An artificial intelligence-based traditional Chinese medicine tongue diagnosis image intelligent analysis system
The TCM tongue diagnosis system, optimized through multimodal correction, deep neural networks, and multicenter validation, solves the problems of subjectivity and accuracy in practical applications of traditional TCM tongue diagnosis, and achieves automated and efficient diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU SHUANGLIU LINSEN TRADITIONAL CHINESE MEDICINE CLINIC CO LTD
- Filing Date
- 2025-10-25
- Publication Date
- 2026-06-19
AI Technical Summary
Traditional Chinese medicine tongue diagnosis relies on manual observation, and the diagnostic results are affected by the doctor's experience and environmental factors. Furthermore, the accuracy and reliability of existing AI systems in actual medical scenarios are insufficient, making it difficult to promote them on a large scale.
The system employs an image acquisition and preprocessing module for multimodal correction, a feature extraction module for extracting tongue features, a classification decision module for diagnosis using a deep neural network, a data augmentation and enhancement module for generating diverse cases, and a validation and optimization module for optimizing the model through cross-validation with multi-center clinical data.
It improves the accuracy and stability of TCM tongue diagnosis, reduces reliance on doctors' experience, enhances the system's reliability in complex medical scenarios, and promotes the objectification and digitalization of TCM tongue diagnosis.
Smart Images
Figure CN122243858A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of combining artificial intelligence with traditional Chinese medicine, specifically to an intelligent image analysis system for tongue diagnosis in traditional Chinese medicine based on artificial intelligence. Background Technology
[0002] Traditional Chinese medicine (TCM) tongue diagnosis relies primarily on the doctor's visual observation. Diagnostic results are often influenced by the doctor's experience and subjective judgment; different doctors may interpret the same tongue appearance differently. Furthermore, the diagnostic process is affected by environmental factors such as light and temperature. Manual tongue diagnosis also requires significant manpower, and with increasing demand for TCM, this method is insufficient to meet practical needs. Therefore, utilizing modern computer technology combined with traditional TCM theory and expert experience to objectify and digitize TCM tongue diagnosis has become a popular research direction. Artificial intelligence-based intelligent image analysis systems for TCM tongue diagnosis have developed against this backdrop.
[0003] After searching, the following technical pain points were found in the existing technology:
[0004] 1. The accuracy of AI-powered TCM tongue diagnosis image analysis systems relies on a large amount of high-quality tongue image data. However, in practice, due to factors such as the shooting environment, equipment performance, and user operation, the acquired image data may contain errors and biases, affecting the training effect of the AI model. In addition, some systems may have incomplete databases, lacking diverse cases from different regions, ages, and physical conditions, resulting in insufficient model recognition ability for certain special cases, thus limiting the accuracy and comprehensiveness of diagnostic results.
[0005] 2. While some research institutions claim that their systems have high diagnostic accuracy, most of these conclusions are drawn in laboratory settings and lack large-scale, multi-center clinical validation. In real-world medical settings, human physiological conditions are complex and variable. Whether the diagnostic results of AI systems can rival traditional medical diagnostic methods requires further verification. This creates a trust barrier in their clinical application and makes them difficult to be widely accepted as a reliable diagnostic basis. Summary of the Invention
[0006] (a) Technical problems to be solved
[0007] To address the shortcomings of existing technologies, this invention provides an intelligent image analysis system for TCM tongue diagnosis based on artificial intelligence, which solves the problem of "low efficiency" in the aforementioned background technologies.
[0008] (II) Technical Solution
[0009] To achieve the above objectives, the present invention provides the following technical solution: an intelligent image analysis system for traditional Chinese medicine tongue diagnosis based on artificial intelligence, comprising:
[0010] The image acquisition and preprocessing module mainly receives raw tongue image data from the tongue image acquisition device through a hardware interface, and performs multimodal correction processing on the raw tongue image data to generate a standardized image.
[0011] The environment adaptation module is mainly connected to the image acquisition and preprocessing module via a data interface. The environment adaptation module monitors at least one environmental parameter of the working environment of the tongue image acquisition device in real time, and generates a correction strategy adjustment instruction based on the environmental parameter to dynamically update the correction parameters of the image acquisition and preprocessing module.
[0012] The feature extraction module mainly receives the standardized image from the image acquisition and preprocessing module via a data interface, and uses a preset feature extraction algorithm to extract a multi-dimensional feature vector containing tongue morphology features, tongue coating color features, and texture distribution features from the standardized image to generate a feature matrix;
[0013] The classification decision module mainly receives the feature matrix from the feature extraction module via a data interface, and processes the feature matrix using a built-in deep neural network model to generate diagnostic results.
[0014] The data augmentation and enhancement module is mainly connected to the classification decision module via a data interface. The data augmentation and enhancement module uses a generative adversarial network to generate diverse simulated case scene images and transmits the simulated case scene images as augmented training data to the classification decision module for training the deep neural network model.
[0015] The validation and optimization module mainly forms a closed-loop connection with the classification and decision module through a data interface. The validation and optimization module performs cross-validation on the diagnostic results generated by the classification and decision module based on a multi-center clinical dataset, and generates weight adjustment factors for model parameters according to the validation results. The weight adjustment factors are then fed back to the classification and decision module to update the model parameters of the deep neural network model.
[0016] Preferably, the image acquisition and preprocessing module includes an RGB channel information extraction unit, a grayscale distribution analysis unit, and a multimodal correction algorithm unit. The RGB channel information extraction unit is used to analyze the color distribution features of the original tongue image data. The grayscale distribution analysis unit is used to extract the overall brightness features of the original tongue image data. The multimodal correction algorithm unit receives the color distribution features and the overall brightness features, and performs correction processing on the original tongue image data.
[0017] The multimodal correction algorithm unit includes an illumination compensation subunit, a color correction subunit, and an angle calibration subunit. The illumination compensation subunit mainly calculates the linear interpolation coefficients of the image pixel brightness values based on the real-time illumination data provided by the environment adaptation module, and adjusts the brightness distribution of the original tongue image data to normalize the brightness to a preset range. The color correction subunit mainly generates a color conversion matrix through a preset color temperature mapping table based on the real-time ambient color temperature data provided by the environment adaptation module, and applies the color conversion matrix to correct the color deviation of the original tongue image data. The angle calibration subunit mainly detects the tongue edge lines in the original tongue image data through the Hough transform algorithm, calculates the tongue tilt angle, and generates a geometric transformation matrix for rotation and translation operations to calibrate the tongue posture.
[0018] Preferably, the environment adaptation module includes an environment parameter monitoring unit and an environment adaptability configuration unit. The environment parameter monitoring unit collects environmental parameters, including light intensity, ambient temperature changes, and air humidity fluctuations, in real time through a sensor integrated with the tongue image acquisition device, and encapsulates the collected environmental parameter data. The environment adaptability configuration unit receives the encapsulated environmental parameter data from the environment parameter monitoring unit and compares the environmental parameter data with a preset benchmark environmental parameter threshold. When any environmental parameter is detected to exceed its corresponding benchmark threshold range, a correction strategy adjustment instruction containing specific correction parameter adjustment values is generated, and the instruction is transmitted to the image acquisition and preprocessing module through the data interface.
[0019] Preferably, the feature extraction module includes a convolutional neural network unit, a color segmentation unit, a texture analysis unit, and a TCM theory diagnostic rule base matching unit. The convolutional neural network unit receives the standardized image and extracts tongue morphological features from the standardized image through multi-layer convolution and pooling operations. The tongue morphological features include the tongue contour area, edge curvature distribution, and tongue surface crack distribution information. The color segmentation unit uses an adaptive threshold segmentation algorithm on the standardized image to accurately separate the tongue coating area from the tongue body area, and further extracts the color features of the tongue coating area. The color features include the dominant hue, color saturation, and brightness values. The texture analysis unit calculates the texture distribution features of the tongue coating surface using a local binary mode algorithm for the tongue coating area separated by the color segmentation unit. The texture distribution features include surface roughness, distribution uniformity, and local contrast index. The TCM theory diagnostic rule base matching unit matches the texture distribution features extracted by the texture analysis unit with the initial diagnostic rules in the rule base to generate initial feature labels and outputs the feature matrix containing the multi-dimensional feature vector and the initial feature labels.
[0020] Preferably, the classification decision module includes:
[0021] The input layer mainly receives the feature matrix and maps the feature matrix to a high-dimensional feature space of a preset dimension through a fully connected network;
[0022] The hidden layer adopts a residual network structure, which contains multiple residual blocks. Each residual block contains a skip connection, which directly passes the input features to the subsequent layers to alleviate the gradient vanishing problem in the training process of deep neural networks and enhance the feature representation ability.
[0023] The output layer mainly receives the feature data processed by the hidden layer and processes the feature data using the softmax activation function to generate a set of probability distributions representing different diagnostic conclusions. The diagnostic conclusion with the highest probability value is determined as the final diagnostic result.
[0024] The classification decision module also uses a multi-level semantic mapping mechanism to fuse low-level features and high-level semantic information of the hidden layer to generate a structured diagnostic report that includes health status assessment and pathological recommendations.
[0025] Preferably, the data augmentation and enhancement module includes a generator unit and a discriminator unit. The generator unit has a deep convolutional network structure and is responsible for receiving random noise vectors and upsampling them to generate diverse simulated case scene images. The discriminator unit also has a deep convolutional network structure and is responsible for receiving the simulated case scene images and real tongue images generated by the generator unit, and for judging the authenticity of the received images. The loss function of the generative adversarial network includes an adversarial training term, which is defined by calculating the distribution difference between the simulated case scene images and real tongue images in the feature space, in order to enhance the robustness of the model during adversarial training. The data augmentation and enhancement module also adopts a progressive resolution improvement strategy, gradually increasing the number of network layers of the generator unit and the discriminator unit during training to gradually improve the resolution and quality of the generated simulated case scene images from low to high.
[0026] Preferably, the data augmentation and enhancement module further includes a data enhancement unit, which mainly performs further data enhancement operations on the highly realistic simulated case scene image output by the generator unit. The data enhancement operations include randomly cropping the image, randomly rotating it within a preset angle range, randomly scaling it within a preset ratio range, and randomly dithering the color in the color space. The enhanced image is then transmitted to the classification decision module through the data interface as input for its continuous training and optimization.
[0027] Preferably, the validation optimization module includes a data sampling subunit, a performance evaluation subunit, and a dynamic weight adjustment subunit. The data sampling subunit extracts a validation sample set from the multi-center clinical dataset using random sampling or stratified sampling strategies, and inputs the feature matrix of the validation sample set into the classification decision module to generate a preliminary diagnosis result. The performance evaluation subunit compares the preliminary diagnosis result with the real clinical label corresponding to each sample in the validation sample set one by one, and calculates the accuracy, recall, and F1 score of the deep neural network model on the current data subset based on the comparison results as performance evaluation indicators. The dynamic weight adjustment subunit generates a weight adjustment factor based on the performance evaluation indicators calculated by the performance evaluation subunit. The value of the weight adjustment factor is calculated by a preset formula, which dynamically configures the weight coefficients of each performance evaluation indicator according to the importance of different data subsets, and applies the generated weight adjustment factor to the parameter update step of the deep neural network model during backpropagation.
[0028] Preferably, an artificial intelligence-based intelligent analysis method for TCM tongue diagnosis images includes the following steps:
[0029] S1: Through the environment adaptation module, at least one environmental parameter of the working environment of the tongue image acquisition device is monitored in real time, and a correction strategy adjustment instruction is generated based on the environmental parameter.
[0030] S2: The image acquisition and preprocessing module receives the original tongue image data and dynamically updates its correction parameters using the correction strategy adjustment instructions. Multimodal correction processing is then performed on the original tongue image data to generate a standardized image.
[0031] S3: Through the feature extraction module, extract a multi-dimensional feature vector containing tongue morphology features, tongue coating color features, and texture distribution features from the standardized image to generate a feature matrix;
[0032] S4: The feature matrix is processed by the classification decision module using a deep neural network model to generate diagnostic results;
[0033] S5: Through the data augmentation and enhancement module, a variety of simulated case scene images are generated using a generative adversarial network, and the simulated case scene images are used as augmented training data to train the deep neural network model;
[0034] S6: Through the validation and optimization module, the diagnostic results are cross-validated based on a multi-center clinical dataset. Based on the validation results, a weight adjustment factor for the model parameters is generated, and the model parameters of the deep neural network model are updated using the weight adjustment factor.
[0035] (III) Beneficial Effects
[0036] This invention provides an intelligent image analysis system for traditional Chinese medicine tongue diagnosis based on artificial intelligence. It has the following beneficial effects:
[0037] (1) This invention reduces image errors caused by shooting environment, equipment, etc. through multimodal correction of image acquisition and preprocessing module, and generates diversified case images to improve dataset by combining data expansion and enhancement module. It effectively solves the problems of poor image quality and insufficient case coverage, reduces reliance on doctors' experience and subjective judgment, reduces the difference in judgment of different diagnostic subjects on the same tongue image, and improves the accuracy and stability of diagnostic results.
[0038] (2) The verification and optimization module of the present invention uses multi-center clinical datasets for cross-validation and continuously optimizes diagnostic performance by dynamically adjusting model parameters. This breaks through the limitations of existing technologies in laboratory environments, making the system's diagnostic results more reliable in complex and ever-changing real medical scenarios. It helps to eliminate trust barriers in clinical promotion and improve the system's widespread acceptance as a diagnostic basis.
[0039] (3) This invention achieves automated acquisition, processing, analysis and diagnosis of tongue images through the collaborative work of various modules, reducing the large amount of manpower required for manual tongue diagnosis, and can quickly respond to people's needs for TCM diagnosis and treatment, adapt to the trend of large-scale development of TCM services, and promote the transformation of TCM tongue diagnosis towards objectivity, digitalization and efficiency. Attached Figure Description
[0040] Figure 1 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation
[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0042] Please see Figure 1This invention provides an intelligent image analysis system for TCM tongue diagnosis based on artificial intelligence. The system architecture includes an image acquisition and preprocessing module, a feature extraction module, a classification decision module, a verification and optimization module, an environment adaptation module, and a data expansion and enhancement module. The modules work together through a clear data flow path.
[0043] The image acquisition and preprocessing module primarily receives raw image data generated by the tongue image acquisition device and performs multimodal correction processing to generate a standardized image. This module connects directly to the tongue image acquisition device via a hardware interface to ensure real-time acquisition of tongue image data. Internally, the module first analyzes the color distribution characteristics of the image using an RGB channel information extraction unit, and simultaneously extracts the overall brightness characteristics using a grayscale distribution analysis unit. These feature data are then transmitted to the multimodal correction algorithm unit, which consists of an illumination compensation subunit, a color correction subunit, and an angle calibration subunit. The illumination compensation subunit calculates the linear interpolation coefficients of the image pixel brightness values based on real-time data provided by the ambient light sensor and adjusts the brightness distribution to normalize the brightness to a preset range. The color correction subunit generates a color conversion matrix using a color temperature mapping table; the matrix elements are determined by the ratio of the current ambient color temperature to the standard color temperature, thereby correcting image color deviations. The angle calibration subunit detects the straight lines of the tongue edge using Hough transform, calculates the tilt angle, and generates a geometric transformation matrix, ultimately completing the image rotation and translation operations. The standardized image, after the above processing, is transmitted to the feature extraction module through the data output interface.
[0044] The feature extraction module receives standardized images from the image acquisition and preprocessing module and extracts feature vectors such as tongue morphology, tongue coating color, and texture distribution through a series of algorithms. The feature extraction module consists of a convolutional neural network unit, a color segmentation unit, and a texture analysis unit, all connected in series via a data flow pipeline. The convolutional neural network unit extracts tongue morphological features from the standardized images, including tongue area, edge curvature, and crack distribution. The color segmentation unit uses a threshold segmentation algorithm to separate tongue coating regions and further extracts tongue coating color features, such as dominant hue, saturation, and brightness. The texture analysis unit calculates the texture distribution features of the tongue coating surface using a local binary mode algorithm, including roughness, uniformity, and local contrast. The extracted feature vectors are input to the TCM theory diagnostic rule base matching unit. The texture analysis unit generates initial feature labels through rule matching and transmits the feature matrix containing multi-dimensional feature information to the classification decision module via a data interface.
[0045] The classification decision module receives the feature matrix and inputs it into a deep neural network model for multi-level semantic mapping. The module consists of an input layer, hidden layers, and an output layer, with data transfer between layers via a fully connected network. The input layer receives the feature matrix and maps it to a high-dimensional feature space through the fully connected layers. The hidden layers employ a residual network structure, using skip connections to mitigate the vanishing gradient problem and enhance feature representation. The output layer uses a softmax function to generate a probability distribution of the diagnostic results, where the category with the highest probability value is determined as the final diagnosis. Furthermore, the multi-level semantic mapping mechanism generates a diagnostic report containing health status assessments and pathological recommendations by fusing low-level features with high-level semantic information. The diagnostic report is transmitted to the validation and optimization module via a data interface.
[0046] The validation and optimization module cross-validates diagnostic results using a multi-center clinical dataset and dynamically adjusts model parameters to optimize diagnostic performance. This module consists of a data sampling subunit, a performance evaluation subunit, and a dynamic weight adjustment subunit. The data sampling subunit randomly selects samples from the multi-center clinical dataset and inputs the sample feature matrix into the classification decision module to generate preliminary diagnostic results. The performance evaluation subunit compares the preliminary diagnostic results with the true labels and calculates the model's precision, recall, and F1 score on different data subsets. The dynamic weight adjustment subunit generates weight adjustment factors based on performance metrics. The factor values are calculated using a formula, where the sum of the first, second, and third weights is 1, and the factors are dynamically configured according to the importance of the data subsets. The weight adjustment factors are applied to the parameter update process of the deep neural network model, generating optimized model parameters and feeding them back to the classification decision module through a data interface.
[0047] The data augmentation and enhancement module simulates diverse case scenarios using a generative adversarial network (GAN) to expand the coverage of the training dataset and improve the model's generalization ability. The module consists of a generator unit, a discriminator unit, and a data augmentation unit. The generator unit generates diverse case scenario images, while the discriminator unit distinguishes between generated and real images. Adversarial training terms are added to the GAN's loss function to enhance the model's robustness by calculating the distribution differences between generated and real images. A progressive resolution enhancement strategy gradually improves the quality of the generated images, ultimately outputting highly realistic and diverse case scenario images. The data augmentation unit further processes the generated images, including random cropping, rotation, scaling, and color dithering. The augmented images are then transmitted to the classification decision module via a data interface for continuous model training and optimization.
[0048] The environment adaptation module monitors the operating environment parameters of the acquisition device in real time and generates an environment adaptability configuration file to adjust the calibration strategy of the image acquisition and preprocessing modules. The environment adaptation module consists of an environment parameter monitoring unit and an environment adaptability configuration unit. The environment parameter monitoring unit collects environmental parameters such as light intensity, temperature changes, and humidity fluctuations in real time through sensors and transmits the parameter data to the environment adaptability configuration unit. The environment adaptability configuration unit generates calibration strategy adjustment instructions based on the environmental parameters and dynamically updates the calibration parameters of the image acquisition and preprocessing modules through a data interface, ensuring stable operation of the system in different environments.
[0049] The connections between the modules are established through data interfaces. Data flows sequentially from the image acquisition and preprocessing module to the feature extraction module, then to the classification decision module, followed by the verification and optimization module, and finally back to the classification decision module, forming a closed loop. The data augmentation and enhancement module and the environment adaptation module connect to the classification decision module and the image acquisition and preprocessing module through independent data interfaces, providing auxiliary support. The collaboration between these modules ensures that the system can efficiently complete tasks such as tongue image acquisition, processing, feature extraction, diagnostic analysis, and optimization verification.
[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. An intelligent image analysis system for traditional Chinese medicine tongue diagnosis based on artificial intelligence, characterized in that, include: The image acquisition and preprocessing module mainly receives raw tongue image data from the tongue image acquisition device through a hardware interface, and performs multimodal correction processing on the raw tongue image data to generate a standardized image. The environment adaptation module is mainly connected to the image acquisition and preprocessing module via a data interface. The environment adaptation module monitors at least one environmental parameter of the working environment of the tongue image acquisition device in real time, and generates a correction strategy adjustment instruction based on the environmental parameter to dynamically update the correction parameters of the image acquisition and preprocessing module. The feature extraction module mainly receives the standardized image from the image acquisition and preprocessing module via a data interface, and uses a preset feature extraction algorithm to extract a multi-dimensional feature vector containing tongue morphology features, tongue coating color features, and texture distribution features from the standardized image to generate a feature matrix; The classification decision module mainly receives the feature matrix from the feature extraction module via a data interface, and processes the feature matrix using a built-in deep neural network model to generate diagnostic results. The data augmentation and enhancement module is mainly connected to the classification decision module via a data interface. The data augmentation and enhancement module uses a generative adversarial network to generate diverse simulated case scene images and transmits the simulated case scene images as augmented training data to the classification decision module for training the deep neural network model. The validation and optimization module mainly forms a closed-loop connection with the classification and decision module through a data interface. The validation and optimization module performs cross-validation on the diagnostic results generated by the classification and decision module based on a multi-center clinical dataset, and generates weight adjustment factors for model parameters according to the validation results. The weight adjustment factors are then fed back to the classification and decision module to update the model parameters of the deep neural network model.
2. The intelligent image analysis system for TCM tongue diagnosis based on artificial intelligence according to claim 1, characterized in that: The image acquisition and preprocessing module includes an RGB channel information extraction unit, a grayscale distribution analysis unit, and a multimodal correction algorithm unit. The RGB channel information extraction unit is used to analyze the color distribution features of the original tongue image data. The grayscale distribution analysis unit is used to extract the overall brightness features of the original tongue image data. The multimodal correction algorithm unit receives the color distribution features and the overall brightness features, and performs correction processing on the original tongue image data. The multimodal correction algorithm unit includes an illumination compensation subunit, a color correction subunit, and an angle calibration subunit. The illumination compensation subunit mainly calculates the linear interpolation coefficients of the image pixel brightness values based on the real-time illumination data provided by the environment adaptation module, and adjusts the brightness distribution of the original tongue image data to normalize the brightness to a preset range. The color correction subunit mainly generates a color conversion matrix through a preset color temperature mapping table based on the real-time ambient color temperature data provided by the environment adaptation module, and applies the color conversion matrix to correct the color deviation of the original tongue image data. The angle calibration subunit mainly detects the tongue edge lines in the original tongue image data through the Hough transform algorithm, calculates the tongue tilt angle, and generates a geometric transformation matrix for rotation and translation operations to calibrate the tongue posture.
3. The intelligent image analysis system for TCM tongue diagnosis based on artificial intelligence according to claim 1, characterized in that: The environment adaptation module includes an environment parameter monitoring unit and an environment adaptability configuration unit. The environment parameter monitoring unit collects environmental parameters in real time, including light intensity, ambient temperature changes, and air humidity fluctuations, through sensors integrated with the tongue image acquisition device, and encapsulates the collected environmental parameter data. The environment adaptability configuration unit receives the encapsulated environmental parameter data from the environment parameter monitoring unit and compares the environmental parameter data with preset benchmark environmental parameter thresholds. When any environmental parameter is detected to exceed its corresponding benchmark threshold range, a correction strategy adjustment instruction containing specific correction parameter adjustment values is generated, and the instruction is transmitted to the image acquisition and preprocessing module through the data interface.
4. The intelligent image analysis system for TCM tongue diagnosis based on artificial intelligence according to claim 1, characterized in that: The feature extraction module includes a convolutional neural network unit, a color segmentation unit, a texture analysis unit, and a TCM theory diagnostic rule base matching unit. The convolutional neural network unit receives the standardized image and extracts tongue morphological features from the standardized image through multi-layer convolution and pooling operations. The tongue morphological features include the tongue contour area, edge curvature distribution, and tongue surface crack distribution information. The color segmentation unit uses an adaptive threshold segmentation algorithm on the standardized image to accurately separate the tongue coating area from the tongue body area, and further extracts the color features of the tongue coating area. The color features include the dominant hue, color saturation, and brightness values. The texture analysis unit calculates the texture distribution features of the tongue coating surface using a local binary mode algorithm for the tongue coating area separated by the color segmentation unit. The texture distribution features include surface roughness, distribution uniformity, and local contrast index. The TCM theory diagnostic rule base matching unit matches the texture distribution features extracted by the texture analysis unit with the initial diagnostic rules in the rule base to generate initial feature labels and outputs the feature matrix containing the multi-dimensional feature vector and the initial feature labels.
5. The intelligent image analysis system for TCM tongue diagnosis based on artificial intelligence according to claim 1, characterized in that: The classification decision module includes: The input layer mainly receives the feature matrix and maps the feature matrix to a high-dimensional feature space of a preset dimension through a fully connected network; The hidden layer adopts a residual network structure, which contains multiple residual blocks. Each residual block contains a skip connection, which directly passes the input features to the subsequent layers to alleviate the gradient vanishing problem in the training process of deep neural networks and enhance the feature representation ability. The output layer mainly receives the feature data processed by the hidden layer and processes the feature data using the softmax activation function to generate a set of probability distributions representing different diagnostic conclusions. The diagnostic conclusion with the highest probability value is determined as the final diagnostic result. The classification decision module also uses a multi-level semantic mapping mechanism to fuse low-level features and high-level semantic information of the hidden layer to generate a structured diagnostic report that includes health status assessment and pathological recommendations.
6. The intelligent image analysis system for TCM tongue diagnosis based on artificial intelligence according to claim 1, characterized in that: The data augmentation and enhancement module includes a generator unit and a discriminator unit. The generator unit has a deep convolutional network structure and is responsible for receiving and upsampling random noise vectors to generate diverse simulated case scene images. The discriminator unit also has a deep convolutional network structure and is responsible for receiving the simulated case scene images and real tongue images generated by the generator unit, and for judging the authenticity of the received images. The loss function of the generative adversarial network includes an adversarial training term, which is defined by calculating the distribution difference between the simulated case scene images and real tongue images in the feature space, in order to enhance the robustness of the model during adversarial training. The data augmentation and enhancement module also adopts a progressive resolution improvement strategy, gradually increasing the number of network layers of the generator unit and the discriminator unit during training to gradually improve the resolution and quality of the generated simulated case scene images from low to high.
7. The intelligent image analysis system for TCM tongue diagnosis based on artificial intelligence according to claim 6, characterized in that: The data augmentation and enhancement module further includes a data enhancement unit, which performs further data enhancement operations on the highly realistic simulated case scene image output by the generator unit. The data enhancement operations include randomly cropping the image, randomly rotating it within a preset angle range, randomly scaling it within a preset ratio range, and randomly dithering the color in the color space. The enhanced image is then transmitted to the classification decision module through the data interface as input for its continuous training and optimization.
8. The intelligent image analysis system for TCM tongue diagnosis based on artificial intelligence according to claim 1, characterized in that: The validation optimization module includes a data sampling subunit, a performance evaluation subunit, and a dynamic weight adjustment subunit. The data sampling subunit extracts a validation sample set from the multi-center clinical dataset using random sampling or stratified sampling strategies, and inputs the feature matrix of the validation sample set into the classification decision module to generate preliminary diagnostic results. The performance evaluation subunit compares the preliminary diagnostic results with the real clinical labels corresponding to each sample in the validation sample set one by one, and calculates the accuracy, recall, and F1 score of the deep neural network model on the current data subset based on the comparison results as performance evaluation indicators. The dynamic weight adjustment subunit generates a weight adjustment factor based on the performance evaluation indicators calculated by the performance evaluation subunit. The value of the weight adjustment factor is calculated by a preset formula, which dynamically configures the weight coefficients of each performance evaluation indicator according to the importance of different data subsets, and applies the generated weight adjustment factor to the parameter update step of the deep neural network model during backpropagation.
9. A method for intelligent analysis of TCM tongue diagnosis images based on artificial intelligence, comprising any one of claims 1-8, characterized in that, Includes the following steps: S1: Through the environment adaptation module, at least one environmental parameter of the working environment of the tongue image acquisition device is monitored in real time, and a correction strategy adjustment instruction is generated based on the environmental parameter. S2: The image acquisition and preprocessing module receives the original tongue image data and dynamically updates its correction parameters using the correction strategy adjustment instructions. Multimodal correction processing is then performed on the original tongue image data to generate a standardized image. S3: Through the feature extraction module, extract a multi-dimensional feature vector containing tongue morphology features, tongue coating color features, and texture distribution features from the standardized image to generate a feature matrix; S4: The feature matrix is processed by the classification decision module using a deep neural network model to generate diagnostic results; S5: Through the data augmentation and enhancement module, a variety of simulated case scene images are generated using a generative adversarial network, and the simulated case scene images are used as augmented training data to train the deep neural network model; S6: Through the validation and optimization module, the diagnostic results are cross-validated based on a multi-center clinical dataset. Based on the validation results, a weight adjustment factor for the model parameters is generated, and the model parameters of the deep neural network model are updated using the weight adjustment factor.