A traditional chinese medicine tongue feature extraction and chronic disease syndrome differentiation auxiliary analysis method

Abstract

The application discloses a traditional Chinese medicine tongue feature extraction and chronic disease syndrome differentiation auxiliary analysis method, comprising the following steps: step one, tongue image preprocessing, step two, lightweight tongue feature extraction, step three, multi-dimensional chronic disease data fusion, step four, chronic disease syndrome differentiation model reasoning, step five, syndrome differentiation result output and model iteration; the feature extraction model and the syndrome differentiation classification model are pruned, quantized and compressed, and the overall volume is controlled within 50MB, which can be directly deployed on the mobile application end such as the WeChat applet and realized local operation, does not need to rely on cloud large computing power support, can quickly complete real-time analysis and syndrome differentiation of tongue image, meets the core needs of liver disease, diabetes and hypertension patients for out-of-hospital self-detection and real-time syndrome differentiation, and breaks the scene limitation that the existing AI tongue diagnosis technology can only be deployed on a professional server end.

Description

Technical Field

[0001] This invention relates to the field of TCM intelligent diagnosis and chronic disease management technology, specifically a method for extracting TCM tongue features and assisting in the analysis of chronic disease syndrome differentiation. Background Technology

[0002] Tongue diagnosis in Traditional Chinese Medicine (TCM), as a core component of "inspection" in the four diagnostic methods, is an important basis for the differentiation and treatment of chronic diseases such as liver disease, diabetes, and hypertension. Different chronic diseases exhibit specific characteristics in their tongue appearance, and changes in tongue appearance can directly reflect the course of the disease, the evolution of its pathogenesis, and the state of organ function. Specifically, the tongue appearance of liver disease is often related to liver dysfunction, qi stagnation and blood stasis, and liver and kidney yin deficiency, commonly characterized by a reddish-purple tongue, a yellow and greasy coating, and cracks or teeth marks on the tongue. The tongue appearance of diabetes is often related to yin deficiency and dryness-heat, phlegm-dampness obstruction, and qi and yin deficiency, commonly characterized by a red tongue with little moisture, a thick and greasy coating or no coating, and a swollen tongue. The tongue appearance of hypertension is often related to liver yang hyperactivity, phlegm and blood stasis, and liver and kidney yin deficiency, commonly characterized by a dark purple tongue with ecchymosis, a yellow and dry coating, and red edges and tip of the tongue.

[0003] Traditional Chinese medicine tongue diagnosis for chronic diseases relies heavily on the physician's clinical experience and subjective judgment. Different physicians may interpret the same tongue appearance differently, and there is a lack of quantitative standards for tongue characteristics and a unified diagnostic system. This makes it difficult to widely apply tongue diagnosis results in primary healthcare institutions and outpatient management of chronic diseases. Furthermore, existing AI-based tongue diagnosis technologies are mostly developed based on large-scale deep learning models. These models have large parameter scales and high hardware computation requirements, meaning they can only be deployed on professional servers and cannot run directly in lightweight mobile applications such as WeChat mini-programs. This makes it difficult to meet the core needs of patients with liver disease, diabetes, and hypertension for self-monitoring and real-time diagnosis outside of hospitals.

[0004] Furthermore, existing AI tongue diagnosis technology suffers from significant shortcomings in interdisciplinary integration. It has not been deeply integrated with the TCM diagnostic systems for liver disease, diabetes, and hypertension. The extraction of tongue features lacks specificity, extracting only general tongue features without focusing on the specific tongue manifestations of these three chronic diseases. At the same time, existing technologies have not achieved multi-dimensional integration of tongue feature data with clinical data of chronic diseases. They rely solely on single tongue data for diagnosis, ignoring the patient's symptoms, TCM constitution information, and laboratory test indicators. This results in insufficient accuracy of the diagnosis results and cannot provide accurate and comprehensive support for the diagnosis of the three types of chronic diseases. This seriously restricts the practical implementation and application of AI tongue diagnosis technology in the management of chronic diseases such as liver disease, diabetes, and hypertension. Summary of the Invention

[0005] The purpose of this invention is to provide a method for extracting tongue features in traditional Chinese medicine and for auxiliary analysis of chronic disease diagnosis, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting tongue features in traditional Chinese medicine and for auxiliary analysis of chronic disease diagnosis, comprising the following steps:

[0007] Step 1: Tongue image preprocessing. The tongue image photos collected and uploaded by the user through the mobile application applet are processed sequentially with illumination correction, region segmentation, and noise removal to extract a standard image of the pure tongue image region of 224×224 pixels.

[0008] Step 2: Lightweight tongue image feature extraction. An improved MobileNet lightweight convolutional neural network is used to extract features from standard tongue images, identify and quantify core tongue image features such as tongue color, coating color, coating thickness, cracks, teeth marks, ecchymosis, tongue body size, and amount of saliva, and calculate the comprehensive index of tongue image feature quantification through a formula.

[0009] Step 3: Multi-dimensional chronic disease data fusion. The quantitative data of tongue appearance features are fused with the chronic disease-related symptom information entered by the user, TCM constitution information, and uploaded laboratory indicator data of liver function, blood sugar, and blood pressure to construct a chronic disease syndrome differentiation dataset.

[0010] Step 4: Chronic disease syndrome differentiation model reasoning. Based on the fused syndrome differentiation dataset, the trained chronic disease syndrome differentiation classification model outputs the syndrome differentiation confidence scores of common syndrome types for liver disease, diabetes, and hypertension, respectively. The syndrome type with the highest confidence score is selected as the main syndrome type. The syndrome differentiation confidence score is calculated using a nonlinear mapping formula.

[0011] Step 5: Outputting the diagnosis results and iterating the model. Based on the reasoning results of the syndrome type, a standardized diagnosis report is generated and pushed to the patient and doctor terminals of the mini-program. At the same time, clinical data and user feedback are collected to incrementally train the model and iteratively optimize the model parameters.

[0012] Furthermore, the illumination correction in step one involves adjusting the pixel values ​​of the RGB three-color channels to achieve color balance and normalizing the average pixel value of each channel to a preset standard range; the region segmentation involves using a semantic segmentation algorithm to extract the complete outline of the tongue and smooth the edges; and the noise removal involves using a Gaussian filtering algorithm to remove shooting noise while retaining the core features of the tongue image.

[0013] Furthermore, the improved MobileNet lightweight convolutional neural network structure in step two is as follows: the input layer is a standard tongue image of 224×224×3, which is sequentially connected to a Conv2d+BN+ReLU convolutional layer, 6 depthwise separable convolutional blocks, and a global average pooling layer, and finally divided into four output branches: tongue color, tongue coating color, tongue shape, and saliva. The Conv2d+BN+ReLU convolutional layer uses a 3×3 convolutional kernel and 32 channels. The depthwise separable convolutional blocks reduce the number of model parameters by about 88%. After pruning and quantization compression, the model size is controlled within 50MB.

[0014] Furthermore, the quantification rules for tongue features in step two are as follows: tongue color is divided into 5 levels: light red, red, crimson, light white, and dark purple, with quantification values ​​of 1-5 respectively; tongue coating color is divided into 5 levels: white coating, yellow coating, gray coating, black coating, and no coating, with quantification values ​​of 1-5 respectively; tongue shape is divided into 5 levels: normal, swollen, thin, cracked, and teeth marks, with quantification values ​​of 0-4 respectively; body fluid is divided into 4 levels: normal, slightly excessive, slightly insufficient, and no body fluid, with quantification values ​​of 0-3 respectively; ecchymosis and thick coating are binary features, with a corresponding feature recorded as 1 and no corresponding feature recorded as 0.

[0015] Furthermore, the formula for calculating the comprehensive index of tongue image feature quantification in step two is as follows: ,in A comprehensive index for quantifying tongue features. The number of tongue features involved in the calculation. For the first Weighting coefficients for individual tongue features For the first Quantitative values ​​of individual tongue features; Based on the TCM syndrome differentiation system for liver disease, diabetes, and hypertension, and trained with clinical tongue diagnosis data, dynamic adjustments are made for these three types of chronic diseases.

[0016] Furthermore, the multi-dimensional data fusion in step three includes: classifying the symptoms of liver disease, diabetes, and hypertension into four levels according to severity: none, mild, moderate, and severe, with quantitative values ​​ranging from 0 to 3; digitizing the nine constitutions in traditional Chinese medicine, with the balanced constitution recorded as 0 and the other constitutions as 1-8, and the degree of constitution imbalance quantified as mild, moderate, and severe, with values ​​ranging from 1 to 3; and normalizing the laboratory indicators of liver function, blood sugar, and blood pressure, mapping the measured values ​​to the 0-1 standard range.

[0017] Furthermore, the formula for calculating the confidence level of chronic disease diagnosis in step four is as follows: ,in For the first The confidence level of dialectical analysis for individual cases It is a non-linear activation function. A comprehensive index for quantifying tongue features. This is the normalized composite value of the laboratory indicators. This is a quantitative value representing the integration of symptoms and physical constitution. , , To integrate the weighting coefficients, different coefficient values ​​were set for liver disease, diabetes, and hypertension.

[0018] Furthermore, the syndrome inference results in step four are as follows: liver disease outputs 5 syndrome types: liver qi stagnation, spleen deficiency with dampness, liver and gallbladder damp heat, liver and kidney yin deficiency, and qi stagnation and blood stasis; diabetes outputs 5 syndrome types: yin deficiency with dryness and heat, phlegm and dampness obstruction, qi and yin deficiency, yin and yang deficiency, and blood stasis in the meridians; hypertension outputs 5 syndrome types: liver yang hyperactivity, phlegm and blood stasis, liver and kidney yin deficiency, yin deficiency with yang hyperactivity, and qi deficiency and blood stasis; the main syndrome type is the syndrome type with the highest confidence among all types of chronic disease syndrome types.

[0019] Furthermore, the standardized diagnosis report in step five includes basic patient information, tongue appearance analysis, chronic disease syndrome type judgment results, chronic disease risk level assessment, and personalized TCM conditioning suggestions. The conditioning suggestions are formulated for different syndrome types of liver disease, diabetes, and hypertension, focusing on diet, rest, emotions, TCM tea, and acupoint massage.

[0020] Furthermore, the model iteration optimization in step five involves collecting tongue image data from the mini-program, user clinical data, AI diagnosis results, and doctor-annotated clinical diagnosis results, incrementally training the lightweight feature extraction model and the chronic disease diagnosis and classification model, and dynamically adjusting the tongue image feature weight coefficient and the data fusion weight coefficient.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] The feature extraction model and syndrome differentiation classification model of this invention have been pruned and optimized through quantization and compression, with the overall size controlled within 50MB. They can be directly deployed on mobile applications such as WeChat mini programs and run locally without relying on cloud computing power. They can quickly complete real-time analysis and syndrome inference of tongue images, meeting the core needs of outpatient self-testing and real-time syndrome differentiation for patients with liver disease, diabetes, and hypertension. This breaks the limitation of existing AI tongue diagnosis technology that can only be deployed on professional servers.

[0023] This invention focuses on the TCM syndrome differentiation system for three types of chronic diseases: liver disease, diabetes, and hypertension. It extracts specific tongue features that are highly correlated with the pathogenesis and syndrome types of these three chronic diseases, rather than general tongue features. At the same time, it assigns dynamically adjusted weight coefficients to the tongue features for different syndrome types of the three chronic diseases, so that the tongue feature extraction results are more in line with the actual needs of clinical syndrome differentiation and provide accurate tongue data support for syndrome type judgment.

[0024] This invention integrates quantitative data on tongue features with patients' chronic disease symptoms, TCM constitution information, and laboratory test indicators from multiple dimensions. This achieves complementary support between tongue features and clinical data, overcoming the shortcomings of existing technologies that rely solely on single data from the tongue for diagnosis. It significantly improves the accuracy of syndrome differentiation for liver disease, diabetes, and hypertension. At the same time, through unified quantification and normalization rules, it achieves standardized processing of tongue features and clinical data, reducing the reliance of traditional tongue diagnosis on physicians' manual experience and improving the standardization level of tongue diagnosis.

[0025] The standardized diagnosis report generated by this invention includes complete content such as tongue appearance feature analysis, syndrome type judgment results, risk level assessment, and personalized conditioning suggestions. It is simultaneously pushed to both the patient and doctor's end. Doctors can use the AI ​​diagnosis results as an important reference for clinical diagnosis, improving the efficiency of chronic disease diagnosis and treatment in primary and specialized medical care. Patients can use the diagnosis report and personalized conditioning suggestions for self-management outside the hospital, which helps in the long-term control of chronic diseases such as liver disease, diabetes, and hypertension. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention;

[0027] Figure 2 This is a schematic diagram of the tongue image preprocessing process of the present invention;

[0028] Figure 3 This is a network structure diagram of the lightweight feature extraction model of the present invention;

[0029] Figure 4 This is a schematic diagram illustrating the application scenario of the present invention. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figure 1-4 This invention provides a method for extracting tongue features in traditional Chinese medicine and for auxiliary analysis of chronic disease diagnosis, comprising the following steps:

[0032] Step 1: Tongue Image Preprocessing

[0033] Standardized preprocessing is performed on tongue images collected and uploaded by users through mobile application mini-programs. The process includes three sub-steps: illumination correction, region segmentation, and noise removal. Background interference is removed and pure tongue image regions are extracted to ensure the uniformity of tongue image quality and the accuracy of subsequent feature extraction.

[0034] Lighting correction: To address the color deviation in tongue images caused by unbalanced ambient lighting during shooting, color balance processing is achieved by adjusting the pixel values ​​of the RGB three-color channels. The average pixel value of each channel is calculated and normalized to a preset standard range, so that the color performance of the tongue image conforms to the visual judgment standards of TCM tongue diagnosis.

[0035] Region segmentation: The semantic segmentation algorithm is used to segment the tongue image at the pixel level, identify and remove non-tongue areas such as lips, face, and shooting background, accurately extract the complete outline of the tongue, and smooth the edges of the tongue to avoid edge pixels interfering with subsequent feature extraction.

[0036] Noise Removal: Gaussian filtering algorithm is used to denoise the segmented tongue image, removing salt and pepper noise and Gaussian noise generated during the shooting process, while retaining the core features of the tongue image such as texture and color. Finally, the preprocessed pure tongue image region image is uniformly adjusted to the standard format of 224×224 pixels, which is used as the input data of the lightweight feature extraction model.

[0037] Step 2: Lightweight tongue image feature extraction

[0038] An improved MobileNet lightweight convolutional neural network is used as the base model to extract features from the standard tongue image obtained in step one. After pruning, quantization, and compression optimization, the model focuses on the specific tongue features of three chronic diseases: liver disease, diabetes, and hypertension. It accurately identifies and quantifies core tongue features such as tongue color, coating color, coating thickness, cracks, teeth marks, ecchymosis, tongue body size, and saliva amount, transforming visual features into computable quantitative data.

[0039] An improved MobileNet model structure is implemented: the input layer is a standard 224×224×3 tongue image, which is sequentially connected to a Conv2d+BN+ReLU convolutional layer (3×3 convolutional kernel, 32 channels), 6 depthwise separable convolutional blocks, and a global average pooling layer. Finally, the model is divided into four output branches: tongue color, tongue coating color, tongue shape, and saliva, which output the quantized results of the corresponding tongue image features. The depthwise separable convolution is the core lightweight structure of the model, which reduces the number of parameters by about 88% compared to traditional convolutional neural networks. After pruning and quantization compression, the model size is controlled within 50MB, which supports local operation on mobile application mini-programs.

[0040] Tongue feature quantification: The tongue features output by the model are graded and quantified. Tongue color is divided into 5 levels: pale red, red, crimson, pale white, and dark purple, with quantization values ​​of 1-5 respectively; tongue coating color is divided into 5 levels: white coating, yellow coating, gray coating, black coating, and no coating, with quantization values ​​of 1-5 respectively; tongue shape is divided into 5 levels: normal, swollen, thin, cracked, and teeth marks, with quantization values ​​of 0-4 respectively; body fluid is divided into 4 levels: normal, excessive, insufficient, and no body fluid, with quantization values ​​of 0-3 respectively; ecchymosis and thick coating are used as binary features, with corresponding features recorded as 1 and no corresponding features recorded as 0.

[0041] Tongue Image Feature Quantitative Comprehensive Index Calculation: To address the diagnostic needs of three types of chronic diseases—liver disease, diabetes, and hypertension—different weighting coefficients are assigned to different tongue image features to calculate a comprehensive index of tongue image features. This index serves as core data support for the diagnosis of these three types of chronic diseases. The calculation formula is as follows:

[0042] (1)

[0043] in, A comprehensive index for quantifying tongue features. The number of tongue features involved in the calculation. For the first Weighting coefficients for individual tongue features For the first Quantitative values ​​of individual tongue features; Based on the TCM diagnostic system for liver disease, diabetes, and hypertension, and trained with clinical tongue diagnosis data, this method dynamically adjusts the specific tongue appearances for these three chronic diseases. The value is the quantized value of the tongue image feature obtained in step 2, which is either the hierarchical quantization number or the binarization number of the corresponding feature.

[0044] Step 3: Multi-dimensional chronic disease data fusion

[0045] Construct a diagnostic dataset for three types of chronic diseases: liver disease, diabetes, and hypertension. Quantify the tongue image features obtained in step two (including quantified values ​​of individual features and...). This system integrates tongue features with information on chronic disease-related symptoms and TCM constitution that users voluntarily enter in the mini-program, as well as laboratory liver function, blood glucose, and blood pressure test data uploaded by users, to achieve complementary support between tongue features and clinical data.

[0046] Symptom information quantification: A standardized symptom scoring table was developed for common TCM symptoms of liver disease, diabetes, and hypertension. Liver disease symptoms include hypochondriac pain, irritability, abdominal distension and loose stools, etc.; diabetes symptoms include dry mouth and excessive thirst, excessive appetite and hunger, frequent urination, etc.; and hypertension symptoms include dizziness and headache, flushed face and red eyes, soreness and weakness of the lower back and knees, etc. Each symptom is divided into four levels according to its severity: none, mild, moderate and severe, with quantification values ​​of 0-3 respectively.

[0047] Constitution information coding: According to the nine constitution classification standards of traditional Chinese medicine, the constitution information entered by the user is digitally coded. The balanced constitution is coded as 0, and the Qi deficiency constitution, Yang deficiency constitution, Yin deficiency constitution, Phlegm-dampness constitution, Damp-heat constitution, Blood stasis constitution, Qi stagnation constitution, and Special constitution are coded as 1-8 respectively. At the same time, the degree of constitution imbalance is quantified, with mild imbalance coded as 1, moderate imbalance coded as 2, and severe imbalance coded as 3.

[0048] Laboratory indicator normalization: The liver function indicators (alanine aminotransferase, aspartate aminotransferase, bilirubin, etc.), blood glucose indicators (fasting blood glucose, 2-hour postprandial blood glucose, glycated hemoglobin, etc.), and blood pressure indicators (systolic blood pressure, diastolic blood pressure) uploaded by users are normalized, and the measured values ​​of each indicator are mapped to the standard range of 0-1 to eliminate the dimensional differences between different indicators.

[0049] Dataset Construction: The quantitative data of tongue appearance features, the quantitative values ​​of symptom information, the encoded values ​​of constitution information, and the normalized values ​​of laboratory indicators are structured and integrated. The three types of chronic disease diagnosis datasets are constructed according to the format of "patient ID-tongue appearance features-symptom data-constitution data-laboratory indicators" to provide complete input data for subsequent diagnosis model reasoning.

[0050] Step 4: Reasoning based on the syndrome differentiation model for chronic diseases

[0051] Based on the fused syndrome differentiation dataset constructed in step three, syndrome inference is performed using a trained chronic disease syndrome differentiation classification model. This model is based on a lightweight deep learning model and optimized by combining traditional Chinese medicine syndrome differentiation rules. It outputs the syndrome differentiation confidence scores for common syndromes of liver disease, diabetes, and hypertension, respectively. The confidence score calculation uses a non-linear mapping formula to map the calculation results of the fused data to the 0-1 interval, intuitively reflecting the reliability of the syndrome differentiation judgment. The calculation formula is as follows:

[0052] (2)

[0053] in, For the first The confidence level of a syndrome type is measured in terms of its range from 0 to 1. The closer the value is to 1, the higher the degree of matching of the syndrome type. For non-linear activation functions, the formula is: This is used to map the calculation results of the fused data to the 0-1 range; The comprehensive index for the tongue image features obtained in step 2.3; The composite value is the normalized value of the laboratory indicators, obtained by weighted summation of the normalized values ​​of each indicator; The quantified value of the integration of symptoms and constitution is obtained by weighted summation of the quantified value of symptom information and the encoded value of constitution information; , , These are the fusion weighting coefficients of tongue appearance characteristics, laboratory indicators, and symptom constitution, respectively. They were obtained by training based on clinical diagnostic data of three types of chronic diseases, and different coefficient values ​​were set for liver disease, diabetes, and hypertension.

[0054] Liver disease syndrome differentiation: The model outputs five common liver disease syndromes and their corresponding patterns: liver qi stagnation, spleen deficiency with dampness, liver and gallbladder damp-heat, liver and kidney yin deficiency, and qi stagnation and blood stasis. ;

[0055] Diabetes Syndrome Differentiation: The model outputs five common diabetes syndromes and their corresponding patterns: Yin Deficiency with Dryness and Heat, Phlegm-Dampness Obstruction, Qi and Yin Deficiency, Yin and Yang Deficiency, and Blood Stasis in the Meridians. ;

[0056] Hypertension Syndrome Differentiation: The model outputs five common hypertension syndromes and their corresponding patterns: Liver Yang Rising, Phlegm and Blood Stasis, Liver and Kidney Yin Deficiency, Yin Deficiency and Yang Rising, and Qi Deficiency and Blood Stasis. ;

[0057] Primary syndrome type determination: The confidence levels of the syndrome types for the three types of chronic diseases are ranked, and the syndrome type with the highest confidence level is selected as the primary syndrome type for the patient. If the highest confidence level is lower than the preset threshold, it indicates that further clinical diagnosis is needed.

[0058] Step 5: Outputting Diagnostic Results and Iterating the Model

[0059] Based on the syndrome differentiation results in step four, a standardized tongue diagnosis report is generated and simultaneously pushed to the patient and doctor ends of the mini-program. At the same time, clinical diagnosis data and user feedback data are continuously collected to iteratively optimize the lightweight feature extraction model and the chronic disease diagnosis and classification model, forming a data closed loop.

[0060] Standardized diagnostic report generation: The report includes basic patient information, tongue feature analysis (including tongue image photos, quantitative results of each feature, etc.) The results include: chronic disease syndrome differentiation (main syndrome type and confidence level), chronic disease risk level assessment (based on syndrome type, laboratory indicators, and symptom comprehensive judgment), and targeted TCM conditioning suggestions (including diet, rest, emotions, Chinese herbal tea, acupoint massage, etc.). Among them, the conditioning suggestions are personalized plans for different syndrome types of the three types of chronic diseases.

[0061] Results push to both patients and doctors: Patients can view the complete diagnosis report and treatment suggestions, while doctors can view the patient's tongue data, integrated clinical data and AI diagnosis results, and are provided with "confirm" and "modify" operation entry points. The diagnosis results modified by doctors will be stored as labeled data in the model training library.

[0062] Model Iteration and Optimization: Continuously collect tongue image data from the mini-program, user clinical data, AI diagnosis results, and doctor-annotated clinical diagnosis results to incrementally train the lightweight feature extraction model and the chronic disease diagnosis and classification model, continuously optimize the model's feature extraction accuracy and syndrome inference accuracy, and dynamically adjust the formula (1) based on clinical feedback. In formula (2) , , This ensures that the model always aligns with the needs of clinical diagnosis.

[0063] The lightweight feature extraction model and chronic disease syndrome differentiation and classification model of this invention have been pruned and optimized by quantization and compression, with the overall size controlled within 50MB. They can run directly on mobile applications locally, realizing real-time analysis of tongue images and syndrome inference without relying on cloud computing power, thus meeting the needs of outpatient self-testing for patients with liver disease, diabetes, and hypertension.

[0064] Example:

[0065] Taking the outpatient diagnosis of diabetes patients as an example, after the patient takes a photo of their tongue via a mini-program, the local lightweight model quickly extracts features and performs diagnostic reasoning. The patient's end instantly displays the diagnosis results for the specific syndrome type and its probability, and pushes a treatment plan including dietary suggestions (such as recommending specific medicinal diets and avoiding specific types of food), exercise guidance (such as brisk walking for a specific duration each day), and acupressure (such as specific acupoints). Simultaneously, the doctor receives quantitative data such as the patient's tongue image, the depth of tongue fissures, and the thickness of the tongue coating, as well as the diagnostic basis generated by the model after integrating clinical data such as the patient's age and fasting blood glucose. Combining this with information from offline consultations, if the doctor believes the patient has a concurrent syndrome type, they can adjust the syndrome type to a complex syndrome type through the "edit" entry and supplement with personalized suggestions. The modified data (including original tongue images, clinical data, AI preliminary judgment results, and doctor's final judgment annotations) was automatically stored in the encrypted training library. The system started incremental training monthly, using an optimizer to update the fully connected layer parameters of the chronic disease syndrome differentiation and classification model, adjusting the weights of features related to specific syndrome types, and adjusting the weight coefficients of different types of data such as tongue image features and symptom data in the syndrome differentiation formula based on recent clinical feedback, thereby improving the model's accuracy in recognizing concurrent syndromes.

[0066] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting tongue features in Traditional Chinese Medicine and for auxiliary analysis in the differentiation of chronic diseases, characterized in that: Includes the following steps: Step 1: Tongue image preprocessing. The tongue image photos collected and uploaded by the user through the mobile application applet are processed sequentially with illumination correction, region segmentation, and noise removal to extract a standard image of the pure tongue image region of 224×224 pixels. Step 2: Lightweight tongue image feature extraction. An improved MobileNet lightweight convolutional neural network is used to extract features from standard tongue images, identify and quantify core tongue image features such as tongue color, coating color, coating thickness, cracks, teeth marks, ecchymosis, tongue body size, and amount of saliva, and calculate the comprehensive index of tongue image feature quantification through a formula. Step 3: Multi-dimensional chronic disease data fusion. The quantitative data of tongue appearance features are fused with the chronic disease-related symptom information entered by the user, TCM constitution information, and uploaded laboratory indicator data of liver function, blood sugar, and blood pressure to construct a chronic disease syndrome differentiation dataset. Step 4: Chronic disease syndrome differentiation model reasoning. Based on the fused syndrome differentiation dataset, the trained chronic disease syndrome differentiation classification model outputs the syndrome differentiation confidence scores of common syndrome types for liver disease, diabetes, and hypertension, respectively. The syndrome type with the highest confidence score is selected as the main syndrome type. The syndrome differentiation confidence score is calculated using a nonlinear mapping formula. Step 5: Outputting the diagnosis results and iterating the model. Based on the reasoning results of the syndrome type, a standardized diagnosis report is generated and pushed to the patient and doctor terminals of the mini-program. At the same time, clinical data and user feedback are collected to incrementally train the model and iteratively optimize the model parameters.

2. The method for extracting TCM tongue features and assisting in the differentiation of chronic diseases according to claim 1, characterized in that: The illumination correction in step one involves adjusting the pixel values ​​of the RGB three-color channels to achieve color balance and normalizing the average pixel value of each channel to a preset standard range; the region segmentation involves using a semantic segmentation algorithm to extract the complete outline of the tongue and smooth the edges; and the noise removal involves using a Gaussian filtering algorithm to remove shooting noise while retaining the core features of the tongue image.

3. The method for extracting TCM tongue features and assisting in the differentiation of chronic diseases according to claim 1, characterized in that: The improved MobileNet lightweight convolutional neural network structure in step two is as follows: the input layer is a standard tongue image of 224×224×3, which is sequentially connected to a Conv2d+BN+ReLU convolutional layer, 6 depthwise separable convolutional blocks, and a global average pooling layer, and finally divided into four output branches: tongue color, tongue coating color, tongue shape, and saliva. The Conv2d+BN+ReLU convolutional layer uses a 3×3 convolutional kernel and 32 channels. The depthwise separable convolutional blocks reduce the number of model parameters by about 88%. After pruning and quantization compression, the model size is controlled within 50MB.

4. The method for extracting TCM tongue features and assisting in the differentiation of chronic diseases according to claim 1, characterized in that: The quantification rules for tongue features in step two are as follows: tongue color is divided into 5 levels: light red, red, crimson, light white, and dark purple, with quantification values ​​of 1-5 respectively; tongue coating color is divided into 5 levels: white coating, yellow coating, gray coating, black coating, and no coating, with quantification values ​​of 1-5 respectively; tongue shape is divided into 5 levels: normal, swollen, thin, cracked, and teeth marks, with quantification values ​​of 0-4 respectively; body fluid is divided into 4 levels: normal, slightly excessive, slightly insufficient, and no body fluid, with quantification values ​​of 0-3 respectively; ecchymosis and thick coating are binary features, with a corresponding feature recorded as 1 and no corresponding feature recorded as 0.

5. The method for extracting TCM tongue features and assisting in the differentiation of chronic diseases according to claim 1, characterized in that: The formula for calculating the comprehensive index of tongue image feature quantification in step two is as follows: ,in A comprehensive index for quantifying tongue features. The number of tongue features involved in the calculation. For the first Weighting coefficients for individual tongue features For the first Quantitative values ​​of individual tongue features; Based on the TCM syndrome differentiation system for liver disease, diabetes, and hypertension, and trained with clinical tongue diagnosis data, dynamic adjustments are made for these three types of chronic diseases.

6. The method for extracting TCM tongue features and assisting in the differentiation of chronic diseases according to claim 1, characterized in that: The multi-dimensional data fusion in step three includes: classifying the symptoms of liver disease, diabetes, and hypertension into four levels of severity: none, mild, moderate, and severe, with quantitative values ​​ranging from 0 to 3; digitizing the nine constitutions in traditional Chinese medicine, with the balanced constitution recorded as 0 and the other constitutions as 1-8, and the degree of constitution imbalance quantified as mild, moderate, and severe, with values ​​ranging from 1 to 3; and normalizing the laboratory indicators of liver function, blood sugar, and blood pressure, mapping the measured values ​​to the 0-1 standard range.

7. The method for extracting TCM tongue features and assisting in the differentiation of chronic diseases according to claim 1, characterized in that: The formula for calculating the confidence level of chronic disease diagnosis in step four is as follows: ,in For the first The confidence level of dialectical analysis for individual cases It is a non-linear activation function. A comprehensive index for quantifying tongue features. This is the normalized composite value of the laboratory indicators. This is a quantitative value representing the integration of symptoms and physical constitution. , , To integrate the weighting coefficients, different coefficient values ​​were set for liver disease, diabetes, and hypertension.

8. The method for extracting tongue features in traditional Chinese medicine and assisting in the diagnosis and analysis of chronic diseases according to claim 1, characterized in that: The syndrome inference results in step four are as follows: liver disease outputs 5 syndrome types: liver qi stagnation, spleen deficiency with dampness, liver and gallbladder damp heat, liver and kidney yin deficiency, and qi stagnation and blood stasis; diabetes outputs 5 syndrome types: yin deficiency with dryness and heat, phlegm and dampness obstruction, qi and yin deficiency, yin and yang deficiency, and blood stasis in the meridians; hypertension outputs 5 syndrome types: liver yang hyperactivity, phlegm and blood stasis, liver and kidney yin deficiency, yin deficiency with yang hyperactivity, and qi deficiency and blood stasis; the main syndrome type is the syndrome type with the highest confidence among all types of chronic disease syndrome types.

9. The method for extracting tongue features in traditional Chinese medicine and assisting in the differentiation of chronic diseases according to claim 1, characterized in that: The standardized diagnosis report in step five includes basic patient information, tongue appearance analysis, chronic disease syndrome type judgment results, chronic disease risk level assessment, and personalized TCM conditioning suggestions. The conditioning suggestions are formulated for different syndrome types of liver disease, diabetes, and hypertension, focusing on diet, rest, emotions, Chinese herbal teas, and acupoint massage.

10. The method for extracting tongue features in traditional Chinese medicine and assisting in the diagnosis and analysis of chronic diseases according to claim 1, characterized in that: The model iteration and optimization in step five involves collecting tongue image data from the mini-program, user clinical data, AI diagnosis results, and doctor-annotated clinical diagnosis results, incrementally training the lightweight feature extraction model and the chronic disease diagnosis and classification model, and dynamically adjusting the tongue image feature weight coefficient and the data fusion weight coefficient.

Images