Oral disease auxiliary evaluation system based on deep learning

Abstract

The invention discloses an oral disease auxiliary evaluation system based on deep learning, which relates to the technical field of oral disease evaluation, and comprises a data preprocessing module, a feature extraction module, a model training and regularization module, a reinforcement learning strategy module and a diagnosis output and feedback module, oral cavity image data from different sources are collected and preprocessed, the data quality and diversity are improved, and the imbalance problem in original data is solved. Through multi-data-source integrated learning and a data enhancement strategy, the model can learn more features from diversified data, the deviation caused by data singleness and imbalance is reduced, and the adaptability to different devices, conditions and patients is improved. In combination with cross validation and an early stop mechanism, overfitting is effectively avoided, use of computing resources is optimized, and it is ensured that the training process is efficient and stable. According to the methods, the generalization ability and accuracy of the model are improved, so that the model is more reliable in actual clinic.

Description

Technical Field

[0001] This invention relates to the field of oral disease assessment technology, and more specifically to an oral disease auxiliary assessment system based on deep learning. Background Technology

[0002] Deep learning-based assisted assessment of oral diseases refers to the use of deep learning technology to automatically identify, diagnose, and assess the state of oral diseases through the analysis of oral images (such as X-rays, CT scans, and dental photographs). Deep learning models, especially convolutional neural networks (CNNs), can learn disease characteristics from large amounts of data and identify minute lesions or abnormal areas in image data. By analyzing this data, the system can provide dentists with accurate assisted diagnosis and treatment suggestions, helping to improve diagnostic efficiency, reduce misdiagnosis rates, and provide patients with more precise treatment plans. The application of this technology in oral medicine can not only improve the quality of clinical diagnosis and treatment but also promote the development of intelligent medical devices and the advancement of oral health management.

[0003] The existing technology has the following shortcomings:

[0004] In existing technologies, deep learning-based oral disease assessment systems may face overfitting, leading to poor model performance on new data. Overfitting refers to a model focusing excessively on details and noise in the training data during training, thus losing its ability to adapt to new or unseen data. Specifically, when a model only memorizes specific features of the training samples during training, ignoring general patterns in the data, it may incorrectly identify lesion features, resulting in inaccurate diagnoses. This is particularly problematic when faced with oral imaging data from diverse sources and of varying quality; overfitting can prevent the model from accurately identifying actual diseases, potentially leading to incorrect diagnoses or overlooking potential problems by physicians. For example, the model might incorrectly diagnose a disease in a healthy area or ignore an actual lesion, thus affecting the patient's treatment decisions. Therefore, ensuring the model has good generalization ability and can adapt to diverse clinical data is a key challenge for deep learning technology in oral disease assessment.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based auxiliary assessment system for oral diseases. Through integrated learning from multiple data sources and data augmentation strategies, the model can learn more features from diverse data, reducing biases caused by data homogeneity and imbalance, and enhancing adaptability to different devices, conditions, and patient groups. Combining cross-validation and early stopping mechanisms avoids overfitting and optimizes the use of computational resources, ensuring the efficiency and stability of the training process. These strategies improve the model's generalization ability and accuracy, making it more reliable and efficient in real-world clinical settings, especially when processing large-scale oral imaging data, thus addressing the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based auxiliary assessment system for oral diseases, comprising a data preprocessing module, a feature extraction module, a model training and regularization module, a reinforcement learning strategy module, and a diagnostic output and feedback module.

[0008] The data preprocessing module collects and preprocesses oral imaging data from different sources to improve data quality and diversity, and to address imbalances in the raw data.

[0009] The feature extraction module uses a convolutional neural network to extract local features from oral images. It uses multi-layer convolution and pooling operations to generate multi-scale feature representations, capturing minute lesions and complex structures of oral diseases.

[0010] The model training and regularization module uses cross-validation and regularization techniques to train deep learning models, avoid overfitting, and ensure that the model has good generalization ability to unseen data.

[0011] The reinforcement learning strategy module introduces reinforcement learning algorithms and adjusts the learning strategy based on the diagnostic feedback of the model, enabling the model to adapt to the characteristics of different oral diseases during the training process.

[0012] The diagnostic output and feedback module outputs diagnostic results through an adaptive convolutional neural network and backpropagation algorithm, and adjusts diagnostic recommendations based on the confidence level of the model predictions to further improve the accuracy and reliability of the assessment.

[0013] Preferably, the specific steps for collecting and preprocessing oral imaging data from different sources to improve data quality and diversity and address imbalances in the original data are as follows:

[0014] First, oral imaging data were collected from multiple data sources;

[0015] To ensure consistency of data from different sources, all images are standardized.

[0016] To reduce the interference of noise on model training, denoising processing is performed;

[0017] By employing data resampling technology, the problem of data imbalance in centralized image data processing can be addressed.

[0018] Preferably, the specific steps for extracting local features from oral images using convolutional neural networks, and generating multi-scale feature representations by employing multi-layer convolution and pooling operations to capture minute lesions and complex structures in oral diseases are as follows:

[0019] After data preprocessing is completed, the first step is to design the structure for the convolutional neural network;

[0020] In the first and subsequent layers of the convolutional neural network, local features in oral images are extracted through convolution operations.

[0021] After the convolution operation extracts local features, the pooling operation is used to further reduce the spatial dimension of the data, thereby reducing the amount of computation and avoiding overfitting;

[0022] After multiple layers of convolution and pooling operations, a convolutional neural network will generate multiple feature maps.

[0023] Preferably, the specific steps for training deep learning models using cross-validation and regularization techniques to avoid overfitting and ensure good generalization ability of the model to unseen data are as follows:

[0024] After feature extraction, the preprocessed data is divided and cross-validated.

[0025] In the process of model training, regularization techniques are introduced to avoid overfitting of deep neural networks when processing oral images;

[0026] During training, cross-validation and regularization are used while the model's hyperparameters are tuned.

[0027] To ensure that the model does not overfit on the training set during training, an early stopping mechanism is introduced.

[0028] Preferably, the specific steps for adjusting the learning strategy based on the diagnostic feedback of the model by introducing a reinforcement learning algorithm, so that the model can adapt to the characteristics of different oral diseases during the training process, are as follows:

[0029] In reinforcement learning, the state and action spaces are first defined, and the model learns the optimal policy through a feedback mechanism. The calculation expression is as follows:

[0030] In the formula, It is the input oral imaging data. It represents the confidence level of the current diagnostic result. It is the first The state at any given moment;

[0031] In reinforcement learning, the reward function is used to provide feedback on the quality of each action. The model adjusts its strategy based on the reward to make accurate decisions in future diagnostic processes. The reward function is defined as follows:

[0032] In the formula, It is the model's predicted action. It is a genuine diagnosis. It is an indicator function; if the model's predicted action... With the real diagnosis If they match, return 1; otherwise, return 0. It is a penalty function for diagnosing errors. It is a moderating factor for positive rewards. It is a negative penalty adjustment factor used to control the intensity of the penalty when the model makes a wrong diagnosis. It is a reward function;

[0033] By introducing reinforcement learning Learning, the model adjusts its behavioral strategy based on the current reward signal. During learning, the model updates its value function. The value update formula is as follows:

[0034] In the formula, The current state and actions The corresponding value function represents the expected return obtained after choosing this action. It's the learning rate, which controls the amount of time spent updating. The adjustment range of the value, It is a discount factor. In the next state Choose the largest The optimal action corresponding to the value.

[0035] Preferably, the specific steps for outputting diagnostic results through an adaptive convolutional neural network and backpropagation algorithm, and adjusting diagnostic recommendations based on the confidence level of the model predictions to further improve the accuracy and reliability of the assessment are as follows:

[0036] First, after multiple convolutional and pooling operations on the oral cavity image, the final layer of the network outputs the diagnostic result, which is presented in the form of classification probabilities. To calculate the probability value for each category, the network uses... Normalizing the function yields a probability distribution. The function converts the score into a probability value, and the calculation expression is as follows:

[0037] In the formula, It is the first The predicted probability of oral disease types, i.e., given input data In this case, the model predicts it belongs to the type of oral disease. The probability, Is the network for the first Types of oral diseases The output score, Is the network for the first Types of oral diseases The output score, It is the natural base. It represents the total number of types of oral diseases. It is the sum of scores from all categories.

[0038] After obtaining the diagnostic results and calculating the predicted probability of each oral disease type, the weights of the convolutional neural network are adjusted using the backpropagation algorithm to minimize the loss function, the calculation formula of which is as follows:

[0039] In the formula, It is the first The true label value of each category, It is the loss function value, which measures the difference between the model's predicted value and the true label;

[0040] According to the loss function The gradient of each parameter is calculated, and then an optimization algorithm is used to update the weights in the network. In this way, the network is gradually optimized so that the predicted disease category probability gets closer and closer to the actual label.

[0041] The weights in the network are updated using an optimization algorithm. The updated weights are adjusted according to the gradient descent method, and the update formula is as follows:

[0042] In the formula, This is the updated weight. It is the weight of the previous step. The loss function with respect to the weights The gradient;

[0043] Finally, after backpropagation is complete and the model weights are updated, the model outputs the final diagnostic result based on the predicted probability of each category. To enhance the accuracy and reliability of the assessment, the final diagnostic recommendation is adjusted according to the model's confidence level. For each disease category, a threshold is introduced to determine whether further investigation is recommended. The adjustment formula is as follows:

[0044] In the formula, This is the final prediction result. It is a preset confidence threshold.

[0045] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0046] This invention employs integrated learning from multiple data sources and data augmentation strategies, enabling the model to learn more features from diverse data and reducing biases caused by data uniformity or imbalance. In this context, the model can not only identify common oral diseases but also effectively handle oral imaging data from different devices, imaging conditions, and patient groups. Through data augmentation and regularization methods, especially Dropout and L2 regularization, the model avoids over-memorizing training data, enhancing its predictive ability for new data. Therefore, the model demonstrates strong adaptability and accuracy regardless of different hospitals, diagnostic environments, or variations in patient oral structures. This effect is particularly important in practical clinical applications because oral imaging data often exhibits significant heterogeneity; ensuring the model's strong generalization ability greatly improves its reliability and practicality in real-world scenarios.

[0047] This invention effectively alleviates overfitting during training by introducing cross-validation and early stopping mechanisms. Cross-validation allows the model to undergo multiple rounds of training and validation on different subsets, enabling a comprehensive evaluation of model performance and ensuring its stability. The early stopping mechanism automatically halts training based on the performance on the validation set, preventing the model from overlearning noise on the training data and ensuring optimal performance on the validation set. These techniques not only reduce overfitting during training but also optimize computational resource usage, as training can be terminated when validation set performance no longer improves, thus avoiding unnecessary computational overhead. Therefore, the entire training process becomes more efficient and stable, improving the model's predictive accuracy and reducing the waste of computational resources from ineffective training. This effect is particularly important for processing large-scale oral imaging data, significantly shortening training time and improving model productivity. Attached Figure Description

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

[0049] Figure 1 This is a schematic diagram of the modules of the deep learning-based oral disease auxiliary assessment system of the present invention. Detailed Implementation

[0050] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0051] This invention provides, for example Figure 1 The deep learning-based oral disease auxiliary assessment system shown includes a data preprocessing module, a feature extraction module, a model training and regularization module, a reinforcement learning strategy module, and a diagnostic output and feedback module.

[0052] The data preprocessing module collects and preprocesses oral imaging data from different sources to improve data quality and diversity, and to address imbalances in the raw data.

[0053] The specific steps for collecting and preprocessing oral imaging data from different sources to improve data quality and diversity, and to address imbalances in the raw data are as follows:

[0054] First, oral imaging data were collected from multiple data sources;

[0055] Data sources include, but are not limited to, hospital clinics, public imaging databases, and clinical diagnostic records. Data sets from various sources can include X-rays, oral CT scans, and dental photographs. These datasets should be diverse, covering patients of different ages, genders, ethnicities, medical histories, and lesion types to ensure the model can learn comprehensive oral disease characteristics. Furthermore, the resolution and quality of the images must meet training requirements to avoid image quality issues affecting subsequent training and evaluation. Integrating diverse data sources can effectively improve the model's adaptability to different oral lesions and individual differences, thereby enhancing the model's generalization ability.

[0056] To ensure consistency of data from different sources, all images are standardized.

[0057] Standardization includes unifying image size, resolution, and color space (e.g., converting RGB to grayscale) to ensure all input data can be processed at the same scale. Simultaneously, image enhancement techniques (such as rotation, scaling, translation, cropping, mirroring, and brightness adjustment) can artificially generate more variant samples. These enhancements not only help expand the dataset and increase the variety of data the model has encountered, but also improve the model's robustness to different poses, angles, and lighting conditions, preventing overfitting on a specific dataset and ensuring its adaptability to a wider range of clinical environments.

[0058] To reduce the interference of noise on model training, denoising processing is performed;

[0059] Common denoising methods include Gaussian filtering, median filtering, and bilateral filtering. These methods effectively remove random noise from images while preserving edge information. Furthermore, edge detection algorithms (such as Canny edge detection) can be used to further highlight important features in the image. The purpose of image preprocessing is to ensure that the data has sufficient clarity and quality before entering the deep learning model, enabling the model to accurately extract valuable features and reduce errors caused by image quality issues.

[0060] By employing data resampling technology, the problem of data imbalance in centralized image data processing can be addressed.

[0061] In oral disease image datasets, there is often a surplus of samples for certain categories (such as oral cancer and dental caries) and a shortage of samples for some rarer diseases (such as certain types of gingivitis or rare oral lesions). This data imbalance can cause the model to favor the more numerous categories during training, reducing its diagnostic ability for rare lesions. To address this issue, data resampling techniques can be employed, such as oversampling (increasing the number of samples for rare categories) or undersampling (reducing the number of samples for common categories), or by synthesizing minority category samples (e.g., using generative adversarial networks (GANs) to generate images of rare lesions). Furthermore, a class-weighted loss function can be introduced to increase the model's learning weight for rare categories, thereby making the model pay more attention to minority categories during training and resolving the bias caused by data imbalance.

[0062] The feature extraction module uses a convolutional neural network to extract local features from oral images. It uses multi-layer convolution and pooling operations to generate multi-scale feature representations, capturing minute lesions and complex structures of oral diseases.

[0063] The specific steps for extracting local features from oral images using convolutional neural networks, and generating multi-scale feature representations by employing multi-layer convolution and pooling operations to capture minute lesions and complex structures in oral diseases are as follows:

[0064] After data preprocessing (such as standardization, data augmentation, and denoising), the first step is to design the structure for the convolutional neural network.

[0065] When constructing a network, considering the complexity of oral images, the network structure typically includes multiple convolutional layers, pooling layers, and fully connected layers. The role of convolutional layers is to extract local features from oral images, while pooling layers help reduce feature dimensionality, improve computational efficiency, and avoid overfitting. When designing the initial convolutional layers, it's crucial to ensure the network can handle oral images of different sizes and scales, using appropriate convolutional kernel sizes to capture lesion regions of varying sizes. At this point, data preprocessing steps (such as data augmentation and denoising) have improved data diversity and quality, enabling the network to better handle images from different lesions, lighting conditions, and angles.

[0066] In the first and subsequent layers of the convolutional neural network, local features in oral images are extracted through convolution operations.

[0067] Convolutional layers perform convolution operations on the input image using multiple filters (convolutional kernels) to generate feature maps. Each convolutional kernel can learn different local features in the image, such as the shape of teeth and the edges of lesion areas. In oral imaging, local features may include the outline of teeth, gingival abnormalities, and minute lesions such as gaps between teeth. These details are crucial for identifying oral diseases, especially for some minute lesions (such as early caries or periodontal disease), which require high-resolution feature extraction. Through multi-layer convolutional networks, multi-level features from local to global can be captured, and through data augmentation techniques (such as rotation and scaling), the model has a strong ability to identify lesion areas at different angles and scales.

[0068] After the convolution operation extracts local features, the pooling operation is used to further reduce the spatial dimension of the data, thereby reducing the amount of computation and avoiding overfitting;

[0069] Pooling layers typically use max pooling or average pooling to downsample each feature map. Max pooling retains the maximum response value in the feature map, thus preserving key feature information while discarding irrelevant details. In this way, the network can extract multi-scale feature representations, especially for complex structures in oral diseases (such as malocclusion and swollen gums), where pooling layers help abstract scale-invariant features. Because the image size is smaller after pooling, the model's computational efficiency is greatly improved. Furthermore, due to the extraction of multi-scale features, the model has better adaptability to lesion regions of different sizes, further enhancing its ability to capture minute lesions and complex oral structures.

[0070] After multiple layers of convolution and pooling operations, a convolutional neural network will generate multiple feature maps.

[0071] These feature maps contain various information from oral images, such as tooth surfaces, gums, and lesion areas. To comprehensively analyze this information, fully connected layers are typically added after the network to flatten and fuse all extracted feature maps. The role of the fully connected layers is to learn weights and map these features to a higher-level feature space, ultimately outputting disease predictions. At this stage, the model has generated multi-level, multi-scale feature representations through data preprocessing, convolution operations, and pooling operations. These features can fully describe various lesions and structures in oral images. Because the data preprocessing stage addresses the data imbalance problem (e.g., through oversampling of minority classes or weighted loss functions), the network's learning can better capture the features of rare lesions and avoid overfitting to common lesions.

[0072] The model training and regularization module uses cross-validation and regularization techniques to train deep learning models, avoid overfitting, and ensure that the model has good generalization ability to unseen data.

[0073] The specific steps for training a deep learning model using cross-validation and regularization techniques to avoid overfitting and ensure good generalization ability on unseen data are as follows:

[0074] After feature extraction, the preprocessed data is divided and cross-validated.

[0075] Cross-validation involves dividing the entire dataset into multiple subsets (typically k subsets), training and validating the model on each subset, and repeating the training process k times to ensure the model's stability across different datasets. A common cross-validation method is K-fold cross-validation, where K is typically set to 5 or 10. Each round of training uses k-1 subsets for training and the remaining subset for validation. The model's generalization ability is then evaluated by averaging the validation results. Through cross-validation, the model avoids overfitting to specific training data, ensuring good performance across different dataset subsets and improving its predictive ability on unseen data.

[0076] In the process of model training, regularization techniques are introduced to avoid overfitting of deep neural networks when processing oral images;

[0077] Common regularization methods include L2 regularization and Dropout. In L2 regularization, a regularization term is added to the loss function to minimize the value of the weight matrix, thereby reducing the model's over-reliance on specific features. This effectively constrains model complexity and avoids overfitting. In Dropout, some neurons in the neural network are randomly "dropped" during each training iteration; their outputs are forced to zero. This reduces the network's over-reliance on specific neurons or features, encouraging the network to learn more generalized features. By introducing regularization techniques, the trained model does not merely remember noise and details from the training data but learns more representative and general features, thus improving its prediction accuracy on unseen data.

[0078] During training, cross-validation and regularization are used while the model's hyperparameters are tuned.

[0079] Examples of suitable hyperparameters include learning rate, batch size, number of network layers, and kernel size. To achieve optimal generalization ability, grid search or random search is typically used to select appropriate combinations of hyperparameters. Grid search exhaustively searches all possible combinations to find the optimal parameter values, while random search randomly selects some parameter combinations, usually finding a better result in a shorter time. Furthermore, choosing a suitable optimization algorithm (such as Adam, SGD, etc.) is crucial to the model's training performance. The Adam optimization algorithm combines the advantages of Momentum and RMSprop, adaptively adjusting the learning rate to achieve faster convergence during training, reduce the risk of overfitting, and help the model perform stably on different oral imaging data.

[0080] During training, an early stopping mechanism is introduced to ensure that the model does not overfit the training set;

[0081] Early stopping is a technique to prevent overfitting. It monitors the performance on the validation set during training. If the loss on the validation set no longer improves within several training epochs, training stops, preventing the model from learning excessive noise on the training set. Early stopping effectively avoids overfitting on the training set while maintaining good performance on the validation set. Furthermore, various evaluation metrics such as F1 score, ROC curve, and confusion matrix can be used to quantitatively assess the model's generalization ability, ensuring that the model not only performs well on training data but also adapts to new and unseen data.

[0082] The reinforcement learning strategy module introduces reinforcement learning algorithms and adjusts the learning strategy based on the diagnostic feedback of the model, enabling the model to adapt to the characteristics of different oral diseases during the training process.

[0083] The specific steps for adjusting the learning strategy based on the model's diagnostic feedback by introducing reinforcement learning algorithms, enabling the model to adapt to the characteristics of different oral diseases during training, are as follows:

[0084] In reinforcement learning, state and action spaces are first defined, allowing the model to learn the optimal strategy through a feedback mechanism. In oral disease diagnosis, the state typically represents the model's current diagnostic result for oral images and its corresponding confidence level, while the action is the decision the model makes for each diagnosis. For example, the model may need to determine whether a lesion exists in a certain area; the action could be "diagnose as normal" or "diagnose as lesion," calculated as follows:

[0085] In the formula, It is the input oral imaging data (such as features extracted through a convolutional neural network). It represents the confidence level of the current diagnosis (based on the output probability predicted by the model). It is the first The state at any given moment indicates the current image. and its corresponding confidence level ;

[0086] state It includes feature representations of the image and diagnostic confidence from the model, providing a foundation for subsequent decision-making. At this point, the model's goal is to base its decisions on the current state. Choose the optimal action .

[0087] In reinforcement learning, the reward function is used to provide feedback on the quality of each action. The model adjusts its strategy based on the reward to make accurate decisions in future diagnostic processes. The design of the reward function depends on the accuracy of the model's diagnosis. For example, if the model's diagnosis is correct, the reward is positive; if the diagnosis is incorrect, the reward is negative. The reward function is defined as follows:

[0088] In the formula, It is the model's predicted action. It is a genuine diagnosis. It is an indicator function; if the model's predicted action... With the real diagnosis If they match, return 1; otherwise, return 0. This is the penalty function for diagnostic errors, representing a measure of the model's error in the current state (e.g., the diagnostic error is calculated by comparing it to the actual diagnostic labels). It is a moderating factor for positive rewards, used to control the weight of the reward given when the model makes a correct prediction. Its function is to determine the strength of the reward signal when the model makes a correct diagnosis. It is a negative penalty adjustment factor used to control the intensity of the penalty when the model makes a wrong diagnosis. It is a reward function;

[0089] The reward function provides feedback based on the model's performance, encouraging it to make more positive decisions in correct diagnostic tasks and reduce erroneous predictions. Through this reward signal, the model gradually adjusts its strategy, enabling it to make more accurate diagnoses for each type of oral disease.

[0090] By introducing reinforcement learning Learning, the model adjusts its behavioral strategy based on the current reward signal. During learning, the model updates its value function. The value update formula is as follows:

[0091] In the formula, The current state and actions The corresponding value function represents the expected return obtained after choosing this action. It's the learning rate, which controls the amount of time spent updating. The adjustment range of the value, It is a discount factor, indicating the importance of future rewards. In the next state Choose the largest The optimal action corresponding to the value.

[0092] In this formula, the model adjusts the value based on the reward and the discount value of the expected future reward each time. The values ​​are updated to optimize the decision-making process. This process allows the model to dynamically adjust its learning strategy based on the different characteristics of oral diseases (such as different lesion types and severity), gradually approaching the optimal strategy.

[0093] The diagnostic output and feedback module outputs diagnostic results through an adaptive convolutional neural network and backpropagation algorithm, and adjusts diagnostic recommendations based on the confidence level of the model prediction, thereby further improving the accuracy and reliability of the assessment.

[0094] The specific steps to improve the accuracy and reliability of the assessment by using an adaptive convolutional neural network and backpropagation algorithm to output diagnostic results and adjust diagnostic recommendations based on the confidence level of the model prediction are as follows:

[0095] First, after multiple convolutional and pooling operations on the oral cavity image, the final layer of the network outputs the diagnostic result, which is presented in the form of classification probabilities. To calculate the probability value for each category, the network uses... Normalizing the function yields a probability distribution. The function converts the score into a probability value, and the calculation expression is as follows:

[0096] In the formula, It is the first The predicted probability of oral disease types, i.e., given input data In this case, the model predicts it belongs to the type of oral disease. The probability, Is the network for the first Types of oral diseases The output score, Is the network for the first Types of oral diseases The output score, It is the natural base. It represents the total number of types of oral diseases. It is the sum of scores from all categories, used for normalization.

[0097] pass The model provides a probability for each oral disease category, representing the confidence level of the occurrence of the disease.

[0098] After obtaining the diagnostic results and calculating the predicted probability of each oral disease type, the weights of the convolutional neural network are adjusted using the backpropagation algorithm to minimize the loss function, the calculation formula of which is as follows:

[0099] In the formula, It is the first The true label values ​​for each category (if the actual category is...) ,but ,otherwise , It is the loss function value, which measures the difference between the model's predicted value and the true label;

[0100] According to the loss function The gradient of each parameter (e.g., convolutional kernel weights) is calculated, and then an optimization algorithm (e.g., the Adam optimizer) is used to update the weights in the network. In this way, the network is gradually optimized so that the predicted disease category probability gets closer and closer to the actual label.

[0101] The weights in the network are updated using an optimization algorithm. The updated weights are adjusted according to the gradient descent method, and the update formula is as follows:

[0102] In the formula, This is the updated weight. It is the weight of the previous step. The learning rate determines the step size for each update. The loss function with respect to the weights The gradient;

[0103] Backpropagation involves continuously adjusting network parameters. This is to optimize the model so that it can make better predictions on new and unseen data.

[0104] Finally, after backpropagation is complete and the model weights are updated, the model outputs the final diagnostic result based on the predicted probability of each category. To enhance the accuracy and reliability of the assessment, the final diagnostic recommendation is adjusted according to the model's confidence level. For each disease category, a threshold is introduced to determine whether further investigation is recommended. The adjustment formula is as follows:

[0105] In the formula, This is the final prediction result (1 indicates high risk, 0 indicates low risk). It is a preset confidence threshold.

[0106] Through the steps described above, the model not only provides predicted probabilities for each disease category but also offers diagnostic suggestions to doctors based on actual confidence levels, further improving the accuracy and reliability of the assessment. Ultimately, doctors can make informed treatment decisions based on the model's output and diagnostic recommendations.

[0107] Implementation Method 1: In practical applications, oral imaging data comes from different hospitals, medical equipment, and patient groups. To ensure that the trained deep learning model has good generalization ability and can effectively cope with the problem of data imbalance, it is necessary to adopt an ensemble learning and data augmentation strategy with multiple data sources. The goal of this implementation method is to enhance the model's performance in different clinical scenarios through diverse datasets and rich data augmentation techniques, making it sufficiently robust.

[0108] First, oral imaging data is collected from various hospitals, clinics, and public imaging databases. Because different hospitals may use different equipment, imaging conditions, and techniques, the quality and characteristics of the imaging data can vary significantly. For example, some hospitals may have high-resolution dental X-ray equipment, while others may use lower-quality imaging equipment. This differentiated imaging data enhances the model's ability to learn more different types of lesion features, including some lesion types that might be difficult to capture on a single dataset. Furthermore, patient data from different ages, sexes, ethnicities, and oral health conditions is also necessary to ensure the model demonstrates strong adaptability when dealing with diverse demographics and oral diseases.

[0109] To extract more information from different data sources, ensemble learning can be used to combine multiple data sources, train multiple models, and then fuse them at the end. Common ensemble learning methods include Bagging (such as Random Forest) and Boosting (such as XGBoost). By integrating multiple models, the performance degradation of a single model due to the bias of certain data sources can be avoided. For example, one data source may be better at identifying specific types of lesions, while other data sources may be stronger at identifying other types of lesions. Ensemble learning can fully utilize this complementary information to further improve the overall performance of the model.

[0110] Oral imaging data is typically limited in quantity and highly variable. To broaden the diversity of training data and reduce overfitting, data augmentation techniques can be employed. Data augmentation not only generates more training samples through standardized techniques (such as rotation, flipping, cropping, and scaling), but also simulates different imaging conditions and disease manifestations through complex transformations (such as color changes, lighting adjustments, and noise addition). In oral imaging, lesion areas are usually small; therefore, data augmentation can help the model learn richer features, improving its ability to identify minute lesions.

[0111] For example, when rotating images, multiple rotation angles can be set, allowing the model to observe lesions from different angles. This helps the model identify minute lesions in patients' mouths when faced with real clinical data. Image scaling and cropping can simulate images from different shooting distances, ensuring the model can handle images with varying resolutions. Furthermore, color adjustments and brightness variations help the model adapt to lighting differences caused by different hospital equipment. All these data augmentation methods, combined with ensemble learning from multiple data sources, can significantly improve the model's adaptability and accuracy for various oral diseases.

[0112] In oral imaging datasets, certain diseases (such as dental caries and periodontitis) typically have a large number of samples, while other diseases (such as oral cancer and rare lesions) have fewer samples. Because deep learning models tend to favor the classes with larger sample sizes during training, they may fail to effectively identify minority disease classes, thus affecting the overall diagnostic accuracy. To address the data imbalance problem, oversampling and undersampling methods can be employed, or generative adversarial networks (GANs) can be used to synthesize minority disease samples. These methods allow minority disease samples to be added to the training set, ensuring that the model, during training, not only focuses on common lesions but also fully learns the characteristics of minority lesions.

[0113] In addition, using a weighted loss function is also an effective way to solve the problem of data imbalance. In the loss function, larger weights are assigned to the minority classes of lesions, so that the model pays more attention to these classes during training and avoids overlearning of common classes during training.

[0114] Implementation Method 2: Deep learning models, especially convolutional neural networks (CNNs), are prone to overfitting when processing oral imaging data. Overfitting means that the model performs well on the training set but deteriorates on the test set or new data, failing to accurately identify oral diseases. To avoid this, regularization techniques and training mechanisms that dynamically adjust the learning rate are needed.

[0115] Regularization limits the complexity of a model by adding an extra penalty term to its loss function, preventing it from over-relying on noise or local features in the training data. Common regularization methods used in training deep learning models include L1 regularization, L2 regularization, and Dropout. L2 regularization (also known as weight decay) limits model complexity by penalizing large weight values, preventing the model from over-relying on certain features during learning and thus enhancing its generalization ability.

[0116] Dropout is a technique that randomly deactivates neurons. During each training session, a portion of neurons are randomly "dropped," preventing the network from relying on any particular neuron or path and effectively avoiding overfitting. Dropout is particularly suitable for deep neural networks because it forces the network to learn more general features, enhancing the model's ability to adapt to unseen data.

[0117] The learning rate is a key hyperparameter affecting model training efficiency and accuracy. In deep learning training, employing an appropriate learning rate strategy can accelerate model convergence and prevent getting trapped in local optima. Traditional fixed learning rate strategies often struggle to adapt to the needs of different training stages; therefore, optimization algorithms that dynamically adjust the learning rate (such as the Adam optimizer) are widely used in deep learning training.

[0118] The Adam optimizer is an adaptive optimization algorithm that combines the advantages of Momentum and RMSprop. It automatically adjusts the learning rate based on the gradient and second moment of each parameter, effectively solving the problems associated with a fixed learning rate. Especially when training deep networks, the Adam optimizer can find the optimal learning rate over a wide range, avoiding overfitting and accelerating the training process. By using the Adam optimizer, the network can converge quickly during training, reducing wasted computational resources and avoiding training instability caused by excessively high or low learning rates.

[0119] Combining regularization techniques with dynamic learning rate adjustment can further improve model training performance. During training, regularization effectively controls model complexity and prevents overfitting, while dynamic learning rate adjustment improves training efficiency and accuracy. Through this comprehensive training mechanism, the model can find a balance between training and validation data, avoiding over-memorization of training data and maintaining good generalization ability.

[0120] Implementation Method 3:

[0121] Cross-validation and early stopping mechanisms are crucial for ensuring the good generalization ability of deep learning models. In the auxiliary assessment of oral diseases, the model's generalization ability directly affects its effectiveness and accuracy in actual clinical practice. Therefore, a comprehensive training scheme combining multiple rounds of cross-validation and early stopping mechanisms can help further improve the model's stability and reliability.

[0122] Cross-validation effectively evaluates a model's generalization ability by dividing the dataset into multiple subsets (usually k subsets) and training the model multiple times. K-fold cross-validation is the most common method, which randomly divides the dataset into k subsets, uses k-1 subsets for training each time, and uses the remaining subset for validation, repeating the training process k times. Through multiple rounds of cross-validation, the performance of the model can be evaluated on different data subsets, reducing the randomness introduced by data partitioning and improving the model's robustness across different datasets.

[0123] Early stopping is a technique to prevent overfitting. It involves monitoring the model's performance on the validation set during training, and stopping training once the loss on the validation set no longer shows significant improvement after several consecutive training iterations. Early stopping effectively avoids overlearning during training, allowing the model to stop training once it reaches optimal performance, preventing it from continuing to fit noisy or irrelevant features to the training data.

[0124] Combining cross-validation with early stopping maximizes model stability across multiple subsets, avoiding over-reliance on any single dataset. The results of each round of cross-validation not only verify model stability but also allow for dynamic adjustments to training strategies (such as learning rate or regularization strength), ensuring the model optimizes for improved generalization in each training iteration. Early stopping prevents overfitting on the validation set, maintaining good performance and making training more efficient. This comprehensive training approach ensures the model performs more accurately and stably when facing unknown data.

[0125] This invention employs integrated learning from multiple data sources and data augmentation strategies, enabling the model to learn more features from diverse data and reducing biases caused by data uniformity or imbalance. In this context, the model can not only identify common oral diseases but also effectively handle oral imaging data from different devices, imaging conditions, and patient groups. Through data augmentation and regularization methods, especially Dropout and L2 regularization, the model avoids over-memorizing training data, enhancing its predictive ability for new data. Therefore, the model demonstrates strong adaptability and accuracy regardless of different hospitals, diagnostic environments, or variations in patient oral structures. This effect is particularly important in practical clinical applications because oral imaging data often exhibits significant heterogeneity; ensuring the model's strong generalization ability greatly improves its reliability and practicality in real-world scenarios.

[0126] This invention effectively alleviates overfitting during training by introducing cross-validation and early stopping mechanisms. Cross-validation allows the model to undergo multiple rounds of training and validation on different subsets, enabling a comprehensive evaluation of model performance and ensuring its stability. The early stopping mechanism automatically halts training based on the performance on the validation set, preventing the model from overlearning noise on the training data and ensuring optimal performance on the validation set. These techniques not only reduce overfitting during training but also optimize computational resource usage, as training can be terminated when validation set performance no longer improves, thus avoiding unnecessary computational overhead. Therefore, the entire training process becomes more efficient and stable, improving the model's predictive accuracy and reducing the waste of computational resources from ineffective training. This effect is particularly important for processing large-scale oral imaging data, significantly shortening training time and improving model productivity.

[0127] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0128] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0129] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0130] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0131] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0132] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0133] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0134] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0135] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0136] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based oral disease auxiliary assessment system, characterized by, The system includes a data preprocessing module, a feature extraction module, a model training and regularization module, an enhanced learning strategy module, and a diagnosis output and feedback module. The data preprocessing module collects and preprocesses oral image data from different sources, improving data quality and diversity, and addressing imbalances in the original data. The feature extraction module uses a convolutional neural network to extract local features from oral images, using multiple layers of convolution and pooling operations to generate multi-scale feature representations that capture subtle lesions and complex structures of oral diseases. The model training and regularization module uses cross-validation and regularization techniques to train deep learning models, avoiding overfitting and ensuring good generalization to unseen data. The enhanced learning strategy module introduces reinforcement learning algorithms to adjust learning strategies based on model diagnostic feedback, allowing the model to adapt to the characteristics of different oral diseases during training. The diagnosis output and feedback module uses adaptive convolutional neural networks and backpropagation algorithms to output diagnostic results and adjust diagnostic recommendations based on the model's predicted confidence, further improving the accuracy and reliability of the evaluation. 2.The deep learning-based oral disease auxiliary assessment system according to claim 1, characterized in that, The specific steps for collecting and preprocessing oral image data from different sources, improving data quality and diversity, and addressing imbalances in the original data are as follows: First, collect oral image data from multiple sources. To ensure consistency across different sources, standardize all images. To reduce noise interference during model training, perform denoising. Use data resampling techniques to address data imbalance in the image dataset. 3.The deep learning-based oral disease auxiliary assessment system according to claim 1, characterized in that, The specific steps for using a convolutional neural network to extract local features from oral images, using multiple layers of convolution and pooling operations to generate multi-scale feature representations that capture subtle lesions and complex structures of oral diseases are as follows: After data preprocessing, first design the structure of the convolutional neural network. In the first layer and subsequent layers of the convolutional neural network, extract local features from oral images through convolution operations. After extracting local features through convolution operations, use pooling operations to further reduce the spatial dimension of the data, thereby reducing computational complexity and avoiding overfitting. After multiple layers of convolution and pooling operations, the convolutional neural network will generate multiple feature maps. 4.The deep learning-based oral disease auxiliary assessment system according to claim 1, characterized in that, The specific steps for using cross-validation and regularization techniques to train deep learning models, avoiding overfitting, and ensuring good generalization to unseen data are as follows: After feature extraction, divide the preprocessed data for cross-validation. During model training, introduce regularization techniques to avoid overfitting of deep neural networks when processing oral images. During training, adjust the model's hyperparameters while using cross-validation and regularization. To ensure that the model does not overfit on the training set, introduce an early stopping mechanism during training. 5.The deep learning-based oral disease auxiliary assessment system according to claim 1, wherein, The specific steps for introducing reinforcement learning algorithms to adjust learning strategies based on model diagnostic feedback, allowing the model to adapt to the characteristics of different oral diseases during training are as follows: In reinforcement learning, first define the state and action spaces, allowing the model to learn the optimal strategy through feedback mechanisms, with the following computational expression: , wherein, is inputted oral image data, is a confidence of a current diagnosis result, is a state at a first moment; In reinforcement learning, the reward function is used to feedback the goodness of each action, and the model adjusts the strategy according to the reward to make accurate decisions in future diagnosis processes. The reward function is defined as follows: wherein, is the predicted action of the model, is the true diagnosis, is an indicator function that returns 1 if the predicted action of the model is consistent with the true diagnosis and 0 otherwise, is a penalty function for a misdiagnosis, is a positive reward adjustment factor, is a negative penalty adjustment factor used to control the strength of the penalty when the model makes a misdiagnosis, is a reward function; By introducing reinforcement learning In supervised learning, the model adjusts its behavior policy based on the current reward signal, in In reinforcement learning, the model updates its value function, The value update formula is as follows: where, is the current state and action the corresponding value function, representing the expected return obtained after choosing this action, is the learning rate, controlling the magnitude of adjustment of the value at each update, is the discount factor, is the optimal action corresponding to the maximum value at the next state . 6.The deep learning-based oral disease auxiliary assessment system according to claim 5, characterized in that, Through the adaptive convolutional neural network and backpropagation algorithm, the output diagnosis result is obtained, and the diagnosis suggestion is adjusted according to the model prediction confidence to further improve the accuracy and reliability of the evaluation. The specific steps are as follows: First, after the multi-layer convolution and pooling operation of the oral cavity image, the last layer of the network outputs the diagnosis result, which is in the form of classification probability. To calculate the probability value of each category, the function is used for normalization to obtain a probability distribution, The function converts the score into a probability value, and the calculation expression is as follows: wherein, is the predicted probability of the dental disease type being the th dental disease type, i.e. given the input data , the model predicts the probability of belonging to the dental disease type , is the output score of the network for the th dental disease type, is the output score of the network for the th dental disease type, is the output score of the network for the th dental disease type, is the natural base, is the total number of dental disease types, is the sum of all class scores After obtaining the diagnosis result and calculating the prediction probability of each oral disease type, the weights of the convolutional neural network are adjusted through the backpropagation algorithm to minimize the loss function, which is calculated as follows: , where, is the true label value of the th class, is the loss function value that measures the difference between the model prediction and the true label. According to the loss function The gradient of each parameter is calculated and then an optimization algorithm is used to update the weights in the network, which is gradually optimized so that the predicted disease class probabilities tend towards the actual labels. 7.The deep learning-based oral disease auxiliary assessment system according to claim 6, characterized in that, The weights in the network are updated using the optimization algorithm, and the updated weights are adjusted according to the gradient descent method, with the update formula as follows: wherein, is the updated weight, is the weight of the previous step, is the gradient of the loss function with respect to the weight ; Finally, after the backpropagation is completed and the model weights are updated, the model outputs the final diagnosis result according to the prediction probability of each category. To enhance the accuracy and reliability of the evaluation, the final diagnosis suggestion is adjusted according to the model's confidence. For each disease category, a threshold is introduced to determine whether further examination is recommended, with the adjustment formula as follows: wherein is the final prediction result, is a preset confidence threshold.

Images