An oral and maxillofacial image assisted detection method based on artificial intelligence

Abstract

The application discloses an oral and maxillofacial image auxiliary detection method based on artificial intelligence, and belongs to the technical field of oral digital processing, and comprises the following steps: acquiring an oral and maxillofacial image to be processed; constructing a multi-task artificial intelligence diagnosis model, and identifying the oral and maxillofacial image to be processed based on the multi-task artificial intelligence diagnosis model to obtain a boundary box of a tooth position, disease characteristics and bone resorption amount of a tooth; binding the segmentation box of the tooth position, the disease characteristics and the bone resorption amount of the tooth to obtain basic information of a patient; calling a structured report generation engine, combining the basic information of the patient and a local disease knowledge base to obtain a preliminary diagnosis report; utilizing a man-machine interactive review interface to review the preliminary diagnosis report, and archiving the reviewed diagnosis report to a structured database. The method solves the problems of lacking intelligent quality control of image quality, lacking standardization and consistency of diagnosis and not fully utilizing image data in existing oral image diagnosis.

Description

Technical Field

[0001] This application relates to the field of oral digital processing technology, and in particular to an artificial intelligence-based method for assisting in the detection of oral and maxillofacial images. Background Technology

[0002] Oral and maxillofacial imaging is an important means of diagnosing oral diseases. Among them, panoramic radiographs can show the entire dentition, periodontal tissues, maxilla and mandible and temporomandibular joint in a single image, and are widely used in clinical screening and diagnosis.

[0003] However, current oral imaging diagnosis faces the following problems: inconsistent image quality: due to differences in imaging equipment and the skill level of technicians, image quality varies, affecting diagnostic accuracy; reliance on physician experience for diagnosis: the lack of unified standards for interpreting images easily leads to missed diagnoses and misdiagnoses; insufficient primary healthcare resources: remote areas or small dental clinics lack professional imaging diagnostic physicians, making it difficult to achieve high-quality imaging diagnosis; low utilization rate of imaging data: massive amounts of imaging data are not stored in a structured manner, making it difficult to use for artificial intelligence training and big data analysis.

[0004] In the existing technology, there are some oral imaging-assisted diagnostic systems, but most of them focus on the detection of single diseases and lack a systematic and standardized end-to-end solution. Furthermore, they are not deeply integrated with the consultation platform, making it difficult to achieve standardization, intelligence, and remoteness in imaging diagnosis. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this application provides an artificial intelligence-based oral and maxillofacial image-assisted detection method that solves the problems of lack of intelligent quality control of image quality, lack of standardization and consistency in diagnosis, and underutilization of image data in existing oral imaging diagnosis.

[0006] To achieve the aforementioned objectives, the technical solution adopted in this application is as follows: This application provides an artificial intelligence-based image-assisted detection method for oral and maxillofacial structures, including: S1: Acquire the oral and maxillofacial images to be processed; S2: Construct a multi-task artificial intelligence diagnostic model, and identify the oral and maxillofacial images to be processed based on the multi-task artificial intelligence diagnostic model to obtain the bounding box of the tooth position, disease features and the amount of bone resorption of the tooth. The multi-task artificial intelligence diagnostic model includes Mask R-CNN model, PP-YOLOv2 model, Faster R-CNN model and U-Net model. S3: Bind the segmentation frame of the tooth position, disease characteristics, and the amount of bone resorption of the tooth to obtain the patient's basic information; S4: Call the structured report generation engine, combine patient basic information and local disease knowledge base to obtain a preliminary diagnosis report; S5: Use the human-computer interaction review interface to review the preliminary diagnostic report and archive the reviewed diagnostic report to the structured database.

[0007] Further, S2 includes: The Mask R-CNN model is used to identify and segment the oral and maxillofacial images to be processed, obtain the contour and segmentation box of each tooth, and encode them to obtain the bounding box of the tooth position. The PP-YOLOv2 model and the Faster R-CNN model were used to identify disease features in the oral and maxillofacial images to be processed. The PP-YOLOv2 model was used to identify dental caries and periapical periodontitis, while the Faster R-CNN model was used to identify wisdom teeth, implants and root canal fillings. The cementoenamel boundary and alveolar ridge of the upper and lower jaws were segmented using the U-Net model. The vertical distance from the cementoenamel boundary to the alveolar ridge along each tooth axis was calculated. The amount of bone resorption was obtained by combining the pixel spacing information.

[0008] Furthermore, the Mask R-CNN model is used to identify and segment the oral and maxillofacial images to be processed, obtaining the contour and segmentation box of each tooth, and encoding them to obtain the bounding box of the tooth position, including: A1: For the oral and maxillofacial images to be processed, multi-scale feature maps are output using the ResNet50-vd-dcn backbone network, where the first... The layer features are:

[0009] In the formula, For the first Layer features, For the first Layer features, For the first The residual blocks of the layer, It is a deformable convolution; The expression for deformable convolution is:

[0010] In the formula, To output the value of the feature map at a certain location, and For location, For the first One sampling point, The number of sampling points. For the first The weight of each sampling point For the input feature map, and For the preset rule sampling positions, and The learnable offset; A2: Generate candidate boxes based on multi-scale feature maps, and align the feature vectors of the candidate boxes to obtain the segmentation boxes for the teeth. The expressions for generating the candidate boxes and aligning the feature vectors of the candidate boxes are as follows:

[0011]

[0012] In the formula, For the candidate box set, For the first One candidate box, and The center coordinates of the candidate box. and For the first The width and height of each candidate box. For the first The feature vectors after aligning the candidate boxes and For network indexing; A3: The contour is output using a segmentation head and optimized using Mask IoU loss. The expression for Mask IoU loss is:

[0013] In the formula, The Mask IoU loss function is... This represents the number of candidate boxes that contain the target object. Let be the intersection-union ratio function. For the first The outline of each candidate box. For a realistic outline, For element-wise multiplication, This is an element-wise addition; A4: The centroid coordinates of the contour are mapped to FDI tooth position codes using a dental arch geometry model, and the FDI tooth position codes are corrected using a classification head. The maximum probability distribution vector of the tooth category is selected as the FDI tooth position code. The expressions for the dental arch geometry model and the correction are as follows:

[0014]

[0015] In the formula, For the first FDI tooth position encoding of each candidate box and For image width and height, and For the first The centroid coordinates of the outline of each candidate bounding box. This is the floor function; Let be the probability distribution vector for tooth categories. For normalized exponential functions, This is the weight matrix of the classification heads. For channel dimension, For the set of real numbers, It is the bias vector; A5: Optimize FDI tooth position coding using coding accuracy, whereby the expression for coding accuracy is:

[0016] In the formula, The accuracy of tooth coding. For the first The actual FDI tooth position code of each candidate box; A6: Match the segmentation box of the tooth with the FDI tooth position code to obtain the bounding box of the tooth position.

[0017] Furthermore, the PP-YOLOv2 model identifies dental caries and periapical periodontitis, including: B1: Multi-scale features of oral and maxillofacial images are extracted using a backbone network to obtain a feature pyramid. The calculation formula is as follows:

[0018] In the formula, and The first feature pyramid Layer features and The output layer of layer features, This is an upsampling operation; B2: The feature pyramid is scanned using the detection head to obtain the bounding boxes, class probabilities, and confidence scores for dental caries and periapical periodontitis; B3: Based on the bounding boxes, class probabilities, and confidence scores of dental caries and periapical periodontitis, a PP-YOLOv2 model is trained using a disease loss function. The trained PP-YOLOv2 model is then used to identify dental caries and periapical periodontitis. The expression for the disease loss function is as follows:

[0019]

[0020]

[0021] In the formula, The disease loss function for the PP-YOLOv2 model. To pinpoint the loss, For confidence loss, For classifying losses, Weights for diseases with ambiguous boundaries For confidence loss weights, For classification loss weights, For generalized intersection-union ratio functions, For bounding box, and The coordinates of the center point of the bounding box. and For the width and height of the bounding box, The actual bounding box to be labeled. For the first The bounding boxes of the candidate boxes, For the first The true bounding boxes of the candidate boxes It is the minimum bounding rectangle. The model predicts the first The confidence score of each candidate box.

[0022] Furthermore, the Faster R-CNN model identifies wisdom teeth, implants, and root canal fillings, including: C1: Candidate boxes are generated using the RPN of the Faster R-CNN model, and RoI Pooling is used to extract features of the wisdom tooth, implant, and root canal filling. The calculation formula is as follows:

[0023] In the formula, For the first Fixed-size feature vectors obtained after pooling candidate boxes This is a max pooling operation; Specifically, for implant extraction, density-sensing weights are introduced, and the calculation formula is as follows:

[0024] In the formula, For density-aware weights, The attenuation coefficient is... For image gradient, for Paradigm; C2: Based on the extracted features of the wisdom tooth, implant, and root canal filling, a Faster R-CNN model is trained using a detection loss function, and recognition is performed based on the trained Faster R-CNN model. The expression for the detection loss function is:

[0025]

[0026] In the formula, To detect the loss function, Locate the loss for RPN. For RPN classification loss, For RCNN, locate the loss. For RCNN classification loss, For the number of regression samples, For smoothing loss function The predicted bounding box parameter offset. This represents the actual bounding box parameter offset.

[0027] Furthermore, the cementoenamel boundary and alveolar ridge of the maxilla and mandible are segmented using the U-Net model, and the vertical distance from the cementoenamel boundary to the alveolar ridge along each tooth axis is calculated. Combined with pixel spacing information, the amount of bone resorption is obtained, including: D1: The cementoenamel boundary and alveolar ridge of the maxilla and mandible are segmented using the U-Net model, and the U-Net model is trained using a hybrid loss function, the expression of which is:

[0028] In the formula, The hybrid loss function for training the U-Net model. For Dice's loss, For balance coefficient, For boundary loss, For pixels The actual label value, For pixels The predicted probability value, This represents the actual boundary region; D2: Based on the cementoenamel boundary and alveolar ridge obtained from the segmentation, calculate the perpendicular distance between the cementoenamel boundary and alveolar ridge along each tooth axis. The calculation formula is as follows:

[0029] In the formula, For the first The average vertical position of the cementoenamel junction in each tooth position For the first The average vertical position of the alveolar ridge in each tooth position. For the dental region, This is a segmentation diagram of the alveolar bone ridge. Diagram showing the cementoenamel boundary; D3: Convert vertical distance to bone resorption and mark the sites of most severe resorption. The calculation formula is as follows:

[0030] In the formula, For the first Bone resorption at each tooth position, This is the pixel pitch conversion factor.

[0031] Further, S3 includes: S301: Based on the disease features of the teeth and the bounding boxes of the tooth positions, the intersection-union ratio (IUU) of the disease features and the bounding boxes of the tooth positions is calculated by iterating through them. The calculation formula is as follows:

[0032] In the formula, The intersection-union ratio of the bounding boxes for disease features and tooth positions. This is the area calculation function. A set of disease characteristics This is the set of bounding boxes for tooth positions. For disease detection boxes, This is a tooth position marker frame. and The center coordinates of the tooth position frame, and The width and height of the tooth positioning frame. According to disease category, The confidence score is... and The coordinates of the center of the disease frame, and The width and height of the disease frame. For tooth position identification; S302: Match the bounding box of the tooth position corresponding to the maximum intersection-union ratio (MUC) with the disease features, thereby binding the disease features of the tooth to the bounding box of the tooth position. The calculation formula is as follows:

[0033] In the formula, The index for the best matching tooth position. For adaptive matching threshold, Based on the threshold, For adjustment coefficients; S303: Based on the bounding box of the tethered tooth position, bone resorption is integrated with disease characteristics to obtain periodontal data for the tooth position, thereby obtaining basic patient information. The calculation formula is as follows:

[0034] In the formula, For the first A complete record of each tooth position. For healthy teeth, For the list of diseases, The threshold for bone resorption. For minimum confidence level, It is an empty set.

[0035] Further, S4 includes: S401: Set up a structured template that includes patient information, quality control ratings, disease lists, and periodontal assessment data; S402: Utilize the local disease knowledge base for mapping to generate disease feature descriptions. The mapping formula is as follows:

[0036] In the formula, For image feature description functions, For location description functions, For size description functions, For density description function, For the image at position The grayscale value at that location; S403: Fill the disease characteristic description and patient basic information into the structured template to obtain a preliminary diagnosis report. The expression of the preliminary diagnosis report is:

[0037] In the formula, This is a preliminary diagnostic report. For structured templates, For patient information, For quality control rating, For diagnostic suggestion functions, This is data for periodontal assessment.

[0038] Furthermore, the review of the preliminary diagnostic report using a human-computer interaction review interface includes: E1: Based on the preliminary diagnostic report, calculate the correction type code using the following formula:

[0039] In the formula, To correct type encoding, Revise the doctor's diagnosis report. For status revision, Additions and deletions to diseases This is a correction vector for bone resorption. E2: Based on the modified type encoding, the diagnostic consistency score and the global consistency score are calculated using the following formula:

[0040]

[0041] In the formula, To diagnose the concordance score, The penalty coefficient is... To report the difference distance function, The score represents the global consistency score. E3: When the global consistency score is less than a preset value, the multi-task AI diagnostic model is retrained using a fine-tuned loss function until the convergence condition is met. The expressions for the fine-tuned loss function and the convergence condition are as follows:

[0042]

[0043] In the formula, To fine-tune the loss function, The original loss function, To correct the loss weights, This is the fine-tuned global consistency score. The global consistency score before fine-tuning. This is the minimum lifting threshold constant; E4: Based on the retrained multi-task AI diagnostic model, a preliminary diagnostic report is generated again and reviewed.

[0044] Furthermore, the reviewed diagnostic report is archived to a structured database, including: Utilize archived diagnostic reports for remote consultations, collaboration, and data sharing; The meeting request function for remote consultation collaboration is as follows:

[0045] In the formula, This is the consultation request function. For the diagnostic report, For identification of consulting experts, For digital signatures, Encrypted using AES-256. For expert public keys; The permission expression for the data sharing is:

[0046] In the formula, For access permission functions, For user identification, Functions for user roles As a dentist As a video expert For patient consent function.

[0047] The beneficial effects of this application are: This application presents an AI-based image-assisted detection method for oral and maxillofacial surgery, achieving intelligent assistance throughout the entire process from image diagnosis to report generation. It deeply embeds AI into the clinical workflow, significantly improving diagnostic efficiency and standardization. Employing a multi-model parallel collaborative architecture, it selects or optimizes the most suitable deep learning model based on the characteristics of different tasks (tooth position, disease, periodontium), improving overall processing speed while ensuring accuracy. Furthermore, it creates a human-machine collaborative model of AI preliminary analysis + physician final review, leveraging AI's advantages in screening, quantification, and standardization while ensuring the physician's core role as the primary diagnostic authority, achieving complementary human and machine strengths. In addition, a structured database and remote collaboration platform are built based on international standards, enabling long-term data value mining and continuous algorithm iteration, and effectively promoting the downward flow and collaboration of high-quality medical resources, demonstrating significant social benefits and promising application prospects. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0049] Figure 1 This is a flowchart illustrating an artificial intelligence-based image-assisted detection method for oral and maxillofacial structures, as provided in an embodiment of this application. Detailed Implementation

[0050] 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 based on this application are within the scope of protection of this application.

[0051] Example 1: This application provides an artificial intelligence-based image-assisted detection method for oral and maxillofacial structures, which can be found in [reference needed]. Figure 1 , Figure 1 The diagram shown is a flowchart illustrating an artificial intelligence-based image-assisted detection method for oral and maxillofacial structures provided in an embodiment of this application, including: S1: Acquire the oral and maxillofacial images to be processed.

[0052] In one embodiment of this application, images can be acquired through two methods: first, physicians can directly upload DICOM format curved tomography films or CBCT images through an integrated web front-end; second, the system can automatically retrieve specified examination images from the hospital's PACS (Picture Archiving and Communication) system through an embedded DICOM gateway. The acquired digital oral and maxillofacial images to be processed include, but are not limited to, curved tomography films and cone-beam CT images. The system supports uploading and reading images in standard formats such as DICOM.

[0053] S2: For images that have passed quality control, the system constructs a multi-task artificial intelligence diagnostic model to analyze the oral and maxillofacial images to be processed.

[0054] In one embodiment of this application, the core of this step is to construct a multi-task parallel inference engine. Its underlying logic is as follows: features are extracted through a shared backbone network, each sub-task model independently processes different diagnostic targets, and finally, low latency (target: ≤500ms / image) is achieved through computational graph parallelization. Let the input image be... ( For the set of real numbers, and (Image width and height, standardized to 2048×1024 after quality control) Quality control passed indicator: (0 indicates failure). The engine output is a multi-task joint feature tensor. ( , The height and width of the feature map, (where the number of channels is the number of channels), and its construction process is as follows: S201: Call the tooth position recognition and segmentation model to segment the precise mask (contour) and bounding box coordinates of all teeth from the image, and automatically encode them according to the FDI tooth position map to obtain the bounding box of the tooth position.

[0055] In one embodiment of this application, a Mask R-CNN model is used to achieve pixel-level tooth segmentation and FDI (International Dental Federation) tooth position encoding (numbers 1-32). The key is to address the boundary blurring problem caused by densely packed teeth. Deformable convolutions enhance geometric adaptability. This Mask R-CNN can effectively perform tasks such as object classification, detection, and semantic segmentation. Different tasks can be accomplished by adding different branches. Mask R-CNN, based on Fast R-CNN with added mask branches, achieves pixel-level object detection. The ROIAlign strategy addresses pixel bias, and the FCN performs semantic segmentation, resulting in high accuracy. Therefore, Mask R-CNN actually consists of four parts: RPN, ROIAlign, Fast R-CNN, and FCN. Specifically, it includes: A1: For the oral and maxillofacial images to be processed, multi-scale feature maps are output using the backbone network ResNet50-vd-dcn. Let the first... The layer features are:

[0056] In the formula, For the first Layer features, For the first Layer features, For the first The residual blocks of the layer, This is deformable convolution, and its expression is:

[0057] In the formula, To output the value of the feature map at a certain location, and For location, For the first One sampling point, The number of sampling points. For the first The weight of each sampling point For the input feature map, and For the preset rule sampling positions, and This is a learnable offset.

[0058] This operation improves robustness to tooth rotation and tilting. The quantification metric uses geometric transformation invariance gain, expressed as:

[0059] In the formula, This represents the average crossover-union ratio (CUIR) of standard convolutional networks on tooth segmentation tasks. This represents the average crossover ratio (CROR) of a deformable convolutional network (DCN) on the tooth segmentation task.

[0060] A2: RPN generates candidate bounding boxes and uses ROIAlign to align features, obtaining the segmentation boxes for the teeth. The expressions for generating the candidate bounding boxes and aligning the features are as follows:

[0061]

[0062] In the formula, For the candidate box set, For the first One candidate box, The total number of candidate boxes. and The center coordinates of the candidate box. and For the first The width and height of each candidate box. For the first The feature vectors after aligning the candidate boxes and For network indexing.

[0063] A3: The contour is output using a segmentation head and optimized using Mask IoU loss. The expression for Mask IoU loss is:

[0064] In the formula, The Mask IoU loss function is... This represents the number of candidate boxes that contain the target object. Let be the intersection-union ratio function. For the first The outline of each candidate box. For a realistic outline, For element-wise multiplication, This is an element-wise addition.

[0065] A4: The centroid coordinates of the dental arch are mapped to FDI tooth position codes using a dental arch geometry model. The expression for the dental arch geometry model is as follows:

[0066] In the formula, For the first FDI tooth position encoding of each candidate box and For image width and height, and For the first The centroid coordinates of the outline of each candidate bounding box. This is the floor function.

[0067] A5: The FDI tooth position code is corrected using a classification head. The maximum probability distribution vector of the tooth category is selected as the FDI tooth position code. The calculation formula is as follows:

[0068] In the formula, Let be the probability distribution vector for tooth categories. For normalized exponential functions, This is the weight matrix of the classification heads. For channel dimension, For the set of real numbers, This is the bias vector. The role of the classification head is to learn subtle differences in tooth features, adapt to individual variations and imaging differences, and correct for errors that may arise from the geometric model: utilizing the aligned features. Representing each tooth region, through the fully connected layer Calculate a score for each tooth category, convert the score into a probability using softmax, and select the category with the highest probability as the final tooth position code.

[0069] A6: Optimize FDI tooth position coding using coding accuracy, the expression for which is:

[0070] In the formula, The accuracy of tooth coding. For the first The actual FDI tooth position code of each candidate box.

[0071] A7: Match the tooth segmentation box with the FDI tooth position code to obtain the tooth position bounding box.

[0072] S202: Call the disease detection model to identify a variety of common diseases in the image.

[0073] In one embodiment of this application, a task-customized detection model is used to handle different disease characteristics (vague boundaries of dental caries / periapical periodontitis and clear morphology of wisdom teeth / implants), and multi-scale features are fused through FPN (Feature Pyramid Network) to solve the problem of missed detection of small targets. Specifically, this includes: using the PP-YOLOv2 model to identify dental caries and periapical periodontitis, and the Faster R-CNN model to identify wisdom teeth, implants, and root canal fillings. PP-YOLO is an algorithm model based on YOLOv3, aiming to achieve a target detector with a relatively balanced effectiveness and efficiency that can be directly applied in practical application scenarios. The YOLO detector consists of three main parts: Backbone, Neck, and Head. Fast-R-CNN is an improvement on the R-CNN algorithm, and the R-CNN algorithm framework is roughly as follows: selective search selects 1000-2000 candidate boxes, constructs an AlexNet convolutional neural network for feature extraction, inputs the features into an SVM classifier for each category for classification, and constructs a regressor to adjust the position of the candidate boxes. In this process, the steps are cumbersome, time-consuming, and the testing speed is slow. Therefore, the improved Fast-R-CNN algorithm is as follows: the last max pooling layer is replaced with an ROI pooling layer; instead of inputting 2000 candidate boxes into the network convolution, the image is normalized to 224×224 and input into the convolutional network to obtain the feature map, effectively avoiding convolution calculations for duplicate candidate boxes; for each candidate box, the corresponding position is found in the feature map according to the proportion, the feature box is cropped, the feature box is divided into 7×7 grids, ROI pooling is performed on each grid, and a 7×7× depth matrix is ​​obtained. The matrix is ​​stretched into a vector and input into the fully connected layer to obtain a fixed-size feature vector, which is then used for softmax classification (using softmax to replace multiple SVM classifiers in RCNN) and candidate box regression.

[0074] In one embodiment of this application, the PP-YOLOv2 model is used to identify dental caries (manifested as a localized low-density area on the crown) and periapical periodontitis (manifested as a low-density shadow in the periapical region), including: B1: Multi-scale features of oral and maxillofacial images are extracted using a backbone network to obtain a feature pyramid. The calculation formula is as follows:

[0075] In the formula, and The first feature pyramid Layer features and The output layer of layer features, This is an upsampling operation.

[0076] B2: The feature pyramid is scanned using a detector head to obtain the bounding boxes for dental caries and periapical periodontitis. Category probability and confidence score ,in and The coordinates of the center point of the bounding box. and This represents the width and height of the bounding box.

[0077] B3: Based on the bounding boxes, class probabilities, and confidence scores of dental caries and periapical periodontitis, a PP-YOLOv2 model is trained using a disease loss function. The trained PP-YOLOv2 model is then used to identify dental caries and periapical periodontitis. The expression for the disease loss function is as follows:

[0078]

[0079]

[0080] In the formula, The disease loss function for the PP-YOLOv2 model. To address the localization loss (GIOU improves the boundary ambiguity problem). For confidence loss, For classifying losses, Weights for diseases with ambiguous boundaries For confidence loss weights, For classification loss weights, For generalized intersection-union ratio functions, For bounding box, and The coordinates of the center point of the bounding box. and For the width and height of the bounding box, The actual bounding box to be labeled. For the first The bounding boxes of the candidate boxes, For the first The true bounding boxes of the candidate boxes It is the minimum bounding rectangle. The model predicts the first The confidence score of each candidate box.

[0081] In one embodiment of this application, the Faster R-CNN model identifies wisdom teeth, (spiral high-density shadows), and root canal fillings (linear high-density shadows), including: C1: Candidate boxes are generated using the RPN of the Faster R-CNN model, and RoI Pooling is used to extract features of the wisdom tooth, implant, and root canal filling. The calculation formula is as follows:

[0082] In the formula, For the first Fixed-size feature vectors obtained after pooling candidate boxes This is a max pooling operation.

[0083] For the extraction of high-density targets (such as implants), density-aware weights are introduced, and the calculation formula is as follows:

[0084] In the formula, For density-aware weights, The attenuation coefficient is... For image gradient, for Paradigm.

[0085] C2: Based on the extracted features of the wisdom tooth, implant, and root canal filling, a Faster R-CNN model is trained using a detection loss function, and recognition is performed based on the trained Faster R-CNN model. The expression for this detection loss function is:

[0086]

[0087] In the formula, To detect the loss function, Locate the loss for RPN. For RPN classification loss, For RCNN, locate the loss. For RCNN classification loss, For the number of regression samples, For smoothing loss function The predicted bounding box parameter offset. This represents the actual bounding box parameter offset.

[0088] To improve the robustness of metal artifact detection, recall is used as a metric, expressed as:

[0089] In the formula, For recall rate, It is a true positive. It is a false negative.

[0090] S203: Call the periodontal condition assessment model (U-Net model) to quantify the degree of alveolar bone resorption in the entire mouth.

[0091] In one embodiment of this application, the U-Net model segments the cementoenamel junction (CEJ) line and alveolar ridge (ABC) line of the maxilla and mandible pixel by pixel on the image. Then, the system calculates the vertical distance from CEJ to ABC along each tooth axis, combines this distance with the pixel spacing information in the DICOM header file, converts it to the amount of bone resorption in millimeters, and marks the sites with the most severe resorption. The segmentation model employs a U-Net encoder-decoder structure, using skip connections to fuse features. The encoder output is set to... The decoder input is , For upsampling operation, for The output feature map of the layer decoder This represents element-wise addition, specifically including: D1: The cementoenamel boundary and alveolar ridge of the maxilla and mandible are segmented using the U-Net model. The U-Net model is then trained using a hybrid loss function to resolve edge blurring. The expression for this hybrid loss function is as follows:

[0092] In the formula, The hybrid loss function for training the U-Net model. For Dice's loss, For balance coefficient, For boundary loss, For pixels The actual label value, For pixels The predicted probability value, This represents the actual boundary region.

[0093] D2: Based on the segmented cementoenamel boundary and alveolar ridge, attachment loss is quantified, and the vertical distance between the cementoenamel boundary and alveolar ridge along each tooth axis is calculated using the following formula:

[0094] In the formula, For the first The average vertical position of the cementoenamel junction in each tooth position For the first The average vertical position of the alveolar ridge in each tooth position. For the dental region, This is a segmentation diagram of the alveolar bone ridge. This is a diagram showing the cementoenamel boundary.

[0095] D3: Will and Converted to bone resorption, and the sites of most severe resorption are marked, the calculation formula is as follows:

[0096] In the formula, For the first Bone resorption at each tooth position, Obtained from DICOM metadata (typical value) ).

[0097] Among these, the error is quantified:

[0098] In the formula, This represents the relative error in bone resorption. This is a relative error. For the first The actual amount of bone resorption in each tooth position.

[0099] In one embodiment of this application, U-Net comprises two parts: the first part is responsible for feature extraction; the second part is an upsampling part, responsible for feature restoration. U-Net converts the regions to be identified into specific "encodings" as class labels according to actual needs. In fact, each object to be identified requires one channel; the number of output channels corresponds to the number of objects to be identified. Finally, a superposition is performed to obtain the desired segmentation result. The segmentation process, based on information provided by the clinical attachment loss map, plots the mesial and distal alveolar ridge crests and cementoenamel junction margins of each tooth in the image, forming two regions: maxillary and mandibular. Each region is labeled to represent a different category, generating a JSON file and a dataset for model training.

[0100] S3: For a "carious cavity" lesion bounding box detected by the disease model, the system iterates through all tooth position bounding boxes and calculates their intersection-over-union (IoU). The lesion bounding box is matched to the tooth with the largest IoU, thus binding the "carious cavity" label to the record of that tooth position. Periodontal data is directly associated with the corresponding tooth position based on its jawbone region. Finally, a JSON structured data is generated as follows, which contains the patient's basic information, specifically including: S301: Based on the disease features of the teeth and the bounding boxes of the tooth positions, the intersection-union ratio (IUU) of the disease features and the bounding boxes of the tooth positions is calculated by iterating through them. The calculation formula is as follows:

[0101] In the formula, The intersection-union ratio of the bounding boxes for disease features and tooth positions. This is the area calculation function. A set of disease characteristics This is the set of bounding boxes for tooth positions. For disease detection boxes, This is a tooth position marker frame. and The center coordinates of the tooth position frame, and The width and height of the tooth positioning frame. According to disease category, The confidence score is... and The coordinates of the center of the disease frame, and The width and height of the disease frame. Used as a tooth position marker.

[0102] S302: Match the bounding box of the tooth position corresponding to the maximum intersection-union ratio (MUC) with the disease features, thereby binding the disease features of the tooth to the bounding box of the tooth position. The calculation formula is as follows:

[0103] In the formula, The index for the best matching tooth position. For adaptive matching threshold, Based on the threshold, This is for adjusting the coefficient.

[0104] S303: Based on the bounding box of the tethered tooth position, bone resorption is integrated with disease characteristics to obtain periodontal data for the tooth position, thereby obtaining basic patient information. The calculation formula is as follows:

[0105] In the formula, For the first A complete record of each tooth position. For healthy teeth, For the list of diseases, The threshold for bone resorption. For minimum confidence level, It is an empty set.

[0106] Here is an example of a JSON structured data: json { "tooth_#36": { "position": "Mandibular Right First Molar", "status": "diseased", "diseases": [ {"type": "caries", "confidence": 0.92}, {"type": "apical_lesion", "confidence": 0.88} ], "bone_loss_mm": 4.2 }, / / ... Other tooth position data } S4: The system automatically populates a preset HTML report template with the generated structured data, patient basic information, and quality control rating. At the same time, the engine queries the local knowledge base for the terms "dental caries" and "periapical periodontitis," inserting standardized recommendations such as "clinical probing is recommended to determine the depth of caries" and "pulp vitality is recommended" into the "health tips" section of the report, forming a preliminary diagnostic report with illustrations.

[0107] In one embodiment of this application, a structured template is first defined. ,in, For patient information (name, ID, etc.) For quality control rating ( , (Rank is 4, with 4 being the optimal). For the list of diseases, For periodontal assessment data. Secondly, utilize the local disease knowledge base ( , According to disease category, For image feature description functions, Mapping the diagnostic suggestion function to generate disease feature descriptions:

[0108] In the formula, For image feature description functions, For location description functions, For size description functions, For density description function, For the image at position The grayscale value at that location.

[0109] Finally, the disease characteristics and basic patient information are populated into the structured template to obtain a preliminary diagnostic report:

[0110] In the formula, This is a preliminary diagnostic report. For structured templates, For patient information, For quality control rating, For diagnostic suggestion functions, This is data for periodontal assessment.

[0111] Among them, the use of integrity indicators Evaluation:

[0112] S5: Use the human-computer interaction review interface to review the preliminary diagnostic report and archive the reviewed diagnostic report to the structured database.

[0113] In one embodiment of this application, a draft report is pushed to the physician's workbench. In the interactive interface, the left side displays the original image, with AI recognition results (such as tooth segmentation contours, disease annotation boxes, and periodontal lines) overlaid with a semi-transparent layer; the right side displays a structured report editor. The physician can click to confirm AI annotations, directly drag to modify the position of the annotation boxes, or add personalized diagnostic opinions in the report text box. All corrections are recorded, and the confirmation process includes: E1: Based on the preliminary diagnostic report, calculate the correction type code using the following formula:

[0114] In the formula, To correct type encoding, Revise the doctor's diagnosis report. This indicates a status correction (normal → diseased). Indicates the addition or deletion of diseases. The basic increase represents Representing disease deletion, This is the correction vector for bone resorption.

[0115] E2: Based on the modified type encoding, the diagnostic consistency score and the global consistency score are calculated using the following formula:

[0116]

[0117] In the formula, To diagnose the concordance score, The penalty coefficient is... To report the difference distance function, The score represents the global consistency score.

[0118] E3: When the global consistency score is less than a preset value (which can be 0.8), the multi-task AI diagnostic model is retrained using a fine-tuned loss function until the convergence condition is met. The expressions for the fine-tuned loss function and the convergence condition are as follows:

[0119]

[0120] In the formula, To fine-tune the loss function, The original loss function, To correct the loss weights, This is the fine-tuned global consistency score. The global consistency score before fine-tuning. This is the minimum lifting threshold constant; E4: Based on the retrained multi-task AI diagnostic model, a preliminary diagnostic report is generated again and reviewed.

[0121] In one embodiment of this application, the final report after review is archived to a structured database, supporting remote consultation collaboration and data sharing. Structured data storage is implemented based on the openEHR standard, supporting FHIR (FastHealthcare Interoperability Resources) protocol transmission. The core features are archetype mapping and privacy protection quantification.

[0122] Data archiving model: Define openEHR archetype DENTAL_EXAM.v1, field mapping:

[0123] in, For the first An openEHR prototype instance for each tooth position. For the tooth serial number, For tooth position marking, For dental health, For the list of diseases, For disease codes, Let be the confidence interval. This represents the amount of attachment loss / bone resorption.

[0124] Data is compressed and stored as follows: This is a complete oral examination record, containing prototype examples of all 32 teeth, with a compression ratio of: .

[0125] The meeting request function in remote consultation collaboration is as follows:

[0126] In the formula, This is the consultation request function. For the diagnostic report, For identification of consulting experts, For digital signatures (SHA-256). Encrypted using AES-256. For expert public keys.

[0127] The permission expression for data sharing is:

[0128] In the formula, For access permission functions, For user identification, Functions for user roles As a dentist As a video expert For patient consent function.

[0129] Among them, privacy protection indicators are adopted. :

[0130] Structured data for AI training, sample gain:

[0131] In the formula, This represents the expected improvement in model accuracy. This is the gain coefficient. The performance of the prediction model is improved by adding more archived samples.

[0132] This application presents an AI-based image-assisted detection method for oral and maxillofacial surgery, achieving intelligent assistance throughout the entire process from image diagnosis to report generation. It deeply embeds AI into the clinical workflow, significantly improving diagnostic efficiency and standardization. Employing a multi-model parallel collaborative architecture, it selects or optimizes the most suitable deep learning model based on the characteristics of different tasks (tooth position, disease, periodontium), improving overall processing speed while ensuring accuracy. Furthermore, it creates a human-machine collaborative model of AI preliminary analysis + physician final review, leveraging AI's advantages in screening, quantification, and standardization while ensuring the physician's core role as the primary diagnostic authority, achieving complementary human and machine strengths. In addition, a structured database and remote collaboration platform are built based on international standards, enabling long-term data value mining and continuous algorithm iteration, and effectively promoting the downward flow and collaboration of high-quality medical resources, demonstrating significant social benefits and promising application prospects.

[0133] It should be noted that those skilled in the art will recognize that the embodiments described herein are for the purpose of helping readers understand the principles of this application, and should be understood as not limiting the scope of protection of this application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this application without departing from the essence of this application, and these modifications and combinations are still within the scope of protection of this application.

Claims

1. An artificial intelligence-based image-assisted detection method for oral and maxillofacial structures, characterized in that, include: S1: Acquire the oral and maxillofacial images to be processed; S2: Construct a multi-task artificial intelligence diagnostic model, and identify the oral and maxillofacial images to be processed based on the multi-task artificial intelligence diagnostic model to obtain the bounding box of the tooth position, disease features and the amount of bone resorption of the tooth. The multi-task artificial intelligence diagnostic model includes Mask R-CNN model, PP-YOLOv2 model, Faster R-CNN model and U-Net model. S3: Bind the segmentation frame of the tooth position, disease characteristics, and the amount of bone resorption of the tooth to obtain the patient's basic information; S4: Call the structured report generation engine, combine patient basic information and local disease knowledge base to obtain a preliminary diagnosis report; S5: Use the human-computer interaction review interface to review the preliminary diagnostic report and archive the reviewed diagnostic report to the structured database.

2. The artificial intelligence-based image-assisted detection method for oral and maxillofacial structures according to claim 1, characterized in that, S2 includes: The Mask R-CNN model is used to identify and segment the oral and maxillofacial images to be processed, obtain the contour and segmentation box of each tooth, and encode them to obtain the bounding box of the tooth position. The PP-YOLOv2 model and the Faster R-CNN model were used to identify disease features in the oral and maxillofacial images to be processed. The PP-YOLOv2 model was used to identify dental caries and periapical periodontitis, while the Faster R-CNN model was used to identify wisdom teeth, implants and root canal fillings. The cementoenamel boundary and alveolar ridge of the upper and lower jaws were segmented using the U-Net model. The vertical distance from the cementoenamel boundary to the alveolar ridge along each tooth axis was calculated. The amount of bone resorption was obtained by combining the pixel spacing information.

3. The artificial intelligence-based image-assisted detection method for oral and maxillofacial structures according to claim 2, characterized in that, The Mask R-CNN model is used to identify and segment the oral and maxillofacial images to be processed, obtaining the contour and segmentation box of each tooth, and encoding them to obtain the bounding box of the tooth position, including: A1: For the oral and maxillofacial images to be processed, multi-scale feature maps are output using the ResNet50-vd-dcn backbone network, where the first... The layer features are: In the formula, For the first Layer features, For the first Layer features, For the first The residual blocks of the layer, It is a deformable convolution; The expression for deformable convolution is: In the formula, To output the value of the feature map at a certain location, and For location, For the first One sampling point, The number of sampling points. For the first The weight of each sampling point For the input feature map, and For the preset rule sampling positions, and The learnable offset; A2: Generate candidate boxes based on multi-scale feature maps, and align the feature vectors of the candidate boxes to obtain the segmentation boxes for the teeth. The expressions for generating the candidate boxes and aligning the feature vectors of the candidate boxes are as follows: In the formula, For the candidate box set, For the first One candidate box, and The center coordinates of the candidate box. and For the first The width and height of each candidate box. For the first The feature vectors after aligning the candidate boxes and For network indexing; A3: The contour is output using a segmentation head and optimized using Mask IoU loss. The expression for Mask IoU loss is: In the formula, The Mask IoU loss function is... This represents the number of candidate boxes that contain the target object. Let be the intersection-union ratio function. For the first The outline of each candidate box. For a realistic outline, For element-wise multiplication, This is an element-wise addition; A4: The centroid coordinates of the contour are mapped to FDI tooth position codes using a dental arch geometry model, and the FDI tooth position codes are corrected using a classification head. The maximum probability distribution vector of the tooth category is selected as the FDI tooth position code. The expressions for the dental arch geometry model and the correction are as follows: In the formula, For the first FDI tooth position encoding of each candidate box and For image width and height, and For the first The centroid coordinates of the outline of each candidate bounding box. This is the floor function; Let be the probability distribution vector for tooth categories. For normalized exponential functions, This is the weight matrix of the classification heads. For channel dimension, For the set of real numbers, It is the bias vector; A5: Optimize FDI tooth position coding using coding accuracy, whereby the expression for coding accuracy is: In the formula, The accuracy of tooth coding. For the first The actual FDI tooth position code of each candidate box; A6: Match the segmentation box of the tooth with the FDI tooth position code to obtain the bounding box of the tooth position.

4. The artificial intelligence-based image-assisted detection method for oral and maxillofacial structures according to claim 3, characterized in that, The PP-YOLOv2 model identifies dental caries and periapical periodontitis, including: B1: Multi-scale features of oral and maxillofacial images are extracted using a backbone network to obtain a feature pyramid. The calculation formula is as follows: In the formula, and The first feature pyramid Layer features and The output layer of layer features, This is an upsampling operation; B2: The feature pyramid is scanned using the detection head to obtain the bounding boxes, class probabilities, and confidence scores for dental caries and periapical periodontitis; B3: Based on the bounding boxes, class probabilities, and confidence scores of dental caries and periapical periodontitis, a PP-YOLOv2 model is trained using a disease loss function. The trained PP-YOLOv2 model is then used to identify dental caries and periapical periodontitis. The expression for the disease loss function is as follows: In the formula, The disease loss function for the PP-YOLOv2 model. To pinpoint the loss, For confidence loss, For classifying losses, Weights for diseases with ambiguous boundaries For confidence loss weights, For classification loss weights, For generalized intersection-union ratio functions, For bounding box, and The coordinates of the center point of the bounding box. and For the width and height of the bounding box, The actual bounding box to be labeled. For the first The bounding boxes of the candidate boxes, For the first The true bounding boxes of the candidate boxes It is the minimum bounding rectangle. The model predicts the first The confidence score of each candidate box.

5. The artificial intelligence-based image-assisted detection method for oral and maxillofacial structures according to claim 4, characterized in that, The Faster R-CNN model identifies wisdom teeth, implants, and root canal fillings, including: C1: Candidate boxes are generated using the RPN of the Faster R-CNN model, and RoI Pooling is used to extract features of the wisdom tooth, implant, and root canal filling. The calculation formula is as follows: In the formula, For the first Fixed-size feature vectors obtained after pooling candidate boxes This is a max pooling operation; Specifically, for implant extraction, density-sensing weights are introduced, and the calculation formula is as follows: In the formula, For density-aware weights, The attenuation coefficient is... For image gradient, for Paradigm; C2: Based on the extracted features of the wisdom tooth, implant, and root canal filling, a Faster R-CNN model is trained using a detection loss function, and recognition is performed based on the trained Faster R-CNN model. The expression for the detection loss function is: In the formula, To detect the loss function, Locate the loss for RPN. For RPN classification loss, For RCNN, locate the loss. For RCNN classification loss, For the number of regression samples, For smoothing loss function The predicted bounding box parameter offset. This represents the actual bounding box parameter offset.

6. The artificial intelligence-based image-assisted detection method for oral and maxillofacial structures according to claim 5, characterized in that, The process involves using the U-Net model to segment the cementoenamel boundary and alveolar ridge of the upper and lower jaws, calculating the vertical distance from the cementoenamel boundary to the alveolar ridge along each tooth axis, and combining this with pixel spacing information to obtain the amount of bone resorption, including: D1: The cementoenamel boundary and alveolar ridge of the maxilla and mandible are segmented using the U-Net model, and the U-Net model is trained using a hybrid loss function, the expression of which is: In the formula, The hybrid loss function for training the U-Net model. For Dice's loss, For balance coefficient, For boundary loss, For pixels The actual label value, For pixels The predicted probability value, This represents the actual boundary region; D2: Based on the cementoenamel boundary and alveolar ridge obtained from the segmentation, calculate the perpendicular distance between the cementoenamel boundary and alveolar ridge along each tooth axis. The calculation formula is as follows: In the formula, For the first The average vertical position of the cementoenamel junction in each tooth position For the first The average vertical position of the alveolar ridge in each tooth position. For the dental region, This is a segmentation diagram of the alveolar bone ridge. Diagram showing the cementoenamel boundary; D3: Convert vertical distance to bone resorption and mark the sites of most severe resorption. The calculation formula is as follows: In the formula, For the first Bone resorption at each tooth position, This is the pixel pitch conversion factor.

7. The artificial intelligence-based image-assisted detection method for oral and maxillofacial structures according to claim 6, characterized in that, The S3 includes: S301: Based on the disease features of the teeth and the bounding boxes of the tooth positions, the intersection-union ratio (IUU) of the disease features and the bounding boxes of the tooth positions is calculated by iterating through them. The calculation formula is as follows: In the formula, The intersection-union ratio of the bounding boxes for disease features and tooth positions. This is the area calculation function. A set of disease characteristics This is the set of bounding boxes for tooth positions. For disease detection boxes, This is a tooth position marker frame. and The center coordinates of the tooth position frame, and The width and height of the tooth positioning frame. According to disease category, The confidence score is... and The coordinates of the center of the disease frame, and The width and height of the disease frame. For tooth position identification; S302: Match the bounding box of the tooth position corresponding to the maximum intersection-union ratio (MUC) with the disease features, thereby binding the disease features of the tooth to the bounding box of the tooth position. The calculation formula is as follows: In the formula, The index for the best matching tooth position. For adaptive matching threshold, Based on the threshold, For adjustment coefficients; S303: Based on the bounding box of the tethered tooth position, bone resorption is integrated with disease characteristics to obtain periodontal data for the tooth position, thereby obtaining basic patient information. The calculation formula is as follows: In the formula, For the first A complete record of each tooth position. For healthy teeth, For the list of diseases, The threshold for bone resorption. For minimum confidence level, It is an empty set.

8. The artificial intelligence-based image-assisted detection method for oral and maxillofacial structures according to claim 7, characterized in that, The S4 includes: S401: Set up a structured template that includes patient information, quality control ratings, disease lists, and periodontal assessment data; S402: Utilize the local disease knowledge base for mapping to generate disease feature descriptions. The mapping formula is as follows: In the formula, For image feature description functions, For location description functions, For size description functions, For density description function, For the image at position The grayscale value at that location; S403: Fill the disease characteristic description and patient basic information into the structured template to obtain a preliminary diagnosis report. The expression of the preliminary diagnosis report is: In the formula, This is a preliminary diagnostic report. For structured templates, For patient information, For quality control rating, For diagnostic suggestion functions, This is data for periodontal assessment.

9. The artificial intelligence-based image-assisted detection method for oral and maxillofacial structures according to claim 8, characterized in that, The review of the preliminary diagnostic report using a human-computer interaction review interface includes: E1: Based on the preliminary diagnostic report, calculate the correction type code using the following formula: In the formula, To correct type encoding, Revise the doctor's diagnosis report. For status revision, Additions and deletions to diseases This is a correction vector for bone resorption. E2: Based on the modified type encoding, the diagnostic consistency score and the global consistency score are calculated using the following formula: In the formula, To diagnose the concordance score, The penalty coefficient is... To report the difference distance function, The score represents the global consistency score. E3: When the global consistency score is less than a preset value, the multi-task AI diagnostic model is retrained using a fine-tuned loss function until the convergence condition is met. The expressions for the fine-tuned loss function and the convergence condition are as follows: In the formula, To fine-tune the loss function, The original loss function, To correct the loss weights, This is the fine-tuned global consistency score. The global consistency score before fine-tuning. This is the minimum lifting threshold constant; E4: Based on the retrained multi-task AI diagnostic model, a preliminary diagnostic report is generated again and reviewed.

10. The artificial intelligence-based image-assisted detection method for oral and maxillofacial structures according to claim 9, characterized in that, The reviewed diagnostic report is archived to a structured database, including: Utilize archived diagnostic reports for remote consultations, collaboration, and data sharing; The meeting request function for remote consultation collaboration is as follows: In the formula, This is the consultation request function. For the diagnostic report, For identification of consulting experts, For digital signatures, Encrypted using AES-256. For expert public keys; The permission expression for the data sharing is: In the formula, For access permission functions, For user identification, Functions for user roles As a dentist As a video expert For patient consent function.

Images