs of children with mixed dentition, this embodiment constructs a multi-relationship heterogeneous graph based on the output of the candidate detection module. The graph construction parameters are set as follows: overlap and competition relationship threshold. , Dental arch adjacency relation nearest neighbor number Nearest neighbor number of the same jaw region Spatial proximity relationship, nearest neighbor number In distance metrics, the longitudinal distance penalty coefficient =3.0, Lateral Distance Penalty Coefficient =5.0, in this embodiment, the candidate node filtering confidence threshold during graph construction can be further set. That is, only when the confidence level of the predicted bounding box is greater than 1. Only then will the predicted bounding box be used as a graph node. Through these parameter settings, the graph structure can better reflect the relative positional relationships and local anatomical structures between teeth in a child's mixed dentition.

[0066] After determining the graph nodes and connecting edges, the initial multi-relation heterogeneous graph of the sample can be obtained.

[0067] For each graph node in the initial multi-relation heterogeneous graph of the samples, when determining the node features, the predicted bounding box, FDI number prediction result and visual features corresponding to the graph node are used as inputs, and the first encoding module is used to determine the node features of the graph node, which is completed by the node feature encoding unit.

[0068] At this point, using the predicted bounding box corresponding to the graph node, the FDI number prediction result, and visual features as input, the first encoding module is used to determine the node features of the graph node, specifically including: (1) Based on the predicted bounding box corresponding to the graph node, determine the geometric features corresponding to the graph node. The geometric features include the center point coordinates, area, aspect ratio and relative dental arch center position encoding of the predicted bounding box corresponding to the graph node.

[0069] (2) Based on the prediction results of the FDI number corresponding to the graph node, determine the prior features corresponding to the graph node. The prior features include the original score and predicted probability of the FDI number of the tooth corresponding to the graph node belonging to each category.

[0070] (3) The visual features, geometric features and prior features corresponding to the graph nodes are encoded and spliced ​​using the first encoding module to obtain the node features of the graph nodes.

[0071] The first encoding module includes a first encoding unit, a second encoding unit, a third encoding unit, and a splicing layer. The outputs of the first encoding unit, the second encoding unit, and the third encoding unit are all connected to the input of the splicing layer. The first encoding unit, the second encoding unit, and the third encoding unit have the same structure. The first encoding unit includes a first linear layer, a first ReLU activation function layer, and a first normalization layer connected in sequence.

[0072] The first encoding unit is used to encode the visual features corresponding to the graph node to obtain the first encoded feature. The second encoding unit is used to encode the geometric features corresponding to the graph node to obtain the second encoded feature. The third encoding unit is used to encode the prior features corresponding to the graph node to obtain the third encoded feature. The concatenation layer is used to concatenate the first encoded feature, the second encoded feature and the third encoded feature to obtain the node feature of the graph node.

[0073] Node features are determined based on visual features, geometric features, and prior features. The node features of a graph node are as follows: ; in, For graph nodes Node characteristics; This is the mapping function for visual features, i.e., the first coding unit; For graph nodes Visual features; This refers to the feature splicing operation, i.e., the splicing layer; The mapping function for prior features is the third coding unit; For graph nodes Prior features; This is a mapping function for geometric features, i.e., the second coding unit; For graph nodes Geometric features.

[0074] The first coding unit, also known as a visual encoder, has a structure consisting of a first linear layer. +First ReLU activation function layer +First Normalization Layer The second coding unit, also known as a geometric encoder, has a structure consisting of the first linear layer. +First ReLU activation function layer +First Normalization Layer The third coding unit, also known as the prior encoder, has a structure consisting of the first linear layer. +First ReLU activation function layer +First Normalization Layer , The number of categories for FDI numbering. The three types of features are encoded by three coding units and then concatenated to form a 384-dimensional node feature, which includes the first coding feature corresponding to 128-dimensional visual features, the second coding feature corresponding to 128-dimensional geometric features, and the third coding feature corresponding to 128-dimensional prior features.

[0075] For each connection edge in the initial multi-relation heterogeneous graph of the samples, when determining the edge attributes, the geometric attributes of the connection edge are determined based on the predicted bounding boxes corresponding to the two graph nodes connected by the connection edge. The edge attributes of the connection edge are determined by the second encoding module using the type and geometric attributes of the connection edge as input, which is completed by the relation edge construction unit.

[0076] At this point, the geometric properties of the connecting edge are determined based on the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge. Using the type and geometric properties of the connecting edge as input, the edge attributes of the connecting edge are determined using the second encoding module, specifically including: (1) Based on the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge, determine the geometric properties of the connecting edge, including center offset, Euclidean distance, scale ratio and overlap.

[0077] The center offset includes the difference in the x-coordinate and y-coordinate of the center points of the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge. The Euclidean distance is the Euclidean distance between the center points of the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge. The scale ratio includes at least one of the width ratio, height ratio, and area ratio of the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge. The overlap is the intersection-union ratio of the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge.

[0078] (2) Use the second encoding module to encode the type and geometric properties of the connecting edge to obtain the edge properties of the connecting edge.

[0079] The second encoding module includes a second linear layer, a second ReLU activation function layer, and a second normalization layer connected in sequence.

[0080] The second encoding module, also known as the edge attribute encoder, consists of two second linear layers. +Second ReLU activation function layer +Second normalization layer .

[0081] After determining the node features and edge attributes, each graph node in the initial multi-relation heterogeneous graph of the sample is assigned node features, and each connecting edge in the initial multi-relation heterogeneous graph of the sample is assigned edge attributes, thus obtaining the intermediate multi-relation heterogeneous graph of the sample.

[0082] This embodiment proposes a graph inference method with edge attribute constraints and a recursive update mechanism. In the graph inference stage, attention weights are calculated for each type of connection edge, enabling the model to dynamically determine whether information should be propagated and the propagation intensity based on the relative position, overlap, and structural relationship between the predicted bounding boxes. The inference is performed recursively. In each step, the prior input for the next step is updated using the current prediction result (i.e., the multi-relation heterogeneous graph of the currently obtained sample is updated and input into the classification module to obtain the intermediate FDI correction result, which is then used as the FDI prediction result to reconstruct the sample intermediate multi-relation heterogeneous graph and infer again). This forms a closed-loop correction process of "initial prediction - relation verification - confidence update - re-inference" to gradually improve the consistency of FDI numbering.

[0083] After obtaining the sample intermediate multi-relation heterogeneous graph, the sample intermediate multi-relation heterogeneous graph is used as input and the graph reasoning module is used to determine the sample multi-relation heterogeneous graph, which is completed by the graph reasoning unit and the recursive update unit.

[0084] The graph inference module includes a graph inference unit and a recursive update unit connected in sequence. The graph inference unit is used for multi-relation message passing, updating the node features of graph nodes in the intermediate multi-relation heterogeneous graph of the sample, and obtaining the updated multi-relation heterogeneous graph of the sample. The recursive update unit is used to determine whether the number of updates has reached the preset number. If not, the updated multi-relation heterogeneous graph of the sample is input into the classification module to obtain the intermediate FDI number correction result. The intermediate FDI number correction result is used as the FDI number prediction result. Then, for each graph node and each connecting edge in the initial multi-relation heterogeneous graph of the sample, the predicted bounding box corresponding to the graph node, the FDI number prediction result, and the... Visual features are used as input. The first encoding module determines the node features of the graph nodes and assigns node features to the graph nodes. Based on the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge, the geometric attributes of the connecting edge are determined. The type and geometric attributes of the connecting edge are used as input. The second encoding module determines the edge attributes of the connecting edge and assigns edge attributes to the connecting edge. The "sample intermediate multi-relation heterogeneous graph" is obtained to construct the sample intermediate multi-relation heterogeneous graph. The constructed sample intermediate multi-relation heterogeneous graph is used as the sample intermediate multi-relation heterogeneous graph for the next update and sent to the graph inference unit for another update. If it is updated, the sample multi-relation heterogeneous graph is used as the sample multi-relation heterogeneous graph.

[0085] The graph reasoning unit includes multiple graph reasoning sub-units connected in sequence. Each graph reasoning sub-unit includes an RGATConv layer (Relational Graph Attention Convolution), an addition layer, and a third normalization layer connected in sequence. The input of the addition layer is also connected to the input of the RGATConv layer.

[0086] As an example, there are two graph inference subunits. In this case, the graph inference unit adopts two RGATConv layers, each layer includes eight attention heads, and each attention head outputs 48 channels. After concatenation, the output dimension of each layer is 384. Each RGATConv layer is followed by a residual connection (i.e., an additive layer) and a third normalization layer LayerNorm. The dropout of the RGATConv layer is 0.2.

[0087] During the graph reasoning phase, conditional message passing is performed by combining node features and edge attributes, for relation types (i.e., the types of connecting edges). The attention weights of the connecting edges below can be expressed as: ; in, For graph nodes Graph Nodes Relationship types The connecting edge below is in the first Attention weight during the next update; For graph nodes The neighborhood under relation type r includes nodes in the graph. By relation type The connecting edge below connects all graph nodes. ; This is the first learnable parameter; This is the second learnable parameter; For graph nodes In the Node characteristics at the time of the next update; For graph nodes In the Node characteristics at the time of the next update; This is the third learnable parameter; For graph nodes Graph Nodes Relationship types The edge properties of the connecting edges below.

[0088] For each graph node, based on attention weights, the node features of all neighboring graph nodes are weighted and aggregated, and then fused with the node features of the node itself to update the node features of the graph node.

[0089] Through multi-round graph reasoning, the node features of graph nodes are iteratively updated, thereby iteratively updating the predicted category of candidate targets.

[0090] After obtaining the sample multi-relationship heterogeneity map, the FDI number correction result of each tooth in the sample image is determined by the classification module using the sample multi-relationship heterogeneity map as input, which is completed by the classification output unit.

[0091] The classification module consists of a third linear layer, a third ReLU activation function layer, a Dropout layer, a fourth linear layer, and a Softmax activation function layer connected in sequence.

[0092] The classification module, also known as a classifier, has a structure consisting of a third linear layer. +Third ReLU activation function layer +Dropout layer +Fourth linear layer (The output at this point is logits) + Softmax activation function layer (The output at this point is the predicted probability vector).

[0093] (iv) Loss Calculation In this embodiment, the total loss is calculated based on the true bounding box, true FDI number, predicted bounding box, and FDI number correction result of each tooth in the sample image.

[0094] The total loss is calculated based on the true bounding box, true FDI number, predicted bounding box, and FDI number correction results for each tooth in the sample image. Specifically, it includes: (1) The detection classification loss is calculated based on the true bounding box, true FDI number, predicted bounding box and FDI number correction results of each tooth in the sample image.

[0095] (2) Based on the true bounding box, true FDI number, predicted bounding box and FDI number correction results of each tooth in the sample image, the corrected classification loss, repeated detection loss and adjacent number conflict loss are calculated respectively. The corrected classification loss, repeated detection loss and adjacent number conflict loss are weighted and summed to obtain the graph reasoning loss.

[0096] (3) Combine the detection classification loss and graph reasoning loss into the total loss.

[0097] Both existing methods have the problem of insufficient preservation of weak targets such as deciduous teeth. In children's mixed dentition, deciduous teeth are often accompanied by physiological root resorption, and the appearance of the root area is unstable, which makes it easier for target detection models based on uniform classification loss to miss deciduous teeth.

[0098] Based on this, this embodiment proposes a category-conditional space-weighted loss to address the problem of missed detection of deciduous teeth. Addressing the issue of unstable root regions in deciduous teeth during mixed dentition in children due to physiological root resorption, leading to frequent missed detections, this embodiment introduces a category-conditional space-weighted loss during the detection phase. This loss only reweights the classification loss for the positive sample regions of the deciduous tooth category. Depending on whether the center of the bounding box is located in the crown region or the root region, it assigns either an increased or decreased weight, thereby enhancing the contribution of stable crown regions to training, reducing gradient interference from unstable root regions, and lowering the probability of missed detection of deciduous teeth.

[0099] To enhance the retention of deciduous teeth in children's mixed dentition, a category-conditional space-weighted loss strategy is introduced during the detection phase. Let the... The ordinate of the upper boundary of each true bounding box is The height is The ordinate of the boundary line between the crown and root is defined as: ; in, The vertical coordinate represents the boundary line between the crown and root of the tooth. This is the preset scaling factor.

[0100] Let the ordinate of the center point of the predicted bounding box be... The weights for predicting the bounding box are then defined as: ; in, For the sample image, the first The weight of each tooth; For the sample image, the first The initial weight of each tooth; This is the weighting coefficient for the crown region, which can be adjusted upwards; The true label is the first element in the sample image. The true probability vector of the tooth, including the first tooth The true probability of a tooth’s FDI number belonging to each category is determined based on the true FDI number. If a category is a true FDI number, then the true probability of that category is 1, and the true probability of other categories is 0. A set of deciduous teeth categories; For the sample image, the first The ordinate of the upper boundary of the predicted bounding box of each tooth; For the sample image, the first The vertical coordinate of the boundary between the crown and root of a tooth; This is the weighting coefficient for the root region, and the weight can be adjusted downwards.

[0101] Introducing the aforementioned weights into the classification loss yields the classification loss function for the detection phase, which is the category conditional space weighted loss (i.e., the detection classification loss). The formula for calculating this detection classification loss is: ; in, To detect classification loss; The normalization coefficient is the normalization term, which is equal to the sum of the weights of all teeth. The number of teeth in the sample image; For the sample image, the first The weight of each tooth; The cross-entropy loss function; For classification output, i.e., the first element in the sample image... The predicted probability vector of each tooth, including the first tooth. The predicted probability that the FDI number of each tooth belongs to each category is determined based on the FDI number correction result; The true label is the first element in the sample image. The true probability vector of the tooth, including the first tooth The true probability that an FDI number of a tooth belongs to each category is determined based on the true FDI number.

[0102] When calculating the loss, it is necessary to match the predicted bounding box and the ground truth bounding box to determine the first... The actual bounding box, actual FDI number, predicted bounding box, and FDI number correction results for each tooth are used. The specific matching process uses existing mature technologies, such as matching based on crossover ratio, which will not be elaborated here.

[0103] In the detection phase, a category conditional space weighted loss is introduced, where the parameters are set to... , , This method can improve the perception and retention of weak targets such as deciduous teeth.

[0104] This embodiment proposes a joint loss optimization strategy for FDI numbering consistency. During the graph inference stage, a joint loss is used for training. The joint loss includes asymmetric Focal Loss, duplicate detection constraints, and adjacent numbering conflict constraints. The duplicate detection constraints suppress highly overlapping bounding boxes from being predicted as the same tooth position, while the adjacent numbering conflict constraints penalize anatomically unreasonable combinations of adjacent tooth position categories based on the FDI-perceived class mutual exclusion matrix. To avoid introducing topological consistency constraints (i.e., duplicate detection constraints and adjacent numbering conflict constraints) too early, which could lead to training instability, this embodiment further employs a warm-up strategy. In the early stages of training, only asymmetric Focal Loss is introduced to learn visual perception capabilities, and then topological consistency constraints are gradually introduced.

[0105] Specifically, to improve the consistency of the final numbering results, this embodiment adopts a training method that combines corrected classification loss and topological consistency constraints for joint optimization. The corrected classification loss is a class-balanced asymmetric Focal Loss (i.e., asymmetric Focal Loss), with the number of classes set to 48. The topological consistency constraints include duplicate detection constraints and adjacent numbering conflict constraints. The topological consistency constraints are introduced in a phased warm-up manner, with the weight coefficients set as follows: 0 for the 1st to 15th epochs (cycles, i.e., iterations), 0.1 for the 16th to 25th epochs, 0.5 for the 26th to 35th epochs, and 0.05 after the 35th epoch. This setting helps to prioritize learning basic visual perception capabilities through corrected classification loss in the early stages of training, and then gradually introduce topological consistency constraints after the model tends to stabilize.

[0106] To improve the consistency of the final numbering results, a joint loss function is used for optimization during training. In this case, the formula for calculating the graph inference loss is: ; in, For graph reasoning loss; To correct for classification loss, i.e., asymmetric Focal Loss; The first weighting coefficient; This is the repeated detection loss, i.e., the repeated detection constraint term; This is the second weighting coefficient; The adjacent number conflict loss is the adjacent number conflict constraint term. When the number of iterations is between 1 and 15, the first weight coefficient and the second weight coefficient are 0; when the number of iterations is between 16 and 25, the first weight coefficient and the second weight coefficient are 0.1; when the number of iterations is between 26 and 35, the first weight coefficient and the second weight coefficient are 0.5; when the number of iterations is greater than 35, the first weight coefficient and the second weight coefficient are 0.05.

[0107] It should be noted that 15, 25, and 35 can also be adjusted according to user needs.

[0108] The classification loss is corrected using asymmetric Focal Loss, calculated as follows: ; in, The number of teeth in the sample image; These are the class parameters used to balance the weights of positive and negative samples; To predict the probability of a sample image belonging to the true class, i.e., the probability of the sample image containing the first element belonging to the true class. The FDI number of each tooth is the predicted probability of the actual FDI number. Based on the FDI number correction results, the category with the highest predicted probability is selected as the FDI number of the tooth. This is an adjustment factor for easy and difficult samples.

[0109] The formula for calculating the repeated detection loss is: ; in, For connecting graph nodes Graph Nodes The attention weights of overlapping competing edges are dynamically learned during the graph reasoning stage, enabling the model to adaptively adjust the penalty intensity based on the reliability of edge relationships. To prevent the smoothing minimum coefficients where the denominator is zero; This is the set of overlapping competing edges; Let i and j be the posterior class probability distribution vectors of graph node i and graph node j, respectively, and then multiply them by a dot product. Similarity of penalty probability distributions The number of categories is C=48. For graph nodes The corresponding predicted probability vector for the teeth, including graph nodes. The predicted probability that the corresponding tooth's FDI number belongs to each category is determined based on the FDI number correction result; For graph nodes The corresponding predicted probability vector for the teeth, including graph nodes. The predicted probability of the corresponding tooth's FDI number belonging to each category is determined based on the FDI number correction result.

[0110] The formula for calculating the penalty for collisions between adjacent numbers is: ; in, For connecting graph nodes Graph Nodes The attention weights of the dental arch adjacent edges are dynamically learned during the graph reasoning stage. It is the set of adjacent edges of the dental arch; The category exclusivity matrix for FDI numbering, , The diagonal elements have a value of 0, and the off-diagonal elements have a value of 1.

[0111] (v) Iterative Updates In this embodiment, the target detection module, the first encoding module, the second encoding module, the graph reasoning module, and the classification module are updated using the total loss, resulting in the updated target detection module, the updated first encoding module, the updated second encoding module, the updated graph reasoning module, and the updated classification module. If the iteration termination condition is met, the updated target detection module, the updated first encoding module, the updated second encoding module, the updated graph reasoning module, and the updated classification module are combined to form an automatic tooth detection and FDI number consistency correction model.

[0112] The target detection module, the first encoding module, the second encoding module, the graph reasoning module, and the classification module are updated using the total loss. Specifically, the target detection module is updated using the detection and classification loss, and the first encoding module, the second encoding module, the graph reasoning module, and the classification module are updated using the graph reasoning loss.

[0113] During the update process, backpropagation is performed using the loss function proposed in this embodiment. The stochastic gradient descent algorithm can be used to optimize all modules. The base learning rate is set to 0.1, the weight decay factor is set to 1e-4, the momentum value is set to 0.9, the maximum number of iterations is set to 200, and the model is saved after training is completed.

[0114] The iteration termination condition is: the value of the loss function tends to stabilize. Training stops only when the value of the loss function stabilizes. After 200 iterations, if the value of the loss function in the last 10 iterations tends to stabilize (for example, the range of the loss function values ​​in the last 10 iterations can be less than a preset threshold), training can stop. Otherwise, additional iterations are needed until the value of the loss function stabilizes. The entire model structure and the values ​​of all parameters are saved. When the stopping condition is met, the model's structure information and all parameter values, including network weights, need to be saved in a parameter file for subsequent model testing and use.

[0115] Through the above training process, model parameters with strong candidate detection capabilities and tooth position relationship correction capabilities are obtained.

[0116] In this embodiment, the training parameters for the candidate detection module and the relation reasoning module can also be designed separately.

[0117] When training the candidate detection module, the optimizer uses SGD with a momentum parameter of 0.9 and a weight decay factor of [missing value]. The initial learning rate is 0.002, which decays to 0.1 times the original value in the 10th and 20th epochs respectively. The batch size is 8, the maximum number of training epochs is 100, and it usually converges around the 30th epoch.

[0118] When training the relational reasoning module, the AdamW optimizer is used, with a weight decay factor of 1. The initial learning rate is The learning rate scheduling strategy uses ReduceLROnPlateau, where mode is max, decay factor is 0.2, patience rounds are 6, and the minimum learning rate is [missing information]. The batch size is 6, the maximum number of training epochs is 200, the early stopping policy patience is 20 epochs, the decision to stop early is based on the clinical scores on the validation set, the gradient pruning limit is 1.0, and the number of inference iterations is 5.

[0119] In summary, the model training and saving process includes the following steps: reading the sample images and their labeled data used for training; preprocessing the sample images; inputting the sample images into the candidate detection module to obtain the predicted bounding boxes, FDI number prediction results, and visual features; constructing graph nodes and multi-relationship connection edges based on the output results of the candidate detection module; inputting the graph structure into the relation inference module for multiple rounds of message passing and recursive updates; calculating the loss in the detection stage and the loss in the graph inference stage respectively; performing backpropagation based on the loss function to update the model parameters; and saving the model parameter file when the validation set performance index reaches the optimal level or the stopping condition is met.

[0120] This embodiment involves not only the construction and training of an artificial intelligence model, but also its specific application in a pediatric mixed dentition panoramic radiograph scenario. Input data includes: sample images, i.e., panoramic radiograph images of pediatric mixed dentition; labeled data, i.e., the true bounding boxes of teeth and their corresponding true FDI numbers; intermediate feature data, i.e., the predicted bounding boxes, visual features, category logits (the original score for each category), category probability (the predicted probability for each category), and category confidence (the confidence level for each category) output by the candidate detection module, where category logits, category probability, and category confidence constitute the FDI number prediction result; and structural relationship data, i.e., the graph structure (including graph nodes and connecting edges) constructed based on the positional, overlapping, and adjacency relationships between the predicted bounding boxes, as well as the node features of the graph nodes and the edge attributes of the connecting edges. The sample images determine the target detection result (i.e., intermediate feature data), the target detection result further determines the construction method of the graph nodes and connecting edges, and the graph nodes and connecting edges jointly determine the final FDI number correction result. The output data includes: the predicted bounding box and FDI number correction result for each tooth. The FDI number correction result includes category logits, category probability, and category confidence. The above output data directly corresponds to the clinical scenario of mixed dentition in children: the predicted bounding box represents the spatial position of the tooth in the curved tomographic image, and the FDI number correction result determines the FDI number, representing the anatomical identity of the tooth in the dentition. Together, they constitute the structured tooth position analysis result. The inherent correlation between input and output is reflected in the following: the candidate detection module determines the candidate tooth position based on the local texture, edge, and tooth morphology information in the input sample image; the relationship reasoning module performs consistency correction on the initially predicted FDI number based on the positional relationship, overlap relationship, and adjacency relationship between candidate targets. The final output FDI number correction result is jointly determined by the local visual information of the candidate target and the overall structural relationship of the dentition, rather than solely by the appearance of a single local area. Therefore, this embodiment outputs tooth localization and FDI number results closely corresponding to the curved tomographic image scenario of mixed dentition in children, which can reflect the actual anatomical relationship of teeth in the specific dentition structure.

[0121] This embodiment can also perform model testing: The trained model parameter file is used to load the parameters, and the trained model parameter file is used to load the model. The panoramic tomographic image of the child's mixed dentition is read, and the image is processed using the same preprocessing method as in the training phase. First, the image is input into the candidate detection module to obtain the predicted bounding box of the tooth, the predicted FDI number, and visual features. Then, the output of the candidate detection module is input into the relationship inference module, and the predicted FDI number is corrected based on the anatomical relationship between the candidate targets. Finally, the predicted bounding box of the tooth and its corresponding corrected FDI number are output. Through the above testing process, automatic detection and numbering of tooth targets in panoramic tomographic images of children's mixed dentition can be achieved.

[0122] The automatic tooth detection and FDI number consistency correction model application method provided in this application embodiment can be applied to, for example... Figure 1 The application environment shown is illustrated. The terminal communicates with the server via a network. A data storage system stores the data the server needs to process. This data storage system can be set up independently, integrated into the server, or located in the cloud or on another server. The terminal can send application requests to be processed to the server. Upon receiving the application request, the server acquires the image to be processed. Using the image as input, it uses an updated object detection module to determine the predicted bounding box, FDI number prediction result, and visual features of each tooth in the image. Each predicted bounding box of a tooth in the image is treated as a graph node. Based on the anatomical structure of the teeth, selectively connecting edges are drawn between two graph nodes to obtain an initial multi-relationship heterogeneous graph. For each graph node and each connecting edge in the initial multi-relationship heterogeneous graph, using the predicted bounding box, FDI number prediction result, and visual features corresponding to the graph node as input, it uses an updated object detection module to determine the predicted bounding box, FDI number prediction result, and visual features of each tooth. The first encoding module determines the node features of the graph nodes and assigns node features to the graph nodes. Based on the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge, the geometric attributes of the connecting edge are determined. Using the type and geometric attributes of the connecting edge as input, the updated second encoding module determines the edge attributes of the connecting edge and assigns edge attributes to the connecting edge, obtaining an intermediate multi-relation heterogeneous graph. Using the intermediate multi-relation heterogeneous graph as input, the updated graph inference module determines the multi-relation heterogeneous graph. Using the multi-relation heterogeneous graph as input, the updated classification module determines the FDI number correction result for each tooth in the image to be processed. Based on the FDI number correction result for each tooth in the image to be processed, the FDI number of each tooth in the image to be processed is determined. The server can feed back the application result of the predicted bounding boxes and FDI numbers of each tooth in the image to be processed for the application request to the terminal.

[0123] Furthermore, in some embodiments, the automatic tooth detection and FDI number consistency correction model application method can also be implemented by the server or the terminal alone. For example, the terminal can directly process the application request to be processed, or the server can obtain the application request to be processed from the data storage system and process it.

[0124] In one exemplary embodiment, such as Figure 3 As shown, a method for applying an automatic tooth detection and FDI number consistency correction model is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 The following steps, T1 to T6, are used as an example to illustrate the process of using a server in the example.

[0125] Step T1: Obtain the image to be processed; the image to be processed is a panoramic tomographic image of a child's mixed dentition.

[0126] Step T2: Using the image to be processed as input, the updated object detection module is used to determine the predicted bounding box, FDI number prediction result, and visual features of each tooth in the image to be processed.

[0127] Step T3: The predicted bounding box of each tooth in the image to be processed is taken as a graph node. Based on the anatomical relationship of the tooth, connection edges are selectively drawn between two graph nodes to obtain the initial multi-relationship heterogeneous graph.

[0128] Step T4: For each graph node and each connecting edge in the initial multi-relation heterogeneous graph, the node features of the graph node are determined using the predicted bounding box, FDI number prediction result, and visual features corresponding to the graph node as input, and node features are assigned to the graph node using the updated first encoding module. The geometric attributes of the connecting edge are determined based on the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge. The edge attributes of the connecting edge are determined using the updated second encoding module as input, and edge attributes are assigned to the connecting edge, thus obtaining the intermediate multi-relation heterogeneous graph.

[0129] Step T5: Using the intermediate multi-relation heterogeneous graph as input, the updated graph inference module determines the multi-relation heterogeneous graph. Using the multi-relation heterogeneous graph as input, the updated classification module determines the FDI number correction result for each tooth in the image to be processed. The updated target detection module, the updated first encoding module, the updated second encoding module, the updated graph inference module, and the updated classification module constitute the automatic tooth detection and FDI number consistency correction model. The automatic tooth detection and FDI number consistency correction model is a model trained using the above-mentioned automatic tooth detection and FDI number consistency correction model training method.

[0130] Step T6: Based on the FDI number correction result of each tooth in the image to be processed, determine the FDI number of each tooth in the image to be processed.

[0131] This embodiment discloses an automatic tooth detection and FDI numbering consistency correction method for curved tomographic images of mixed dentition in children. The method divides the automatic tooth detection and FDI numbering tasks into a detection phase and a graph inference phase. In the detection phase, a target detection model is used to generate predicted bounding boxes and FDI number predictions for high-recall teeth, and the retention rate of primary teeth candidates is improved through category conditional space weighted loss. In the graph inference phase, the predicted bounding boxes are constructed into a multi-relational heterogeneous graph that integrates visual features, prior features, and geometric features. A recursive graph inference mechanism with edge attribute constraints is used to model the overlap competition, arch adjacency, same-maxillary neighborhood, and spatial proximity relationships between predicted bounding boxes. Furthermore, a joint loss based on FDI perception is combined to suppress duplicate detections and adjacent numbering conflicts, thereby outputting more consistent FDI numbering correction results. This embodiment can improve the accuracy and consistency of tooth position numbering in complex scenarios of mixed dentition in children.

[0132] This embodiment proposes a two-stage automatic tooth detection and numbering method for curved tomographic radiographs of children with mixed dentition. This method divides the automatic tooth detection and FDI numbering process into a detection stage and a graph reasoning stage. First, candidate teeth are obtained, and then the numbering of the candidate teeth is corrected by combining the dentition structure relationship, thereby improving the accuracy and consistency of tooth numbering in children with mixed dentition. It is an automatic tooth detection and FDI numbering method based on the combination of candidate detection and relationship correction. That is, the candidate detection module generates candidate tooth results, and the relationship reasoning module further corrects the FDI numbering consistency of the candidate tooth results. This paper proposes a detection optimization mechanism suitable for weak target retention in children's mixed dentition. Addressing the issue of missed detection of deciduous teeth in curved tomographic images due to root resorption and incomplete local morphology, a category-conditional space-weighted loss strategy is introduced to enhance the target detection model's ability to perceive and retain deciduous teeth, providing more complete candidate inputs for subsequent numbering inference. This is a detection loss optimization method for deciduous teeth in children's mixed dentition, which differentiates the weighting of the loss function in the detection stage based on tooth category characteristics and candidate region spatial distribution characteristics to improve the problem of missed detection of deciduous teeth. Furthermore, a tooth position consistency inference mechanism integrating anatomical structural relationships is proposed. Based on candidate tooth targets, it combines overlapping, adjacency, and spatial relationships between candidate tooth targets to model relationships, and optimizes the candidate numbering results through iterative inference and consistency constraints, thereby reducing duplicate detection and confusion between adjacent tooth positions. This is a tooth position numbering inference method based on multi-relationship structure modeling, which uses candidate bounding boxes as graph nodes and combines various anatomical structural relationships between graph nodes for information transmission, result updating, and consistency constraints, thus achieving a technical solution for tooth numbering optimization.

[0133] Compared with related technologies, this embodiment has at least the following beneficial effects: (1) To address the problem that tooth detection and numbering in children’s mixed dentition panoramic radiographs are easily affected by changes in deciduous tooth morphology, complex dentition structure and local artifacts, a technical solution combining candidate detection and relation correction is proposed, which can effectively improve the accuracy and consistency of automatic tooth numbering results.

[0134] (2) By optimizing the detection stage, the ability to perceive and retain weak target teeth in children's mixed dentition has been enhanced, providing a more complete and reliable candidate basis for subsequent numbering and correction.

[0135] (3) By introducing the anatomical relationship of the dental arch for reasoning optimization, it is possible to effectively reduce problems such as repeated testing, confusion of adjacent tooth positions and local misjudgment, thereby improving the stability and practicality of automatic tooth analysis results in complex clinical scenarios.

[0136] (4) It can provide technical support for the structured analysis, tooth position interpretation and related clinical applications of panoramic radiographs of mixed dentition in children, and has good application value.

[0137] The model in this embodiment can also be a deep learning model.

[0138] This application also provides an application scenario in which the above-described automatic tooth detection and FDI number consistency correction model application method is applied. Specifically, the automatic tooth detection and FDI number consistency correction model application method provided in this embodiment can be applied in a tooth detection scenario. The tooth detection scenario includes a detection stage and a display stage. The detection stage is used to determine the predicted bounding box and FDI number of the tooth, and the display stage is used to show the predicted bounding box and FDI number of the tooth to the dentist for use. The automatic tooth detection and FDI number consistency correction model application method provided in this embodiment belongs to the detection stage.

[0139] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 4 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements either a method for training an automatic tooth detection and FDI number consistency correction model or an application method for the same model.

[0140] Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0141] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0142] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0143] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0144] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations and be authorized by the owner of the corresponding device.

[0145] Any references to memory, databases, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0146] The databases involved in the various embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases.

[0147] The processors involved in the various embodiments provided in this application may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited thereto.

[0148] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0149] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for training a tooth automatic detection and FDI number consistency correction model, characterized in that, include: Acquire multiple sample images and the corresponding annotation data for each sample image; The sample images are panoramic tomographic images of children with mixed dentition. The annotation data includes the true bounding box and true FDI number of each tooth in the sample image. Using sample images as input, the target detection module determines the predicted bounding box, FDI number prediction result, and visual features of each tooth in the sample image. Each predicted bounding box of a tooth in the sample image is treated as a graph node. Based on the anatomical relationship of the tooth, connection edges are selectively drawn between two graph nodes to obtain the initial multi-relationship heterogeneous graph of the sample. The types of connection edges include overlapping competition edges, dental arch adjacency edges, same-jaw neighborhood edges, and spatial proximity edges. For each graph node and each connecting edge in the initial multi-relation heterogeneous graph of the sample, the predicted bounding box corresponding to the graph node, the prediction result of the FDI number, and the visual features are used as inputs. The first encoding module is used to determine the node features of the graph node and assign node features to the graph node. Based on the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge, the geometric attributes of the connecting edge are determined. The type and geometric attributes of the connecting edge are used as inputs. The second encoding module is used to determine the edge attributes of the connecting edge and assign edge attributes to the connecting edge, thus obtaining the intermediate multi-relation heterogeneous graph of the sample. Using the sample intermediate multi-relation heterogeneity graph as input, the graph reasoning module is used to determine the sample multi-relation heterogeneity graph. Using the sample multi-relation heterogeneity graph as input, the classification module is used to determine the FDI number correction result of each tooth in the sample image. The total loss is calculated based on the true bounding box, true FDI number, predicted bounding box, and FDI number correction results for each tooth in the sample image. The target detection module, the first encoding module, the second encoding module, the graph reasoning module, and the classification module are updated using the total loss, resulting in updated target detection module, updated first encoding module, updated second encoding module, updated graph reasoning module, and updated classification module. If the iteration termination condition is met, the updated target detection module, updated first encoding module, updated second encoding module, updated graph reasoning module, and updated classification module are combined to form an automatic tooth detection and FDI number consistency correction model.

2. The tooth automatic detection and FDI number consistency correction model training method according to claim 1, characterized in that, Based on the anatomical relationships of teeth, selectively draw connecting edges between two graph nodes, specifically including: For each graph node, perform the following steps: The graph node is designated as the first graph node, and all other graph nodes are designated as the second graph nodes. For each second graph node, based on the predicted bounding box corresponding to the first graph node and the predicted bounding box corresponding to the second graph node, the intersection-union ratio (IUR) between the first graph node and the second graph node is calculated, and the second graph node with an IUR greater than a preset IUR threshold is selected as the first target graph node. Overlapping competing edges are drawn between the first graph node and the first target graph node. Second graph nodes that are not connected to first graph nodes via overlapping competing edges are denoted as third graph nodes. For each third graph node, based on the predicted bounding boxes corresponding to the first and third graph nodes, the first distance between the first and third graph nodes is calculated. The third graph nodes are then sorted in ascending order of the first distance, and the top-ranked nodes are selected. A third graph node is used as a second target graph node, and dental arch adjacent edges are drawn between the first graph node and the second target graph node; Based on the predicted bounding boxes corresponding to the nodes in the first image, the jawbone labels of the first image nodes are determined. For each node in the second image, based on the predicted bounding boxes corresponding to the second image nodes, the jawbone labels of the second image nodes are determined. The second image nodes whose jawbone labels are the same as those of the first image nodes are selected as the fourth image nodes. For each fourth image node, based on the predicted bounding boxes corresponding to the first and fourth image nodes, the second distance between the first and fourth image nodes is calculated. The fourth image nodes are then sorted in ascending order of the second distance, and the top-ranked nodes are selected. The fourth graph node is used as the third target graph node, and the same-jaw neighborhood edge is drawn between the first graph node and the third target graph node; For each node in the second graph, based on the predicted bounding boxes corresponding to the nodes in the first and second graphs, the third distance between the first and second graph nodes is calculated. The second graph nodes are then sorted in ascending order of this third distance, and the top nodes are selected. The second graph node is used as the fourth target graph node, and spatial proximity edges are drawn between the first graph node and the fourth target graph node.

3. The method for training a tooth automatic detection and FDI number consistency correction model according to claim 1, characterized in that, Using the predicted bounding box, FDI number prediction result, and visual features corresponding to the graph nodes as input, the first encoding module is used to determine the node features of the graph nodes, specifically including: Based on the predicted bounding box corresponding to the graph node, the geometric features corresponding to the graph node are determined; the geometric features include the center point coordinates, area, aspect ratio and relative dental arch center position encoding of the predicted bounding box corresponding to the graph node. Based on the prediction results of the FDI number corresponding to the graph node, the prior features corresponding to the graph node are determined; the prior features include the original score and predicted probability of the FDI number of the tooth corresponding to the graph node belonging to each category. The first encoding module is used to encode and concatenate the visual features, geometric features and prior features corresponding to the graph nodes to obtain the node features of the graph nodes; The first encoding module includes a first encoding unit, a second encoding unit, a third encoding unit, and a splicing layer. The outputs of the first encoding unit, the second encoding unit, and the third encoding unit are all connected to the input of the splicing layer. The first encoding unit, the second encoding unit, and the third encoding unit have the same structure. The first encoding unit includes a first linear layer, a first ReLU activation function layer, and a first normalization layer connected in sequence. The first encoding unit is used to encode the visual features corresponding to the graph nodes to obtain the first encoded features; the second encoding unit is used to encode the geometric features corresponding to the graph nodes to obtain the second encoded features; the third encoding unit is used to encode the prior features corresponding to the graph nodes to obtain the third encoded features; the concatenation layer is used to concatenate the first encoded features, the second encoded features and the third encoded features to obtain the node features of the graph nodes.

4. The method for training a tooth automatic detection and FDI number consistency correction model according to claim 1, characterized in that, The geometric properties of the connecting edge are determined based on the predicted bounding boxes corresponding to the two graph nodes connected by the edge. Using the type and geometric properties of the connecting edge as input, the edge attributes are determined using a second encoding module, specifically including: Based on the predicted bounding boxes corresponding to the two graph nodes connected by the connecting edge, the geometric properties of the connecting edge are determined; the geometric properties include center offset, Euclidean distance, scale ratio, and overlap. The type and geometric properties of the connecting edges are encoded using the second encoding module to obtain the edge properties of the connecting edges; The second encoding module includes a second linear layer, a second ReLU activation function layer, and a second normalization layer connected in sequence.

5. The method for training a tooth automatic detection and FDI number consistency correction model according to claim 1, characterized in that, The graph reasoning module includes a graph reasoning unit and a recursive update unit connected in sequence. The graph reasoning unit is used to perform multi-relation message passing, update the node features of graph nodes in the sample intermediate multi-relation heterogeneous graph, and obtain the sample updated multi-relation heterogeneous graph. The recursive update unit is used to determine whether the number of updates has reached the preset number. If not, the sample updated multi-relation heterogeneous graph is input into the classification module to obtain the intermediate FDI number correction result. The intermediate FDI number correction result is used as the FDI number prediction result, and the sample intermediate multi-relation heterogeneous graph is reconstructed. The constructed sample intermediate multi-relation heterogeneous graph is used as the sample intermediate multi-relation heterogeneous graph for the next update and sent to the graph reasoning unit for another update. If yes, the sample updated multi-relation heterogeneous graph is used as the sample multi-relation heterogeneous graph. The graph reasoning unit includes multiple graph reasoning sub-units connected in sequence. Each graph reasoning sub-unit includes an RGATConv layer, an addition layer, and a third normalization layer connected in sequence. The input of the addition layer is also connected to the input of the RGATConv layer. The classification module consists of a third linear layer, a third ReLU activation function layer, a Dropout layer, a fourth linear layer, and a Softmax activation function layer connected in sequence.

6. The tooth automatic detection and FDI number consistency correction model training method according to claim 1, characterized in that, Based on the true bounding box, true FDI number, predicted bounding box, and FDI number correction results for each tooth in the sample image, the total loss is calculated, specifically including: Based on the true bounding box, true FDI number, predicted bounding box, and FDI number correction results for each tooth in the sample image, the detection and classification loss is calculated. Based on the true bounding box, true FDI number, predicted bounding box, and FDI number correction results for each tooth in the sample image, the corrected classification loss, duplicate detection loss, and adjacent number conflict loss are calculated respectively. The corrected classification loss, duplicate detection loss, and adjacent number conflict loss are then weighted and summed to obtain the graph reasoning loss. The detection classification loss and the graph reasoning loss are combined to form the total loss; The formula for calculating the classification loss is as follows: ； in, To detect classification loss; These are the normalization coefficients; The number of teeth in the sample image; For the sample image, the first The weight of each tooth; The cross-entropy loss function; For the sample image, the first The predicted probability vector of each tooth, including the first tooth. The predicted probability that the FDI number of each tooth belongs to each category is determined based on the FDI number correction result; For the sample image, the first The true probability vector of the tooth, including the first tooth The true probability that an FDI number of a tooth belongs to each category is determined based on the true FDI number; ； in, For the sample image, the first The initial weight of each tooth; The weighting factor for the crown region; A set of deciduous tooth categories; For the sample image, the first The ordinate of the upper boundary of the predicted bounding box of each tooth; For the sample image, the first The vertical coordinate of the boundary between the crown and root of a tooth; is the weighting coefficient for the root region.

7. The tooth automatic detection and FDI number consistency correction model training method according to claim 6, characterized in that, The formula for calculating graph inference loss is: ； in, For graph reasoning loss; To correct for classification loss; The first weighting coefficient; Loss due to repeated detection; This is the second weighting coefficient; The loss is calculated for adjacent number conflicts. When the number of iterations is between 1 and 15, the first and second weight coefficients are 0; when the number of iterations is between 16 and 25, the first and second weight coefficients are 0.1; when the number of iterations is between 26 and 35, the first and second weight coefficients are 0.5; when the number of iterations is greater than 35, the first and second weight coefficients are 0.

05. ； in, The number of teeth in the sample image; These are the class parameters used to balance the weights of positive and negative samples; For the sample image, the first The predicted probability that the FDI number of a tooth is the true FDI number is determined based on the FDI number correction result; Adjustment factor for easy and difficult samples; ； in, For connecting graph nodes Graph Nodes Attention weights for overlapping competing edges; To prevent the smoothing minimum coefficients where the denominator is zero; This is the set of overlapping competing edges; For graph nodes The corresponding predicted probability vector for the teeth, including graph nodes. The predicted probability that the corresponding tooth's FDI number belongs to each category is determined based on the FDI number correction result; For graph nodes The corresponding predicted probability vector for the teeth, including graph nodes. The predicted probability that the corresponding tooth's FDI number belongs to each category is determined based on the FDI number correction result; ； in, For connecting graph nodes Graph Nodes Attention weights for adjacent edges of the dental arch; It is the set of adjacent edges of the dental arch; The category exclusivity matrix for FDI numbering.

8. A method for applying an automatic tooth detection and FDI number consistency correction model, characterized in that, include: Obtain the image to be processed; The images to be processed are panoramic tomographic images of mixed dentition in children; Using the image to be processed as input, the updated object detection module is used to determine the predicted bounding box, FDI number prediction result and visual features of each tooth in the image to be processed. The predicted bounding box of each tooth in the image to be processed is taken as a graph node. Based on the anatomical relationship of the tooth, connection edges are selectively drawn between two graph nodes to obtain an initial multi-relationship heterogeneous graph. For each graph node and each connecting edge in the initial multi-relation heterogeneous graph, the node features of the graph node are determined by the updated first encoding module, and node features are assigned to the graph node, based on the predicted bounding boxes of the two graph nodes connected by the connecting edge. The geometric attributes of the connecting edge are determined by the updated second encoding module, based on the type and geometric attributes of the connecting edge, and edge attributes are assigned to the connecting edge, thus obtaining the intermediate multi-relation heterogeneous graph. Using an intermediate multi-relation heterogeneous graph as input, the updated graph inference module determines the multi-relation heterogeneous graph. Using the multi-relation heterogeneous graph as input, the updated classification module determines the FDI number correction result for each tooth in the image to be processed. The updated target detection module, the updated first encoding module, the updated second encoding module, the updated graph inference module, and the updated classification module constitute an automatic tooth detection and FDI number consistency correction model. The automatic tooth detection and FDI number consistency correction model is a model trained using the automatic tooth detection and FDI number consistency correction model training method described in any one of claims 1-7. Based on the FDI number correction results of each tooth in the image to be processed, the FDI number of each tooth in the image to be processed is determined.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the automatic tooth detection and FDI number consistency correction model training method according to any one of claims 1-7 or the automatic tooth detection and FDI number consistency correction model application method according to claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the training method for the automatic tooth detection and FDI number consistency correction model as described in any one of claims 1-7, or the application method of the automatic tooth detection and FDI number consistency correction model as described in claim 8.

Images