An ai-enhanced real-time occlusion detection system based on intraoral images

Abstract

The application relates to the field of oral medicine diagnosis and discloses an AI-enhanced real-time occlusion detection system based on intraoral images, which comprises a multi-source intraoral image acquisition module, which is used for acquiring dynamic oral image data through a high-definition oral endoscope, recording the three-dimensional structure of a tooth surface in combination with a laser scanning device, and collecting the force distribution of an occlusion contact point through a pressure sensor array; the multi-source intraoral image acquisition module comprises a time stamp synchronization unit, which performs multi-source synchronization on the oral image frame sequence and the occlusion force distribution data by using a unified time mark; through the multi-source intraoral image acquisition module, the synchronous acquisition of dynamic images, three-dimensional structure data and occlusion force distribution is realized, and the time sequence consistency of the data is ensured through the time stamp mark. Compared with the acquisition mode of a traditional single data source, multi-source acquisition can comprehensively record the dynamic characteristics in the occlusion process, provide rich and accurate input for subsequent AI analysis, and avoid data omission or acquisition deviation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oral medical diagnosis, in particular to an AI-enhanced real-time occlusion detection system based on intraoral images. BACKGROUND

[0002] Oral health is one of the important factors affecting human overall health, and the abnormality of occlusion is closely related to temporomandibular joint disease, bruxism and various oral chronic pain diseases. Clinically, occlusion detection and analysis are of great significance for predicting temporomandibular joint disorder, developing orthodontic treatment plan, and evaluating orthodontic treatment effect. However, occlusion detection usually needs to combine dynamic images, three-dimensional structure data and mechanical distribution, etc. Therefore, it puts forward higher requirements for technology and equipment. At the same time, the traditional occlusion detection method often depends on the experience of doctors, and lacks objective, real-time and high-precision analysis tools, which greatly limits the diagnosis efficiency and treatment quality of oral medicine.

[0003] In the prior art, the equipment and method for occlusion detection mainly include digital pressure detection system and oral three-dimensional scanner. These tools can provide pressure distribution information and tooth surface geometry structure when the patient bites, respectively. However, these systems are usually single-function, and can only collect a certain type of data, which is difficult to record the dynamic changes in the occlusion process comprehensively. In addition, although some existing intelligent detection systems integrate multiple data sources, the collection and analysis process of multi-source data often exists the problems of time sequence asynchronization and information isolation, which cannot realize real-time and high-precision comprehensive diagnosis. At the same time, the real-time performance and deployment flexibility of the existing system are poor, which is difficult to adapt to the resource-limited scene.

[0004] In the prior art, there is a key problem: it is difficult to realize efficient collection and real-time fusion of multi-source data, and it is difficult to provide comprehensive and accurate occlusion detection and diagnosis results. This problem is particularly evident in the current insufficient development of multi-modal data analysis technology and limited multi-device collaboration capability. In order to solve this problem, a system capable of realizing synchronous collection, efficient fusion and intelligent diagnosis of multi-source data is needed, so as to provide real-time and accurate occlusion detection support for clinics. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides an AI-enhanced real-time occlusion detection system based on intraoral images, which solves the problem that multi-source data cannot be efficiently collected and fused in real time, and it is difficult to provide comprehensive and accurate occlusion detection and diagnosis results.

[0006] In order to achieve the above object, the present application is realized by the following technical scheme: An AI enhanced real-time occlusion detection system based on intraoral image, comprising a multi-source intraoral image acquisition module, which is used to acquire dynamic oral image data through a high-definition oral endoscope, and record the three-dimensional structure of the tooth surface in combination with a laser scanning device, and simultaneously acquire the force distribution of the occlusal contact point through a pressure sensor array;

[0007] The multi-source intraoral image acquisition module comprises a timestamp synchronization unit, which performs multi-source synchronization on the oral image frame sequence and the occlusal force distribution data by using a unified time marker;

[0008] A data preprocessing and enhancement module is connected with the multi-source intraoral image acquisition module, and is used to perform noise filtering, illumination equalization and geometric distortion calibration on the acquired image data, and generate multi-angle samples through data enhancement means such as random rotation, cropping and mirror flipping;

[0009] An AI-based occlusion detection module is connected with the data preprocessing and enhancement module, and is used to perform feature extraction, occlusal point detection and dynamic change modeling on the processed data through a deep learning model;

[0010] A real-time dynamic feedback and abnormality analysis module is connected with the AI-based occlusion detection module, and is used to generate an occlusal point distribution map and a force distribution heat map, and analyze the occlusal data;

[0011] A multi-modal data fusion module is connected with the AI-based occlusion detection module and the real-time dynamic feedback and abnormality analysis module, respectively, and is used to fuse image data, occlusal force data and patient medical record information;

[0012] A lightweight edge computing and system deployment module is connected with the above modules, and is used to optimize the AI model calculation efficiency and support the collaborative calculation of edge devices and the cloud.

[0013] Preferably, the high-definition oral endoscope is used to acquire dynamic image data with a resolution greater than 2K and a frame rate ≥60fps;

[0014] The laser scanning device is used to capture the three-dimensional structure and surface texture of the teeth;

[0015] The micro pressure sensor array is used to acquire the force distribution of the occlusal contact point;

[0016] The timestamp synchronization unit is used to add a unified time marker to the dynamic image and the force distribution data.

[0017] Preferably, a noise reduction processing unit is connected with the multi-source intraoral image acquisition module, and is used to reduce noise of the image data based on a non-local mean filtering technology and a spectral domain anti-reflection technology;

[0018] An image correction unit is configured to adjust the brightness and contrast of the image based on an adaptive light equalization algorithm.

[0019] A geometric calibration unit is configured to correct the geometric distortion of the image through a feature point matching algorithm.

[0020] A data enhancement unit is configured to randomly rotate, crop and mirror flip the corrected image data to generate diversified image samples.

[0021] Preferably, a deep learning model is connected to the data preprocessing and enhancement module and is configured to extract multi-scale features through a feature pyramid network.

[0022] An attention mechanism module is embedded in a convolution block attention module to enhance the detection capability of small targets.

[0023] A dynamic modeling unit is configured to perform time modeling on the image frame sequence based on a bidirectional LSTM and a time convolution network.

[0024] A classification and positioning unit is configured to optimize the class distribution of the bite point detection based on a focal loss and optimize the positioning of the target bounding box through an IoU loss.

[0025] Preferably, a transfer learning module is configured to be pre-trained on a general object detection dataset and adapted to oral image data through fine-tuning.

[0026] An adaptive hyperparameter optimization module is configured to adjust the learning rate, batch size and regularization parameter through a Bayesian optimization algorithm.

[0027] Preferably, a real-time feedback unit is connected to the AI-based bite detection module and is configured to generate a bite point distribution map and a force distribution heat map.

[0028] An anomaly analysis unit is configured to extract bite abnormality features from the detection results and generate a diagnosis report.

[0029] A visualization interface unit is configured to dynamically display three-dimensional bite data and historical data comparison analysis.

[0030] Preferably, a data alignment unit is connected to the AI-based bite detection module and is configured to uniformly encode image data and mechanical data through a coordinate channel attention mechanism.

[0031] A joint modeling unit is configured to interactively model image data, mechanical features and medical record data based on a Transformer framework.

[0032] A fusion analysis unit is configured to generate comprehensive bite diagnosis results and correction suggestions.

[0033] Preferably, a model compression unit is connected with the AI-based occlusion detection module, for compressing the model through pruning and quantization techniques;

[0034] A knowledge distillation module is used to optimize the lightweight student model through a teacher model;

[0035] An edge-cloud collaborative unit is used to complete real-time inference on the edge device and migrate complex computing tasks to the cloud.

[0036] Preferably, the detection function of the system is realized through the connection of the multi-source intraoral image acquisition module, the data preprocessing and enhancement module, and the AI-based occlusion detection module to complete the real-time detection of dynamic bite points and contact points.

[0037] Preferably, the system is connected with the hospital information system through an API interface to realize the sharing and synchronization of occlusion detection data and support cross-device diagnosis record updating.

[0038] The present application provides an AI-enhanced real-time occlusion detection system based on intraoral images.

[0039] The present application has the following advantages:

[0040] 1. The present application realizes the synchronous acquisition of dynamic images, three-dimensional structure data and occlusion force distribution through the multi-source intraoral image acquisition module, and ensures the time sequence consistency of the data through timestamp marking. Compared with the traditional single data source acquisition method, multi-source acquisition can record the dynamic characteristics in the occlusion process comprehensively, provide rich and accurate input for subsequent AI analysis, and avoid data omission or acquisition deviation.

[0041] 2. The present application uses the data preprocessing and enhancement module to perform noise reduction, light balance, geometric correction and other optimization processing on the collected data, and uses data enhancement technology to expand the diversity of samples. These technologies greatly improve the quality of input data and the generalization ability of the model, so that the system can accurately identify the bite points and contact points in the complex oral environment, and improve the robustness of the system in different scenes.

[0042] 3. The AI-based occlusion detection module of the present application uses the YOLOv8 deep learning model combined with dynamic modeling technology to detect the bite points and contact points with high precision, and captures the time sequence change characteristics in the occlusion process through dynamic modeling. Compared with traditional manual analysis, this module can quickly and accurately identify occlusion abnormalities and provide reliable data support for real-time diagnosis.

[0043] 4、The real-time dynamic feedback and abnormal analysis module of the application realizes the dynamic display of the occlusion process by generating occlusion point distribution graph, force distribution heat map and dynamic change curve, and identifies problems such as eccentric occlusion, overload and contact abnormality by combining rule analysis and AI detection results. This intuitive feedback method not only improves the doctor's diagnosis efficiency, but also provides visual occlusion health assessment for patients.

[0044] 5、The multi-modal data fusion module of the application generates comprehensive diagnosis results, including occlusion abnormality score, abnormal mode classification and diagnosis suggestion, by deep fusion of dynamic images, force distribution and medical record data through joint modeling. Compared with single data source analysis, multi-modal fusion can capture deeper correlation features and provide doctors with more comprehensive diagnosis basis, improving the intelligent level of the system.

[0045] 6、Through the lightweight edge computing and system deployment module, the application uses model compression and hardware acceleration technology to realize real-time inference on embedded devices, and at the same time, through the edge-cloud collaborative mechanism, complex tasks are allocated to the cloud for processing. This design not only guarantees the real-time and stability of the system, but also can run efficiently in resource-limited devices, suitable for clinical and mobile medical scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is the main framework diagram of the application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the application will be described below with reference to the drawings in the specification of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0048] Please refer to the drawings in the specification of the application Figure 1 The embodiment of the application provides an AI enhanced real-time occlusion detection system based on intraoral images, which comprises:

[0049] Multi-source intraoral image acquisition module

[0050] For the multi-source intraoral image acquisition module, the core goal of the system is to realize comprehensive data acquisition of the patient's occlusion dynamics through the cooperative work of multiple hardware devices, including dynamic oral image, three-dimensional structure of teeth and occlusion force distribution and other core information. The acquisition of multi-source data is the basis for subsequent processing and AI analysis of the system, and the time synchronization mechanism is used to ensure the time sequence consistency between different data sources. The design of the multi-source intraoral image acquisition module needs to consider many factors, including the accuracy of data acquisition, synchronization requirements and the reliability of hardware, in order to meet the needs of real-time occlusion detection.

[0051] Module composition and function

[0052] In this embodiment, the multi-source intraoral image acquisition module includes the following components:

[0053] High-definition intraoral camera: used to acquire high-resolution image data of patient's occlusion dynamics;

[0054] Laser scanning device: used to record the three-dimensional structural features of the patient's tooth surface;

[0055] Pressure sensor array: used to acquire real-time force distribution data of the occlusion contact point;

[0056] Timestamp synchronization unit: used to add uniform time markers to the acquired dynamic image data, three-dimensional structure data and mechanical data.

[0057] Data acquisition process

[0058] In this embodiment, the working steps of the multi-source intraoral image acquisition module are as follows:

[0059] The high-definition intraoral camera acquires real-time dynamic images of the patient's oral cavity, with a resolution of 2K or above and a frame rate of 60fps. This image data is mainly used to capture the dynamic changes in the patient's occlusion process, including the formation and disappearance of tooth contact points. The endoscope is designed to be handheld for easy positioning of the patient's key oral areas by the doctor.

[0060] The laser scanning device acquires the three-dimensional geometry of the patient's teeth through a non-contact laser scanning method. The scanning resolution of the device reaches 50μm, which is used to record the detailed features of the tooth surface, such as the position of the occlusion point and its surrounding geometric relationship. During the three-dimensional scanning process, the device automatically calibrates the laser angle and focal length to ensure scanning accuracy.

[0061] The pressure sensor array is arranged on the occlusion contact surface of the patient's teeth to acquire the force distribution of the occlusion contact point. The single-point detection accuracy of each sensor is 1N, and the sampling frequency is 500Hz. The pressure sensor sends the real-time acquired data to the processing module through the data transmission unit, ensuring that the mechanical information can be updated synchronously.

[0062] The timestamp synchronization unit uses a unified time reference to synchronously mark the above three types of data (dynamic image, three-dimensional structure and force distribution), ensuring the time sequence consistency of the multi-source data. The timestamp information is recorded in units of microseconds, ensuring that the data can accurately correspond to the same time point of the occlusion action in subsequent processing.

[0063] High-definition intraoral camera:

[0064] In this embodiment, the endoscope enhances imaging clarity through a multi-layer anti-reflection coating lens and integrates a high-sensitivity CMOS sensor for high-definition imaging in low-light environments. The endoscope has a built-in ring-shaped LED light source that provides uniform illumination, eliminating image interference caused by local shadows. The endoscope's video output is connected to a pre-processing module through a high-speed data interface (such as USB 3.0), supporting real-time data transmission.

[0065] Laser scanning device:

[0066] In this embodiment, the laser scanning device uses point cloud scanning technology to accumulate a complete three-dimensional structure of the teeth through multiple scans. The device has a built-in adaptive focal length adjustment system that adjusts the laser focal point position according to the depth of the patient's teeth, avoiding data loss or overlap. Laser scanning data is compressed through point cloud encoding and transmitted to the data processing module to reduce transmission bandwidth occupancy.

[0067] Pressure sensor array:

[0068] In this embodiment, the sensor array is made of piezoelectric ceramic material, with high sensitivity and fast response capability. The arrangement of the sensor array is adjusted according to the actual occlusion form of the patient's teeth to ensure that the force distribution data covers all contact areas. Sensor data is transmitted synchronously through a multi-channel data acquisition unit with a sampling frequency of 500Hz, supporting real-time monitoring of force distribution changes.

[0069] Timestamp synchronization unit:

[0070] In this embodiment, the timestamp synchronization unit provides a unified time reference through a high-precision crystal oscillator and aligns the data streams of the endoscope, laser scanning device, and pressure sensor through a synchronization protocol. The timestamp records with a precision of microseconds, and each frame of image data, each scanning point, and each set of force distribution data is accompanied by a timestamp, ensuring the synchronization of the data.

[0071] Data output

[0072] In this embodiment, the output data of the multi-source intraoral image acquisition module includes:

[0073] Dynamic image data: records the patient's occlusion dynamics at a frequency of 60 frames per second, in the format of high-resolution video stream;

[0074] Three-dimensional structure data: outputs the three-dimensional geometric information of the tooth surface in point cloud data format, with a data storage precision of 50μm;

[0075] Occlusal force distribution data: outputs the force value of each occlusal contact point in the form of a force matrix, with units of Newton (N);

[0076] Timestamp information: each set of data is accompanied by a timestamp to mark the time point of acquisition.

[0077] In this embodiment, the output data of the multi-source intraoral image acquisition module is directly transmitted to the data preprocessing and enhancement module. High-definition dynamic image data is used for image noise reduction and geometric calibration; three-dimensional structure data is used for point cloud optimization; bite force distribution data is used for generating pressure mapping. Timestamp synchronization information is used in the preprocessing stage to time-align the data stream, ensuring that the subsequent processing modules can accurately process multi-source data.

[0078] Data preprocessing and enhancement module

[0079] For the data preprocessing and enhancement module, its main function is to normalize the data output by the multi-source intraoral image acquisition module, including noise reduction, light balance, geometric correction and data enhancement, to optimize data quality and provide high-quality input data for subsequent AI detection. The design of this module needs to combine the characteristics of multi-source data and the downstream analysis requirements to ensure that the processed data meets the accuracy requirements of bite detection in terms of spatial consistency, temporal continuity and content integrity.

[0080] Image preprocessing: mainly includes noise filtering, light balance and geometric distortion correction. These processes are to improve image quality, reduce unnecessary noise, improve the accuracy of model training and avoid overfitting.

[0081] Data enhancement: including random rotation, cropping, mirror flipping and other means, aiming to expand the training set through diversified training data and increase the generalization ability of the model.

[0082] In this embodiment, the data preprocessing and enhancement module receives dynamic image data, three-dimensional structure data and force distribution data from the multi-source intraoral image acquisition module, and processes these data accordingly, including:

[0083] Noise reduction, light balance and geometric correction of image data;

[0084] Point cloud optimization and data compression of three-dimensional structure data;

[0085] Normalization of force distribution data;

[0086] Data enhancement techniques generate samples with multiple angles and features.

[0087] The goal of this module is to improve the quality and adaptability of data through a series of normalization and enhancement techniques, providing more accurate input data for AI models.

[0088] Noise reduction processing

[0089] In this embodiment, to eliminate random noise in dynamic images, the Non-Local Means (NLM) filtering algorithm is used. This algorithm analyzes the similarity of adjacent pixels to remove random noise while preserving key details. In addition, for optical artifacts in dynamic images (such as reflection points), a spectral domain anti-reflection technique is used. By analyzing the spectral frequency distribution of the image, abnormal high-frequency reflection signals are located and removed, thereby optimizing the visual quality of the image.

[0090] Illumination equalization

[0091] In this embodiment, to solve the local brightness difference caused by uneven illumination in dynamic images, an adaptive illumination equalization algorithm is used. The specific method is as follows:

[0092] By calculating the local histogram of the image, the brightness range of pixel distribution is obtained;

[0093] Apply histogram equalization technology to optimize the contrast of local regions;

[0094] Introduce Gaussian smoothing operation to prevent over-sharpening phenomenon in the image after illumination adjustment.

[0095] This technique uniformly processes the overall brightness of the image while preserving the texture and morphological details of the tooth surface.

[0096] Geometric correction

[0097] In this embodiment, to handle geometric distortion caused by imaging angle or device jitter, the SIFT (Scale-Invariant Feature Transform) algorithm and the RANSAC (Random Sample Consensus) algorithm are combined:

[0098] Use the SIFT algorithm to extract key feature points in the image;

[0099] Robustly optimize the matching results of feature points by the RANSAC algorithm to remove incorrectly matched feature points;

[0100] According to the matching results, generate a geometric correction model to perform overall geometric adjustment on the image to maintain spatial consistency.

[0101] Geometric correction preserves the accurate correspondence between each frame of image in the dynamic image and the actual tooth structure.

[0102] Data augmentation

[0103] In this embodiment, to improve the learning ability of the AI model for diversified occlusion features, the following data augmentation techniques are used:

[0104] Random Rotation: Randomly rotate the dynamic images in the vertical direction by a certain angle (±15°) to simulate possible head micro-movements during patient occlusion;

[0105] Mirror Flip: Perform horizontal mirror processing on the image data to expand sample diversity;

[0106] Random Cropping and Scaling: Randomly select a local area of the image for cropping and adjust the scaling ratio to preserve key occlusion areas;

[0107] Generative Adversarial Network (GAN): Based on the original image data collected, generate simulated images under multiple angles and multiple lighting environments through GAN technology to further expand the data set.

[0108] Processing of three-dimensional structure data

[0109] In this embodiment, for the three-dimensional point cloud data output by the laser scanning device, the following optimizations are performed:

[0110] Point cloud denoising: Remove outliers based on neighborhood density algorithms to optimize the integrity of three-dimensional data;

[0111] Data compression: Use Octree Encoding technology to compress point cloud data into voxel representation, reducing transmission bandwidth and storage requirements.

[0112] Processing of force distribution data

[0113] In this embodiment, the force distribution data collected by the pressure sensor array is processed as follows:

[0114] Data normalization: Map the force distribution value to the 0-1 range according to the maximum and minimum force value range, facilitating subsequent AI model processing;

[0115] Time series smoothing: Use weighted moving average algorithm to perform time series smoothing on the force distribution data to remove transient noise in the mechanical sampling.

[0116] Data output

[0117] In this embodiment, the output data of the data preprocessing and enhancement module includes:

[0118] Dynamic image data after denoising, light equalization and geometric correction processing;

[0119] Optimized three-dimensional point cloud data output in compressed format;

[0120] Normalized force distribution data output in matrix form.

[0121] All data is timestamped before output to ensure synchronization of multi-source data with occlusion time.

[0122] In this embodiment, the data preprocessing and enhancement module receives the raw data output by the multi-source intraoral image acquisition module, and after preprocessing, it transmits the processed image data, three-dimensional structure data, and force distribution data to the AI-based occlusion detection module. This module not only ensures the synchronization of multi-source data through timestamps, but also expands the training samples of the AI model through data enhancement techniques, providing high-quality input for subsequent occlusion point detection and mechanical analysis.

[0123] AI-based occlusion detection module

[0124] For the AI-based occlusion detection module, its main task is to receive preprocessed multi-source data (dynamic images, three-dimensional structure data, force distribution data), detect occlusion points and contact points through deep learning algorithms, analyze dynamic changes during occlusion, and identify abnormal situations. This module is the core analysis module of the system, relying on artificial intelligence technology to achieve high-precision, real-time occlusion feature extraction, providing basic data support for the subsequent feedback and fusion module.

[0125] In this embodiment, the AI-based occlusion detection module uses deep learning technology to extract features, detect, and dynamically model multi-source data. Specific functions include:

[0126] Multi-scale extraction of dynamic image features;

[0127] Positioning and classification of occlusion points and contact points;

[0128] Time series modeling during occlusion to extract dynamic change features;

[0129] In the dynamic change modeling part, the main goal is to process time series data to capture long-term and short-term dependencies of occlusion points and dynamic changes. To achieve this goal, LSTM (Long Short-Term Memory Network) is used, which is a deep learning model widely used in time series modeling and sequence data prediction. Its core advantage is to capture long-term dependency information in time series and overcome the gradient vanishing problem of traditional neural networks in handling long-term time dependency problems.

[0130] LSTM network model architecture

[0131] LSTM is a variant of Recurrent Neural Network (RNN) that maintains a memory of past information while processing current input data at each time step. It consists of three important gating mechanisms: Forget Gate, Input Gate, and Output Gate, which control the forgetting, storing, and outputting of information, respectively.

[0132] Forget Gate: controls how much historical information to forget from the memory cell.

[0133] Input Gate: decides how much new information to store in the memory cell.

[0134] Output Gate: decides which information in the memory cell to base the current output on.

[0135] This mechanism allows LSTM to effectively handle long-term dependencies in sequences, especially when dealing with complex, non-linear dynamic changes, such as the time-varying bite point.

[0136] Modeling Process

[0137] Data Preparation:

[0138] The input data is in time series format, typically including multiple image frames and corresponding mechanical data. Each image frame contains information about the bite point's position, angle, and mechanical stress.

[0139] The data is first cleaned and standardized by a data preprocessing module (such as noise filtering, illumination balancing, etc.) to ensure the quality of input data.

[0140] Feature Extraction:

[0141] Features are extracted from image data using techniques such as Convolutional Neural Networks (CNN), such as the position, shape, and motion trajectory of the bite point in the image.

[0142] Mechanical data (such as applied pressure, strain, etc.) is converted into time series form by a preprocessing module and input into the LSTM network along with image features.

[0143] LSTM Network Modeling:

[0144] The time series data (including image features and mechanical data) is input into the LSTM model. The LSTM network continuously updates the memory cell over multiple time steps and outputs the predicted value (such as position, mechanical stress, etc.) of the current bite point state.

[0145] The modeling effect can be further enhanced by using Bidirectional LSTM (Bidirectional LSTM), which allows the network to consider both past and future time dependencies.

[0146] Prediction and Modeling:

[0147] The output of the LSTM network will be used to predict the state changes of the bite points at future time steps. This is particularly important for dynamic modeling, as bite points can change continuously over time (such as changes in bite force or adjustments in oral structure).

[0148] Through the output of the LSTM, we can predict the dynamic evolution of the bite points, including predicting their exact positions and trends at a future time.

[0149] Dynamic Change Analysis:

[0150] Through the prediction results of the LSTM model, we can model and analyze the dynamic changes of the bite points. For example, we can analyze the amplitude of changes between different time steps, trends, and whether there are certain periodic changes (such as changes in bite rhythm).

[0151] Subsequent Processing:

[0152] For complex dynamic data, we can combine temporal convolution networks (TCN) or other time series analysis methods to further optimize model performance, especially when dealing with high-frequency changes and complex nonlinear changes.

[0153] Optimization algorithms are used to train and adjust the model, improving detection accuracy and generalization ability.

[0154] By passing the detection results to subsequent modules, this module achieves seamless connection from data analysis to bite state detection.

[0155] 1. Model Design and Architecture

[0156] In this embodiment, the AI-based bite detection module uses the YOLOv8 (You Only Look Once Version 8) framework, combined with various optimization techniques, with the following specific design:

[0157] Model Base Framework:

[0158] YOLOv8 is a single-stage target detection framework that uses deep convolutional neural networks for feature extraction, detection, and classification. This embodiment selects YOLOv8 as the base framework, mainly considering its performance advantages in small target detection.

[0159] Feature Pyramid Network (FPN):

[0160] In the feature extraction part of YOLOv8, the feature pyramid network (FPN) is integrated, which is used to extract multi-scale features such as bite points and contact points in dynamic images, ensuring that the detection targets can be recognized at different resolutions.

[0161] Attention mechanism embedding:

[0162] In the backbone of YOLOv8, CBAM (Convolutional Block Attention Module) is embedded to enhance the attention ability to small targets (such as contact points) and key features. CBAM guides the model to preferentially learn important feature regions through channel attention and spatial attention modules.

[0163] 2. Feature extraction and bite point detection

[0164] In this embodiment, the feature extraction and detection process of the bite detection model is as follows:

[0165] Feature extraction:

[0166] The input is pre-processed dynamic image data, and the model extracts local and global features in the image layer by layer through a deep convolutional neural network, combining FPN to extract multi-scale features.

[0167] Target positioning and classification:

[0168] Using the detection head of YOLOv8, the extracted features are positioned for bite points and contact points, and the target bounding box and class confidence are output. The classification task divides the target into three categories: bite point, non-bite point, and abnormal contact point.

[0169] Loss function optimization:

[0170] Using focal loss to handle the class imbalance problem, improve the detection performance of rare classes (such as abnormal contact points);

[0171] Using IoU loss to optimize the positioning of the target bounding box, reducing errors.

[0172] 3. Dynamic modeling

[0173] In this embodiment, in order to capture the time series features in the bite process, the following dynamic modeling techniques are used:

[0174] Bi-directional LSTM (Bi-LSTM):

[0175] Temporal modeling of image frame sequences, capturing the before-and-after relationship of dynamic changes in bite points through bidirectional information flow, generating bite feature descriptions in the time dimension.

[0176] Temporal convolutional network (Temporal CNN):

[0177] A one-dimensional time convolution network is used to model the bite force distribution data, extract the dynamic change rule of mechanical characteristics, and realize multi-modal feature fusion combined with image features.

[0178] 4. Model training and optimization

[0179] In this embodiment, the following training and optimization strategies are used to improve the model detection performance:

[0180] Transfer learning:

[0181] The YOLOv8 model is pre-trained using general object detection datasets such as COCO, and then fine-tuned on the oral image dataset to adapt to the bite point detection task in the oral scene.

[0182] Hyperparameter optimization:

[0183] The Bayesian optimization algorithm is used to automatically adjust the hyperparameters of the model, including learning rate, batch size (BatchSize), and regularization strength, to improve the training effect of the model.

[0184] Data augmentation:

[0185] Data augmentation techniques such as rotation, cropping, and mirror flipping are used to expand the training data, and generative adversarial networks (GAN) are used to generate diverse simulated image data to improve the generalization ability of the model.

[0186] 5. Output results

[0187] In this embodiment, the output results of the AI-based bite detection module include:

[0188] Bite point distribution map: shows the detected bite point position and classification results;

[0189] Contact point bounding box: marks the position information of the contact point in coordinate form;

[0190] Force distribution dynamic diagram: combines dynamic force distribution characteristics to generate time series analysis results;

[0191] Abnormal contact point marking: outputs possible abnormal bite characteristics.

[0192] All output results are transmitted to the real-time dynamic feedback module in a standardized format.

[0193] In this embodiment, the AI-based bite detection module receives the standardized data output by the data preprocessing and enhancement module, completes the feature extraction and detection task, and transmits the analysis results to the real-time dynamic feedback and anomaly analysis module. Time stamp information is used to realize time alignment of data, ensuring the collaborative work between multiple sources of data.

[0194] Real-time dynamic feedback and anomaly analysis module

[0195] For the real-time dynamic feedback and anomaly analysis module, its main role is to receive the occlusion point distribution and force distribution data output by the AI-based occlusion detection module, and provide immediate occlusion diagnosis support for doctors through data visualization, real-time feedback, and anomaly feature analysis. Through efficient data processing and intuitive visualization means, this module realizes dynamic tracking of occlusion state and accurate identification of abnormal patterns, providing important basis for subsequent correction treatment or further diagnosis.

[0196] In this embodiment, the functions of the real-time dynamic feedback and anomaly analysis module include:

[0197] Real-time generation of occlusion point distribution graph and force distribution heat map to visually display the dynamic changes in the occlusion process;

[0198] Anomaly pattern recognition of occlusion data, including detection of eccentric bite, overload, and contact abnormality;

[0199] Provide an interactive three-dimensional visualization interface to dynamically replay the patient's occlusion process and compare historical data.

[0200] Through real-time feedback and anomaly analysis, this module realizes the conversion of detection data to diagnostic results, ensuring the efficiency and practicality of the system in clinical use.

[0201] 1. Real-time feedback generation

[0202] In this embodiment, the real-time dynamic feedback module generates the following feedback content by receiving the output data of the occlusion detection module:

[0203] Occlusion point distribution graph:

[0204] According to the detected occlusion point coordinates and category information, generate an occlusion point distribution graph to visually display the position of each occlusion point and its type (such as normal occlusion point, abnormal contact point, etc.). Use different colors or symbols to mark each category of occlusion point to facilitate the doctor to quickly identify the key area.

[0205] Force distribution heat map:

[0206] Map the force distribution data collected by the pressure sensor onto the three-dimensional tooth model to generate a force distribution heat map. The heat map uses a color gradient (such as from blue to red) to represent the force size, clearly reflecting the force distribution state and trend in the occlusion process.

[0207] Dynamic change curve:

[0208] Draw the changes of occlusion force and contact points over time as a dynamic curve graph to facilitate the doctor to understand the time sequence characteristics in the occlusion process.

[0209] Data processing uses GPU acceleration technology to ensure that feedback content can be calculated and displayed within 50ms, achieving real-time dynamic feedback.

[0210] Abnormal pattern recognition

[0211] In this embodiment, to detect abnormal patterns during occlusion, the module uses a combination of rule analysis and model analysis:

[0212] Bias occlusion detection:

[0213] By analyzing the spatial distribution of occlusion points, the number of occlusion points on the left and right sides and the balance of force distribution are calculated. If the unilateral occlusion point or force distribution proportion exceeds the preset threshold, it is marked as bias occlusion.

[0214] Overload detection:

[0215] According to the pressure values of each contact point in the force distribution heat map, the area where the force distribution is abnormally concentrated is identified. If the pressure value of a certain contact point is significantly higher than that of the surrounding area (more than a certain proportion, for example, more than 2 times), it is marked as an overload area.

[0216] Contact anomaly detection:

[0217] Combined with the classification results output by the AI detection module, abnormal contact points (such as contact points that should not exist or points that do not meet the occlusion specifications) are identified, and doctors are prompted to pay attention through visual markers.

[0218] Time series analysis:

[0219] Using dynamic change curves, the occurrence time and characteristics of abnormal force fluctuations during occlusion are detected, such as rapid changes in force values or abnormal increases and decreases in the number of contact points within a short period of time.

[0220] 3. Data visualization

[0221] In this embodiment, the real-time dynamic feedback module provides an interactive three-dimensional visualization interface to display the real-time changes of the patient's occlusion and historical data comparison:

[0222] Three-dimensional model display:

[0223] Based on the three-dimensional model of teeth generated by laser scanning equipment, the occlusion point distribution and force distribution heat map are superimposed on the three-dimensional model, supporting model rotation, scaling, and cross-section viewing functions.

[0224] Dynamic playback:

[0225] By loading image frame sequences, combining force distribution data and timestamp information, the patient's occlusion process is dynamically played back, showing the formation, disappearance of occlusion points, and dynamic changes in force distribution.

[0226] Historical data comparison:

[0227] Compare the current occlusion data with the patient's historical data to generate a comparison report, helping doctors assess treatment effectiveness or disease progression.

[0228] Data output

[0229] In this embodiment, the output content of the real-time dynamic feedback module includes:

[0230] occlusion point distribution graph and force distribution heat map;

[0231] abnormal pattern analysis results, including the marking of eccentric bite, overload and contact abnormality;

[0232] dynamic change curve graph;

[0233] Three-dimensional model superposition data for visual display.

[0234] All output data is transmitted to the multi-modal data fusion module in a standard format, supporting further comprehensive diagnosis.

[0235] In this embodiment, the real-time dynamic feedback and abnormal analysis module receives the output data of the AI-based occlusion detection module through the interface and generates feedback content. After combining the feedback results with the data of the multi-modal data fusion module, it supports more complex comprehensive diagnosis tasks. In addition, the visualization results of this module can be directly presented to the doctor, realizing real-time interaction and diagnosis support.

[0236] Multi-modal data fusion module

[0237] For the multi-modal data fusion module, its main task is to unify and deeply analyze various heterogeneous data such as dynamic image data, occlusion force distribution data and patient medical record data. By introducing coordinate channel attention mechanism and multi-modal joint modeling technology, spatial and temporal alignment and feature interaction of data are realized, and comprehensive diagnosis results are generated. This module is an important link for the system to derive from single data analysis to overall diagnosis, closely connected with the previous modules and subsequent diagnosis suggestions.

[0238] In this embodiment, the functions of the multi-modal data fusion module include:

[0239] Data alignment and standardization: spatial and temporal alignment of dynamic image data, three-dimensional structure data and force distribution data;

[0240] Multi-modal joint modeling: combine Transformer architecture and multi-source feature interaction modeling method to extract cross-modal correlation;

[0241] Comprehensive diagnosis generation: analyze the occlusion characteristics of patients through fused multi-modal data and generate comprehensive diagnosis suggestions.

[0242] This module provides global diagnostic support for doctors through deep fusion of multi-modal data, enhancing the intelligent level of the system.

[0243] 1. Data alignment and normalization

[0244] In this embodiment, to realize the integration of multi-source data, first, the input dynamic image, bite force distribution and medical record data are aligned and normalized:

[0245] Time alignment:

[0246] Using the time markers provided by the timestamp synchronization unit, align the dynamic image frame sequence with the bite force distribution data.

[0247] The precision of time alignment reaches millisecond level, ensuring that the image and force distribution data at the same time point are consistent.

[0248] Spatial alignment:

[0249] Map the bite point position in the dynamic image data to the geometric coordinates of the three-dimensional structure data to generate one-to-one bite point position information.

[0250] The force distribution data is superimposed on the three-dimensional tooth model through coordinate projection technology to construct the spatial distribution mapping of bite force.

[0251] Medical record data integration:

[0252] Associate the patient's basic information, medical history and treatment records with the real-time collected data.

[0253] Data is formatted through a pre-defined medical record structured template for subsequent model analysis.

[0254] 2. Multi-modal joint modeling

[0255] In this embodiment, the multi-modal data fusion module uses feature interaction and joint modeling technology to deeply integrate multi-source data:

[0256] Transformer framework:

[0257] Use the Transformer architecture to jointly model dynamic image features, force distribution features and medical record features.

[0258] Each modality data first generates its own weighted features through self-attention mechanism, and then realizes cross-modal feature interaction through multi-head attention mechanism.

[0259] Encoding and weighting process

[0260] To better explain how to encode and weight the data sources, you can choose one of the following two processing schemes, or combine them:

[0261] Scheme One: Encode and weight each data source separately

[0262] Image Data Encoding: First, use a Convolutional Neural Network (CNN) to encode the image data and extract spatial feature information. Each frame of image is processed into a feature vector, representing the features of each region in the image.

[0263] Mechanical Data Encoding: At the same time, use a fully connected network or convolutional layer to encode the mechanical data (such as force, pressure, strain, etc.) and convert it into a corresponding feature vector, representing the mechanical features at different time steps.

[0264] Coordinate Channel Attention Mechanism: For each modality's feature vector, apply a coordinate channel attention mechanism to calculate the respective weighting coefficients. This coefficient will be based on the spatial location and importance of the features to weight different channels.

[0265] Weighted Processing: Weight the encoded feature vectors to highlight important features and suppress irrelevant information.

[0266] Scheme Two: First, perform data fusion, then perform weighted processing

[0267] Feature Fusion: First, fuse the image data and mechanical data into a joint feature vector. You can use concatenation or additive fusion to combine the two, forming a unified feature representation.

[0268] Coordinate Channel Attention Mechanism Weighting: Apply the coordinate channel attention mechanism to the fused feature vector, assigning weights to each feature through the attention mechanism. This mechanism will focus on the importance of spatial features and channel features in the data, dynamically adjusting the weights.

[0269] Weighted Feature Processing: Pass the weighted features to the subsequent network for decision-making and prediction.

[0270] Coordinate Channel Attention Mechanism:

[0271] Embed the coordinate channel attention module in the dynamic image feature extraction, guiding the model to accurately model the spatial location features of the occlusion points;

[0272] Combine the contact point information in the force distribution data to strengthen the correlation of cross-modal features;

[0273] The coordinate channel attention mechanism aims to handle the fusion of data from different modalities (image data, mechanical data, etc.), ensuring that different types of data can be reasonably weighted to optimize the performance of the model.

[0274] Coordinate channel refers to the mutual relationship between the spatial position (coordinates) of data and the feature dimension (channel). This mechanism further improves the attention of the network to different modalities by focusing on the specific position of each data source and its corresponding feature channel.

[0275] This mechanism can automatically calculate and selectively weight the features of different data sources during feature learning, helping the model highlight the most informative parts of different modalities and avoid interference from irrelevant features.

[0276] Feature fusion strategy:

[0277] Dynamic images and force distribution features are based on time dimension, and temporal convolutional network (Temporal CNN) is used to realize the time series fusion of multi-modal features.

[0278] Medical record data is embedded through a fully connected network and combined with other modal features to form a multi-modal feature vector of the patient's occlusion characteristics.

[0279] 3. Comprehensive diagnosis generation

[0280] In this embodiment, the fused multi-modal features are used to generate a comprehensive diagnosis result, which specifically includes:

[0281] Occlusion abnormality comprehensive score:

[0282] According to the joint analysis of dynamic images, force distribution and medical record data, a comprehensive score reflecting the occlusion health status of the patient is generated, with a score range of 0-100, and the lower the score, the higher the risk of occlusion abnormalities.

[0283] Abnormal pattern classification:

[0284] Based on the feature vector output by the model, it is determined whether the patient has occlusion problems such as crossbite, overload or contact abnormalities;

[0285] Generate specific abnormal classification labels and attach corresponding diagnosis reasons.

[0286] Diagnosis suggestion generation:

[0287] Combined with the past records of medical record data, preliminary diagnosis suggestions are generated, such as suggestions for adjusting occlusion, treatment correction or further imaging examination.

[0288] 4. Data output

[0289] In this embodiment, the output content of the multi-modal data fusion module includes:

[0290] Comprehensive diagnosis report, including occlusion abnormality score, classification label and diagnosis suggestion;

[0291] Three-dimensional occlusion model with dynamic image data and force distribution mapping;

[0292] Cross-modal feature vectors for subsequent depth analysis or storage.

[0293] In this embodiment, the multi-modal data fusion module receives analysis data and medical record data from the real-time dynamic feedback module, and generates comprehensive diagnostic results through feature fusion and joint modeling. The output data is transmitted to the subsequent diagnosis suggestion module to provide support for doctor's decision-making. At the same time, the module is connected with the data storage unit of the system to ensure that all processing results can be recorded and archived.

[0294] Lightweight edge computing and system deployment module

[0295] For the lightweight edge computing and system deployment module, its main task is to ensure that the system can run stably on real-time, efficient and resource-limited embedded devices through model optimization, edge computing deployment and edge-cloud collaborative mechanism. This module optimizes through model compression, hardware acceleration and task allocation strategy, realizes efficient collaborative computing between embedded devices and cloud, provides technical support for real-time occlusion detection of the system, and guarantees multi-platform adaptability.

[0296] In this embodiment, the functions of the lightweight edge computing and system deployment module include:

[0297] Model compression and optimization: reduce model calculation complexity and storage requirements through pruning, quantization and knowledge distillation technology;

[0298] Edge computing deployment: realize real-time inference of deep learning model on embedded devices;

[0299] Edge-cloud collaborative mechanism: divide lightweight tasks and heavy computing tasks to realize efficient collaboration between edge devices and cloud;

[0300] System multi-platform adaptation: support flexible deployment and performance optimization in different hardware environments.

[0301] Through module design, the system can balance real-time performance and computing performance, providing efficient and convenient technical support for clinical operations.

[0302] 1. Model compression and optimization

[0303] In this embodiment, in order to run deep learning model on resource-limited embedded devices, the following optimization methods are adopted:

[0304] Model pruning:

[0305] Prune the convolutional layer weights of deep learning model based on importance score, remove the neuron connections that contribute less to the inference result;

[0306] The pruning ratio is adjusted through layer-by-layer optimization, which significantly reduces the computational load and memory occupation while maintaining the model accuracy.

[0307] Model quantization:

[0308] Using INT8 quantization technology, the model weights and activation values are compressed from 32-bit floating-point numbers (FP32) to 8-bit integers (INT8), significantly reducing the model storage requirements and computational complexity;

[0309] Introducing symmetric quantization strategy, through post-quantization training (Post-Training Quantization, PTQ) to restore the model accuracy.

[0310] Knowledge distillation:

[0311] Using a high-precision teacher model (Teacher Model) to guide the training process of a lightweight student model (Student Model), and preserving the knowledge structure of the teacher model through a distillation loss function;

[0312] Combined with transfer learning technology, to speed up the convergence speed of the student model on small-scale data sets.

[0313] 2. Edge computing deployment

[0314] In this embodiment, the optimized lightweight model is deployed on embedded devices (such as portable oral endoscopes, mobile tablet devices), and the specific implementation includes:

[0315] Hardware adaptation:

[0316] Embedded devices use embedded GPUs (such as NVIDIA Jetson series) or NPUs (such as ARM Mali or Google Edge TPU) to accelerate the inference process of the model;

[0317] The model running environment designed for low-power devices, such as TensorRT or ONNX Runtime.

[0318] Real-time inference optimization:

[0319] Using batch inference (Batch Inference) mechanism to process multiple frames of data in parallel, improving throughput;

[0320] Introducing dynamic memory allocation strategy to reduce resource waste and improve inference efficiency.

[0321] Local cache and response:

[0322] Storing the key inference results of the model to the local cache, supporting doctors to quickly view the intermediate results of the bite detection;

[0323] Data interaction with other hardware devices (such as hospital terminals) through low-latency interfaces (such as USB or Bluetooth).

[0324] 3. Edge-Cloud Collaboration Mechanism

[0325] In this embodiment, to achieve efficient allocation of computing tasks and resource utilization, an edge-cloud collaborative computing mechanism is designed:

[0326] Task partitioning strategy:

[0327] High real-time detection tasks (such as bite point positioning and force distribution analysis) are assigned to edge devices for execution;

[0328] High-complexity tasks (such as dynamic modeling and multi-modal fusion) are assigned to cloud processing.

[0329] Edge-Cloud Data Interaction:

[0330] Efficient data transmission between edge devices and the cloud is achieved through 5G networks or low-latency communication protocols (such as MQTT);

[0331] Data interaction uses an asynchronous communication mode, where edge devices continue to execute subsequent tasks while receiving cloud processing results, reducing latency.

[0332] Cloud Task Processing:

[0333] The cloud server performs heavy computing tasks based on high-performance GPUs or TPUs;

[0334] Provides data storage and backup services, long-term saving of patients' bite detection records and diagnosis results.

[0335] 4. System Multi-Platform Adaptation

[0336] In this embodiment, to achieve efficient operation of the system on multiple hardware platforms, the following adaptation schemes are designed:

[0337] Cross-Platform Model Format Conversion:

[0338] Use ONNX format to export models from development frameworks (such as PyTorch or TensorFlow) and adapt them to different inference engines (such as TensorRT, CoreML, TFLite);

[0339] Adjust quantization parameters and optimization strategies according to the characteristics of different hardware platforms.

[0340] Device-side Dynamic Configuration:

[0341] The system dynamically adjusts the running parameters of the model (such as resolution, batch size, etc.) according to the computing power and memory size of the target hardware device;

[0342] Support fast switching and uniform deployment between devices through configuration files.

[0343] 5. Data output and interaction

[0344] In this embodiment, the output of the lightweight edge computing and system deployment module includes:

[0345] Real-time inference results of the edge device (such as bite point distribution, force distribution heat map);

[0346] Comprehensive diagnostic report and multi-modal analysis result generated by the cloud;

[0347] Data interaction record, including task allocation status and processing delay information.

[0348] Through standardized interfaces, module output data is directly transmitted to the real-time feedback module and multi-modal fusion module, ensuring closed-loop cooperation of system functions.

[0349] In this embodiment, the lightweight edge computing and system deployment module receives data from the preprocessing module and AI detection module, and transmits the results to the real-time feedback module after completing the inference task. At the same time, collaborative interaction with the cloud module ensures the processing capacity of complex tasks, forming an efficient integration of edge and cloud computing resources.

[0350] Working principle: First, the multi-source intraoral image acquisition module acquires core data such as dynamic images, three-dimensional structures, and bite force distribution, and achieves time sequence synchronization of multi-source data through time stamping; the collected data is processed in the preprocessing and enhancement module to optimize data quality through noise reduction, light balance, geometric correction, and data enhancement. The processed data enters the AI-based bite detection module, extracts bite point and contact point features through a deep learning model, and realizes abnormal detection of the bite process through dynamic modeling. Subsequently, the real-time dynamic feedback and abnormal analysis module generates visualization content such as bite point distribution map and force distribution heat map based on the detection results, and identifies problems such as eccentric bite and overload through abnormal pattern analysis, providing intuitive diagnostic support for doctors. At the same time, the multi-modal data fusion module deeply fuses dynamic images, force distribution data, and medical record information to generate a comprehensive diagnostic report and provide treatment recommendations. The lightweight edge computing and system deployment module of the system supports real-time operation on embedded devices and cloud processing of complex tasks through model optimization and edge-cloud collaboration mechanism, ensuring efficient and stable operation of the system in resource-limited environments. The orderliness of data flow between modules, the interactivity of functions, and the edge-cloud collaboration design together form a complete technical chain of the invention in the field of real-time bite detection, which can effectively meet the precise diagnostic needs of oral medicine.

[0351] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. An AI-enhanced real-time occlusal detection system based on intraoral imaging, characterized in that, include: The multi-source intraoral imaging module is used to acquire dynamic oral imaging data through a high-definition oral endoscope, and to record the three-dimensional structure of the tooth surface in combination with a laser scanning device. At the same time, it collects the force distribution at the occlusal contact point through a pressure sensor array. The multi-source intraoral image acquisition module includes a timestamp synchronization unit, which uses a unified timestamp to synchronize the oral image frame sequence and occlusal force distribution data from multiple sources. The data preprocessing and enhancement module is connected to the multi-source intra-port image acquisition module. It is used to perform noise filtering, illumination equalization and geometric distortion calibration on the acquired image data, and generate multi-angle samples through random rotation, cropping and mirror flipping data enhancement methods. The AI-based bite detection module is connected to the data preprocessing and enhancement module and is used to perform feature extraction, bite point detection and dynamic change modeling on the processed data through a deep learning model. The real-time dynamic feedback and anomaly analysis module is connected to the AI-based occlusion detection module to generate occlusion point distribution maps and force distribution heatmaps, and to analyze occlusion data. The multimodal data fusion module is connected to the AI-based occlusal detection module and the real-time dynamic feedback and anomaly analysis module, respectively, and is used to fuse image data, occlusal force data and patient medical record information; The lightweight edge computing and system deployment module connects to the above modules to optimize the computational efficiency of AI models and support collaborative computing between edge devices and the cloud.

2. The AI-enhanced real-time occlusal detection system based on intraoral imaging according to claim 1, characterized in that, The multi-source intraoral image acquisition module includes: High-definition oral endoscope, used to acquire dynamic image data with a resolution greater than 2K and a frame rate of ≥60fps; Laser scanning equipment is used to capture the three-dimensional structure and surface texture of teeth; A miniature pressure sensor array is used to collect the force distribution at the biting contact point; The timestamp synchronization unit is used to add a unified time stamp to dynamic images and force distribution data.

3. The AI-enhanced real-time occlusal detection system based on intraoral imaging according to claim 1, characterized in that, The data preprocessing and enhancement module includes: The noise reduction processing unit is connected to the multi-source intra-port image acquisition module and is used to reduce noise in image data based on nonlocal mean filtering technology and spectral domain dereflection technology. An optical correction unit is used to adjust the brightness and contrast of an image based on an adaptive illumination equalization algorithm; The geometric calibration unit is used to correct geometric distortion of the image using a feature point matching algorithm. The data augmentation unit is used to randomly rotate, crop, and mirror the corrected image data to generate diverse image samples.

4. The AI-enhanced real-time occlusal detection system based on intraoral imaging according to claim 1, characterized in that, The AI-based bite detection module includes: A deep learning model, connected to the data preprocessing and enhancement module, is used to extract multi-scale features through a feature pyramid network. An attention mechanism module is embedded, which incorporates a convolutional block attention module to enhance the detection capability for small targets; The dynamic modeling unit is used to perform temporal modeling of image frame sequences based on bidirectional LSTM and temporal convolutional networks. The classification and localization unit is used to optimize the category distribution of bite point detection based on focal loss and optimize the localization of the target bounding box through IoU loss.

5. The AI-enhanced real-time occlusal detection system based on intraoral imaging according to claim 1, characterized in that, The AI-based bite detection module also includes: The transfer learning module is used for pre-training on a general object detection dataset and fine-tuning it to adapt to oral imaging data. The adaptive hyperparameter optimization module is used to adjust the learning rate, batch size, and regularization parameters using a Bayesian optimization algorithm.

6. The AI-enhanced real-time occlusal detection system based on intraoral imaging according to claim 1, characterized in that, The real-time dynamic feedback and anomaly analysis module includes: The real-time feedback unit is connected to the AI-based bite detection module and is used to generate bite point distribution maps and force distribution heat maps. The anomaly analysis unit is used to extract occlusal abnormality features from the detection results and generate a diagnostic report; The visualization interface unit is used to dynamically display three-dimensional occlusal data and compare and analyze historical data.

7. The AI-enhanced real-time occlusal detection system based on intraoral imaging according to claim 1, characterized in that, The multimodal data fusion module includes: The data alignment unit, connected to the AI-based bite detection module, is used to uniformly encode image data and mechanical data through a coordinate channel attention mechanism; The joint modeling unit is used for interactive modeling of image data, mechanical features and medical record data based on the Transformer framework; The fusion analysis unit is used to generate comprehensive occlusal diagnostic results and orthodontic recommendations.

8. The AI-enhanced real-time occlusal detection system based on intraoral imaging according to claim 1, characterized in that, The lightweight edge computing and system deployment module includes: The model compression unit, connected to the AI-based bite detection module, is used to compress the model using pruning and quantization techniques. The knowledge distillation module is used to optimize the lightweight student model using the teacher model; The edge-cloud collaboration unit is used to perform real-time inference on edge devices and migrate complex computing tasks to the cloud.

9. The AI-enhanced real-time occlusal detection system based on intraoral imaging according to claim 1, characterized in that, The system's detection function achieves real-time detection of dynamic occlusal points and contact points by connecting a multi-source intraoral image acquisition module, a data preprocessing and enhancement module, and an AI-based occlusal detection module.

10. The AI-enhanced real-time occlusal detection system based on intraoral imaging according to claim 1, characterized in that, The system connects to the hospital information system via an API interface to enable the sharing and synchronization of occlusal test data and supports cross-device diagnostic record updates.

Images