Medical-grade oral cavity cleaning method and device based on AI assistance

Abstract

The invention belongs to the technical field of oral cavity medical treatment, and particularly relates to a medical-grade oral cavity cleaning method and device based on AI assistance, and the method comprises a fine-grained oral cavity state detection method which is used for recognizing and positioning an oral cavity state image shot by an oral cavity image collection system, accurately recognizing tiny dental plaque and dental calculus, and detecting the oral cavity state image. Guidance and reference are provided for oral cavity cleaning; the intelligent analysis report generation method is used for generating a personalized and customized analysis report according to a result given by the fine-grained oral cavity state detection method; according to the intelligent personalized cleaning scheme generation method, a personalized and customized medical-grade oral cleaning scheme is generated according to key information input by a user, and an oral cleaning doctor is guided to take effective and safe oral cleaning operation on the oral cavity of a patient; the invention also provides a medical oral cavity cleaning device based on AI assistance, which comprises an oral cavity image acquisition system, an image processing subsystem, an imaging display, a report generation subsystem and an oral cavity cleaning subsystem, so that the medical oral cavity cleaning accuracy and efficiency are effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of oral medical treatment, and particularly relates to an AI-assisted medical-grade oral cleaning method and device. BACKGROUND

[0002] With the rapid development of artificial intelligence (AI) technology, its application in the medical field is becoming more and more extensive, and the oral medical field is no exception. Traditional oral cleaning methods mainly rely on the clinical experience of dentists and the self-maintenance of patients, but such methods have many limitations. For example, manual tooth brushing is difficult to completely remove dental plaque, especially in areas that are difficult to reach; and the early diagnosis of oral diseases and the design of individualized treatment plans are also limited by the experience and time of doctors. Therefore, the research on AI-assisted medical-grade oral cleaning methods and devices has emerged as the times require, aiming to improve the effect and efficiency of oral cleaning and improve the level of prevention and treatment of oral diseases through intelligent means.

[0003] The application of AI in the field of oral medical treatment can be traced back to the breakthroughs in multiple technologies in recent years, including deep learning algorithms, three-dimensional imaging technology, and robot technology. The integration of these technologies has shown great potential for AI in the early diagnosis of oral diseases, individualized treatment, and the development of intelligent devices. In particular, in the field of oral cleaning, AI-assisted devices and methods can accurately identify dental plaque, develop individualized cleaning plans, and monitor cleaning effects in real time, thereby greatly improving the intelligent level of oral cleaning.

[0004] Firstly, the application of AI in the field of oral medical treatment benefits from its powerful data processing and analysis capabilities. AI algorithms can process massive amounts of oral health data, including oral images, patient lifestyle habits, genetic information, etc., thereby achieving accurate assessment of oral health. In the field of oral cleaning, AI can identify and analyze the distribution of bacteria in the oral cavity and the formation of dental plaque, etc., providing scientific basis for developing individualized cleaning plans; Secondly, AI-assisted oral cleaning devices usually combine high-precision sensors and intelligent control systems. These devices can monitor the cleaning conditions in the oral cavity in real time and make adjustments as needed. For example, some intelligent toothbrushes and oral irrigators can analyze brushing force, frequency, and coverage range in real time through built-in sensors and AI algorithms, thereby providing individualized cleaning recommendations. In addition, AI can also assist in designing more accurate oral cleaning paths to ensure that every corner is thoroughly cleaned.

[0005] In medical-grade oral cleaning methods, the application of AI technology also lies in the prediction and diagnosis of oral diseases. Through learning and analyzing a large number of oral cases, AI can predict the oral diseases that patients may have and develop prevention and treatment strategies in advance. This not only improves the effect of oral cleaning, but also effectively prevents the occurrence of oral diseases.

[0006] In addition, the innovation of AI technology in the field of oral medicine also promotes the development of robot technology. Oral robots can realize examination, diagnosis and treatment of patients' oral cavity through high-precision sensors and control systems. In terms of oral cleaning, robots can simulate human tooth brushing actions to clean teeth and gums more evenly and deeply. This robot-assisted oral cleaning method not only improves cleaning efficiency, but also reduces the difficulty and risk of manual operation.

[0007] Currently, the significance of the research on AI-assisted medical-grade oral cleaning methods and devices includes: 1. Improve the accuracy and efficiency of oral cleaning Traditional oral cleaning methods often rely on patients' self-tooth brushing habits and dentists' clinical experience, which leads to large individual differences in cleaning effect. AI-assisted oral cleaning devices can accurately identify the distribution of dental plaque in patients' oral cavity through deep learning algorithms and three-dimensional imaging technology, and develop personalized cleaning programs. For example, some devices can use cameras and indicator sprayers to obtain image data of users' oral cavity, determine dental plaque in target areas, and adjust cleaning intensity and duration accordingly, to ensure the accuracy and efficiency of cleaning.

[0008] 2. Realize early prevention of oral diseases Another important application of AI technology in the field of oral medicine is early diagnosis of diseases. By analyzing patients' oral scan images and data, AI can identify early symptoms and potential problems, such as early signs of dental caries and periodontal disease. This not only helps patients receive effective treatment in time and avoid further deterioration of the disease, but also reduces treatment costs and improves treatment effect. In terms of oral cleaning, AI-assisted devices can monitor changes in oral environment, such as bacterial content and pH value, to help patients discover potential oral problems in time and take preventive measures.

[0009] 3. Promote the development of personalized treatment programs The oral health status and needs of each patient are unique, so it is particularly important to provide personalized treatment plans. AI technology, through big data analysis and machine learning, can tailor oral cleaning and treatment plans for each patient. For example, in the field of orthodontics, AI can recommend the most suitable type of orthodontic device and wearing plan based on the patient's tooth morphology, bite situation, and other data. In the field of dental implants, AI can accurately calculate the best position and angle of the implant, improving the success rate and aesthetics of the implant. Similarly, in the field of oral cleaning, AI-assisted devices can develop personalized cleaning plans based on the patient's oral image data and cleaning history to maximize cleaning effectiveness.

[0010] 4. Optimizing the allocation of oral medical resources The application of AI technology in the field of oral medicine can also optimize the allocation of medical resources. Traditional oral medical services are often limited by the number and experience level of doctors, resulting in long patient waiting times and low treatment efficiency. However, with AI-assisted remote medical and intelligent diagnosis systems, doctors can provide remote diagnosis and consultation services to patients, thereby alleviating the imbalance of medical resources to some extent. In addition, AI can be used to optimize the design and data analysis of clinical trials, accelerating the development process of new drugs and new therapies, and bringing more innovation and development to the field of oral medicine.

[0011] 5. Enhancing patient experience and satisfaction AI-assisted oral cleaning devices and methods not only improve cleaning effectiveness and treatment efficiency, but also enhance patient experience and satisfaction. For example, some smart oral cleaning devices have built-in AI algorithms that can automatically adjust the motion mode and intensity of the brush head based on the user's brushing habits and oral health status, and provide immediate feedback and improvement suggestions through a mobile application. This real-time intelligent guidance not only improves the effectiveness of brushing, but also effectively prevents gum damage caused by excessive force. In addition, AI technology can be used to develop intelligent oral care devices such as smart electric toothbrushes and smart oral cameras, which can monitor changes in the oral environment in real time and provide personalized care recommendations, further enhancing patient experience and satisfaction.

[0012] In summary, the research on AI-assisted medical-grade oral cleaning methods and devices has far-reaching significance. It not only improves the accuracy and efficiency of oral cleaning, enables early prevention of oral diseases, promotes the development of personalized treatment plans, optimizes the allocation of oral medical resources, but also enhances patient experience and satisfaction. With the continuous development and expansion of AI technology, we have reason to believe that AI-assisted oral cleaning methods and devices will play an increasingly important role in the future of oral medicine, bringing more well-being to human oral health.

[0013] An oral cleaning device and method are disclosed in prior art CN118453165A, comprising an electric toothbrush including a handle and a detachable toothbrush head, and a flushing mechanism and a liquid suction mechanism are arranged on the electric toothbrush; the flushing mechanism includes flushing pipes located on both sides of the toothbrush head, and flushing holes are arranged on the flushing pipes; the liquid suction mechanism includes liquid suction pipes located on the inner side of the flushing pipes, and liquid suction holes are arranged on the liquid suction pipes. The invention is designed for patients who cannot take care of themselves. The invention uses an electric toothbrush, vacuum negative pressure liquid suction and flushing liquid to clean the oral cavity of patients using the method of brushing teeth. It is simple and convenient to use, has strong cleaning strength, and can be flushed, brushed and sucked. It is more suitable for patients who cannot actively cooperate with the suction action. However, the invention mainly realizes the most basic oral cleaning from the perspective of mechanical design, and the cleaning effect and efficiency are poor. In prior art CN17598816A, a method and device for preventing tooth decay based on an oral cleaning device and a storage medium are disclosed, comprising obtaining the historical cleaning record of the oral cleaning device; marking the caries risk tooth area according to the historical cleaning record; if the current cleaning tooth area is the caries risk tooth area, increasing the frequency and / or duty cycle of the oral cleaning device. In this way, the historical cleaning record of the user using the oral cleaning device can be used to more accurately configure and optimize the oral cleaning needs of each user. When the oral cleaning device cleans the tooth area marked as a caries risk, the operating frequency and / or duty cycle of the device can be increased to enhance its cleaning effect, thereby more effectively preventing the occurrence of caries and improving the quality of life of the user. However, this method is based on the historical cleaning record of the oral cleaning device, and has poor generalization ability and flexibility. It cannot effectively identify, mark, clean and remove newly occurring caries. SUMMARY

[0014] To solve the above problems, the present application proposes a medical-grade oral cleaning method and device based on AI assistance to solve the problems existing in the above-mentioned prior art and improve the accuracy, efficiency and generalization ability of oral cleaning.

[0015] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a medical-grade oral cleaning method based on AI assistance applied to a medical-grade oral cleaning device based on AI assistance, comprising an oral state intelligent recognition and analysis method and an intelligent individualized cleaning scheme generation method.

[0016] Further, the AI-assisted medical-grade oral cleaning method includes an oral state intelligent detection and analysis method combined with an oral visual sensor and an artificial intelligence algorithm to accurately detect and identify the tooth arrangement state, tooth growth condition, tooth spacing, dental plaque distribution, dental calculus distribution, and gum state in the oral cavity, and provide a tooth condition analysis and evaluation report according to the identification condition. The method specifically includes a fine-grained oral state detection method and an intelligent analysis report generation method.

[0017] Further, the AI-assisted medical-grade oral cleaning method includes a fine-grained oral state detection method mainly based on a fine-grained image recognition neural network model to identify and locate the oral state images captured by the oral image acquisition system, accurately identify the tiny dental plaque and dental calculus, and provide guidance and reference for oral cleaning. The specific steps include: S10: The oral image acquisition system captures the oral state information of the patient from multiple angles and in all directions, mainly including tooth state information and gum state information; S11: The captured images are subjected to image space transformation, including geometric transformation operations such as translation, transposition, mirroring, rotation, and scaling, to correct the system error of the image acquisition system and the random error of the instrument position (such as imaging angle, perspective relationship, and even the lens itself); S12: During image space transformation, a gray scale interpolation algorithm is used to process the transformed images to prevent the phenomenon that the pixels of the output image are mapped to non-integer coordinates of the input image; S13: The image after image space transformation is subjected to low-illumination image enhancement operation to eliminate image shadows, improve brightness, and enhance the visibility and intelligibility of the image; S14: The enhanced image is subjected to restoration operation, including eliminating noise, deviation, and shadow interference; S15: The image after step S14 processing is subjected to image registration operation to calibrate the images captured at different times and from different angles to the same coordinate system, facilitating subsequent neural network feature recognition and processing; S16: The two-dimensional plane image after step S15 processing is used to build a three-dimensional oral model to restore a real, three-dimensional visual oral state model; S17: The three-dimensional visual oral model after modeling is compared with a standard oral three-dimensional model, and the mean absolute error (MAE) index is used to measure the gap between the modeled oral model and the standard oral model. The comparison parameters include the number of teeth, tooth spacing, tooth length, tooth width, and tooth curvature degree. The mean absolute error calculation formula is: ; wherein i is the tooth number, is the i th tooth generated by modeling, the i th tooth corresponding to the standard tooth model; S18: Label the image processed in step S15 by image labeling software, including using different color labeling boxes to frame the positions of plaque, calculus, and gingival swelling, and giving corresponding text labeling information, and on this basis, constructing a training data set and a test data set in a ratio of 7:3, the training data set including the labeled oral image data, and the test set including unlabeled image data; S19: Construct a fine-grained image detection neural network model based on the improved YOLOv8, on the basis of the YOLOv8 neural network model, use a pyramid convolution layer instead of a normal convolution layer to fuse features of different scales, so as to improve the detection ability of the model on multi-scale targets; add a multi-head attention mechanism in the network to strengthen the model's ability to capture fine-grained features; and add a fine-grained classifier to further improve the accuracy of identification and classification of fine-grained small targets such as calculus, plaque, and gingival swelling; S20: Test the neural network constructed in step S19 on the training data set constructed in step S18, select the Adam optimizer, and design a multi-dimensional loss function for classification and positioning as the loss function, including two parts of classification loss and positioning loss; The classification loss selects a cross-entropy loss function, which is used to measure the accuracy of the neural network model in target classification, and the calculation formula is: ; wherein, y represents the true label of the sample, y represents the label predicted by the model. When y = 1, it means that the sample belongs to the positive class; when y = 0, it means that the sample belongs to the negative class;

[0018] The positioning loss is a Smooth L1 loss function, which is used to measure the accuracy of the model in target positioning, and the calculation formula is: ; wherein, x is the coordinate difference between the predicted bounding box and the true bounding box, and the classification loss function and the positioning loss function are combined as the optimization objective function for training the neural network model; S21: Test the neural network model trained in step 20 on the test set, and select the precision index as the test index, and the calculation formula is: ; wherein, tp is the number of instances that the model correctly predicts as positive samples, fn is the number of instances that the model incorrectly predicts as positive samples.

[0019] In one preferred embodiment, the gray scale interpolation algorithm of step S12 includes the following specific steps: S120: Determine the type of image transformation according to actual needs, such as scaling, rotation, translation; S121: Calculate the corresponding transformation matrix according to the transformation type. In scaling transformation, the transformation matrix is a diagonal matrix, and the elements on the diagonal are scaling factors. In rotation transformation, the transformation matrix is a rotation matrix; S122: Select the appropriate gray scale interpolation method according to the requirements, including nearest neighbor interpolation, bilinear interpolation, and cubic interpolation; S123: For each pixel point in the target image, calculate its corresponding coordinates in the original image according to the transformation matrix; S124: Determine the neighborhood pixels around the target image pixel according to its corresponding coordinates in the original image. The gray scale values of these neighborhood pixels will be used for interpolation calculation; S125: According to the selected interpolation method, select the corresponding interpolation formula, and use the selected interpolation formula and the gray scale values of the neighborhood pixels to calculate the gray scale value of the target pixel; S126: Repeat the mapping and interpolation calculation steps for each pixel point in the target image until the gray scale values of all pixel points are calculated. Combine the calculated target pixel gray scale values to generate the transformed complete image; S127: Perform smoothing processing on the transformed image to reduce noise and artifacts that may be generated in the interpolation process.

[0020] Another preferred embodiment, the low-illumination image enhancement of step S13 is a low-illumination image enhancement method based on wavelet transform, the specific steps include: S130: Decompose the original image using wavelet transform algorithm. This step usually involves decomposing the image through low-pass filter (scale function) and high-pass filter (wavelet function), thereby obtaining sub-band images of different frequency bands. Specifically, wavelet decomposition can decompose the image into a low-frequency sub-band ( ) and three high-frequency sub-bands ( 、 、 ), which represent different frequency components of the image; S131: Nonlinear enhancement, perform nonlinear enhancement processing on the sub-band images obtained by decomposition to improve the contrast of the image and more effectively control the signal-to-noise ratio. The specific methods include adjusting the gray scale values of the sub-band images and applying nonlinear functions; S132: Detail extraction and enhancement, for high-frequency sub-band ( 、 、 ), further detail extraction and enhancement can be performed; S133: Reconstructing the image, after completing the processing of the sub-band image, the processed sub-band image is reconstructed into an enhanced image by using an inverse wavelet transform algorithm; the inverse wavelet transform is the inverse process of the wavelet transform, which combines the processed sub-band image to generate the final enhanced image.

[0021] Further, the AI-assisted medical-grade oral cleaning method, the intelligent analysis report generation method generates a personalized customized analysis report based on the image generation text idea according to the oral state image captured by the oral image acquisition system and the result given by the fine-grained oral state detection method, assists the oral doctor in making a judgment, and gives an effective oral cleaning and treatment scheme, and the specific steps include: S30: Constructing a medical-grade oral cleaning special data set according to prior expert experience, the data set includes labeled oral image data and corresponding expert text description labels; the expert text description label is the expert diagnosis opinion given by the oral medicine expert after observing and researching the oral image, which is divided into a training set and a test set according to a proportion; S31: Constructing a conditional generative adversarial network model, first defining a generator model, the generator network is based on a convolutional neural network architecture, the input is random noise, image features and image condition information, the image condition information is the parameter deviation of the modeled three-dimensional visual oral model and the standard oral three-dimensional model calculated in step S17; the goal of the generator is to generate a corresponding text description according to the input image features; the output of the generator is a text sequence, which can be characters, words or sentences; S32: Designing a discriminator network, the discriminator network is also based on a convolutional neural network model, the input of the discriminator network is the text description of the expert diagnosis opinion and its corresponding image features, the goal of the discriminator is to judge whether the input text matches the image features, that is, to judge whether the text is real or generated, the output of the discriminator is a probability value, indicating the degree of matching between the input text and the image features; S33: Initialize the parameters of the generator and the discriminator, initialize the parameters of the generator and the discriminator using a random initialization method before the training starts; S34: Training the generator network, extracting a batch of images and corresponding real texts from the training set; taking these images and real texts as the input of the discriminator and calculating the probability value output by the discriminator; using a cross-entropy loss function to measure the difference between the probability value output by the discriminator and the real label (i.e. whether the text is real); updating the parameters of the discriminator through a backpropagation algorithm to minimize the loss function; S35: Train the discriminator network. Extract a batch of images from the training set and use the generator to generate corresponding text. Input these generated texts and corresponding images into the discriminator and calculate the probability value of the discriminator output. Use the cross-entropy loss function to measure the difference between the text generated by the generator and the real text (indirectly measured through the output of the discriminator). Update the parameters of the generator through the backpropagation algorithm to maximize the probability of the discriminator's incorrect judgment of the generated text (i.e., make the discriminator think that the generated text is real). S36: Cycle update. During the training process, update the parameters of the generator once every few times the parameters of the discriminator are updated. Through this adversarial training, the performance of the generator and discriminator will continue to improve, and this cycle will continue until the model reaches a state of convergence. S37: After the model training reaches convergence, evaluate the performance of the conditional generative adversarial network model on the test set, and select the neural network model weight parameters with excellent performance for deployment and use.

[0022] In one preferred embodiment, the performance indicator evaluation in step S37 combines manual evaluation and objective indicator evaluation. The manual evaluation method invites experts to score or evaluate the generated text, and the objective evaluation indicator uses the BLEU text generation evaluation indicator to evaluate the quality of the generated text.

[0023] Further, the intelligent personalized cleaning scheme generation method based on AI assistance is based on a generative AI large model. According to the key information input by the user, a reasonable and personalized medical-grade oral cleaning scheme is generated to guide the oral cleaning physician to perform effective and safe oral cleaning operations on the patient's oral cavity. The specific steps include: S40: According to the actual needs, select a suitable language generation large model for local deployment. Download the selected language generation large model from the official website or model library, and after downloading and installing, verify the integrity and correctness of the model to ensure that the model can run normally. S41: Source code deployment. According to the model running requirements, configure the development environment, copy the model's source code and dependencies to the local development environment, and compile and install according to the instructions. After configuring the parameters and configuration files required by the model, start the model service. S42: Application deployment. Install and configure according to the tool instructions, import the model file into the tool, start the tool service, and interact with the model through the provided interface or interface. S43: Function test. Test whether the basic functions of the model in the local environment are normal, including input and output, response speed. S44: Performance test. Evaluate the performance of the model in the local environment, including processing speed, resource occupation. S45: Optimization, according to the test results, optimize the model, including adjusting parameters, optimizing data processing flow; S46: Continuous monitoring, after the model is deployed, continuously monitor the running situation and performance indicators of the model, and timely discover and solve problems; S47: Permission setting, ensure that only authorized users can access and operate the model service; S48: Data encryption, encrypt sensitive data to ensure data security; S49: Log recording, record the running log and abnormal information of the model, which is convenient for subsequent analysis and troubleshooting; S50: Fine-tuning of large models, using fine-tuning tools to load selected pre-trained models, keeping the model architecture unchanged during the loading process, only adjusting part of the parameters or layers S51: Set training hyperparameters, including learning rate, batch size, and training epoch; S52: Preprocess the input data, which includes real oral cavity photos, key information, and corresponding oral cavity cleaning plans formulated by experts. Real oral cavity photos and key information are inputs of the large model, and oral cavity cleaning plans formulated by experts are outputs of the large model. The above data is preprocessed by data cleaning, normalization, and encoding; S53: Input the preprocessed data into the large model for small-epoch fine-tuning training; S54: Deploy the fine-tuned large model to the system for use, and input the oral cavity state key information to automatically generate a customized oral cavity cleaning plan.

[0024] On the other hand, an AI-assisted medical-grade oral cleaning device is provided, which is applied to any one of the AI-assisted medical-grade oral cleaning methods. The AI-assisted medical-grade oral cleaning device comprises:

[0025] Oral image acquisition system: The oral image acquisition system is composed of a high-definition rotatable oral camera. The high-definition rotatable oral camera can realize 360-degree oral imaging according to needs, and can select video recording and image shooting according to needs, providing intuitive data basis for subsequent oral state recognition and analysis.

[0026] Image processing subsystem: The image processing subsystem carries image preprocessing algorithms, and performs preprocessing operations on the oral images captured by the high-definition oral camera to obtain high-definition effective oral images.

[0027] Image display: The image display takes a liquid crystal display as a carrier and can display the captured oral images in high definition.

[0028] Big computing server: the big computing server contains multiple big computing graphics cards, which can support the training and testing of image recognition neural network and generative adversarial network model at the same time, and also support the operation of other image preprocessing algorithms.

[0029] Compared with the prior art, the beneficial effects of the present application are:

[0030] 1. The present application accurately detects and identifies the tooth arrangement state, tooth growth condition, tooth spacing, dental plaque distribution, dental calculus distribution and gum state in the oral cavity through the oral cavity state intelligent detection and analysis method, and gives a tooth condition analysis and evaluation report according to the identification condition, greatly improving the effect and efficiency of oral cleaning.

[0031] 2. The present application generates a reasonable and personalized medical-grade oral cleaning scheme based on the generated AI big model according to the key information input by the user, and guides the oral cleaning physician to take effective and safe oral cleaning operation on the patient's oral cavity. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The present application is based on an AI-assisted medical-grade oral cleaning method flowchart.

[0033] Figure 2 The present application is based on an AI-assisted medical-grade oral cleaning device structure diagram. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0035] In the description of the present application, the terms "first", "second" are used only for description purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0036] In the description of the present application, the term "for example" is used to mean "serving as an instance, example or illustration". Any embodiment described as "for example" in the present application is not necessarily to be construed as preferred or advantageous over other embodiments. The following description is presented to enable any person skilled in the art to make and use the application. In the following description, for the purpose of explanation, details are set forth. It is apparent to those skilled in the art that the present application can be practiced without the use of these specific details. In other instances, well-known structures and processes are not elaborated in order not to obscure the description of the present application with unnecessary details. Thus, the present application is not intended to be limited by the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0037] In an embodiment of the present application, a medical-grade oral cleaning method based on AI assistance is provided, which can be specifically seen from the accompanying Figure 1 application to an AI-assisted medical-grade oral cleaning device, which includes a dental oral state intelligent recognition and analysis method and an intelligent personalized cleaning scheme generation method.

[0038] The oral state intelligent detection and analysis method includes a fine-grained oral state detection method and an intelligent analysis report generation method, which, in combination with an oral visual sensor and an artificial intelligence algorithm, accurately detects and identifies the tooth arrangement state, tooth growth condition, tooth spacing, dental plaque distribution, dental calculus distribution and gum state in the oral cavity, and gives a tooth condition analysis and evaluation report according to the identification.

[0039] The intelligent personalized cleaning scheme generation method is based on a generative AI large model, which generates a reasonable and personalized medical-grade oral cleaning scheme according to the key information input by the user, to guide the oral cleaning physician to take effective and safe oral cleaning operations on the patient's oral cavity.

[0040] Referring to the accompanying Figure 2 The present application also provides an AI-assisted medical-grade oral cleaning device, which is applied to any one of the AI-assisted medical-grade oral cleaning methods, and includes: Oral image acquisition system: The oral image acquisition system is composed of a high-definition rotatable oral camera, which can realize 360-degree oral imaging according to needs, and can select video recording and image shooting according to needs, to provide an intuitive data basis for subsequent oral state recognition and analysis.

[0041] Image processing subsystem: The image processing subsystem carries an image preprocessing algorithm, which performs preprocessing operations on the oral images captured by the high-definition oral camera, to obtain high-definition effective oral images.

[0042] Image display: The image display uses a liquid crystal display as a carrier and can display the captured oral image in high definition.

[0043] Report generation subsystem: The report generation subsystem is built-in with intelligent analysis report generation methods and intelligent personalized cleaning solution generation methods, and is loaded on a high-performance computer device. The user interacts with the report generation subsystem through the display, mouse, keyboard, and upper computer functional interface to realize operation and information exchange.

[0044] Oral cleaning subsystem: The oral cleaning subsystem is equipped with an intelligent mechanical arm and a sensing device. The front end of the mechanical arm is installed with cleaning instruments required for tooth cleaning, and the oral cavity can be effectively cleaned and protected according to the needs.

[0045] High-performance server: The high-performance server contains multiple high-performance computing graphics cards, which can simultaneously support the training and testing of image recognition neural networks and generative adversarial network models, and can also support the operation of other image preprocessing algorithms.

[0046] In specific embodiments, the dental nurse captures the oral cavity state pictures of the patient from multiple angles and in all directions through the oral image acquisition system, and realizes the spatial transformation, denoising, and enhancement of the image through the image preprocessing method built-in the image processing subsystem, to obtain high-resolution and high-quality oral cavity images in all directions. Then, the dental nurse accurately detects, identifies, and locates the positions of dental plaque, dental calculus, and gingival swelling through the fine-grained oral state detection method, and accurately calculates the state deviation between the patient's teeth and the standard tooth model. The labeled image and the calculation result are displayed on the image display. Then, the dental nurse generates a preliminary detection analysis report according to the intelligent analysis report generation method to guide the subsequent oral cleaning work. The dental nurse inputs the keywords into the report generation subsystem upper computer interface according to the generated preliminary analysis detection report, and automatically generates a personalized cleaning solution through the functional buttons, and controls the oral cleaning subsystem according to the personalized cleaning solution to realize medical-grade oral cleaning of the patient, including eliminating dental plaque, dental calculus, and replacing gingival swelling.

[0047] Example 1

[0048] In one embodiment, the dental nurse detects the oral cavity state through the granularity oral cavity state detection method, and identifies and locates the oral cavity state image captured by the oral cavity image acquisition system based on a fine-grained image recognition neural network model, accurately identifies tiny dental plaque and dental calculus, and provides guidance and reference for oral cavity cleaning. First, the dental nurse captures the oral cavity state information of the patient through the oral cavity image acquisition system from multiple angles and in all directions, obtains the tooth state information and the gum state information, and then performs image space transformation on the collected images, that is, performs geometric transformation operations including translation, transposition, mirroring, rotation, and scaling, to correct the system error of the image acquisition system and the random error of the instrument position (such as the imaging angle, the perspective relationship, and even the lens itself). When performing image space transformation, a gray scale interpolation algorithm is used to process the transformed image to prevent the phenomenon that the pixels of the output image are mapped to non-integer coordinates of the input image. The specific execution steps include: first, determining the type of image transformation according to actual needs, such as scaling, rotation, and translation; calculating the corresponding transformation matrix according to the transformation type, wherein in scaling transformation, the transformation matrix is a diagonal matrix, and the elements on the diagonal line are scaling factors; in rotation transformation, the transformation matrix is a rotation matrix; selecting a suitable gray scale interpolation method according to the requirements, including nearest neighbor interpolation, bilinear interpolation, and cubic interpolation; for each pixel point in the target image, calculating its corresponding coordinates in the original image according to the transformation matrix; determining the neighborhood pixels around the corresponding coordinates of the target image pixel in the original image, and the gray scale values of these neighborhood pixels will be used for interpolation calculation; according to the selected interpolation method, selecting the corresponding interpolation formula, and using the selected interpolation formula and the gray scale values of the neighborhood pixels to calculate the gray scale value of the target pixel; repeating the mapping and interpolation calculation steps for each pixel point in the target image until the gray scale values of all pixel points are calculated, combining the calculated target pixel gray scale values to generate the complete transformed image; and performing smoothing processing on the transformed image to reduce the noise and artifacts that may be generated in the interpolation process.

[0049] The image after image space transformation is subjected to low-illumination image enhancement operation to eliminate image shadows, improve brightness, and enhance the visibility and intelligibility of the image; the enhanced image is subjected to restoration operation, including eliminating noise, deviation, and shadow interference; the processed image is subjected to image registration operation to calibrate the images captured at different times and from different angles to the same coordinate system, facilitating subsequent neural network feature recognition and processing; and a three-dimensional oral cavity model is established based on the processed two-dimensional plane image to restore a real, three-dimensional visual oral cavity state model. Then, the dental nurse compares the three-dimensional visualized dental model after modeling with the standard dental three-dimensional model, uses the mean absolute error (MAE) index to measure the gap between the dental model after modeling and the standard dental model, and the comparison parameters include the number of teeth, the tooth spacing, the tooth length, the tooth width, and the tooth bending degree, and the mean absolute error calculation formula is: ; wherein i is the tooth number, is the ith tooth generated by modeling, is the ith tooth corresponding to the standard tooth model; At the same time, the processed images are labeled by image labeling software, including using different color labeling boxes to frame the positions of plaque, calculus, and gingival swelling, and giving corresponding text labeling information, and on this basis, a training data set and a test data set are constructed in a ratio of 7:3, the training data set includes labeled dental image data, and the test set includes unlabeled image data. A fine-grained image detection neural network model based on improved YOLOv8 is constructed, on the basis of the YOLOv8 neural network model, a pyramid convolution layer is used instead of an ordinary convolution layer, different scale features are fused, so as to improve the detection ability of the model to multi-scale targets; a multi-head attention mechanism is added to the network to strengthen the model's ability to capture fine-grained features; and a fine-grained classifier is added to further improve the accuracy of identification and classification of fine-grained small targets such as calculus, plaque, and gingival swelling. The constructed neural network is tested on the constructed training data set, an Adam optimizer is selected, and a multi-dimensional loss function combined with classification and positioning is designed as the loss function, including two parts of classification loss and positioning loss. The classification loss selects a cross-entropy loss function, which is used to measure the accuracy of the neural network model in target classification, and the calculation formula is: ; wherein, y represents the true label of the sample, y represents the label predicted by the model. When y = 1, it means that the sample belongs to the positive class; when y = 0, it means that the sample belongs to the negative class. The positioning loss is a Smooth L1 loss function, which is used to measure the accuracy of the model in target positioning, and the calculation formula is: ; wherein x is the coordinate difference between the predicted bounding box and the real bounding box, and the classification loss function and the positioning loss function are combined as the optimization objective function of the neural network model training; The trained neural network model is tested on the test set, and the precision index is selected as the test index, and the calculation formula is: ; wherein, is the number of instances correctly predicted by the model as positive samples, is the number of instances where the model incorrectly predicted negative samples as positive samples.

[0050] The trained neural network model is deployed in the system to detect and identify the actual oral images of patients and output the oral detection results.

[0051] Optionally, the low-light image enhancement is a low-light image enhancement method based on wavelet transform, and the specific steps include: The original image is decomposed using the wavelet transform algorithm. This step usually involves decomposing the image through a low-pass filter (scaling function) and a high-pass filter (wavelet function) to obtain sub-band images of different frequency bands. Specifically, wavelet decomposition can decompose the image into low-frequency sub-bands ( ) and three high-frequency sub-bands ( 、 、 ), representing different frequency components of the image; Nonlinear enhancement: Perform nonlinear enhancement processing on the decomposed sub-band image to improve the contrast of the image and more effectively control the signal-to-noise ratio. Specific methods include adjusting the grayscale value of the sub-band image and applying nonlinear functions; Detail extraction and enhancement, for high frequency sub-bands ( 、 、 ), further detail extraction and enhancement can be performed; Reconstruct the image. After completing the processing of the sub-band image, the inverse wavelet transform algorithm is used to reconstruct the processed sub-band image into the enhanced image. The inverse wavelet transform is the inverse process of the wavelet transform. It generates the final enhanced image by merging the processed sub-band images.

[0052] Example 2

[0053] In one embodiment, based on the oral condition test results, the dental nurse generates a preliminary test analysis report using an intelligent analysis report generation method. The specific steps include: First, a medical-grade oral hygiene dataset was constructed based on prior expert experience. The dataset includes annotated oral images and their corresponding expert text description labels. The expert text description labels are the expert diagnostic opinions given by oral medicine experts after observing and studying oral images. The dataset is divided into training and test sets according to the ratio. Secondly, a conditional generative adversarial network model is constructed. First, a generator model is defined. The generator network is based on a convolutional neural network architecture. The input includes random noise, image features, and image condition information. The image condition information is the parameter deviation between the three-dimensional visualized oral model obtained after modeling in step S17 and the standard oral three-dimensional model. The goal of the generator is to generate a corresponding text description based on the input image features. The output of the generator is a text sequence, which can be characters, words, or sentences. Then, a discriminator network is designed. The discriminator network is also based on a convolutional neural network model. The input of the discriminator network includes the text description of the expert diagnosis and the corresponding image features. The goal of the discriminator is to determine whether the input text matches the image features, i.e., whether the text is real or generated. The output of the discriminator is a probability value representing the degree of matching between the input text and the image features. Furthermore, the parameters of the generator and the discriminator are initialized. Before the start of training, the parameters of the generator and the discriminator are initialized using a random initialization method. Then, the generator network is trained. A batch of images and corresponding real texts are extracted from the training set. These images and real texts are used as the input of the discriminator, and the probability value output by the discriminator is calculated. The cross-entropy loss function is used to measure the difference between the probability value output by the discriminator and the true label (i.e., whether the text is real). The parameters of the discriminator are updated through the backpropagation algorithm to minimize the loss function. At the same time, the discriminator network is trained. A batch of images are extracted from the training set, and the corresponding texts are generated using the generator. These generated texts and corresponding images are used as the input of the discriminator, and the probability value output by the discriminator is calculated. The cross-entropy loss function is used to measure the difference between the generated text and the real text (indirectly measured through the output of the discriminator). The parameters of the generator are updated through the backpropagation algorithm to maximize the probability of the discriminator making a mistake in judging the generated text (i.e., making the discriminator think that the generated text is real). Then, the parameters are updated in a loop. During the training process, the parameters of the generator are updated once every few times the parameters of the discriminator are updated. Through this adversarial training method, the performance of the generator and the discriminator will continuously improve. This loop continues until the model converges. Finally, after the model training converges, the conditional generative adversarial network model is evaluated on the test set using a combination of human evaluation and objective indicator evaluation. The neural network model weight parameters with excellent performance are selected for deployment and use, and the preliminary detection and analysis report of the patient's oral condition is output.

[0054] Optionally, the human evaluation method invites experts to score or evaluate the generated text. The objective evaluation indicator selects the BLEU text generation evaluation indicator to evaluate the quality of the generated text.

[0055] Example 3

[0056] In one embodiment, the dental nurse generates a medical-grade oral cleaning plan based on the intelligent personalized cleaning plan generation method, generates a reasonable and personalized medical-grade oral cleaning plan under the assistance of a generative AI large model, according to the key information input by the user, and guides the dental nurse to perform effective and safe oral cleaning operations on the patient's oral cavity. The specific steps include:

[0057] According to the actual needs, select the appropriate language generation model for local deployment, download the selected language generation model from the official website or model library, and after downloading and installing, verify the integrity and correctness of the model to ensure that the model can run normally; Source deployment, configure the development environment according to the model running needs, copy the source code and dependencies of the model to the local development environment, and compile and install according to the instructions; after configuring the parameters and configuration files required by the model, start the model service; Application deployment, install and configure according to the tool instructions, import the model file into the tool, start the service of the tool, and interact with the model through the provided interface or interface; Function test, test whether the basic functions of the model in the local environment are normal, including input and output, response speed; Performance testing, evaluate the performance of the model in the local environment, including processing speed, resource occupation; Optimization, according to the test results, optimize the model, including adjusting parameters, optimizing data processing flow; Continuous monitoring, after the model is deployed, continuously monitor the running situation and performance indicators of the model, and timely find and solve problems; Permission settings, ensure that only authorized users can access and operate the model service; Data encryption, encrypt sensitive data to ensure data security; Log recording, record the running log and abnormal information of the model, which is convenient for subsequent analysis and troubleshooting; Large model fine-tuning, use the fine-tuning tool to load the selected pre-trained model, keep the model architecture unchanged during the loading process, and only adjust part of the parameters or layers, Set training hyperparameters, including learning rate, batch size, and training epoch; Preprocess the input data, which includes real oral cavity photos, key information, and corresponding oral cleaning plans formulated by experts. Real oral cavity photos and key information are inputs of the large model, and oral cleaning plans formulated by experts are outputs of the large model. Preprocess the above data through data cleaning, normalization, and encoding preprocessing; The pre-processed data is input into the large model for small algebra fine-tuning training; The fine-tuned large model is deployed to the system for use, and the user inputs the key information of the oral state, and a customized oral cleaning scheme can be automatically generated.

[0058] Although preferred embodiments of the application have been described herein, changes and modifications can be suggested to one skilled in the art, and it is intended that the application encompass such changes and modifications as fall within the scope of the appended claims.

[0059] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A medical-grade oral cleaning method based on AI assistance, including an intelligent oral status recognition and analysis method and an intelligent personalized cleaning plan generation method, characterized in that: The oral condition intelligent detection and analysis method combines oral visual sensors and artificial intelligence algorithms to accurately detect and identify the tooth arrangement, tooth growth, tooth spacing, dental plaque distribution, dental calculus distribution and gum condition in the oral cavity, and provides a dental condition analysis and evaluation report based on the identification results. Specifically, it includes a fine-grained oral condition detection method and an intelligent analysis report generation method. The fine-grained oral state detection method is based on a fine-grained image recognition neural network model to identify and locate oral state images captured by an oral image acquisition system, accurately identifying tiny dental plaque and calculus, and providing guidance and reference for oral cleaning. The specific steps of the fine-grained oral state detection method include: S10: Use the oral image acquisition system to capture the patient's oral status information from multiple angles and in all directions, mainly including the status of teeth and gums; S11: Performing image space transformation on the acquired image, including geometric transformation operations such as translation, transposition, mirroring, rotation, and scaling, to correct the systematic error of the image acquisition system and the random error of the instrument position; S12: When performing image space transformation, a grayscale interpolation algorithm is used to process the transformed image to prevent pixels of the output image from being mapped to non-integer coordinates of the input image; S13: performing a low-light image enhancement operation on the image after the image space transformation to eliminate image shadows, increase brightness, and enhance image visibility and intelligibility; S14: performing restoration operations on the enhanced image, including eliminating noise, deviation, and shadow interference; S15: Performing image registration on the images processed in step S14 to calibrate the images taken at different times and angles to the same coordinate system, so as to facilitate feature recognition and processing by the subsequent neural network; S16: Performing three-dimensional oral cavity modeling based on the two-dimensional plane image processed in step S15 to restore a realistic, three-dimensional visualized oral cavity state model; S17: The modeled 3D visualization oral model is compared with the standard 3D oral model. The mean absolute error (MAE) is used to measure the difference between the modeled oral model and the standard oral model. The comparison parameters include the number of teeth, tooth spacing, tooth length, tooth width, and tooth curvature. The MEA calculation formula is: ;in, i Number the teeth, The first i teeth, The first i teeth; S18: Annotating the image processed in step S15 using image annotation software, including using different colored annotation frames to mark the locations of dental plaque, dental calculus, and gingival redness and swelling, and providing corresponding text annotation information. On this basis, a training dataset and a test dataset are constructed in a ratio of 7:

3. The training dataset includes the annotated oral image data, and the test dataset includes the unannotated image data. S19: Build a fine-grained image detection neural network model based on the improved YOLOv8. Based on the YOLOv8 neural network model, use pyramid convolution layers instead of ordinary convolution layers to fuse features of different scales to improve the model's ability to detect multi-scale targets. Add a multi-head attention mechanism to the network to enhance the model's ability to capture fine-grained features. Add a fine-grained classifier to further improve the accuracy of recognition and classification of small, fine-grained targets such as tartar, plaque, and gingival redness. S20: The neural network constructed in step S19 is tested on the training data set constructed in step S18. The Adam optimizer is selected, and the loss function is designed as a multi-dimensional loss function combining classification and positioning, including classification loss and positioning loss. The classification loss uses the cross entropy loss function to measure the accuracy of the neural network model in target classification. The calculation formula is: ;in, represents the true label of the sample, Represents the label predicted by the model. y = 1, it means the sample belongs to the positive class; when y = 0, indicating that the sample belongs to the negative class; The positioning loss is the Smooth L1 loss function, which is used to measure the accuracy of the model in target positioning. The calculation formula is: ;in, x It is the coordinate difference between the predicted bounding box and the true bounding box. The classification loss function and the positioning loss function are combined as the optimization objective function for neural network model training; S21: Test the neural network model trained in step 20 on the test set. The test indicator is the accuracy indicator, and the calculation formula is: ;in, is the number of instances correctly predicted by the model as positive samples, is the number of instances where the model incorrectly predicted negative samples as positive samples.

2. The AI-assisted medical-grade oral cleaning method according to claim 1, characterized in that: The specific steps of the intelligent analysis report generation method include: S30: Construct a medical-grade oral hygiene dataset based on prior expert experience. The dataset includes annotated oral image data and corresponding expert text description labels. The expert text description labels are expert diagnostic opinions given by oral medicine experts after observing and studying oral images. The dataset is divided into a training set and a test set according to the ratio. S31: Construct a conditional generative adversarial network model. First, define a generator model. The generator network is based on a convolutional neural network architecture. The input is random noise, image features, and image condition information. The image condition information is the parameter deviation between the modeled 3D visualization oral model and the standard oral 3D model calculated in step S17. The goal of the generator is to generate a corresponding text description based on the input image features. The output of the generator is a text sequence, which can be a character, word, or sentence. S32: Design a discriminator network. This network is also based on a convolutional neural network model. The input to the discriminator network is the text description of the expert's diagnosis and its corresponding image features. The discriminator's goal is to determine whether the input text matches the image features, that is, to determine whether the text is real or generated. The output of the discriminator is a probability value, indicating the degree of match between the input text and the image features. S33: Initialize the parameters of the generator and discriminator. Before training begins, use the random initialization method to initialize the parameters of the generator and discriminator. S34: Train the generator network and extract a batch of images and corresponding real text from the training set; use these images and real text as input to the discriminator and calculate the probability value of the discriminator output; use the cross-entropy loss function to measure the difference between the probability value output by the discriminator and the true label (i.e., whether the text is real); update the discriminator parameters through the backpropagation algorithm to minimize the loss function; S35: Train the discriminator network, extract a batch of images from the training set, and use the generator to generate corresponding text; use these generated texts and corresponding images as input to the discriminator, and calculate the probability value of the discriminator output; use the cross-entropy loss function to measure the difference between the text generated by the generator and the real text; update the parameters of the generator through the backpropagation algorithm to maximize the probability of the discriminator misjudging the generated text; S36: Cyclic update. During the training process, after updating the discriminator parameters several times, the generator parameters are updated again. Through this adversarial training method, the performance of the generator and discriminator will continue to improve. This cycle continues until the model reaches convergence. S37: After the model training reaches convergence, the performance index evaluation of the conditional generative adversarial network model is completed on the test set, and the weight parameters of the neural network model with excellent performance are selected for deployment.

3. The AI-assisted medical-grade oral cleaning method according to claim 1, characterized in that: The specific steps of the intelligent personalized cleaning solution include: S40: Select the appropriate language based on actual needs to generate a large model for local deployment. Download the selected language from the official website or model library. After downloading and installing, verify the integrity and correctness of the model to ensure that it can operate normally. S41: Source code deployment: Configure the development environment according to the model's running requirements, copy the model's source code and dependencies to the local development environment, and compile and install them according to the instructions. After configuring the parameters and configuration files required by the model, start the model service. S42: Application deployment: Install and configure the tool according to its instructions, import the model file into the tool, start the tool's services, and interact with the model through the provided interface or screen; S43: Functional testing, testing whether the basic functions of the model in the local environment are normal, including input and output, and response speed; S44: Performance testing, evaluating the performance of the model in the local environment, including processing speed and resource usage; S45: Optimization: Optimize the model based on the test results, including adjusting parameters and optimizing data processing procedures; S46: Continuous monitoring: After the model is deployed, the model's operation and performance indicators are continuously monitored to detect and resolve problems in a timely manner. S47: Permission settings ensure that only authorized users can access and operate model services; S48: Data encryption, encrypt sensitive data to ensure data security; S49: Logging, recording the model's operation log and exception information to facilitate subsequent analysis and troubleshooting; S50: Fine-tuning a large model. Use the fine-tuning tool to load the selected pre-trained model. During the loading process, keep the model architecture unchanged and only adjust some parameters or layers. S51: Set training hyperparameters, including learning rate, batch size, and training generations; S52: Preprocessing input data. The input data includes real oral images, key information, and corresponding oral hygiene plans developed by experts. The real oral images and key information are inputs to the large model, and the oral hygiene plans developed by experts are outputs of the large model. The above data is preprocessed by data cleaning, normalization, and encoding. S53: Input the preprocessed data into the large model for small algebraic fine-tuning training; S54: The fine-tuned large model is deployed to the system for use. The user inputs key information about the oral condition, and a customized oral cleaning plan can be automatically generated.

4. An AI-assisted medical-grade oral cleaning device, characterized in that: An AI-assisted medical-grade oral cleaning method according to any one of claims 1 to 3, the device comprising: Oral image acquisition system: The oral image acquisition system consists of a high-definition rotatable oral camera, which can achieve 360-degree all-round oral photography as needed. Video recording and image capture can be selected as needed, providing an intuitive data basis for subsequent oral status identification and analysis.

5. The AI-assisted medical-grade oral cleaning device according to claim 4, characterized in that: The oral cleaning device further comprises: Image processing subsystem: The image processing subsystem carries the image preprocessing algorithm, performs preprocessing operations on the oral images captured by the high-definition oral camera, and obtains high-definition and effective oral images; Imaging display: The imaging display uses a liquid crystal display as a carrier and can display the captured oral images in high definition.

6. The AI-assisted medical-grade oral cleaning device according to claim 4, characterized in that: The oral cleaning device further comprises: Report generation subsystem: The report generation subsystem has built-in intelligent analysis report generation methods and intelligent personalized cleaning plan generation methods. It is installed on high-performance computer equipment. Users can operate and interact with the report generation subsystem through the display, mouse, keyboard and host computer function interface; Oral cleaning subsystem: The oral cleaning subsystem is equipped with an intelligent robotic arm and sensing equipment. The front end of the robotic arm is equipped with cleaning instruments required for teeth cleaning, which can effectively clean and protect the oral cavity according to needs.

Images