Tooth anomaly recognition method, training method and device of oral panoramic image recognition model, computer device and storage medium

Abstract

The application provides a tooth abnormality recognition method, a training method and device of an oral cavity panoramic image recognition model, a computer device and a storage medium. The tooth abnormality recognition method comprises the following steps: performing tooth detection on a to-be-tested oral cavity panoramic image to divide a plurality of tooth image regions; identifying an abnormal image region from the plurality of tooth image regions according to the image features of the tooth image regions; performing magnification processing on a target abnormal image region of uncertain abnormality to obtain a magnified image region; if the certainty of the target abnormal image region cannot be determined through the magnification processing, a target tooth image region that is in a jaw symmetrical mirror image relationship with the target abnormal image region is selected from the plurality of tooth image regions; and the image features of the target tooth image region and the target abnormal image region are compared to generate an abnormality recognition result of the target abnormal image region. The application can reliably distinguish subtle abnormalities and effectively reduce the missed diagnosis rate and misdiagnosis rate of subtle abnormalities.

Description

Technical Field

[0001] This application relates to the field of image processing technology, specifically to a method for identifying dental abnormalities, a training method, apparatus, computer equipment, and storage medium for a panoramic oral image recognition model. Background Technology

[0002] Orthopantomogram (OPG) is a basic imaging tool for screening dental abnormalities, which can fully display the anatomical structures of the upper and lower jaws, all teeth, and periodontal tissues in a single image.

[0003] Current deep learning-based oral image analysis typically employs a model that performs a single inference on the entire panoramic image and directly outputs anomaly detection results. However, some anomalous images often exhibit low contrast and blurred boundaries, and are susceptible to anatomical overlap and noise interference. In a single, global observation, the model struggles to reliably distinguish these poorly characterized anomalous images, failing to determine whether they are definitive anomalies (i.e., obvious anomalies, such as implants or impacted teeth) or indeterminate anomalies (i.e., subtle anomalies requiring further verification, such as bone resorption, dental caries, or periapical periodontitis). This results in a high rate of missed and misdiagnosed anomalies with unclear image representations. Summary of the Invention

[0004] In view of this, this application proposes a method for identifying dental abnormalities, a training method for a panoramic oral image recognition model, an apparatus, a computer device, and a storage medium to solve the problem that in the prior art, a single global recognition of panoramic oral images cannot reliably distinguish subtle abnormalities, resulting in a high rate of missed diagnoses and misdiagnoses of subtle abnormalities.

[0005] A first aspect of this application provides a method for identifying dental abnormalities, the method comprising: A dental examination is performed on the panoramic oral image to be tested, and several dental image regions are obtained. Abnormal image regions are identified from the plurality of dental image regions based on the image features of the dental image regions; the abnormality types of the abnormal image regions include deterministic abnormalities and uncertain abnormalities; A magnified image region of a target anomalous image region of uncertain anomaly type is obtained by performing magnification processing on the target anomalous image region, and a deterministic anomaly of the target anomalous image region is determined based on the image characteristics of the magnified image region. If the definite abnormality of the target abnormal image region cannot be determined by magnification, then a target tooth image region that is a symmetrical mirror image of the target abnormal image region in the jaw is selected from the plurality of tooth image regions. A comparative analysis is performed on the image features of the target tooth image region and the target abnormal image region to generate anomaly identification results for the target abnormal image region.

[0006] In this embodiment of the application, selecting a target tooth image region that is a symmetrical mirror image of the target abnormal image region from the plurality of tooth image regions includes: Determine the facial midline of the panoramic oral cavity image to be tested; Based on the facial midline, a horizontal mirror transformation is performed on the target abnormal image region to determine the target tooth image region.

[0007] In this embodiment of the application, a comparative analysis is performed based on the image features of the target tooth image region and the target abnormal image region, including: The image features of the target tooth image region and the target abnormal image region are compared side by side. Based on the grayscale differences, structural symmetry differences, and typical anomaly imaging features of the two images, anomaly identification results for the target anomaly image region are generated.

[0008] The second aspect of this application provides a method for training a panoramic oral image recognition model, the model being used to perform the dental abnormality recognition method described in the first aspect above, the training method comprising: The training sample set based on panoramic oral images is used to adjust the reasoning logic of the visual language model, so that the visual language model can learn abnormal reasoning logic and the calling rules of image tools. The image tools include a local magnification tool and a symmetrical mirroring tool. The calling rules include the timing and method of calling the image tools. The training sample set includes multiple panoramic oral images and the reasoning trajectory corresponding to each panoramic oral image. The reasoning trajectory includes the reasoning content, tool calling action, and reasoning conclusion. A hybrid reward function containing three types of sub-reward functions is constructed, and the adjusted visual language model is optimized based on the hybrid reward function to output a panoramic oral image recognition model; the three types of sub-reward functions include a multi-dimensional scoring reward function, a conditional reward function, and a structural verification reward function.

[0009] In this embodiment of the application, the inference logic of the visual language model is adjusted based on the training sample set of panoramic oral images, including:

[0010] in, This indicates the total loss from monitoring and fine-tuning. This represents the temporal steps within the reasoning trajectory. This represents the total number of temporal steps within the inference trajectory. This represents the conditional probability distribution of the visual language model output. The parameters of the language module, visual encoder, and projector in the visual language model are frozen and not updated. This represents the inference trajectory in the training sample set. This indicates an anomaly detection command. This indicates a panoramic training image of the oral cavity. Indicates the first The trajectory content of all time sequences prior to the step.

[0011] In this embodiment of the application, the multi-dimensional scoring reward function is used to score the completeness of anomaly recognition, the accuracy of dental image region localization, the accuracy of anomaly recognition, and the standardization of terminology, and to obtain a comprehensive score by weighted summation of the scores of multiple dimensions. The conditional reward function is used to issue positive rewards for the invocation of the local magnification tool and the symmetry mirroring tool when the comprehensive score output by the multi-dimensional scoring reward function is higher than the preset scoring threshold. The structure verification reward function is used to verify whether the reasoning trajectory output by the model follows a preset hierarchical structure, which includes reasoning content, tool invocation actions, and reasoning conclusions.

[0012] In this embodiment, the anomaly recognition completeness refers to the degree of matching between the abnormal image region detected by the oral panoramic image recognition model and the set of real anomalies in the image; the tooth image region localization accuracy refers to the degree of matching between the tooth image region and the standard tooth frame; the anomaly recognition accuracy refers to the accuracy rate of the oral panoramic image recognition model in recognizing anomalies in the tooth image region; the terminology standardization refers to the degree of matching between the inference conclusion output by the oral panoramic image recognition model based on the target oral panoramic training image and the target inference conclusion in the target inference trajectory, wherein the target inference trajectory refers to the inference trajectory in the training sample set corresponding to the target oral panoramic training image.

[0013] In this embodiment of the application, the conditional reward function includes:

[0014] in, This represents the inference trajectory of a single output of the model. Representing the reasoning trajectory The corresponding conditional exploration reward scalar value, This is an indicator function that outputs 1 if the condition within the parentheses is true and 0 if it is false. This indicates the preset scoring threshold. This represents the comprehensive score output by the multi-dimensional scoring reward function. This represents the reward intensity scaling factor, used to control the magnitude of the conditional exploration reward scalar value to balance the training gradient. This indicates the maximum exploration saturation limit. Representation and input query The corresponding current exploration saturation, enter the query. This includes panoramic oral cavity training images and anomaly detection instructions. The current exploration saturation refers to the number of times the model has invoked the image tools on the panoramic oral cavity training images. This indicates a tool indicator variable, with a value of , Representing the reasoning trajectory Execute image tool call, Representing the reasoning trajectory The image tool call was not executed.

[0015] An embodiment of the third aspect of this application provides a dental abnormality detection device, comprising: The dental image region segmentation module is used to perform dental detection on the panoramic oral image to be tested and to segment it into several dental image regions. An anomaly preliminary identification module is used to identify abnormal image regions from the plurality of dental image regions based on the image features of the dental image regions; the anomaly types of the abnormal image regions include deterministic anomalies and uncertain anomalies; The magnification processing module is used to perform magnification processing on the target abnormal image region whose anomaly type is uncertain anomaly, to obtain the magnified image region, and to determine the deterministic anomaly of the target abnormal image region based on the image characteristics of the magnified image region. The mirror region filtering module is used to filter out a target tooth image region that is symmetrically mirrored with the target abnormal image region from the plurality of tooth image regions if the definite abnormality of the target abnormal image region cannot be determined by magnification processing. The comparison analysis module is used to perform comparative analysis based on the image features of the target tooth image region and the target abnormal image region to generate anomaly identification results for the target abnormal image region.

[0016] An embodiment of the fourth aspect of this application provides a training apparatus for a panoramic oral image recognition model, characterized in that the model is used to perform the method described in the embodiment of the first aspect above, and the apparatus includes: The model fine-tuning module is used to adjust the inference logic of the visual language model based on the training sample set of panoramic oral images, so that the visual language model can learn abnormal inference logic and the calling rules of image tools; the image tools include a local magnification tool and a symmetrical mirroring tool, and the calling rules include the timing and method of calling the image tools; the training sample set includes multiple panoramic oral images and the inference trajectory corresponding to each panoramic oral image, and the inference trajectory includes the inference thinking content, tool calling action, and inference conclusion; The model reward optimization module is used to construct a hybrid reward function containing three types of sub-reward functions, and optimize the adjusted visual language model based on the hybrid reward function to output a panoramic oral image recognition model; the three types of sub-reward functions include a multi-dimensional scoring reward function, a conditional reward function, and a structural verification reward function.

[0017] An embodiment of the fifth aspect of this application provides a computer device including a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method described in the first or second aspect above.

[0018] An embodiment of the sixth aspect of this application provides a computer-readable storage medium storing computer instructions that are used to cause a computer to perform the methods described in the first or second aspect above.

[0019] The technical solutions provided by the embodiments of the present invention have at least the following beneficial effects: (1) By dividing the abnormal image area into deterministic abnormalities and uncertain abnormalities, and performing iterative local magnification and symmetrical mirror comparison for uncertain abnormalities, the dynamic diagnostic process of clinicians “global browsing → local detailed observation → contralateral verification” is effectively simulated, which significantly reduces the rate of missed diagnosis and misdiagnosis of subtle abnormalities. (2) By explicitly utilizing the symmetrical mirror relationship of the jaw, a reliable self-calibration reference is provided for uncertain anomalies, enabling the model to make more robust judgments based on the gray-scale differences, structural symmetry differences and typical radiographic manifestations of the bilateral images. (3) Through a two-stage training strategy (supervised fine-tuning + reinforcement learning), the model not only learns the correct tool usage norms, but also performs meaningful re-examination when the diagnostic confidence is sufficient and the exploration is not saturated, thus avoiding the problems of uncontrolled exploration behavior or reward abuse in existing technologies. (4) By using a hybrid reward function that includes multi-dimensional scoring rewards, conditional rewards and structure verification rewards, continuous and fine-grained clinical feedback is provided, which stabilizes the training process of long-cycle multi-round reasoning. Attached Figure Description

[0020] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A schematic flowchart of a method for identifying dental abnormalities according to an embodiment of this application is shown; Figure 2 A schematic diagram of the inference trajectory provided in one embodiment of this application is shown; Figure 3 A flowchart illustrating a training method for a panoramic oral image recognition model provided in an embodiment of this application is shown. Figure 4 This illustration shows a structural schematic diagram of a dental abnormality identification device according to an embodiment of this application; Figure 5 This illustration shows a schematic diagram of the structure of a computer device according to an embodiment of this application; Figure 6 A schematic diagram of a storage medium provided in one embodiment of this application is shown. Detailed Implementation

[0021] Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0022] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0023] Example 1 According to an embodiment of this application, a method for identifying dental abnormalities is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0024] This embodiment provides a method for identifying dental abnormalities. Figure 1 This is a flowchart of a method for identifying dental abnormalities according to an embodiment of this application, such as... Figure 1As shown, the process includes the following steps: Step S101: Perform tooth detection on the panoramic oral image to be tested to divide it into several tooth image regions.

[0025] Specifically, this step can use a tooth number detection model (such as a target detection model based on DETR or YOLO architecture) to locate teeth in the panoramic image of the oral cavity to be tested, identify the bounding boxes of all teeth in the image, and number the detected tooth regions based on the FDI tooth position numbering system (for example, 32 teeth can be numbered as 11-18, 21-28, 31-38, 41-48). Each tooth image region corresponds to a bounding box coordinate, representing the spatial position of the tooth in the panoramic image.

[0026] Step S102: Identify abnormal image regions from the plurality of tooth image regions based on the image features of the tooth image regions.

[0027] Specifically, this step can perform preliminary anomaly identification based on the image features of each tooth image region (such as grayscale distribution, texture features, edge features, etc.). Anomalies are classified into deterministic anomalies and uncertain anomalies.

[0028] Deterministic anomalies: These are anomalies that are distinctive and easily identifiable in panoramic images, including implants, orthodontic appliances, surgical instruments, restorations, and impacted teeth. These anomalies typically have well-defined geometric shapes and high image contrast, and can be confirmed without further verification.

[0029] Uncertain anomalies: These refer to anomalies in panoramic images that are blurry and require further verification. They include subtle lesions (such as dental caries, periapical periodontitis, bifurcation lesions, root resorption, and root fragments) and skeletal lesions (such as bone resorption). These anomalies typically appear as low-contrast shadows, poorly defined radiolucent areas, or slight changes in bone density, making reliable judgments difficult based on global observation alone.

[0030] Step S103: Perform magnification processing on the target abnormal image region with an abnormality type of uncertain abnormality to obtain magnified image region, and determine the deterministic abnormality of the target abnormal image region based on the image characteristics of the magnified image region.

[0031] Specifically, when a tooth image region is identified as an uncertain abnormality, it is designated as the target abnormal image region, and a local magnification tool is invoked to perform magnification processing on this region. The magnification processing includes: based on the bounding box coordinates of the target abnormal image region, performing fill and cropping at a preset magnification (preferably 1.5x) to obtain the locally magnified image region. This operation simulates the behavior of a clinician using a magnifying glass to observe a suspicious area in detail.

[0032] More specifically, based on the image features of the magnified image region, the abnormality type of the target abnormal image region is re-evaluated. If the image features of the magnified image region are sufficient to clearly determine the definite abnormality of the target abnormal image region (such as confirming dental caries, periapical periodontitis, etc.), then the abnormality identification result of the target abnormal image region is directly generated.

[0033] In some specific embodiments, generating anomaly identification results includes the following steps: Step a1: Generate a detailed visual summary for the magnified image area based on the visual description agent, including: radiographic features (such as transillumination dark areas, changes in bone density, etc.), visual characteristics (boundary clarity, morphological regularity, etc.), location information (specific tooth position, regional hierarchy), and lesion pattern (the degree of matching with typical lesions). Step a2: Combine clinical definitions to make diagnostic inferences. For example, if the magnified image shows "a translucent dark area in the enamel or dentin, triangular in shape with the tooth surface as the base", then it is confirmed as a definite abnormality - dental caries based on the definition of caries; if the magnified image shows "a translucent dark area around the periapical region, with widening of the periodontal ligament space", then it is confirmed as a definite abnormality - periapical periodontitis based on the definition of periapical periodontitis.

[0034] Step a3 integrates the diagnostic reasoning process into a standardized reasoning trajectory, including: reasoning content: describing the key evidence observed from the magnified images and the reasoning logic; tool call actions: recording the executed "magnification" operations and their parameters (boundary box coordinates); and reasoning conclusions: clear abnormality identification results (abnormality type, tooth position, lesion description).

[0035] Step S104: If the definite abnormality of the target abnormal image region cannot be determined by magnification processing, then a target tooth image region that is a symmetrical mirror image of the target abnormal image region in the jaw is selected from the plurality of tooth image regions.

[0036] Specifically, if the image features of the magnified image area are insufficient to clearly determine the definite abnormality of the target abnormal image area (such as confirming dental caries, periapical periodontitis, etc.), then the target abnormal image area is considered to be an uncertain abnormality that cannot be eliminated by magnification. In this case, a symmetrical mirror comparison method is adopted to determine the definite abnormality of the target abnormal image area.

[0037] More specifically, if the image features of the magnified image area are still blurry and have low contrast, making it impossible to make a definitive judgment, then a target tooth image area that has a symmetrical mirror relationship with the target abnormal image area is selected from the plurality of tooth image areas, as shown in the following steps S1041-S1042: Step S1041: Determine the facial midline of the panoramic oral cavity image to be tested.

[0038] Specifically, the facial midline is the baseline that divides the panoramic image of the oral cavity into two symmetrical parts along the vertical direction, and it is usually located in the center of the image.

[0039] Step S1042: Based on the facial midline, perform a horizontal mirror transformation on the target abnormal image region to determine the target tooth image region.

[0040] Specifically, let the width of the panoramic image of the oral cavity to be tested be W, and the bounding box coordinates of the target abnormal image region be [x1,x2]×[y1, y2]. Then, the bounding box coordinates of the target tooth image region in the symmetrical mirror area of ​​the jaw are [W-x2, W-x1]×[y1, y2]. This operation performs a horizontal mirroring of the target abnormal image region along the midline of the face to obtain the tooth image region at the corresponding position on the opposite side of the jaw.

[0041] Step S105: Based on the image features of the target tooth image region and the target abnormal image region, a comparative analysis is performed to generate an anomaly recognition result for the target abnormal image region.

[0042] Specifically, a magnified image of the target abnormal region is displayed side-by-side with an image of the target tooth region (the contralateral symmetrical region), forming a dual-view comparative analysis. The comparative analysis includes: Gray-scale difference analysis: Compare the gray-scale value distribution of corresponding positions in bilateral images. If the target abnormal image area has a significant low-density radiolucent area or a high-density hardened area, while the contralateral symmetrical area has no corresponding manifestation, it suggests a real lesion.

[0043] Structural symmetry difference analysis: Compare the anatomical symmetry of bilateral images, such as alveolar bone height, periodontal ligament space width, periapical bone density, etc. If there is obvious asymmetry, it suggests the possibility of lesion.

[0044] Analysis of typical abnormal imaging features: Combine the imaging features of the target abnormal imaging area to determine whether they conform to the typical manifestations of a specific lesion (such as the triangular radiolucent shadow of caries, the periapical radiolucent area of ​​periapical periodontitis, etc.) to comprehensively judge the type of abnormality.

[0045] Based on the above comparative analysis, the final anomaly identification result of the target anomaly image region is generated, and its anomaly type is updated from uncertain anomaly to deterministic anomaly.

[0046] In some specific embodiments, after step S105, the method further includes: integrating all dental image regions identified as definitive abnormalities (including definitive abnormalities identified in the initial screening and uncertain abnormalities confirmed after magnification / mirror comparison), and generating a complete dental abnormality conclusion report (i.e., reasoning trajectory), such as... Figure 2As shown. The report includes the FDI number of the abnormal tooth, the type of abnormality, a description of the lesion location, and the diagnostic confidence level.

[0047] In some specific embodiments, if the anomaly type of the target abnormal image region is still uncertain after the comparative analysis in step S105, steps S103-S105 can be repeated, that is, the magnification process and / or symmetrical mirror comparison can be performed again until the region is determined to be a deterministic anomaly, or the preset maximum number of iterations (such as 5 rounds) is reached.

[0048] The technical solutions provided by the embodiments of the present invention have at least the following beneficial effects: (1) By dividing the abnormal image area into deterministic abnormalities and uncertain abnormalities, and performing iterative local magnification and symmetrical mirror comparison for uncertain abnormalities, the dynamic diagnostic process of clinicians “global browsing → local detailed observation → contralateral verification” is effectively simulated, which significantly reduces the rate of missed diagnosis and misdiagnosis of subtle abnormalities. (2) By explicitly utilizing the symmetrical mirror relationship of the jaw, a reliable self-calibration reference is provided for uncertain abnormalities, enabling the model to make more robust judgments based on the gray-scale differences, structural symmetry differences and typical radiographic manifestations of the bilateral images.

[0049] Example 2: This application also provides a method for training a panoramic oral image recognition model to execute the dental abnormality recognition method described in Embodiment 1 above. Figure 3 As shown, this training method includes the following two stages: Phase One: Monitoring and Fine-tuning S201, the inference logic of the visual language model is adjusted based on the training sample set of panoramic oral images, so that the visual language model learns the abnormal inference logic and the calling rules of the image tools.

[0050] Specifically, the visual language model used was Qwen2.5-VL-3B / 7B-Instruct, and the training sample set included multiple panoramic oral images and the inference trajectory corresponding to each panoramic oral image. The inference trajectory included inference thinking content, tool calling actions, tool observation, and inference conclusion, reflecting the complete process of clinicians conducting multi-round diagnoses based on panoramic images.

[0051] More specifically, the process of constructing the reasoning trajectory includes the following steps b1-b5: Step b1: Perform tooth detection and lesion labeling alignment on the panoramic oral training images.

[0052] Specifically, a tooth count detection model is used to locate tooth regions in training images and align them with lesion labels annotated by experts.

[0053] Step b2: Classify lesions into deterministic and indeterminate abnormalities based on their degree of prominence.

[0054] Step b3: Cluster the dental imaging regions corresponding to the uncertain abnormalities to generate candidate examination regions.

[0055] Specifically, the K-Means clustering algorithm is used to cluster the tooth-level bounding boxes corresponding to uncertain anomalies, generating multiple candidate examination regions. Each region represents a vague or difficult-to-diagnose part that needs to be analyzed in detail.

[0056] Step b4: Construct multi-round diagnostic trajectories based on rule templates.

[0057] The multi-round diagnostic trajectory includes an initial global examination round and subsequent local examination rounds: Round 1 (Global Inspection): The model receives user queries (such as "Is there any abnormality in this panoramic image?"), outputs global observation results, identifies obvious lesions, and marks uncertain abnormal areas that require further examination.

[0058] Round 2 and beyond (local examination): For areas of uncertainty and abnormality, the model sequentially calls local magnification tools or symmetric mirroring tools to obtain supplementary image information and gradually refine the diagnostic conclusions.

[0059] Step b5: Optimize multi-round diagnostic trajectories through multi-agent collaboration.

[0060] Multi-agent systems include at least: Judgment Agent: Assess whether the current local view is sufficient to support diagnostic decisions; if insufficient, trigger the symmetry mirror tool to obtain contralateral comparison information; if the symmetry quality is unreliable (e.g., severe asymmetry or low quality of mirror evidence), redirect the reasoning to the local magnification operation.

[0061] Visual description agent: Generates region-level visual summaries for images invoked by each tool, describing imagery features, visual characteristics, location, and patterns.

[0062] Rewrite Agent: Integrates Wikipedia-style clinical radiology definitions and visual descriptions into a visual summary generated by the agent, and rewrites the reasoning steps in a clinically standardized manner to ensure clinical consistency and diversity in language expression.

[0063] The specific implementation of monitoring and fine-tuning: Fully parameter-supervised fine-tuning is performed on the language module of the visual language model, while the visual encoder and projector are frozen to ensure the stability of visual features. The training objective is to maximize the likelihood probability of true inference trajectories.

[0064] in, This indicates the total loss from monitoring and fine-tuning. This represents the temporal steps within the reasoning trajectory. This represents the total number of temporal steps within the inference trajectory. This represents the conditional probability distribution of the visual language model output. The parameters of the language module, visual encoder, and projector in the visual language model are frozen and not updated. This represents the inference trajectory in the training sample set. This indicates an anomaly detection command. This indicates a panoramic training image of the oral cavity. Indicates the first The trajectory content of all time sequences prior to the step.

[0065] The second stage involves reinforcement learning optimization based on a hybrid reward function: S202, Construct a hybrid reward function containing three types of sub-reward functions, the three types of sub-reward functions being: (1) Multi-dimensional scoring reward function

[0066] It is used to score the completeness of anomaly identification, the accuracy of dental image region localization, the accuracy of anomaly identification, and the standardization of terminology, and the scores of multiple dimensions are weighted and summed to obtain a comprehensive score.

[0067] Anomaly recognition completeness refers to the degree of matching between the abnormal image regions detected by the oral panoramic image recognition model and the actual set of anomalies in the image. For example, if there are a total of 10 anomalies, but the model only detects 6 anomalies, the anomaly recognition completeness is 60% or 0.6. Dental imaging region positioning accuracy refers to the degree of matching between the dental imaging region and the standard dental frame, such as whether the dental imaging region corresponding to tooth 1 is the same as the position of tooth 1 in the standard dental frame. Anomaly recognition accuracy refers to the accuracy rate of the oral panoramic image recognition model in recognizing anomalies in the dental image region; The terminology standardization refers to the degree of matching between the inference conclusions output by the oral panoramic image recognition model based on the target oral panoramic training image and the target inference conclusions in the target inference trajectory; the target inference trajectory refers to the inference trajectory in the training sample set corresponding to the target oral panoramic training image.

[0068] (2) Conditional reward function

[0069] This function is used to issue positive rewards for calling the local magnification tool and the symmetry mirroring tool when the comprehensive score output by the multi-dimensional scoring reward function is higher than a preset scoring threshold. The expression for the conditional reward function is:

[0070] in, This represents the inference trajectory of a single output of the model. Representing the reasoning trajectory The corresponding conditional exploration reward scalar value, This is an indicator function that outputs 1 if the condition within the parentheses is true and 0 if it is false. This indicates the preset scoring threshold. This represents the comprehensive score output by the multi-dimensional scoring reward function. This represents the reward intensity scaling factor, used to control the magnitude of the conditional exploration reward scalar value to balance the training gradient. This indicates the maximum exploration saturation limit. Representation and input query The corresponding current exploration saturation, enter the query. This includes panoramic oral cavity training images and anomaly detection instructions. The current exploration saturation refers to the number of times the model has invoked the image tools on the panoramic oral cavity training images. This indicates a tool indicator variable, with a value of , Representing the reasoning trajectory Execute image tool call, Representing the reasoning trajectory The image tool call was not executed.

[0071] The core design principle of this conditional reward function is: it only applies when the diagnosis already has a certain level of reliability ( And explore unsaturated ( Only when the confidence level of the diagnosis is low should the model be encouraged to make tool calls to improve the diagnosis; if the confidence level of the diagnosis is insufficient or the exploration has been over-explored, additional tool calls should be suppressed to avoid the abuse of rewards and ineffective exploration.

[0072] (3) Structural verification reward function

[0073] This is used to verify whether the reasoning trajectory output by the model follows a preset hierarchical structure, which includes reasoning content, tool invocation actions, and reasoning conclusions.

[0074] Specifically, if the reasoning trajectory output by the model strictly follows the standard format of "thinking → tools → observation → conclusion", a positive reward will be given; if the format is disordered or necessary components are missing, a penalty will be given.

[0075] S203 optimizes the adjusted visual language model based on the hybrid reward function and outputs a panoramic oral cavity image recognition model.

[0076] Specifically, the GRPO (Group Relative Policy Optimization) algorithm is used for optimization. The GRPO algorithm estimates the relative advantage of multiple trajectories (within-group samples) sampled for the same query, eliminates baseline variance through within-group comparison, and updates the magnitude using a pruned importance ratio constraint strategy to ensure training stability.

[0077] More specifically, the total reward of the hybrid reward function is:

[0078] The advantage value of each trajectory is calculated based on the total reward, and the model parameters are updated through the GRPO objective function until convergence, finally outputting the oral panoramic image recognition model.

[0079] Figure 3 In ( , , , ): represents the reward / reward score, derived from various scoring metrics of the mixed reward; ( 、 、 、 ): Represents advantage / advantage value, a core concept in reinforcement learning (such as the PPO algorithm), indicating the degree to which a certain action or decision is superior or inferior to the expected baseline, used to guide model policy updates. That is... R This is the original rating. A These are the values ​​used for actual training optimization after advantage calculation.

[0080] The technical solutions provided by the embodiments of the present invention have at least the following beneficial effects: (1) Through a two-stage training strategy (supervised fine-tuning + reinforcement learning), the model not only learns the correct tool usage norms, but also performs meaningful re-examination when the diagnostic confidence is sufficient and the exploration is not saturated, thus avoiding the problems of uncontrolled exploration behavior or reward abuse in existing technologies. (2) By using a hybrid reward function that includes multi-dimensional scoring rewards, conditional rewards and structure verification rewards, continuous and fine-grained clinical feedback is provided, which stabilizes the training process of long-cycle multi-round reasoning.

[0081] Example 3: Corresponding to the above implementation of the dental anomaly recognition method, this application also provides a dental anomaly recognition device for executing the dental anomaly recognition method described in Embodiment 1. Figure 4As shown, the dental abnormality detection device includes: The dental image region segmentation module is used to perform dental detection on the panoramic oral image to be tested and to segment it into several dental image regions. An anomaly preliminary identification module is used to identify abnormal image regions from the plurality of dental image regions based on the image features of the dental image regions; the anomaly types of the abnormal image regions include deterministic anomalies and uncertain anomalies; The magnification processing module is used to perform magnification processing on the target abnormal image region whose anomaly type is uncertain anomaly, to obtain the magnified image region, and to determine the deterministic anomaly of the target abnormal image region based on the image characteristics of the magnified image region. The mirror region filtering module is used to filter out a target tooth image region that is symmetrically mirrored with the target abnormal image region from the plurality of tooth image regions if the definite abnormality of the target abnormal image region cannot be determined by magnification processing. The comparison analysis module is used to perform comparative analysis based on the image features of the target tooth image region and the target abnormal image region to generate anomaly identification results for the target abnormal image region.

[0082] Optionally, the mirror region filtering module is further configured to determine the facial midline of the panoramic oral image to be tested; based on the facial midline, perform a horizontal mirror transformation on the target abnormal image region to determine the target tooth image region.

[0083] Optionally, the comparison analysis module is also used to compare the image features of the target tooth image region with those of the target abnormal image region side by side; and to generate anomaly recognition results for the target abnormal image region based on the grayscale differences, structural symmetry differences, and typical abnormal radiographic features of the two images.

[0084] The dental abnormality recognition device and the dental abnormality recognition method provided in the above embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0085] Example 4: This application embodiment also provides a training device for a panoramic oral image recognition model, used to execute the training method for the panoramic oral image recognition model described in Embodiment 2 above. The training device includes: The model fine-tuning module is used to adjust the inference logic of the visual language model (i.e., Qwen2.5-VL-3B / 7B-Instruct) based on the training sample set of panoramic oral images, so that the visual language model can learn abnormal inference logic and the calling rules of image tools; the image tools include a local magnification tool and a symmetrical mirroring tool, and the calling rules include the timing and method of calling the image tools; the training sample set includes multiple panoramic oral images and the inference trajectory corresponding to each panoramic oral image, and the inference trajectory includes the inference thinking content, tool calling action, and inference conclusion; The model reward optimization module is used to construct a hybrid reward function containing three types of sub-reward functions, and optimize the adjusted visual language model based on the hybrid reward function to output a panoramic oral image recognition model; the three types of sub-reward functions include a multi-dimensional scoring reward function, a conditional reward function, and a structural verification reward function.

[0086] Optionally, the multi-dimensional scoring reward function is used to score the completeness of anomaly recognition, the accuracy of dental image region localization, the accuracy of anomaly recognition, and the standardization of terminology, and to obtain a comprehensive score by weighted summation of the scores of multiple dimensions; the conditional reward function is used to issue positive rewards for the invocation of the local magnification tool and the symmetry mirroring tool when the comprehensive score output by the multi-dimensional scoring reward function is higher than a preset scoring threshold; the structure verification reward function is used to verify whether the inference trajectory output by the model follows a preset hierarchical structure, which includes the inference thinking content, tool invocation actions, and inference conclusions.

[0087] The training device for the oral panoramic image recognition model provided in the above embodiments of this application and the training method for the oral panoramic image recognition model provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application stored therein.

[0088] This application also provides a computer device for executing the above-described method for identifying dental abnormalities or training a panoramic oral image recognition model. Please refer to... Figure 5 This illustrates a schematic diagram of a computer device provided by some embodiments of this application. For example... Figure 5 As shown, the computer device 5 includes: a processor 500, a memory 501, a bus 502, and a communication interface 503. The processor 500, the communication interface 503, and the memory 501 are connected via the bus 502. The memory 501 stores a computer program that can run on the processor 500. When the processor 500 runs the computer program, it executes the tooth abnormality recognition method or the training method of the panoramic oral image recognition model provided in any of the foregoing embodiments of this application.

[0089] The memory 501 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 503 (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0090] Bus 502 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. Memory 501 is used to store programs. After receiving an execution instruction, processor 500 executes the program. The tooth abnormality recognition method or the training method for the panoramic oral image recognition model disclosed in any of the foregoing embodiments can be applied to processor 500, or implemented by processor 500.

[0091] The processor 500 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 500 or by instructions in software form. The processor 500 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 501. The processor 500 reads the information in memory 501 and, in conjunction with its hardware, completes the steps of the above method.

[0092] The computer device provided in this application embodiment and the training method for the tooth abnormality recognition method or the oral panoramic image recognition model provided in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.

[0093] This application also provides a computer-readable storage medium corresponding to the training method for the tooth abnormality recognition method or the panoramic oral image recognition model provided in the foregoing embodiments. Please refer to [link / reference]. Figure 6 The computer-readable storage medium shown is an optical disc 30, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it executes the training method for the tooth abnormality recognition method or the panoramic oral image recognition model provided in any of the foregoing embodiments.

[0094] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.

[0095] The computer-readable storage medium provided in the above embodiments of this application and the training method for the tooth abnormality recognition method or the oral panoramic image recognition model provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0096] It should be noted that: Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known structures and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0097] Similarly, it should be understood that, in order to simplify this application and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of this application, various features of this application are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosure should not be construed as reflecting a schematic diagram in which the claimed application requires more features than expressly recited in each claim. Rather, as reflected in the claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0098] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features but not others included in other embodiments, combinations of features from different embodiments are intended to be within the scope of this application and form different embodiments. For example, in the claims, any one of the claimed embodiments can be used in any combination.

[0099] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying dental abnormalities, characterized in that, The method includes: A dental examination is performed on the panoramic oral image to be tested, and several dental image regions are obtained. Abnormal image regions are identified from the plurality of dental image regions based on the image features of the dental image regions; the abnormality types of the abnormal image regions include deterministic abnormalities and uncertain abnormalities; A magnified image region of a target anomalous image region of uncertain anomaly type is obtained by performing magnification processing to obtain magnified image region, and a deterministic anomaly of the target anomalous image region is determined based on the image characteristics of the magnified image region. If the definite abnormality of the target abnormal image region cannot be determined by magnification, then a target tooth image region that is a symmetrical mirror image of the target abnormal image region in the jaw is selected from the plurality of tooth image regions. A comparative analysis is performed on the image features of the target tooth image region and the target abnormal image region to generate anomaly identification results for the target abnormal image region.

2. The method according to claim 1, characterized in that, From the plurality of dental image regions, target dental image regions that are symmetrically mirrored with the target abnormal image region in the jaw region are selected, including: Determine the facial midline of the panoramic oral cavity image to be tested; Based on the facial midline, a horizontal mirror transformation is performed on the target abnormal image region to determine the target tooth image region.

3. The method according to claim 1 or 2, characterized in that, A comparative analysis is performed based on the image features of the target tooth image region and the target abnormal image region, including: The image features of the target tooth image region and the target abnormal image region are compared side by side. Based on the grayscale differences, structural symmetry differences, and typical anomaly imaging features of the two images, anomaly identification results for the target anomaly image region are generated.

4. A training method for a panoramic oral cavity image recognition model, characterized in that, The model is used to perform the method as described in any one of claims 1-3, wherein the training method includes: The training sample set based on panoramic oral images is used to adjust the reasoning logic of the visual language model, so that the visual language model can learn abnormal reasoning logic and the calling rules of image tools. The image tools include a local magnification tool and a symmetrical mirroring tool. The calling rules include the timing and method of calling the image tools. The training sample set includes multiple panoramic oral images and the reasoning trajectory corresponding to each panoramic oral image. The reasoning trajectory includes the reasoning content, tool calling action, and reasoning conclusion. A hybrid reward function containing three types of sub-reward functions is constructed, and the adjusted visual language model is optimized based on the hybrid reward function to output a panoramic oral image recognition model; the three types of sub-reward functions include a multi-dimensional scoring reward function, a conditional reward function, and a structural verification reward function.

5. The method according to claim 4, characterized in that, The inference logic of the visual language model is adjusted based on the training sample set of panoramic oral images, including: in, This indicates the total loss from monitoring and fine-tuning. This represents the temporal steps within the reasoning trajectory. This represents the total number of temporal steps within the inference trajectory. This represents the conditional probability distribution of the visual language model output. The parameters of the language module, visual encoder, and projector in the visual language model are frozen and not updated. This represents the inference trajectory in the training sample set. This indicates an anomaly detection command. This indicates a panoramic training image of the oral cavity. Indicates the first The trajectory content of all time sequences prior to the step.

6. The method according to claim 4, characterized in that, The multi-dimensional scoring reward function is used to score the completeness of anomaly recognition, the accuracy of dental image region localization, the accuracy of anomaly recognition, and the standardization of terminology, and the scores of multiple dimensions are weighted and summed to obtain a comprehensive score. The conditional reward function is used to issue positive rewards for the invocation of the local magnification tool and the symmetry mirroring tool when the comprehensive score output by the multi-dimensional scoring reward function is higher than the preset scoring threshold. The structure verification reward function is used to verify whether the reasoning trajectory output by the model follows a preset hierarchical structure, which includes reasoning content, tool invocation actions, and reasoning conclusions.

7. The method according to claim 6, characterized in that, The anomaly recognition completeness refers to the degree of matching between the abnormal image region detected by the oral panoramic image recognition model and the set of real anomalies in the image; the tooth image region localization accuracy refers to the degree of matching between the tooth image region and the standard tooth frame; the anomaly recognition accuracy refers to the accuracy rate of the oral panoramic image recognition model in recognizing anomalies in the tooth image region; the terminology standardization refers to the degree of matching between the inference conclusion output by the oral panoramic image recognition model based on the target oral panoramic training image and the target inference conclusion in the target inference trajectory, where the target inference trajectory refers to the inference trajectory in the training sample set corresponding to the target oral panoramic training image.

8. The method according to claim 6, characterized in that, The conditional reward function includes: in, This represents the inference trajectory of a single output of the model. Representing the reasoning trajectory The corresponding conditional exploration reward scalar value, This is an indicator function that outputs 1 if the condition within the parentheses is true and 0 if it is false. This indicates the preset scoring threshold. This represents the comprehensive score output by the multi-dimensional scoring reward function. This represents the reward intensity scaling factor, used to control the magnitude of the conditional exploration reward scalar value to balance the training gradient. This indicates the maximum exploration saturation limit. Representation and input query The corresponding current exploration saturation, enter the query. This includes panoramic oral cavity training images and anomaly detection instructions. The current exploration saturation refers to the number of times the model has invoked the image tools on the panoramic oral cavity training images. This indicates a tool indicator variable, with a value of , Representing the reasoning trajectory Execute image tool call, Representing the reasoning trajectory The image tool was not invoked.

9. A device for detecting dental abnormalities, characterized in that, The device includes: The dental image region segmentation module is used to perform dental detection on the panoramic oral image to be tested and to segment it into several dental image regions. An anomaly preliminary identification module is used to identify abnormal image regions from the plurality of dental image regions based on the image features of the dental image regions; the anomaly types of the abnormal image regions include deterministic anomalies and uncertain anomalies; The magnification processing module is used to perform magnification processing on the target abnormal image region whose anomaly type is uncertain anomaly, to obtain the magnified image region, and to determine the deterministic anomaly of the target abnormal image region based on the image features of the magnified image region. The mirror region filtering module is used to filter out a target tooth image region that is symmetrically mirrored with the target abnormal image region from the plurality of tooth image regions if the definite abnormality of the target abnormal image region cannot be determined by magnification processing. The comparison analysis module is used to perform comparative analysis based on the image features of the target tooth image region and the target abnormal image region to generate anomaly identification results for the target abnormal image region.

10. A training device for a panoramic oral image recognition model, characterized in that, The model is used to perform the method as described in any one of claims 1-3, and the apparatus includes: The model fine-tuning module is used to adjust the inference logic of the visual language model based on the training sample set of panoramic oral images, so that the visual language model can learn abnormal inference logic and the calling rules of image tools; the image tools include a local magnification tool and a symmetrical mirroring tool, and the calling rules include the timing and method of calling the image tools; the training sample set includes multiple panoramic oral images and the inference trajectory corresponding to each panoramic oral image, and the inference trajectory includes the inference thinking content, tool calling action, and inference conclusion; The model reward optimization module is used to construct a hybrid reward function containing three types of sub-reward functions, and optimize the adjusted visual language model based on the hybrid reward function to output a panoramic oral image recognition model; the three types of sub-reward functions include a multi-dimensional scoring reward function, a conditional reward function, and a structural verification reward function.

11. A computer device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 8.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the method of any one of claims 1 to 8.

Images