A child oral cavity detection method and system based on image recognition

Abstract

The present application relates to a kind of children oral cavity detection method and system based on image recognition, belong to medical detection field.Adopt the strategy network of imitation learning algorithm training strategy, oral cavity detection process is modeled as sequential decision problem;Through the detection trajectory data of expert doctor is collected, combined with behavior cloning, generative adversarial imitation learning and DAGGER iteration optimization, make strategy network learn optimal detection path and action logic.System with current oral cavity image feature map, observed area mask and image quality score as state, dynamic output next observation position and scaling factor, control acquisition equipment adaptive acquisition local high-definition image;Abnormal recognition module carries out pixel-level segmentation to local image, outputs abnormal area such as dental caries, gingivitis;Diagnosis suggestion generation module integrates detection path and identification result, outputs structured natural language text.Realize the closed-loop adaptive control of children oral cavity detection, improve detection efficiency and abnormal detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of medical testing technology, specifically to a method and system for pediatric oral cavity testing based on image recognition. Background Technology

[0002] Oral health is an important component of overall health in children, and dental caries (cavities) and gingivitis are the most common oral diseases in this age group. According to the World Health Organization, approximately 60% to 90% of school-aged children worldwide have dental caries, and early symptoms are often subtle. If not detected and treated promptly, it can lead to pain, chewing dysfunction, and even affect the development of permanent teeth. Therefore, regular oral health checkups for children are of significant clinical importance.

[0003] Currently, pediatric oral examinations primarily rely on doctors' visual examination and experience-based judgment. This method suffers from drawbacks such as high subjectivity, poor consistency, long processing time, and high requirements for children's cooperation. In recent years, with the development of deep learning technology, some researchers have attempted to apply image recognition to oral examinations. For example, Chinese patent application CN119494820A discloses a pediatric oral examination method and system based on image recognition, which uses a convolutional neural network to automatically identify and label abnormal areas such as cavities and gingivitis in oral images. However, this approach is essentially static supervised learning, only classifying or segmenting single images, lacking the ability to dynamically plan the examination process, and unable to adaptively adjust the observation strategy according to individual differences in children and real-time image quality. More importantly, this approach fails to learn from the examination path selection logic and clinical decision-making experience of expert doctors in actual diagnosis and treatment, i.e., it does not introduce imitation learning or strategy optimization mechanisms, resulting in fixed examination paths and templated diagnostic suggestions, making it difficult to cope with complex scenarios such as uncooperative children and diverse oral structures. Furthermore, existing technologies do not model oral examination as a sequential decision problem, lack effective utilization of historical observation information (such as the mask of observed areas), and cannot achieve closed-loop adaptive control.

[0004] Therefore, how to learn optimal detection strategies from expert demonstrations using methods such as imitation learning, transforming the pediatric oral examination process into a Markov decision process, and achieving a leap from static image recognition to dynamic path planning, while generating personalized and interpretable diagnostic suggestions, is a technical problem that urgently needs to be solved in this field. This invention is proposed against this background. Summary of the Invention

[0005] To address the problems of the prior art, this invention provides a method and system for detecting children's oral cavity based on image recognition.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: Firstly, a method for detecting pediatric oral cavity based on image recognition, comprising the following steps: Obtain initial image sequences of the inside of a child's oral cavity; The image sequence is preprocessed, including denoising, enhancement, cropping, and normalization; The preprocessed image features and the current detection state are input into a policy network trained using an imitation learning algorithm, and the policy network outputs the next detection action. Based on the next detection action, the image acquisition device is automatically controlled to adjust the shooting angle, focus position, or zoom level to acquire new local oral cavity images; Anomaly identification is performed on the new local oral cavity image to segment and classify abnormal regions in the oral cavity; Repeat the above strategy steps of network inference, device control, image acquisition and anomaly recognition until the detection termination condition is met; By combining all anomaly identification results and the detection paths generated by the policy network, diagnostic suggestions are generated and output.

[0007] In one specific implementation of the first aspect, the imitation learning algorithm includes behavioral cloning, specifically: The operation trajectory data of several pediatric dentists were collected. Each trajectory includes a state sequence and a corresponding expert action sequence. The status includes: the feature map of the current oral cavity image region, the mask of the observed region, and the image quality score; The observed region mask is a binary matrix of the same size as the panoramic image of the oral cavity, where 1 indicates that the pixel position has been observed and 0 indicates that it has not been observed. The image quality score uses Laplacian variance to calculate image sharpness, and the calculated value is normalized to... The range indicates that the higher the value, the clearer the image; The actions include: the center coordinates of the next observation area, the zoom level, whether to mark an anomaly, and whether to generate suggestion text; The loss function minimizes the difference between the policy network's output action and the expert's action through supervised learning.

[0008] in, The total number of time steps for a trajectory; For the first The state of the step; For the first Step-by-step actions performed by experts; For the policy network in state Down Output Action The probability of is given by , and its range is . ; It is the natural logarithm.

[0009] In one specific implementation of the first aspect, the imitation learning algorithm further employs a generative adversarial imitation learning (GAIL) optimization policy network: Constructing a discriminator network Its input is the state. and actions The output is a scalar value, representing The range of probabilities derived from expert trajectories is . ; Reinforcement learning is used to optimize the policy network to maximize the authenticity score of the discriminator output while minimizing the policy entropy; The optimization objective is:

[0010] in: This represents the policy network to be optimized. Indicate expert strategy; Representation Strategy The expected trajectory of the generated state and action; This represents the expected value of the expert's trajectory; For discriminator output; Here is the entropy regularization coefficient, which takes the value of 1. ; Let be the entropy of the strategy.

[0011] In one specific implementation of the first aspect, the training of the policy network further employs the DAGGER algorithm: The initial policy is trained using behavior cloning; Trajectories generated by the policy network are collected in a real-world environment, and human experts provide corrected optimal actions for each state in the trajectory. Add the corrected state-action pairs to the training set and retrain the policy network; Iterate until the policy network achieves an action accuracy of over 95% on the validation set.

[0012] In one specific implementation of the first aspect, the anomaly recognition employs a segmentation model based on a deep convolutional neural network, wherein the segmentation model is Mask R-CNN or U-Net, outputting the category probability of each pixel, and the categories include healthy teeth, cavities, gingivitis, and tartar; the training loss function is the sum of cross-entropy loss and Dice loss, as shown in the formula:

[0013] in: This represents the total number of pixels in the image. The total number of categories,

[0014] For pixel index, Indexed by category; For the first The pixel belongs to the first The actual label of the class is encoded using one-hot encoding; The model predicts the first The pixel belongs to the first The probability of a class.

[0015] In one specific implementation of the first aspect, the diagnostic suggestion is generated using a Transformer-based natural language generation model. The input is the detection path and anomaly identification results output by the policy network, and the output is a structured Chinese diagnostic suggestion text, including a problem description, possible causes, treatment plan, and follow-up time suggestion.

[0016] In one specific implementation of the first aspect, the detection termination condition is any of the following: The policy network outputs a "End detection" action; The observed area accounts for a proportion of the total area of ​​the panoramic oral cavity image that exceeds a preset threshold. This threshold ranges from [value missing]. ; The confidence level of the output of three consecutive actions is lower than Wherein, confidence level is the probability value corresponding to the action output by the policy network. .

[0017] Secondly, an image recognition-based pediatric oral cavity detection system is provided for performing an image recognition-based pediatric oral cavity detection method, comprising: Image acquisition module: Used to acquire image sequences of the inside of a child's mouth, including images with a resolution of at least [resolution not specified]. High-definition cameras or oral endoscopes, automatic light source adjustment devices, and acquisition and control units that receive action commands from the policy network; Image preprocessing unit: performs noise reduction, enhancement, cropping, and normalization on the image; Imitation learning module: Stores and runs the policy network trained using the imitation learning algorithm, and outputs the next detection action based on the current state; Anomaly detection module: Segmentation and classification of abnormal regions in oral cavity images based on convolutional neural networks; Diagnostic suggestion generation module: Generates natural language diagnostic suggestions based on comprehensive detection results and detection paths; User interface: used to display test results, test path animation, diagnostic suggestions, and historical records.

[0018] In one specific implementation of the second aspect, the user interface further includes a path visualization module, used for: The heatmap displays the regions and intensity of interest that the policy network has focused on, where the heat value is proportional to the probability of the policy network outputting an action at that location. .

[0019] Thirdly, a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the steps of a method.

[0020] The beneficial effects of this invention are as follows: 1. This invention models the pediatric oral cavity detection process as a Markov decision process and employs an imitation learning algorithm to learn the optimal detection strategy from the real-world trajectories of expert physicians. Through a multi-stage training framework combining behavioral cloning, generative adversarial imitation learning, and DAGGER iterative optimization, the policy network can autonomously generate detection paths comparable to those of experts, including dynamically selecting observation positions, adaptively adjusting scaling factors, and determining the timing of anomaly labeling. This mechanism achieves a leap from static image classification to sequential decision-making, enabling the system to possess strategy optimization capabilities. It effectively solves the problems of fixed detection paths and the inability to adaptively adjust according to differences in children's oral cavity structure and varying levels of cooperation in existing technologies, significantly improving the coverage integrity and detection efficiency of abnormal areas.

[0021] 2. This invention, within an imitation learning framework, utilizes observed region masks and image quality scores as state feedback to form a closed-loop adaptive control strategy. Simultaneously, action commands output by the strategy network drive the acquisition device, achieving real-time optimization of the detection process. Building upon this, a pixel-level anomaly detection module combined with a natural language-based diagnostic suggestion module transforms the detection path and recognition results into structured, interpretable clinical suggestions, enhancing the interpretability of the diagnostic process and human-machine collaboration capabilities. Compared to existing static supervised learning-based schemes, this invention is no longer limited to anomaly detection of single images but extracts generalization rules from expert experience. While ensuring the accuracy of anomaly detection, it reduces reliance on manual operation by experienced physicians, minimizing human error, and is particularly suitable for early screening and long-term health management of pediatric oral diseases. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall system flow of the present invention.

[0023] Figure 2This is a schematic diagram of the training and reasoning process of the imitation learning module of the present invention.

[0024] Figure 3 This is a schematic diagram of the strategy network structure of the present invention.

[0025] Figure 4 This is a schematic diagram of the user interface flow of the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0027] like Figures 1 to 4 This paper presents a method and system for detecting children's oral cavity based on image recognition.

[0028] I. System Overall Architecture This embodiment provides a pediatric oral cavity detection system based on image recognition, including an image acquisition module, an image preprocessing unit, an imitation learning module, an anomaly recognition module, a diagnostic suggestion generation module, and a user interface. The data flow between the modules is as follows: Figure 1 As shown.

[0029] The core innovation of the system lies in modeling the process of children's oral examination as a sequential decision problem and using an imitation learning algorithm to train a policy network. This network automatically plans the next observation position, zoom level, and other actions based on the current state of the oral image, achieving adaptive and intelligent oral examination and generating diagnostic suggestions in natural language.

[0030] II. Image Acquisition Module Step 1: Equipment Preparation A high-definition oral endoscope with a resolution of 1920×1080 pixels (model: DCI-1080P) is selected, equipped with an adjustable LED ring light source, with a brightness adjustable range of 0~1000 lux in 10 lux increments. The camera is mounted on a motorized pan-tilt head that can rotate in two dimensions, with a horizontal rotation range of ±90° and a pitch range of ±45°, driven by a stepper motor with a control accuracy of 0.5°.

[0031] Step 2: Environmental and Patient Preparation The examination room is brightly and evenly lit with a neutral gray background. Children are placed in a semi-reclining position with their heads secured to soft headrests and their chins resting on adjustable height supports. Disposable sterile oral endoscope covers are used to prevent cross-infection.

[0032] Step 3: Image Acquisition Control Initially, the system automatically aligns the endoscope directly in front of the child's mouth, at a distance of approximately 5cm, and captures the first panoramic image. Subsequently, the imitation learning module outputs the next action command based on the current image state, including: : The coordinates of the center of the next observation area, relative to the panoramic image coordinate system, in pixels; Scaling factor, range of values Step length ; : Boolean value, whether to mark the current region as abnormal; : Boolean value, indicating whether to immediately generate intermediate suggestion text.

[0033] The data acquisition and control unit will This is converted into the rotation angle and focus adjustment of the gimbal, driving the motor to move the camera and automatically focus. After each acquisition, the image is saved in JPEG format, with the file naming convention: Child ID_Date_Time_Region Number.jpg.

[0034] III. Image Preprocessing Unit The preprocessing steps are as follows: Cropping: The effective areas of the oral cavity wall, teeth, and gums are automatically cropped using a pre-trained oral region detector (based on YOLOv5, trained on 3000 labeled oral images), while irrelevant areas such as the lips and face are removed.

[0035] Grayscale conversion: For color images, the standard luminance formula is used.

[0036]

[0037] Convert to a single-channel grayscale image.

[0038] Noise reduction: Apply a Gaussian filter with a kernel size of 5×5 and a standard deviation σ=1.0. The formula is as follows:

[0039] Suppress high-frequency noise.

[0040] Contrast Enhancement: Histogram equalization is used to stretch the grayscale dynamic range to... .

[0041] Edge sharpening: Extracting horizontal gradients using the Sobel operator and vertical gradient Calculate the gradient magnitude This image is overlaid on the original image to enhance tooth boundaries and cavity edges.

[0042] Scaling: All images are scaled uniformly to 512×512 pixels using bilinear interpolation.

[0043] Normalization: Converting pixel values ​​from Linear mapping to ,formula

[0044] in .

[0045] The preprocessed image is input into the policy network as part of the state.

[0046] IV. Imitation Learning Module Expert trajectory collection We are recruiting 5 associate chief physicians with over 10 years of experience in pediatric dentistry to perform standard oral examinations on 200 children (aged 3-12, including healthy children, those with cavities, gingivitis, and other conditions) using the system's endoscopic equipment in a real clinical setting. The system records the status and actions at every moment. state : : CNN feature map of the current local image. 2048-dimensional feature vectors are extracted using a pre-trained ResNet-50 (with the classification layer removed).

[0047] The observed area mask is a binary matrix of the same size as the panoramic image (512×512). It is initially all 0s. After observing each area, the pixels in the corresponding circular area (radius 50 pixels) are set to 1.

[0048] Image quality scoring uses Laplacian variance to calculate sharpness, normalized to... Calculation method: Calculate the Laplacian operator on the image, calculate the variance Var, and then... Normalization.

[0049] action : The next point the doctor will observe is the center coordinates (pixel value).

[0050] The magnification of the endoscope actually used by the doctor (1.0~3.0).

[0051] : Whether the doctor marks the current area as abnormal (0 / 1).

[0052] Did the doctor give a diagnosis recommendation verbally or in writing? (0 / 1)

[0053] Simultaneously, the complete diagnostic report text generated by the doctor was recorded. A total of 15,000 valid trajectory data were collected, with each trajectory containing an average of 12 state-action pairs.

[0054] Policy Network Structure The structure of the policy network π is as follows: Input layer: Receives three types of input Feature map F_t: 7×7×2048 dimensions (after downsampling by ResNet-50), which is transformed into a 2048-dimensional vector through a global average pooling layer.

[0055] The mask M_t is 512×512. After passing through three convolutional layers (16, 32, and 64 channels, kernel size 3, stride 2, ReLU activation), it is flattened to obtain a 4096-dimensional vector.

[0056] Quality score q_t: a 1-dimensional scalar.

[0057] Fusion layer: The above three vectors are concatenated to obtain a total dimension of 2048 + 4096 + 1 = 6145.

[0058] Fully connected layers: The first layer has 256 neurons, and the second layer has 128 neurons, both using ReLU activation and Dropout (0.3).

[0059] Output layer: divided into four branches coordinate Output via sigmoid Multiply by the image width and height. .

[0060] zoom: Maps to [1, 3] by multiplying 2 by sigmoid and adding 1.

[0061] mark and suggest: Output probability via sigmoid, binarized with a threshold of 0.5.

[0062] Behavioral Cloning (BC) Training Training dataset: 12,000 trajectories; validation dataset: 3,000 trajectories. Loss function:

[0063] Using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 50 training epochs, the action prediction accuracy on the validation set reached 78% (coordinate error <10 pixels is considered correct, scaling error <0.2 is considered correct, and label and suggestion accuracy >85%).

[0064] GAIL optimization To further enhance the robustness and generalization ability of the strategy, generative adversarial imitation learning is employed. A discriminator D(s, a) is constructed: its structure consists of three fully connected layers (256-128-1), outputting a scalar representation. The probability of the trajectory derived from the expert is used. The original behavioral cloning loss is retained as a regularization term. Joint training is performed alternately. With a fixed policy π, update the discriminator D: maximize With D fixed, update strategy π: maximize The PPO algorithm is used for optimization.

[0065] PPO hyperparameter settings: Cutting parameters ; Generalized dominance estimation (GAE) parameters ; Learning rate 0.0003, batch size 256, update rounds 10; Entropy coefficient (Same as λ in GAIL); After 100 training rounds, the policy network's action accuracy on the validation set improved to 89%.

[0066] DAGGER Iterative Correction Because GAIL may introduce biases outside the state distribution, the DAGGER algorithm was further employed: detection trajectories were automatically generated on 50 new children using the current policy π, with expert physicians providing corrective actions after each step. The corrected (s, a_expert) values ​​were added to the dataset, and the policy was retrained. After three iterations, the action accuracy reached 96%, meeting the requirements for clinical application (>95%).

[0067] The final strategy network is frozen and deployed in the imitation learning module of the system.

[0068] V. Anomaly Detection Module Anomaly detection employs a pre-trained U-Net segmentation network. The network structure consists of a 4-layer encoder (64-128-256-512 channels) and a 4-layer symmetrical decoder with skip connections. The input is a pre-processed 512×512 local image (RGB three channels), and the output is a pixel-level classification of four categories: background, healthy teeth, cavities, and gingivitis.

[0069] Training dataset: Contains 10,000 expert-annotated local images of the oral cavity, with areas of tooth decay and gingivitis marked as pixel-level polygons in each image. Image enhancement includes random rotation (±15°), scaling (0.8 to 1.2 times), and brightness and contrast adjustment (brightness adjustment range 0.8 to 1.2, contrast adjustment range 0.8 to 1.2).

[0070] The loss function is the sum of multi-class cross-entropy loss and multi-class Dice loss:

[0071] in: This represents the total number of pixels in the image. Number of categories; The true label (one-hot encoding) for the i-th pixel belonging to class c; Let be the probability predicted by the model that the i-th pixel belongs to the c-th class; It is the natural logarithm.

[0072] The Adam optimizer was used for training, with a learning rate of 0.0001, a batch size of 16, and 80 training epochs, ultimately achieving an mIoU of 0.87.

[0073] During inference, the model outputs the category of each pixel for each newly acquired local image, and outputs the contour and confidence score of the abnormal area (cavity or gingivitis).

[0074] VI. Detailed Implementation of the Diagnostic Suggestion Generation Module This module is fine-tuned based on the GPT-2 Chinese pre-trained model (124M parameters). The input is constructed as follows: Detection path Observation sequence: (120, 240, 1.2) - (300, 180, 2.0) - (450, 230, 2.5) - end; Abnormal area: Area 1: Tooth decay, confidence level 0.95, location: occlusal surface of the left mandibular first primary molar; Region 2: Gingivitis, confidence level 0.88, location: gingival margin of the maxillary anterior teeth region; Child's age: 6 years old, gender: female, past medical history: two previous episodes of tooth decay.

[0075] The output is a structured text: [Main Problems] Superficial caries on the occlusal surface of the left lower first deciduous molar; redness and swelling of the upper anterior gums, with bleeding on probing.

[0076] [Possible causes] Frequent consumption of sugary foods and poor oral hygiene habits.

[0077] [Treatment Recommendations] 1. Preventive resin filling for cavities; 2. Topical medication for the gums and instruction on proper brushing techniques; 3. Follow-up examination every three months.

[0078] [Home Care] Use fluoride toothpaste, brush your teeth twice a day, and reduce your intake of sweets.

[0079] Training data: 5000 real pediatric dental records were used, from which detection paths (simulated generation) and doctor-written diagnostic reports were extracted as supervised pairs. Standard cross-entropy loss was employed, using the AdamW optimizer with a learning rate of 5e-5, 5 training epochs, and a batch size of 4. The generated text achieved a BLEU-4 score of 0.72, and the average score from human evaluation (3 doctors independently scoring, out of 5) was 4.3, with 89% considering it clinically acceptable.

[0080] VII. User Interface Implementation The user interface is based on web technology (React framework) and can be accessed by doctors or parents via tablets or PCs. Main functions: Login authentication: Use hospital employee ID / mobile phone number + dynamic verification code.

[0081] Main interface: Displays a list of children awaiting testing today, system notifications, and a summary of the latest test results.

[0082] Detection and Control: After clicking "Start Detection", the system automatically starts the acquisition process controlled by the policy network. The endoscope image is displayed in real time, and the attention distribution of the current policy network is shown in the form of a heatmap (the heatmap value is proportional to the probability density of the coordinate output in π(a|s)).

[0083] Results Display: After the test, a panoramic image of the oral cavity is displayed, with areas of tooth decay highlighted in red and gingivitis highlighted in yellow. Doctors can click on any area to view a high-resolution local image and abnormal confidence levels.

[0084] Historical records: Displays past test reports on a timeline, supporting comparison of abnormal area changes in the same child at different times.

[0085] Report Export: Generates a diagnostic report in PDF format, including images, diagnostic recommendations, and treatment plans.

[0086] Help and Support: Provides operation videos, frequently asked questions and online customer service.

[0087] VIII. Examples Example: Oral examination of a 6-year-old girl Child Information: Zhang, 6 years old, sought medical attention for toothache. No systemic diseases.

[0088] Testing process: Initial panoramic acquisition: The system controls the endoscope to automatically capture a panoramic image of the front of the oral cavity; the image is clear (q=0.92). The pre-processed image is then input into the policy network.

[0089] First round policy output: State s1: F1 is the global feature map, M1 is all 0, q=0.92. Policy network output action: coordinates (450, 230) (lower right region), zoom=2.0, mark=0, suggest=0.

[0090] Acquiring a local image: The device automatically moves the camera to (450, 230), zooms in by 2x, and obtains a clear image (q=0.97). The anomaly detection module did not detect any anomalies in this local image.

[0091] Update status: Set the circular region (450, 230) in mask M2 (radius 50 pixels) to 1. The policy network infers again.

[0092] Second round output: Action (120, 310) (lower left posterior teeth), zoom=2.5, mark=0, suggest=0. Anomaly identification after data collection revealed a brown patch on the occlusal surface of the lower left first deciduous molar, with a caries confidence level of 0.92. The system automatically marked this area as abnormal.

[0093] Third-round output: Based on the observed areas and anomaly detection, the policy network outputs an action (300, 180) (upper anterior teeth region), zoom=1.5, mark=0, suggest=0. Anomaly detection after data collection revealed gingival redness and swelling, with a confidence level of 0.85 for gingivitis.

[0094] Fourth round output: The policy network determines that the observed area accounts for 89% (threshold set at 90%), and outputs the "End detection" action.

[0095] Comprehensive Diagnosis: After summarizing all abnormal areas, input them into the diagnostic suggestion generation module, and output: Main problems: superficial caries in the lower left first deciduous molar; mild gingivitis in the upper anterior region.

[0096] Possible cause: Incomplete brushing, food gets stuck.

[0097] Treatment recommendations: preventative resin filling for cavities; basic periodontal treatment for gingivitis, and guidance on proper brushing techniques.

[0098] Follow-up examination recommendation: Follow-up examination in three months.

[0099] Report presentation: Doctors confirm the report on the interface, and parents obtain an electronic version by scanning a QR code with their mobile phones. The entire testing process takes approximately 2 minutes, and doctors only need to initiate the process and make final confirmations; no manual operation of the endoscope is required.

[0100] Comparison of results: Traditional manual examinations for the same condition typically take 5-8 minutes and rely heavily on the doctor's experience. This system automates and standardizes the detection path planning, making it particularly suitable for situations where children are uncooperative (the system can quickly adapt to minute head movements and compensate by adjusting the movements).

[0101] Summary of parameter ranges and explanation of their rationality ; This technical solution can also be implemented as a computer-readable storage medium storing a computer program. When executed by a processor, the program can implement all the steps of the method. The storage medium includes, but is not limited to, non-transitory media such as ROM, RAM, hard disk, solid-state drive, USB flash drive, CD-ROM, DVD, and cloud storage.

[0102] In a clinical trial involving 100 children (aged 3-12), the diagnostic results of this system were compared with those of three senior pediatric dentists: Sensitivity for detecting cavities: 97.2% vs. dentist's average 94.5%; Specificity for detecting gingivitis: 96.8% vs. 95.2% on average for doctors; Average testing time: 2'15" vs. 6'30" for manual testing by a doctor; Parent satisfaction rating (out of 10): 9.4 vs. traditional check-up 8.2.

[0103] The above data fully demonstrates the effectiveness, efficiency, and clinical applicability of this system.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting pediatric oral cavity based on image recognition, characterized in that, Includes the following steps: Obtain initial image sequences of the inside of a child's oral cavity; The image sequence is preprocessed, including denoising, enhancement, cropping, and normalization; The preprocessed image features and the current detection state are input into a policy network trained using an imitation learning algorithm, and the policy network outputs the next detection action. Based on the next detection action, the image acquisition device is automatically controlled to adjust the shooting angle, focus position, or zoom level to acquire new local oral cavity images; Anomaly identification is performed on the new local oral cavity image to segment and classify abnormal regions in the oral cavity; Repeat the above strategy steps of network inference, device control, image acquisition and anomaly recognition until the detection termination condition is met; By combining all anomaly identification results and the detection paths generated by the policy network, diagnostic suggestions are generated and output.

2. The method for detecting pediatric oral cavity based on image recognition according to claim 1, characterized in that: The imitation learning algorithm includes behavior cloning, specifically: The operation trajectory data of several pediatric dentists were collected. Each trajectory includes a state sequence and a corresponding expert action sequence. The status includes: the feature map of the current oral cavity image region, the mask of the observed region, and the image quality score; The observed region mask is a binary matrix of the same size as the panoramic image of the oral cavity, where 1 indicates that the pixel position has been observed and 0 indicates that it has not been observed. The image quality score uses Laplacian variance to calculate image sharpness, and the calculated value is normalized to... The range indicates that the higher the value, the clearer the image; The actions include: the center coordinates of the next observation area, the zoom level, whether to mark an anomaly, and whether to generate suggestion text; The loss function minimizes the difference between the policy network's output action and the expert's action through supervised learning. in, The total number of time steps for a trajectory; For the first The state of the step; For the first Step-by-step actions performed by experts; For the policy network in state Down Output Action The probability of is given by , and its range is . ; It is the natural logarithm.

3. The method for detecting pediatric oral cavity based on image recognition according to claim 2, characterized in that: The imitation learning algorithm further employs Generative Adversarial Imitation Learning (GAIL) to optimize the policy network: Constructing a discriminator network Its input is the state. and actions The output is a scalar value, representing The range of probabilities derived from expert trajectories is . ; Reinforcement learning is used to optimize the policy network to maximize the authenticity score of the discriminator output while minimizing the policy entropy; The optimization objective is: in: This represents the policy network to be optimized. Indicate expert strategy; Representation Strategy The expected trajectory of the generated state and action; This represents the expected value of the expert's trajectory; For discriminator output; Here, is the entropy regularization coefficient, taking a value of . ; Let be the entropy of the strategy.

4. The method for detecting pediatric oral cavity based on image recognition according to claim 1, characterized in that: The policy network is also trained using the DAGGER algorithm: The initial policy is trained using behavior cloning; Trajectories generated by the policy network are collected in a real-world environment, and human experts provide corrected optimal actions for each state in the trajectory. Add the corrected state-action pairs to the training set and retrain the policy network; Iterate until the policy network achieves an action accuracy of over 95% on the validation set.

5. The method for detecting pediatric oral cavity based on image recognition according to claim 1, characterized in that: The anomaly detection employs a segmentation model based on a deep convolutional neural network, such as Mask R-CNN or U-Net, which outputs the category probability for each pixel. Categories include healthy teeth, cavities, gingivitis, and tartar. The training loss function is the sum of cross-entropy loss and Dice loss, as shown in the formula: in: This represents the total number of pixels in the image. The total number of categories, For pixel index, Indexed by category; For the first The pixel belongs to the first The actual label of the class is encoded using one-hot encoding; The model predicts the first The pixel belongs to the first The probability of a class.

6. The method for detecting pediatric oral cavity based on image recognition according to claim 1, characterized in that: The diagnostic recommendations are generated using a Transformer-based natural language generation model. The input consists of the detection path and anomaly identification results output by the policy network, and the output is a structured Chinese diagnostic recommendation text, including a problem description, possible causes, treatment plan, and suggested follow-up time.

7. The method for detecting pediatric oral cavity based on image recognition according to claim 1, characterized in that: The detection termination condition is any of the following: The policy network outputs the "End detection" action; The observed area accounts for a proportion of the total area of ​​the panoramic oral cavity image that exceeds a preset threshold. This threshold ranges from [value missing]. ; The confidence level of the output of three consecutive actions is lower than Wherein, confidence level is the probability value corresponding to the action output by the policy network. .

8. A pediatric oral cavity detection system based on image recognition, used to perform the pediatric oral cavity detection method based on image recognition as described in any one of claims 1 to 7, characterized in that, include: Image acquisition module: Used to acquire image sequences of the inside of a child's mouth, including images with a resolution of at least [resolution not specified]. High-definition cameras or oral endoscopes, automatic light source adjustment devices, and acquisition and control units that receive action commands from the policy network; Image preprocessing unit: performs noise reduction, enhancement, cropping, and normalization on the image; Imitation learning module: Stores and runs the policy network trained using the imitation learning algorithm, and outputs the next detection action based on the current state; Anomaly detection module: Segmentation and classification of abnormal regions in oral cavity images based on convolutional neural networks; Diagnostic suggestion generation module: Generates natural language diagnostic suggestions based on comprehensive detection results and detection paths; User interface: used to display test results, test path animation, diagnostic suggestions, and historical records.

9. A pediatric oral cavity detection system based on image recognition according to claim 8, characterized in that, The user interface also includes a path visualization module, used for: The heatmap displays the regions and intensity of interest that the policy network has focused on, where the heat value is proportional to the probability of the policy network outputting an action at that location. .

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images